Stata Survey Data Reference Manual: Release 14

STATA SURVEY DATA REFERENCE
MANUAL
RELEASE 14
A Stata Press Publication

StataCorp LP
College Station, Texas
®
Copyright c 1985–2015 StataCorp LP
All rights reserved
Version 14
Published by Stata Press, 4905 Lakeway Drive, College Station, Texas 77845
Typeset in TEX
ISBN-10: 1-59718-167-6
ISBN-13: 978-1-59718-167-9
This manual is protected by copyright. All rights are reserved. No part of this manual may be reproduced, stored
in a retrieval system, or transcribed, in any form or by any means—electronic, mechanical, photocopy, recording, or
otherwise—without the prior written permission of StataCorp LP unless permitted subject to the terms and conditions
of a license granted to you by StataCorp LP to use the software and documentation. No license, express or implied,
by estoppel or otherwise, to any intellectual property rights is granted by this document.
StataCorp provides this manual “as is” without warranty of any kind, either expressed or implied, including, but
not limited to, the implied warranties of merchantability and fitness for a particular purpose. StataCorp may make
improvements and/or changes in the product(s) and the program(s) described in this manual at any time and without
notice.
The software described in this manual is furnished under a license agreement or nondisclosure agreement. The software
may be copied only in accordance with the terms of the agreement. It is against the law to copy the software onto
DVD, CD, disk, diskette, tape, or any other medium for any purpose other than backup or archival purposes.
The automobile dataset appearing on the accompanying media is Copyright
c 1979 by Consumers Union of U.S.,
Inc., Yonkers, NY 10703-1057 and is reproduced by permission from CONSUMER REPORTS, April 1979.
Stata, , Stata Press, Mata, , and NetCourse are registered trademarks of StataCorp LP.
Stata and Stata Press are registered trademarks with the World Intellectual Property Organization of the United Nations.
NetCourseNow is a trademark of StataCorp LP.
Other brand and product names are registered trademarks or trademarks of their respective companies.
For copyright information about the software, type help copyright within Stata.
The suggested citation for this software is

StataCorp. 2015. Stata: Release 14 . Statistical Software. College Station, TX: StataCorp LP.
Contents
intro . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Introduction to survey data manual 1

survey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Introduction to survey commands 2
bootstrap options . . . . . . . . . . . . . . . . . . . . . More options for bootstrap variance estimation 23

brr options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . More options for BRR variance estimation 24
direct standardization . . . . . . . . . . . Direct standardization of means, proportions, and ratios 25
estat . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Postestimation statistics for survey data 29
jackknife options . . . . . . . . . . . . . . . . . . . . . . More options for jackknife variance estimation 51
ml for svy . . . . . . . . . . . . . . . . . . . . . Maximum pseudolikelihood estimation for survey data 52
poststratification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Poststratification for survey data 54
sdr options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . More options for SDR variance estimation 58

subpopulation estimation . . . . . . . . . . . . . . . . . . . . Subpopulation estimation for survey data 59
svy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . The survey prefix command 65
svy bootstrap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Bootstrap for survey data 74
svy brr . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Balanced repeated replication for survey data 82
svy estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Estimation commands for survey data 90
svy jackknife . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jackknife estimation for survey data 102
svy postestimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Postestimation tools for svy 110
svy sdr . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Successive difference replication for survey data 127
svy: tabulate oneway . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . One-way tables for survey data 133
svy: tabulate twoway . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Two-way tables for survey data 140
svydescribe . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Describe survey data 161
svymarkout . . . . . . . Mark observations for exclusion on the basis of survey characteristics 167
svyset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Declare survey design for dataset 168
variance estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Variance estimation for survey data 186
i
ii Contents
Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
Subject and author index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205

Cross-referencing the documentation
When reading this manual, you will find references to other Stata manuals. For example,
[U] 26 Overview of Stata estimation commands
[R] regress
[D] reshape
The first example is a reference to chapter 26, Overview of Stata estimation commands, in the User’s
Guide; the second is a reference to the regress entry in the Base Reference Manual; and the third
is a reference to the reshape entry in the Data Management Reference Manual.
All the manuals in the Stata Documentation have a shorthand notation:
[GSM] Getting Started with Stata for Mac

[GSU] Getting Started with Stata for Unix
[GSW] Getting Started with Stata for Windows
[U] Stata User’s Guide
[R] Stata Base Reference Manual
[BAYES] Stata Bayesian Analysis Reference Manual
[D] Stata Data Management Reference Manual
[FN] Stata Functions Reference Manual
[G] Stata Graphics Reference Manual
[IRT] Stata Item Response Theory Reference Manual
[XT] Stata Longitudinal-Data/Panel-Data Reference Manual
[ME] Stata Multilevel Mixed-Effects Reference Manual
[MI] Stata Multiple-Imputation Reference Manual
[MV] Stata Multivariate Statistics Reference Manual
[PSS] Stata Power and Sample-Size Reference Manual
[P] Stata Programming Reference Manual
[SEM] Stata Structural Equation Modeling Reference Manual
[SVY] Stata Survey Data Reference Manual
[ST] Stata Survival Analysis Reference Manual
[TS] Stata Time-Series Reference Manual
[TE] Stata Treatment-Effects Reference Manual:
Potential Outcomes/Counterfactual Outcomes
[I] Stata Glossary and Index
[M] Mata Reference Manual
iii
Title
intro — Introduction to survey data manual
Description Also see
Description
This manual documents the survey data commands and is referred to as [SVY] in references.
After this entry, [SVY] survey provides an overview of the survey commands. This manual is
arranged alphabetically. If you are new to Stata’s survey data commands, we recommend that you
read the following sections first:
[SVY] survey Introduction to survey commands

[SVY] svyset Declare survey design for dataset
[SVY] svydescribe Describe survey data
[SVY] svy estimation Estimation commands for survey data
[SVY] svy postestimation Postestimation tools for svy
Stata is continually being updated, and Stata users are continually writing new commands. To
find out about the latest survey data features, type search survey after installing the latest official
updates; see [R] update.
Also see
[U] 1.3 What’s new
[R] intro — Introduction to base reference manual
1
Title
survey — Introduction to survey commands
Description Remarks and examples Acknowledgments References Also see
Description
The Survey Data Reference Manual is organized alphabetically, making it easy to find an individual
entry if you know the name of a command. This overview organizes and presents the commands
conceptually, that is, according to the similarities in the functions they perform.
Survey design tools

[SVY] svyset Declare survey design for dataset
[SVY] svydescribe Describe survey data
Survey data analysis tools

[SVY] svy The survey prefix command
[SVY] svy estimation Estimation commands for survey data
[SVY] svy: tabulate oneway One-way tables for survey data
[SVY] svy: tabulate twoway Two-way tables for survey data
[SVY] svy postestimation Postestimation tools for svy
[SVY] estat Postestimation statistics for survey data, such as design effects
[SVY] svy bootstrap Bootstrap for survey data
[SVY] bootstrap options More options for bootstrap variance estimation
[SVY] svy brr Balanced repeated replication for survey data
[SVY] brr options More options for BRR variance estimation
[SVY] svy jackknife Jackknife estimation for survey data
[SVY] jackknife options More options for jackknife variance estimation
[SVY] svy sdr Successive difference replication for survey data
[SVY] sdr options More options for SDR variance estimation
Survey data concepts
[SVY] variance estimation Variance estimation for survey data
[SVY] subpopulation estimation Subpopulation estimation for survey data
[SVY] direct standardization Direct standardization of means, proportions, and ratios
[SVY] poststratification Poststratification for survey data
Tools for programmers of new survey commands

[SVY] ml for svy Maximum pseudolikelihood estimation for survey data
[SVY] svymarkout Mark observations for exclusion on the basis of survey
characteristics
2
survey — Introduction to survey commands 3
Remarks and examples

Remarks are presented under the following headings:
Introduction
Survey design tools
Video examples
Introduction
Stata’s facilities for survey data analysis are centered around the svy prefix command. After you
identify the survey design characteristics with the svyset command, prefix the estimation commands
in your data analysis with “svy:”. For example, where you would normally use the regress command
to fit a linear regression model for nonsurvey data, use svy: regress to fit a linear regression model
for your survey data.
Why should you use the svy prefix command when you have survey data? To answer this question,
we need to discuss some of the characteristics of survey design and survey data collection because
these characteristics affect how we must perform our analysis if we want to get it right.
Survey data are characterized by the following:
• Sampling weights, also called probability weights — pweights in Stata’s terminology
• Cluster sampling
• Stratification
These features arise from the design and details of the data collection procedure. Here’s a brief
description of how these design features affect the analysis of the data:
• Sampling weights. In sample surveys, observations are selected through a random process,
but different observations may have different probabilities of selection. Weights are equal to
(or proportional to) the inverse of the probability of being sampled. Various postsampling
adjustments to the weights are sometimes made, as well. A weight of wj for the j th observation
means, roughly speaking, that the j th observation represents wj elements in the population
from which the sample was drawn.
Omitting weights from the analysis results in estimates that may be biased, sometimes seriously
so. Sampling weights also play a role in estimating standard errors.
• Clustering. Individuals are not sampled independently in most survey designs. Collections of
individuals (for example, counties, city blocks, or households) are typically sampled as a group,
known as a cluster.
There may also be further subsampling within the clusters. For example, counties may be
sampled, then city blocks within counties, then households within city blocks, and then finally
persons within households. The clusters at the first level of sampling are called primary sampling
units (PSUs) — in this example, counties are the PSUs. In the absence of clustering, the PSUs
are defined to be the individuals, or, equivalently, clusters, each of size one.
Cluster sampling typically results in larger sample-to-sample variability than sampling individuals
directly. This increased variability must be accounted for in standard error estimates, hypothesis
testing, and other forms of inference.
4 survey — Introduction to survey commands
• Stratification. In surveys, different groups of clusters are often sampled separately. These groups
are called strata. For example, the 254 counties of a state might be divided into two strata, say,
urban counties and rural counties. Then 10 counties might be sampled from the urban stratum,
and 15 from the rural stratum.
Sampling is done independently across strata; the stratum divisions are fixed in advance. Thus
strata are statistically independent and can be analyzed as such. When the individual strata
are more homogeneous than the population as a whole, the homogeneity can be exploited to
produce smaller (and honestly so) estimates of standard errors.
To put it succinctly: using sampling weights is important to get the point estimates right. We must
consider the weighting, clustering, and stratification of the survey design to get the standard errors
right. If our analysis ignores the clustering in our design, we would probably produce standard errors
that are smaller than they should be. Stratification can be used to get smaller standard errors for a
given overall sample size.
For more detailed introductions to complex survey data analysis, see Cochran (1977); Heeringa,
West, and Berglund (2010); Kish (1965); Levy and Lemeshow (2008); Scheaffer et al.; (2012);
Skinner, Holt, and Smith (1989); Stuart (1984); Thompson (2012); and Williams (1978).
Survey design tools

Before using svy, first take a quick look at [SVY] svyset. Use the svyset command to specify
the variables that identify the survey design characteristics and default method for estimating standard
errors. Once set, svy will automatically use these design specifications until they are cleared or
changed or a new dataset is loaded into memory.
As the following two examples illustrate, svyset allows you to identify a wide range of complex
sampling designs. First, we show a simple single-stage design and then a complex multistage design.
Example 1: Survey data from a one-stage design

A commonly used single-stage survey design uses clustered sampling across several strata, where
the clusters are sampled without replacement. In a Stata dataset composed of survey data from this
design, the survey design variables identify information about the strata, PSUs (clusters), sampling
weights, and finite population correction. Here we use svyset to specify these variables, respectively
named strata, su1, pw, and fpc1.
. use http://www.stata-press.com/data/r14/stage5a
. svyset su1 [pweight=pw], strata(strata) fpc(fpc1)
pweight: pw
VCE: linearized
Single unit: missing
Strata 1: strata
SU 1: su1
FPC 1: fpc1
In addition to the variables we specified, svyset reports that the default method for estimating
standard errors is Taylor linearization and that svy will report missing values for the standard errors
when it encounters a stratum with one sampling unit (also called singleton strata).
Example 2: Multistage survey data

We have (fictional) data on American high school seniors (12th graders), and the data were collected
according to the following multistage design. In the first stage, counties were independently selected
within each state. In the second stage, schools were selected within each chosen county. Within each
chosen school, a questionnaire was filled out by every attending high school senior. We have entered
all the information into a Stata dataset called multistage.dta.
The survey design variables are as follows:
• state contains the stratum identifiers.
• county contains the first-stage sampling units.
• ncounties contains the total number of counties within each state.
• school contains the second-stage sampling units.
• nschools contains the total number of schools within each county.
• sampwgt contains the sampling weight for each sampled individual.
Here we load the dataset into memory and use svyset with the above variables to declare that
these data are survey data.
. use http://www.stata-press.com/data/r14/multistage
. svyset county [pw=sampwgt], strata(state) fpc(ncounties) || school,
> fpc(nschools)
pweight: sampwgt
VCE: linearized
Strata 1: state
SU 1: county
FPC 1: ncounties
Strata 2: <one>
SU 2: school
FPC 2: nschools
. save highschool
file highschool.dta saved
We saved the svyset dataset to highschool.dta. We can now use this new dataset without having
to worry about respecifying the design characteristics.
. clear
. describe
Contains data
obs: 0
vars: 0
size: 0
Sorted by:
. use highschool
. svyset
pweight: sampwgt
VCE: linearized
Strata 1: state
SU 1: county
FPC 1: ncounties
Strata 2: <one>
SU 2: school
FPC 2: nschools
After the design characteristics have been svyset, you should also look at [SVY] svydescribe. Use
svydescribe to browse each stage of your survey data; svydescribe reports useful information
on sampling unit counts, missing data, and singleton strata.
Example 3: Survey describe

Here we use svydescribe to describe the first stage of our survey dataset of sampled high school
seniors. We specified the weight variable to get svydescribe to report on where it contains missing
values and how this affects the estimation sample.
. svydescribe weight
Survey: Describing stage 1 sampling units
pweight: sampwgt
VCE: linearized
Strata 1: state
SU 1: county
FPC 1: ncounties
Strata 2: <one>
SU 2: school
FPC 2: nschools
#Obs with #Obs with #Obs per included Unit
#Units #Units complete missing
Stratum included omitted data data min mean max
1 2 0 92 0 34 46.0 58
2 2 0 112 0 51 56.0 61
3 2 0 43 0 18 21.5 25
4 2 0 37 0 14 18.5 23
5 2 0 96 0 38 48.0 58
(output omitted )
46 2 0 115 0 56 57.5 59
47 2 0 67 0 28 33.5 39
48 2 0 56 0 23 28.0 33
49 2 0 78 0 39 39.0 39
50 2 0 64 0 31 32.0 33
50 100 0 4,071 0 14 40.7 81
4,071
From the output, we gather that there are 50 strata, each stratum contains two PSUs, the PSUs vary
in size, and the total sample size is 4,071 students. We can also see that there are no missing data
in the weight variable.

Stata’s suite of survey data commands is governed by the svy prefix command; see [SVY] svy and
[SVY] svy estimation. svy runs the supplied estimation command while accounting for the survey
design characteristics in the point estimates and variance estimation method. The available variance
estimation methods are balanced repeated replication (BRR), the bootstrap, the jackknife, successive
difference replication, and first-order Taylor linearization. By default, svy computes standard errors
by using the linearized variance estimator — so called because it is based on a first-order Taylor series
linear approximation (Wolter 2007). In the nonsurvey context, we refer to this variance estimator as
the robust variance estimator, otherwise known in Stata as the Huber/White/sandwich estimator; see
[P] robust.
Example 4: Estimating a population mean

Here we use the svy prefix with the mean command to estimate the average weight of high school
seniors in our population.
. svy: mean weight
(running mean on estimation sample)
Survey: Mean estimation
Number of strata = 50 Number of obs = 4,071
Number of PSUs = 100 Population size = 8,000,000
Design df = 50
Linearized
Mean Std. Err. [95% Conf. Interval]
weight 160.2863 .7412512 158.7974 161.7751
In its header, svy reports the number of strata and PSUs from the first stage, the sample size, an
estimate of population size, and the design degrees of freedom. Just like the standard output from
the mean command, the table of estimation results contains the estimated mean and its standard error
as well as a confidence interval.
Example 5: Survey regression

Here we use the svy prefix with the regress command to model the association between weight
and height in our population of high school seniors.
. svy: regress weight height
(running regress on estimation sample)
Survey: Linear regression
Design df = 50
F( 1, 50) = 593.99
Prob > F = 0.0000
R-squared = 0.2787
Linearized
weight Coef. Std. Err. t P>|t| [95% Conf. Interval]
height .7163115 .0293908 24.37 0.000 .6572784 .7753447

_cons -149.6183 12.57265 -11.90 0.000 -174.8712 -124.3654
In addition to the header elements we saw in the previous example using svy: mean, the command
svy: regress also reports a model F test and estimated R2 . Although many of Stata’s model-fitting
commands report Z statistics for testing coefficients against zero, svy always reports t statistics and
uses the design degrees of freedom to compute p-values.
The svy prefix can be used with many estimation commands in Stata; see [SVY] svy estimation
for a list of estimation commands that support the svy prefix.
Example 6: Cox’s proportional hazards model

Suppose that we want to model the incidence of lung cancer by using three risk factors: smoking
status, sex, and place of residence. Our dataset comes from a longitudinal health survey: the First
National Health and Nutrition Examination Survey (NHANES I) (Miller 1973; Engel et al. 1978) and its
1992 Epidemiologic Follow-up Study (NHEFS) (Cox et al. 1997); see the National Center for Health
Statistics website at http://www.cdc.gov/nchs/. We will be using data from the samples identified by
NHANES I examination locations 1–65 and 66–100; thus we will svyset the revised pseudo-PSU and
strata variables associated with these locations. Similarly, our pweight variable was generated using
the sampling weights for the nutrition and detailed samples for locations 1–65 and the weights for
the detailed sample for locations 66–100.
. use http://www.stata-press.com/data/r14/nhefs
. svyset psu2 [pw=swgt2], strata(strata2)
pweight: swgt2
VCE: linearized
Strata 1: strata2
SU 1: psu2
FPC 1: <zero>
The lung cancer information was taken from the 1992 NHEFS interview data. We use the participants’
ages for the time scale. Participants who never had lung cancer and were alive for the 1992 interview
were considered censored. Participants who never had lung cancer and died before the 1992 interview
were also considered censored at their age of death.
. stset age_lung_cancer [pw=swgt2], fail(lung_cancer)
failure event: lung_cancer != 0 & lung_cancer < .
obs. time interval: (0, age_lung_cancer]
exit on or before: failure
weight: [pweight=swgt2]
14407 total observations

5126 event time missing (age_lung_cancer>=.) PROBABLE ERROR
9281 observations remaining, representing

83 failures in single-record/single-failure data
599691 total analysis time at risk and under observation
at risk from t = 0
earliest observed entry t = 0
last observed exit t = 97
Although stset warns us that it is a “probable error” to have 5,126 observations with missing event
times, we can verify from the 1992 NHEFS documentation that there were indeed 9,281 participants
with complete information.
For our proportional hazards model, we pulled the risk factor information from the NHANES I and
1992 NHEFS datasets. Smoking status was taken from the 1992 NHEFS interview data, but we filled
in all but 132 missing values by using the general medical history supplement data in NHANES I.
Smoking status is represented by separate indicator variables for former smokers and current smokers;
the base comparison group is nonsmokers. Sex was determined using the 1992 NHEFS vitality data
and is represented by an indicator variable for males. Place-of-residence information was taken from
the medical history questionnaire in NHANES I and is represented by separate indicator variables for
rural and heavily populated (more than 1 million people) urban residences; the base comparison group
is urban residences with populations of fewer than 1 million people.
. svy: stcox former_smoker smoker male urban1 rural

(running stcox on estimation sample)
Survey: Cox regression
Design df = 70
F( 5, 66) = 14.07
Prob > F = 0.0000
Linearized
_t Haz. Ratio Std. Err. t P>|t| [95% Conf. Interval]
former_smoker 2.788113 .6205102 4.61 0.000 1.788705 4.345923

smoker 7.849483 2.593249 6.24 0.000 4.061457 15.17051
male 1.187611 .3445315 0.59 0.555 .6658757 2.118142
urban1 .8035074 .3285144 -0.54 0.594 .3555123 1.816039
rural 1.581674 .5281859 1.37 0.174 .8125799 3.078702
From the above results, we can see that both former and current smokers have a significantly
higher risk for developing lung cancer than do nonsmokers.
svy: tabulate can be used to produce one-way and two-way tables with survey data and can
produce survey-adjusted tests of independence for two-way contingency tables; see [SVY] svy: tabulate
oneway and [SVY] svy: tabulate twoway.
Example 7: Two-way tables for survey data

With data from the Second National Health and Nutrition Examination Survey (NHANES II)
(McDowell et al. 1981), we use svy: tabulate to produce a two-way table of cell proportions along
with their standard errors and confidence intervals (the survey design characteristics have already
been svyset). We also use the format() option to get svy: tabulate to report the cell values
and marginals to four decimal places.
. use http://www.stata-press.com/data/r14/nhanes2b
. svy: tabulate race diabetes, row se ci format(%7.4f)
(running tabulate on estimation sample)
Design df = 31
1=white,
2=black, diabetes, 1=yes, 0=no
3=other 0 1 Total
White 0.9680 0.0320 1.0000

(0.0020) (0.0020)
[0.9638,0.9718] [0.0282,0.0362]
Black 0.9410 0.0590 1.0000

(0.0061) (0.0061)
[0.9271,0.9523] [0.0477,0.0729]
Other 0.9797 0.0203 1.0000

(0.0076) (0.0076)
[0.9566,0.9906] [0.0094,0.0434]
Total 0.9658 0.0342 1.0000

(0.0018) (0.0018)
[0.9619,0.9693] [0.0307,0.0381]
Key: row proportion

(linearized standard error of row proportion)
[95% confidence interval for row proportion]
Pearson:
Uncorrected chi2(2) = 21.3483
Design-based F(1.52, 47.26) = 15.0056 P = 0.0000
svy: tabulate has many options, such as the format() option, for controlling how the table
looks. See [SVY] svy: tabulate twoway for a discussion of the different design-based and unadjusted
tests of association.
All the standard postestimation commands (for example, estimates, lincom, margins, nlcom,
test, testnl) are also available after svy.
Example 8: Comparing means

Going back to our high school survey data in example 2, we estimate the mean of weight (in
pounds) for each subpopulation identified by the categories of the sex variable (male and female).
. use http://www.stata-press.com/data/r14/highschool
. svy: mean weight, over(sex)
Design df = 50
male: sex = male
female: sex = female
Linearized
Over Mean Std. Err. [95% Conf. Interval]
weight
male 175.4809 1.116802 173.2377 177.7241
female 146.204 .9004157 144.3955 148.0125
Here we use the test command to test the hypothesis that the average male is 30 pounds heavier
than the average female; from the results, we cannot reject this hypothesis at the 5% level.
. test [weight]male - [weight]female = 30
Adjusted Wald test
( 1) [weight]male - [weight]female = 30
F( 1, 50) = 0.23
Prob > F = 0.6353
estat has specific subroutines for use after svy; see [SVY] estat.
• estat svyset reports the survey design settings used to produce the current estimation results.
• estat effects and estat lceffects report a table of design and misspecification effects
for point estimates and linear combinations of point estimates, respectively.
• estat size reports a table of sample and subpopulation sizes after svy: mean, svy: pro-
portion, svy: ratio, and svy: total.
• estat sd reports subpopulation standard deviations on the basis of the estimation results from
mean and svy: mean.
• estat strata reports the number of singleton and certainty strata within each sampling stage.
• estat cv reports the coefficient of variation for each coefficient in the current estimation results.
• estat gof reports a goodness-of-fit test for binary response models using survey data.
Example 9: Design effects

Here we use estat effects to report the design effects DEFF and DEFT for the mean estimates
from the previous example.
. estat effects
male: sex = male
female: sex = female
Linearized
Over Mean Std. Err. DEFF DEFT
weight
male 175.4809 1.116802 2.61016 1.61519
female 146.204 .9004157 1.7328 1.31603
Note: Weights must represent population totals for deff to

be correct when using an FPC; however, deft is
invariant to the scale of weights.
Now we use estat lceffects to report the design effects DEFF and DEFT for the difference of
the mean estimates from the previous example.
. estat lceffects [weight]male - [weight]female
( 1) [weight]male - [weight]female = 0
Mean Coef. Std. Err. DEFF DEFT
(1) 29.27691 1.515201 2.42759 1.55768
Note: Weights must represent population totals for deff to

be correct when using an FPC; however, deft is
invariant to the scale of weights.
The svy brr prefix command produces point and variance estimates by using the BRR method;
see [SVY] svy brr. BRR was first introduced by McCarthy (1966, 1969a, and 1969b) as a method of
variance estimation for designs with two PSUs in every stratum. The BRR variance estimator tends to
give more reasonable variance estimates for this design than the linearized variance estimator, which
can result in large values and undesirably wide confidence intervals.
The svy jackknife prefix command produces point and variance estimates by using the jackknife
replication method; see [SVY] svy jackknife. The jackknife is a data-driven variance estimation
method that can be used with model-fitting procedures for which the linearized variance estimator is
not implemented, even though a linearized variance estimator is theoretically possible to derive (Shao
and Tu 1995).
To protect the privacy of survey participants, public survey datasets may contain replicate-weight
variables instead of variables that identify the PSUs and strata. These replicate-weight variables can be
used with the appropriate replication method for variance estimation instead of the linearized variance
estimator; see [SVY] svyset.
The svy brr and svy jackknife prefix commands can be used with those commands that may
not be fully supported by svy but are compatible with the BRR and the jackknife replication methods.
They can also be used to produce point estimates for expressions of estimation results from a prefixed
command.
The svy bootstrap and svy sdr prefix commands work only with replicate weights. Both assume
that you have obtained these weight variables externally.
The svy bootstrap prefix command produces variance estimates that have been adjusted for
bootstrap sampling. Bootstrap sampling of complex survey has become more popular in recent years
and is the variance-estimation method used in the National Population Health Survey conducted by
Statistics Canada; see [SVY] svy bootstrap and [SVY] variance estimation for more details.
The svy sdr prefix command produces variance estimates that implement successive difference
replication (SDR), first introduced by Fay and Train (1995) as a method for annual demographic
supplements to the Current Population Survey. This method is typically applied to systematic samples
where the observed sampling units follow a natural order; see [SVY] svy sdr and [SVY] variance
estimation for more details.
Example 10: BRR and replicate-weight variables

The survey design for the NHANES II data (McDowell et al. 1981) is specifically suited to BRR:
there are two PSUs in every stratum.
. use http://www.stata-press.com/data/r14/nhanes2
. svydescribe
pweight: finalwgt
VCE: linearized
Strata 1: strata
SU 1: psu
FPC 1: <zero>
#Obs per Unit
Stratum #Units #Obs min mean max
1 2 380 165 190.0 215

2 2 185 67 92.5 118
3 2 348 149 174.0 199
4 2 460 229 230.0 231
5 2 252 105 126.0 147
(output omitted )
29 2 503 215 251.5 288
30 2 365 166 182.5 199
31 2 308 143 154.0 165
32 2 450 211 225.0 239
31 62 10,351 67 167.0 288
Here is a privacy-conscious dataset equivalent to the one above; all the variables and values
remain, except that strata and psu are replaced with BRR replicate-weight variables. The BRR
replicate-weight variables are already svyset, and the default method for variance estimation is
vce(brr).
. use http://www.stata-press.com/data/r14/nhanes2brr
. svyset
pweight: finalwgt
VCE: brr
MSE: off
brrweight: brr_1 brr_2 brr_3 brr_4 brr_5 brr_6 brr_7 brr_8 brr_9 brr_10
brr_11 brr_12 brr_13 brr_14 brr_15 brr_16 brr_17 brr_18 brr_19
brr_29 brr_30 brr_31 brr_32
Strata 1: <one>
SU 1: <observations>
FPC 1: <zero>
Suppose that we were interested in the population ratio of weight to height. Here we use total
to estimate the population totals of weight and height and the svy brr prefix to estimate their ratio
and variance; we use total instead of ratio (which is otherwise preferable here) to show how to
specify an expression when using svy: brr.
. svy brr WtoH = (_b[weight]/_b[height]): total weight height
(running total on estimation sample)
BRR replications (32)
1 2 3 4 5
................................
BRR results Number of obs = 10,351
Population size = 117,157,513
Replications = 32
Design df = 31
command: total weight height
WtoH: _b[weight]/_b[height]
BRR
Coef. Std. Err. t P>|t| [95% Conf. Interval]
WtoH .4268116 .0008904 479.36 0.000 .4249957 .4286276

The variance estimation methods that Stata uses are discussed in [SVY] variance estimation.
Subpopulation estimation involves computing point and variance estimates for part of the population.
This method is not the same as restricting the estimation sample to the collection of observations
within the subpopulation because variance estimation for survey data measures sample-to-sample
variability, assuming that the same survey design is used to collect the data. Use the subpop() option
of the svy prefix to perform subpopulation estimation, and use if and in only when you need to
make restrictions on the estimation sample; see [SVY] subpopulation estimation.
Example 11: Subpopulation estimation

Here we will use our svyset high school data to model the association between weight and
height in the subpopulation of male high school seniors. First, we describe the sex variable to
determine how to identify the males in the dataset. We then use label list to verify that the variable
label agrees with the value labels.
. describe sex
storage display value
variable name type format label variable label
sex byte %9.0g sex 1=male, 2=female

. label list sex
sex:
1 male
2 female
Here we generate a variable named male so that we can easily identify the male high school
seniors. We specified if !missing(sex); doing so will cause the generated male variable to contain
a missing value at each observation where the sex variable does. This is done on purpose (although it
is not necessary if sex is free of missing values) because missing values should not be misinterpreted
to imply female.
. generate male = sex == 1 if !missing(sex)
Now we specify subpop(male) as an option to the svy prefix in our model fit.
. svy, subpop(male): regress weight height
Subpop. no. obs = 1,938
Subpop. size = 3,848,021
Design df = 50
F( 1, 50) = 225.38
Prob > F = 0.0000
R-squared = 0.2347
Linearized
height .7632911 .0508432 15.01 0.000 .6611696 .8654127

_cons -168.6532 22.5708 -7.47 0.000 -213.988 -123.3184
Although the table of estimation results contains the same columns as earlier, svy reports some
extra subpopulation information in the header. Here the extra header information tells us that 1,938
of the 4,071 sampled high school seniors are male, and the estimated number of male high school
seniors in the population is 3,848,021.
Direct standardization is an estimation method that allows comparing rates that come from different
frequency distributions; see [SVY] direct standardization. In direct standardization, estimated rates
(means, proportions, and ratios) are adjusted according to the frequency distribution of a standard
population. The standard population is partitioned into categories, called standard strata. The stratum
frequencies for the standard population are called standard weights. In the standardizing frequency
distribution, the standard strata are most commonly identified by demographic information such as
age, sex, and ethnicity. The standardized rate estimate is the weighted sum of unadjusted rates, where
the weights are the relative frequencies taken from the standardizing frequency distribution. Direct
standardization is available with svy: mean, svy: proportion, and svy: ratio.
Example 12: Standardized rates

Table 3.12-6 of Korn and Graubard (1999, 156) contains enumerated data for two districts of
London for the years 1840–1841. The age variable identifies the age groups in 5-year increments,
bgliving contains the number of people living in the Bethnal Green district at the beginning of
1840, bgdeaths contains the number of people who died in Bethnal Green that year, hsliving
contains the number of people living in St. George’s Hanover Square at the beginning of 1840, and
hsdeaths contains the number of people who died in Hanover Square that year.
. use http://www.stata-press.com/data/r14/stdize, clear
. list, noobs sep(0) sum
age bgliving bgdeaths hsliving hsdeaths
0-5 10739 850 5738 463

5-10 9180 76 4591 55
10-15 8006 38 4148 28
15-20 7096 37 6168 36
20-25 6579 38 9440 68
25-30 5829 51 8675 78
30-35 5749 51 7513 64
35-40 4490 56 5091 78
40-45 4385 47 4930 85
45-50 2955 66 2883 66
50-55 2995 74 2711 77
55-60 1644 67 1275 55
60-65 1835 64 1469 61
65-70 1042 64 649 55
70-75 879 68 619 58
75-80 366 47 233 51
80-85 173 39 136 20
85-90 71 22 48 15
90-95 21 6 10 4
95-100 4 2 2 1
unknown 50 1 124 0
Sum 74088 1764 66453 1418
We can use svy: ratio to compute the death rates for each district in 1840. Because this
dataset is identified as census data, we will create an FPC variable that will contain a sampling
rate of 100%. This method will result in zero standard errors, which are interpreted to mean no
variability—appropriate because our point estimates came from the entire population.
. gen fpc = 1
. svyset, fpc(fpc)
pweight: <none>
VCE: linearized
Strata 1: <one>
FPC 1: fpc
. svy: ratio (Bethnal: bgdeaths/bgliving) (Hanover: hsdeaths/hsliving)
(running ratio on estimation sample)
Survey: Ratio estimation
Number of strata = 1 Number of obs = 21
Number of PSUs = 21 Population size = 21
Design df = 20
Bethnal: bgdeaths/bgliving
Hanover: hsdeaths/hsliving
Linearized
Ratio Std. Err. [95% Conf. Interval]
Bethnal .0238095 0 . .
Hanover .0213384 0 . .
Note: Zero standard errors because of 100% sampling rate

detected for FPC in the first stage.
The death rates are 2.38% for Bethnal Green and 2.13% for St. George’s Hanover Square. These
observed death rates are not really comparable because they come from two different age distributions.
We can standardize based on the age distribution from Bethnal Green. Here age identifies our standard
strata and bgliving contains the associated population sizes.
. svy: ratio (Bethnal: bgdeaths/bgliving) (Hanover: hsdeaths/hsliving),
> stdize(age) stdweight(bgliving)
N. of std strata = 21 Design df = 20
Linearized
Bethnal .0238095 0 . .
Hanover .0266409 0 . .

The standardized death rate for St. George’s Hanover Square, 2.66%, is larger than the death rate
for Bethnal Green.
Poststratification is a method for adjusting the sampling weights, usually to account for under-
represented groups in the population; see [SVY] poststratification. This method usually results in
decreasing bias because of nonresponse and underrepresented groups in the population. It also tends to
result in smaller variance estimates. Poststratification is available for all survey estimation commands
and is specified using svyset; see [SVY] svyset.
Example 13: Poststratified mean

Levy and Lemeshow (2008, sec. 6.6) give an example of poststratification by using simple survey
data from a veterinarian’s client list. The data in poststrata.dta were collected using simple
random sampling (SRS) without replacement. The totexp variable contains the total expenses to the
client, type identifies the cats and dogs, postwgt contains the poststratum sizes (450 for cats and
850 for dogs), and fpc contains the total number of clients (850 + 450 = 1300).
. use http://www.stata-press.com/data/r14/poststrata, clear
. svyset, poststrata(type) postweight(postwgt) fpc(fpc)
pweight: <none>
VCE: linearized
Poststrata: type
Postweight: postwgt
Strata 1: <one>
FPC 1: fpc
. svy: mean totexp
Number of PSUs = 50 Population size = 1,300
N. of poststrata = 2 Design df = 49
Linearized
totexp 40.11513 1.163498 37.77699 42.45327
The mean total expenses is $40.12 with a standard error of $1.16. In the following, we omit the
poststratification information from svyset, resulting in mean total expenses of $39.73 with standard
error $2.22. The difference between the mean estimates is explained by the facts that expenses tend
to be larger for dogs than for cats and that the dogs were slightly underrepresented in the sample
(850/1,300 ≈ 0.65 for the population; 32/50 = 0.64 for the sample). This reasoning also explains why
the variance estimate from the poststratified mean is smaller than the one that was not poststratified.
. svyset, fpc(fpc)
pweight: <none>
VCE: linearized
Strata 1: <one>
FPC 1: fpc
. svy: mean totexp
Design df = 49
Linearized
totexp 39.7254 2.221747 35.26063 44.19017

The ml command can be used to fit a model by the method of maximum likelihood. When the
svy option is specified, ml performs maximum pseudolikelihood, applying sampling weights and
design-based linearization automatically; see [R] ml and Gould, Pitblado, and Poi (2010).
Example 14
The ml command requires a program that computes likelihood values to perform maximum
likelihood. Here is a likelihood evaluator used in Gould, Pitblado, and Poi (2010) to fit linear
regression models using the likelihood from the normal distribution.
program mynormal_lf
version 14
args lnf mu lnsigma
quietly replace ‘lnf’ = ln(normalden($ML_y1,‘mu’,exp(‘lnsigma’)))
end
Back in example 5, we fit a linear regression model using the high school survey data. Here we
use ml and mynormal lf to fit the same survey regression model.
. ml model lf mynormal_lf (mu: weight = height) /lnsigma, svy
. ml max
initial: log pseudolikelihood = -<inf> (could not be evaluated)
feasible: log pseudolikelihood = -7.301e+08
rescale: log pseudolikelihood = -51944380
rescale eq: log pseudolikelihood = -47565331
Iteration 0: log pseudolikelihood = -47565331
Iteration 1: log pseudolikelihood = -41226725 (not concave)
Iteration 6: log pseudolikelihood = -39379181 (backed up)
Design df = 50
F( 1, 50) = 593.99
Prob > F = 0.0000
Linearized
mu
height .716311 .0293908 24.37 0.000 .6572778 .7753442
_cons -149.6181 12.57266 -11.90 0.000 -174.871 -124.3652
lnsigma
_cons 3.372153 .0180777 186.54 0.000 3.335843 3.408464
svymarkout is a programmer’s command that resets the values in a variable that identifies the
estimation sample, dropping observations for which any of the survey characteristic variables contain
missing values. This tool is most helpful for developing estimation commands that use ml to fit
models using maximum pseudolikelihood directly, instead of relying on the svy prefix.
Video examples
Basic introduction to the analysis of complex survey data in Stata
How to download, import, and merge multiple datasets from the NHANES website
How to download, import, and prepare data from the NHANES website
Acknowledgments
Many of the svy commands were developed in collaboration with John L. Eltinge of the Bureau
of Labor Statistics. We thank him for his invaluable assistance.
We thank Wayne Johnson of the National Center for Health Statistics for providing the NHANES II
dataset.
We thank Nicholas Winter of the Politics Department at the University of Virginia for his diligent
efforts to keep Stata up to date with mainstream variance estimation methods for survey data, as well
as for providing versions of svy brr and svy jackknife.

William Gemmell Cochran (1909–1980) was born in Rutherglen, Scotland, and educated at the
Universities of Glasgow and Cambridge. He accepted a post at Rothamsted before finishing his
doctorate. Cochran emigrated to the United States in 1939 and worked at Iowa State, North
Carolina State, Johns Hopkins, and Harvard. He made many major contributions across several
fields of statistics, including experimental design, the analysis of counted data, sample surveys,
and observational studies, and was author or coauthor (with Gertrude M. Cox and George W.
Snedecor) of various widely used texts.
Leslie Kish (1910–2000) was born in Poprad, Hungary, and entered the United States with his
family in 1926. He worked as a lab assistant at the Rockefeller Institute for Medical Research and
studied at the College of the City of New York, fighting in the Spanish Civil War before receiving
his first degree in mathematics. Kish worked for the Bureau of the Census, the Department of
Agriculture, the Army Air Corps, and the University of Michigan. He carried out pioneering
work in the theory and practice of survey sampling, including design effects, BRR, response
errors, rolling samples and censuses, controlled selection, multipurpose designs, and small-area
estimation.

References
Cochran, W. G. 1977. Sampling Techniques. 3rd ed. New York: Wiley.
Cox, C. S., M. E. Mussolino, S. T. Rothwell, M. A. Lane, C. D. Golden, J. H. Madans, and J. J. Feldman. 1997.
Plan and operation of the NHANES I Epidemiologic Followup Study, 1992. In Vital and Health Statistics, series 1,
no. 35. Hyattsville, MD: National Center for Health Statistics.
Engel, A., R. S. Murphy, K. Maurer, and E. Collins. 1978. Plan and operation of the HANES I augmentation survey
of adults 25–74 years: United States 1974–75. In Vital and Health Statistics, series 1, no. 14. Hyattsville, MD:
National Center for Health Statistics.
Fay, R. E., and G. F. Train. 1995. Aspects of survey and model-based postcensal estimation of income and poverty
characteristics for states and counties. In Proceedings of the Government Statistics Section, 154–159. American
Statistical Association.
Gould, W. W., J. S. Pitblado, and B. P. Poi. 2010. Maximum Likelihood Estimation with Stata. 4th ed. College
Station, TX: Stata Press.
Heeringa, S. G., B. T. West, and P. A. Berglund. 2010. Applied Survey Data Analysis. Boca Raton, FL: Chapman
& Hall/CRC.
Kish, L. 1965. Survey Sampling. New York: Wiley.
Korn, E. L., and B. I. Graubard. 1999. Analysis of Health Surveys. New York: Wiley.
Kreuter, F., and R. Valliant. 2007. A survey on survey statistics: What is done and can be done in Stata. Stata Journal
7: 1–21.
Levy, P. S., and S. A. Lemeshow. 2008. Sampling of Populations: Methods and Applications. 4th ed. Hoboken, NJ:
Wiley.
McCarthy, P. J. 1966. Replication: An approach to the analysis of data from complex surveys. In Vital and Health
Statistics, series 2. Hyattsville, MD: National Center for Health Statistics.
. 1969a. Pseudoreplication: Further evaluation and application of the balanced half-sample technique. In Vital
and Health Statistics, series 2. Hyattsville, MD: National Center for Health Statistics.
. 1969b. Pseudo-replication: Half-samples. Revue de l’Institut International de Statistique 37: 239–264.
McDowell, A., A. Engel, J. T. Massey, and K. Maurer. 1981. Plan and operation of the Second National Health and
Nutrition Examination Survey, 1976–1980. Vital and Health Statistics 1(15): 1–144.
Miller, H. W. 1973. Plan and operation of the Health and Nutrition Examination Survey: United States 1971–1973.
Hyattsville, MD: National Center for Health Statistics.
Scheaffer, R. L., W. Mendenhall, III, R. L. Ott, and K. G. Gerow. 2012. Elementary Survey Sampling. 7th ed.
Boston: Brooks/Cole.
Shao, J., and D. Tu. 1995. The Jackknife and Bootstrap. New York: Springer.
Skinner, C. J., D. Holt, and T. M. F. Smith, ed. 1989. Analysis of Complex Surveys. New York: Wiley.
Stuart, A. 1984. The Ideas of Sampling. 3rd ed. New York: Griffin.
Thompson, S. K. 2012. Sampling. 3rd ed. Hoboken, NJ: Wiley.
Williams, B. 1978. A Sampler on Sampling. New York: Wiley.
Wolter, K. M. 2007. Introduction to Variance Estimation. 2nd ed. New York: Springer.
Also see
[SVY] svyset — Declare survey design for dataset
[SVY] svy — The survey prefix command
[SVY] svy estimation — Estimation commands for survey data
[P] robust — Robust variance estimates
Title
bootstrap options — More options for bootstrap variance estimation
Description Syntax Options Also see
Description
svy accepts more options when performing bootstrap variance estimation. See [SVY] svy bootstrap
for a complete discussion.
Syntax
bootstrap options Description
SE
mse use MSE formula for variance
nodots suppress replication dots
bsn(#) bootstrap mean-weight adjustment
saving( filename, . . .) save results to filename
verbose display the full table legend
noisily display any output from command
trace trace command
title(text) use text as the title for results
nodrop do not drop observations
reject(exp) identify invalid results
saving, verbose, noisily, trace, title(), nodrop, and reject() are not shown in the dialog boxes for estimation
commands.
Options

SE
mse specifies that svy compute the variance by using deviations of the replicates from the observed
value of the statistics based on the entire dataset. By default, svy computes the variance by using
deviations of the replicates from their mean.
nodots suppresses display of the replication dots. By default, one dot character is printed for each
successful replication. A red ‘x’ is displayed if command returns with an error, and ‘e’ is displayed
if at least one of the values in the exp list is missing.
bsn(#) specifies that # bootstrap replicate-weight variables were used to generate each bootstrap
mean-weight variable specified in the bsrweight() option of svyset. The bsn() option of
bootstrap overrides the bsn() option of svyset; see [SVY] svyset.
saving(), verbose, noisily, trace, title(), nodrop, reject(); see [SVY] svy bootstrap.
Also see
[SVY] svy bootstrap — Bootstrap for survey data
23
Title
brr options — More options for BRR variance estimation

Description
svy accepts more options when performing BRR variance estimation. See [SVY] svy brr for a
complete discussion.
Syntax
brr options Description
SE
hadamard(matrix) Hadamard matrix
fay(#) Fay’s adjustment
trace trace command
saving(), verbose, noisily, trace, title(), nodrop, and reject() are not shown in the dialog boxes for
estimation commands.
Options

SE
hadamard(matrix) specifies the Hadamard matrix to be used to determine which PSUs are chosen
for each replicate.
fay(#) specifies Fay’s adjustment. This option overrides the fay(#) option of svyset; see [SVY] svy-
set.
saving(), verbose, noisily, trace, title(), nodrop, reject(); see [SVY] svy brr.
Also see
[SVY] svy brr — Balanced repeated replication for survey data
24
Title
direct standardization — Direct standardization of means, proportions, and ratios
Description Remarks and examples Methods and formulas

References Also see
Description
Direct standardization is an estimation method that allows comparing rates that come from
different frequency distributions. The mean, proportion, and ratio commands can estimate means,
proportions, and ratios by using direct standardization.
See [SVY] poststratification for a similar estimation method given population sizes for strata not
used in the sampling design.

In direct standardization, estimated rates (means, proportions, and ratios) are adjusted according
to the frequency distribution of a standard population. The standard population is partitioned into
categories, called standard strata. The stratum frequencies for the standard population are called
standard weights. In the standardizing frequency distribution, the standard strata are most commonly
identified by demographic information such as age, sex, and ethnicity.
Stata’s mean, proportion, and ratio estimation commands have options for estimating means,
proportions, and ratios by using direct standardization. The stdize() option takes a variable that
identifies the standard strata, and the stdweight() option takes a variable that contains the standard
weights.
The standard strata (specified using stdize()) from the standardizing population are not the same
as the strata (specified using svyset’s strata() option) from the sampling design. In the output
header, “Number of strata” is the number of strata in the first stage of the sampling design, and
“N. of std strata” is the number of standard strata.
In the following example, we use direct standardization to compare the death rates between two
districts of London in 1840.
Example 1: Standardized rates

Table 3.12-6 of Korn and Graubard (1999, 156) contains enumerated data for two districts of
London for the years 1840–1841. The age variable identifies the age groups in 5-year increments,
bgliving contains the number of people living in the Bethnal Green district at the beginning of
1840, bgdeaths contains the number of people who died in Bethnal Green that year, hsliving
contains the number of people living in St. George’s Hanover Square at the beginning of 1840, and
hsdeaths contains the number of people who died in Hanover Square that year.
25
26 direct standardization — Direct standardization of means, proportions, and ratios
. use http://www.stata-press.com/data/r14/stdize
. list, noobs sep(0) sum
age bgliving bgdeaths hsliving hsdeaths
0-5 10739 850 5738 463

5-10 9180 76 4591 55
10-15 8006 38 4148 28
15-20 7096 37 6168 36
20-25 6579 38 9440 68
25-30 5829 51 8675 78
30-35 5749 51 7513 64
35-40 4490 56 5091 78
40-45 4385 47 4930 85
45-50 2955 66 2883 66
50-55 2995 74 2711 77
55-60 1644 67 1275 55
60-65 1835 64 1469 61
65-70 1042 64 649 55
70-75 879 68 619 58
75-80 366 47 233 51
80-85 173 39 136 20
85-90 71 22 48 15
90-95 21 6 10 4
95-100 4 2 2 1
unknown 50 1 124 0
Sum 74088 1764 66453 1418
We can use svy: ratio to compute the death rates for each district in 1840. Because this
dataset is identified as census data, we will create an FPC variable that will contain a sampling
rate of 100%. This method will result in zero standard errors, which are interpreted to mean no
variability—appropriate because our point estimates came from the entire population.
. gen fpc = 1
. svyset, fpc(fpc)
pweight: <none>
VCE: linearized
Strata 1: <one>
FPC 1: fpc
. svy: ratio (Bethnal: bgdeaths/bgliving) (Hanover: hsdeaths/hsliving)
Design df = 20
Linearized
Bethnal .0238095 0 . .
Hanover .0213384 0 . .

direct standardization — Direct standardization of means, proportions, and ratios 27
The death rates are 2.38% for Bethnal Green and 2.13% for St. George’s Hanover Square. These
observed death rates are not really comparable because they come from two different age distributions.
We can standardize based on the age distribution from Bethnal Green. Here age identifies our standard
strata and bgliving contains the associated population sizes.
. svy: ratio (Bethnal: bgdeaths/bgliving) (Hanover: hsdeaths/hsliving),
> stdize(age) stdweight(bgliving)
N. of std strata = 21 Design df = 20
Linearized
Bethnal .0238095 0 . .
Hanover .0266409 0 . .

The standardized death rate for St. George’s Hanover Square, 2.66%, is larger than the death rate
for Bethnal Green.
For this example, we could have used dstdize to compute the death rates; however, dstdize will
not compute the correct standard errors for survey data. Furthermore, dstdize is not an estimation
command, so test and the other postestimation commands are not available.
Technical note
The values in the variable supplied to the stdweight() option are normalized so that (1) is true;
see Methods and formulas. Thus the stdweight() variable can contain either population sizes or
population proportions for the associated standard strata.
Methods and formulas

The following discussion assumes that you are already familiar with the topics discussed in
[SVY] variance estimation.
In direct standardization, a weighted sum of the point estimates from the standard strata is used to
produce an overall point estimate for the population. This section will show how direct standardization
affects the ratio estimator. The mean and proportion estimators are special cases of the ratio estimator.
Suppose that you used a complex survey design to sample m individuals from a population of
size M . Let Dg be the set of individuals in the sample that belong to the g th standard stratum, and
let IDg (j) indicate if the j th individual is in standard stratum g , where
(
1, if j ∈ Dg
IDg (j) =
0, otherwise
28 direct standardization — Direct standardization of means, proportions, and ratios
Also let LD be the number of standard strata, and let πg be the proportion of the population that
belongs to standard stratum g .
LD
X
πg = 1 (1)
g=1
In subpopulation estimation, πg is set to zero if none of the individuals in standard stratum g are in
the subpopulation. Then the standard stratum proportions are renormalized.
Let yj and xj be the items of interest and wj be the sampling weight for the j th sampled individual.
The estimator for the standardized ratio of R = Y /X is
LD
bD =
X Ybg
R πg
g=1 Xbg
where
m
X
Ybg = IDg (j) wj yj
j=1
with X
bg similarly defined.
For replication-based variance estimation, replicates of the standardized values are used in the
variance formulas.
The score variable for the linearized variance estimator of the standardized ratio is
LD bg yj − Ybg xj
bD ) =
X X
zj (R πg IDg (j)
Xb2
g=1 g
This score variable was derived using the method described in [SVY] variance estimation and is a
direct result of the methods described in Deville (1999), Demnati and Rao (2004), and Shah (2004).
For the mean and proportion commands, the mean estimator is a ratio estimator with the
denominator variable equal to one (xj = 1) and the proportion estimator is the mean estimator with
an indicator variable in the numerator (yj ∈ {0, 1}).
References
Demnati, A., and J. N. K. Rao. 2004. Linearization variance estimators for survey data. Survey Methodology 30:
17–26.
Deville, J.-C. 1999. Variance estimation for complex statistics and estimators: Linearization and residual techniques.
Survey Methodology 25: 193–203.
Shah, B. V. 2004. Comment [on Demnati and Rao (2004)]. Survey Methodology 30: 29.
Also see
[SVY] poststratification — Poststratification for survey data
[SVY] survey — Introduction to survey commands
Title
estat — Postestimation statistics for survey data
Description Quick start Menu Syntax

Options Remarks and examples Stored results Methods and formulas
References Also see
Description
estat svyset reports the survey design characteristics associated with the current estimation
results.
estat effects displays a table of design and misspecification effects for each estimated parameter.
estat lceffects displays a table of design and misspecification effects for a user-specified linear
combination of the parameter estimates.
estat size displays a table of sample and subpopulation sizes for each estimated subpopulation
mean, proportion, ratio, or total. This command is available only after svy: mean, svy: proportion,
svy: ratio, and svy: total; see [R] mean, [R] proportion, [R] ratio, and [R] total.
estat sd reports subpopulation standard deviations based on the estimation results from mean
and svy: mean; see [R] mean. estat sd is not appropriate with estimation results that used direct
standardization or poststratification.
estat strata displays a table of the number of singleton and certainty strata within each
sampling stage. The variance scaling factors are also displayed for estimation results where
singleunit(scaled) was svyset.
estat cv reports the coefficient of variation (CV) for each coefficient in the current estimation
results. The CV for coefficient b is
SE(b)
CV(b) = × 100%
|b|
estat gof reports a goodness-of-fit test for binary response models using survey data. This
command is available only after svy: logistic, svy: logit, and svy: probit; see [R] logistic,
[R] logit, and [R] probit.
estat vce displays the covariance or correlation matrix of the parameter estimates of the previous
model. See [R] estat vce for examples.
Quick start
Design effects for each parameter in current estimation results after a command using the svy: prefix
estat effects
Design effects for the sum of parameter estimates for variables v1 and v2
estat lceffects v1 + v2
As above, but add misspecification effects
estat lceffects v1 + v2, deff deft meff meft
29
30 estat — Postestimation statistics for survey data
Number of observations used and subpopulation size for each parameter

estat size
Estimate of subpopulation standard deviation based on estimation results from svy: mean
estat sd
Compute standard deviation using an estimate of SRS variance for sampling within a subpopulation
estat sd, srssubpop
Display the number of singleton and certainty strata within each sampling stage
estat strata
Coefficient of variation for each parameter in current estimation results
estat cv
Goodness-of-fit test for binary response models using survey data and grouping data into quintiles
estat gof, group(5)
Variance–covariance matrix of parameter estimates from the most recent model
estat vce
As above, but display a correlation matrix
estat vce, correlation
Menu
Statistics > Survey data analysis > DEFF, MEFF, and other statistics
estat — Postestimation statistics for survey data 31
Syntax
Survey design characteristics
estat svyset
Design and misspecification effects for point estimates

estat effects , estat effects options
Design and misspecification effects for linear combinations of point estimates

estat lceffects exp , estat lceffects options
Subpopulation sizes

estat size , estat size options
Subpopulation standard-deviation estimates

estat sd , estat sd options
Singleton and certainty strata

estat strata
Coefficient of variation for survey data

estat cv , estat cv options
Goodness-of-fit test for binary response models using survey data

estat gof if in , estat gof options
Display covariance matrix estimates

estat vce , estat vce options
estat effects options Description

deff report DEFF design effects
deft report DEFT design effects
srssubpop report design effects, assuming SRS within subpopulation
meff report MEFF design effects
meft report MEFT design effects
display options control spacing and display of omitted variables and base and
empty cells
estat lceffects options Description

deff report DEFF design effects
deft report DEFT design effects
srssubpop report design effects, assuming SRS within subpopulation
meff report MEFF design effects
meft report MEFT design effects
estat size options Description

obs report number of observations (within subpopulation)
size report subpopulation sizes
estat sd options Description

variance report subpopulation variances instead of standard deviations
srssubpop report standard deviation, assuming SRS within subpopulation
estat cv options Description

nolegend suppress the table legend
display options control spacing and display of omitted variables and base and
empty cells
estat gof options Description

group(#) compute test statistic using # quantiles
total compute test statistic using the total estimator instead of the mean
estimator
all execute test for all observations in the data
estat vce options Description

covariance display as covariance matrix; the default
correlation display as correlation matrix
equation(spec) display only specified equations
block display submatrices by equation
diag display submatrices by equation; diagonal blocks only
format(% fmt) display format for covariances and correlations
nolines suppress lines between equations
display options control display of omitted variables and base and empty cells
Options
Options are presented under the following headings:
Options for estat effects
Options for estat lceffects
Options for estat size
Options for estat sd
Options for estat cv
Options for estat gof
Options for estat vce
Options for estat effects

deff and deft request that the design-effect measures DEFF and DEFT be displayed. This is the
default, unless direct standardization or poststratification was used.
The deff and deft options are not allowed with estimation results that used direct standardization
or poststratification. These methods obscure the measure of design effect because they adjust the
frequency distribution of the target population.
srssubpop requests that DEFF and DEFT be computed using an estimate of simple random sampling
(SRS) variance for sampling within a subpopulation. By default, DEFF and DEFT are computed using
an estimate of the SRS variance for sampling from the entire population. Typically, srssubpop is
used when computing subpopulation estimates by strata or by groups of strata.
meff and meft request that the misspecification-effect measures MEFF and MEFT be displayed.
display options: noomitted, vsquish, noemptycells, baselevels, allbaselevels; see [R] es-
timation options.
Options for estat lceffects

deff and deft request that the design-effect measures DEFF and DEFT be displayed. This is the
default, unless direct standardization or poststratification was used.
or poststratification. These methods obscure the measure of design effect because they adjust the
frequency distribution of the target population.
srssubpop requests that DEFF and DEFT be computed using an estimate of simple random sampling
(SRS) variance for sampling within a subpopulation. By default, DEFF and DEFT are computed using
an estimate of the SRS variance for sampling from the entire population. Typically, srssubpop is
used when computing subpopulation estimates by strata or by groups of strata.
meff and meft request that the misspecification-effect measures MEFF and MEFT be displayed.
Options for estat size

obs requests that the number of observations used to compute the estimate be displayed for each row
of estimates.
size requests that the estimate of the subpopulation size be displayed for each row of estimates. The
subpopulation size estimate equals the sum of the weights for those observations in the estimation
sample that are also in the specified subpopulation. The estimated population size is reported when
a subpopulation is not specified.
Options for estat sd

variance requests that the subpopulation variance be displayed instead of the standard deviation.
srssubpop requests that the standard deviation be computed using an estimate of SRS variance for
sampling within a subpopulation. By default, the standard deviation is computed using an estimate
of the SRS variance for sampling from the entire population. Typically, srssubpop is given when
computing subpopulation estimates by strata or by groups of strata.
Options for estat cv

nolegend prevents the table legend identifying the subpopulations from being displayed.
display options: noomitted, vsquish, noemptycells, baselevels, allbaselevels; see [R] es-
timation options.
Options for estat gof

group(#) specifies the number of quantiles to be used to group the data for the goodness-of-fit test.
The minimum allowed value is group(2). The maximum allowed value is group(df ), where df
is the design degrees of freedom (e(df r)). The default is group(10).
total requests that the goodness-of-fit test statistic be computed using the total estimator instead of
the mean estimator.
all requests that the goodness-of-fit test statistic be computed for all observations in the data, ignoring
any if or in restrictions specified with the model fit.
Options for estat vce

covariance displays the matrix as a variance–covariance matrix; this is the default.
correlation displays the matrix as a correlation matrix rather than a variance–covariance matrix.
rho is a synonym for correlation.
equation(spec) selects the part of the VCE to be displayed. If spec is eqlist, the VCE for the listed
equations is displayed. If spec is eqlist1 \ eqlist2, the part of the VCE associated with the equations
in eqlist1 (rowwise) and eqlist2 (columnwise) is displayed. If spec is *, all equations are displayed.
equation() implies block if diag is not specified.
block displays the submatrices pertaining to distinct equations separately.
diag displays the diagonal submatrices pertaining to distinct equations separately.
format(% fmt) specifies the number format for displaying the elements of the matrix. The default is
format(%10.0g) for covariances and format(%8.4f) for correlations. See [U] 12.5 Formats:
Controlling how data are displayed for more information.
nolines suppresses lines between equations.
display options: noomitted, noemptycells, baselevels, allbaselevels; see [R] estimation
options.

Example 1
Using data from the Second National Health and Nutrition Examination Survey (NHANES II)
(McDowell et al. 1981), let’s estimate the population means for total serum cholesterol (tcresult)
and for serum triglycerides (tgresult).
. svy: mean tcresult tgresult
Design df = 31
Linearized
tcresult 211.3975 1.252274 208.8435 213.9515

tgresult 138.576 2.071934 134.3503 142.8018
We can use estat svyset to remind us of the survey design characteristics that were used to
produce these results.
. estat svyset
pweight: finalwgt
VCE: linearized
Strata 1: strata
SU 1: psu
FPC 1: <zero>
estat effects reports a table of design and misspecification effects for each mean we estimated.
. estat effects, deff deft meff meft
Linearized
Mean Std. Err. DEFF DEFT MEFF MEFT
tcresult 211.3975 1.252274 3.57141 1.88982 3.46105 1.86039

tgresult 138.576 2.071934 2.35697 1.53524 2.32821 1.52585
estat size reports a table that contains sample and population sizes.
. estat size
Linearized
Mean Std. Err. Obs Size
tcresult 211.3975 1.252274 5,050 56,820,832

tgresult 138.576 2.071934 5,050 56,820,832
estat size can also report a table of subpopulation sizes.

. svy: mean tcresult, over(sex)
(output omitted )
. estat size
Male: sex = Male
Female: sex = Female
Linearized
Over Mean Std. Err. Obs Size
tcresult
Male 210.7937 1.312967 4,915 56,159,480
Female 215.2188 1.193853 5,436 60,998,033
estat sd reports a table of subpopulation standard deviations.

. estat sd
Male: sex = Male
Over Mean Std. Dev.
tcresult
Male 210.7937 45.79065
Female 215.2188 50.72563
estat cv reports a table of coefficients of variations for the estimates.

. estat cv
Male: sex = Male
Linearized
Over Mean Std. Err. CV (%)
tcresult
Male 210.7937 1.312967 .622868
Female 215.2188 1.193853 .554716
Example 2: Design effects with subpopulations

When there are subpopulations, estat effects can compute design effects with respect to one
of two different hypothetical SRS designs. The default design is one in which SRS is conducted
across the full population. The alternate design is one in which SRS is conducted entirely within the
subpopulation of interest. This alternate design is used when the srssubpop option is specified.
Deciding which design is preferable depends on the nature of the subpopulations. If we can imagine
identifying members of the subpopulations before sampling them, the alternate design is preferable.
This case arises primarily when the subpopulations are strata or groups of strata. Otherwise, we may
prefer to use the default.
Here is an example using the default with the NHANES II data.

. svy: mean iron, over(sex)
(output omitted )
. estat effects
Male: sex = Male
Linearized
iron
Male 104.7969 .557267 1.36097 1.16661
Female 97.16247 .6743344 2.01403 1.41916
Thus the design-based variance estimate is about 36% larger than the estimate from the hypothetical
SRS design including the full population. We can get DEFF and DEFT for the alternate SRS design by
using the srssubpop option.
. estat effects, srssubpop
Male: sex = Male
Linearized
iron
Male 104.7969 .557267 1.348 1.16104
Female 97.16247 .6743344 2.03132 1.42524
Because the NHANES II did not stratify on sex, we think it problematic to consider design effects
with respect to SRS of the female (or male) subpopulation. Consequently, we would prefer to use the
default here, although the values of DEFF differ little between the two in this case.
For other examples (generally involving heavy oversampling or undersampling of specified sub-
populations), the differences in DEFF for the two schemes can be much more dramatic.
Consider the NMIHS data (Gonzalez, Krauss, and Scott 1992), and compute the mean of birthwgt
over race:
. use http://www.stata-press.com/data/r14/nmihs
. svy: mean birthwgt, over(race)
(output omitted )
. estat effects
nonblack: race = nonblack
black: race = black
Linearized
birthwgt
nonblack 3402.32 7.609532 1.44376 1.20157
black 3127.834 6.529814 .172041 .414778
. estat effects, srssubpop

black: race = black
Linearized
birthwgt
nonblack 3402.32 7.609532 .826842 .909308
black 3127.834 6.529814 .528963 .727298
Because the NMIHS survey was stratified on race, marital status, age, and birthweight, we believe it
reasonable to consider design effects computed with respect to SRS within an individual race group.
Consequently, we would recommend here the alternative hypothetical design for computing design
effects; that is, we would use the srssubpop option.
Example 3: Misspecification effects

Misspecification effects assess biases in variance estimators that are computed under the wrong
assumptions. The survey literature (for example, Scott and Holt 1982, 850; Skinner 1989) defines
misspecification effects with respect to a general set of “wrong” variance estimators. estat effects
considers only one specific form: variance estimators computed under the incorrect assumption that
our observed sample was selected through SRS.
The resulting “misspecification effect” measure is informative primarily when an unweighted point
estimator is approximately unbiased for the parameter of interest. See Eltinge and Sribney (1996a)
for a detailed discussion of extensions of misspecification effects that are appropriate for biased point
estimators.
Note the difference between a misspecification effect and a design effect. For a design effect, we
compare our complex-design–based variance estimate with an estimate of the true variance that we
would have obtained under a hypothetical true simple random sample. For a misspecification effect,
we compare our complex-design–based variance estimate with an estimate of the variance from fitting
the same model without weighting, clustering, or stratification.
estat effects defines MEFF and MEFT as
MEFF = Vb /Vbmsp
√
MEFT = MEFF
where Vb is the appropriate design-based estimate of variance and Vbmsp is the variance estimate
computed with a misspecified design—ignoring the sampling weights, stratification, and clustering.
Here we request that the misspecification effects be displayed for the estimation of mean zinc
levels from our NHANES II data.
. svy: mean zinc, over(sex)
(output omitted )
. estat effects, meff meft
Male: sex = Male
Linearized
Over Mean Std. Err. MEFF MEFT
zinc
Male 90.74543 .5850741 6.28254 2.5065
Female 83.8635 .4689532 6.32648 2.51525
If we run ci without weights, we get the standard errors that are (Vbmsp )1/2 .
. sort sex
. ci zinc if sex == "Male":sex
Variable Obs Mean Std. Err. [95% Conf. Interval]
zinc 4,375 89.53143 .2334228 89.0738 89.98906

. display [zinc]_se[Male]/r(se)
2.5064994
. display ([zinc]_se[Male]/r(se))^2
6.2825393
. ci zinc if sex == "Female":sex
Variable Obs Mean Std. Err. [95% Conf. Interval]
zinc 4,827 83.76652 .186444 83.40101 84.13204

. display [zinc]_se[Female]/r(se)
2.515249
. display ([zinc]_se[Female]/r(se))^2
6.3264774
Example 4: Design and misspecification effects for linear combinations

Let’s compare the mean of total serum cholesterol (tcresult) between men and women in the
NHANES II dataset.
. svy: mean tcresult, over(sex)
Design df = 31
Male: sex = Male
Linearized
tcresult
Male 210.7937 1.312967 208.1159 213.4715
Female 215.2188 1.193853 212.784 217.6537
We can use estat lceffects to report the standard error, design effects, and misspecification effects
of the difference between the above means.
. estat lceffects [tcresult]Male - [tcresult]Female, deff deft meff meft
( 1) [tcresult]Male - [tcresult]Female = 0
Mean Coef. Std. Err. DEFF DEFT MEFF MEFT
(1) -4.425109 1.086786 1.31241 1.1456 1.27473 1.12904
Example 5: Using survey data to determine Neyman allocation

Suppose that we have partitioned our population into L strata and stratum h contains Nh individuals.
Also let σh represent the standard deviation of a quantity we wish to sample from the population.
According to Cochran (1977, sec. 5.5), we can minimize the variance of the stratified mean estimator,
for a fixed sample size n, if we choose the stratum sample sizes according to Neyman allocation:
Nh σ h
nh = n PL (1)
i=1 Ni σi
We can use estat sd with our current survey data to produce a table of subpopulation standard-
deviation estimates. Then we could plug these estimates into (1) to improve our survey design for
the next time we sample from our population.
Here is an example using birthweight from the NMIHS data. First, we need estimation results from
svy: mean over the strata.
. svyset [pw=finwgt], strata(stratan)
pweight: finwgt
VCE: linearized
Strata 1: stratan
FPC 1: <zero>
. svy: mean birthwgt, over(stratan)
(output omitted )
Next we will use estat size to report the table of stratum sizes. We will also generate matrix
p obs to contain the observed percent allocations for each stratum. In the matrix expression, r( N)
is a row vector of stratum sample sizes and e(N) contains the total sample size. r( N subp) is a
row vector of the estimated population stratum sizes.
. estat size
1: stratan = 1
2: stratan = 2
3: stratan = 3
4: stratan = 4
5: stratan = 5
6: stratan = 6
Linearized
Over Mean Std. Err. Obs Size
birthwgt
1 1049.434 19.00149 841 18,402.98161
2 2189.561 9.162736 803 67,650.95932
3 3303.492 7.38429 3,578 579,104.6188
4 1036.626 12.32294 710 29,814.93215
5 2211.217 9.864682 714 153,379.07445
6 3485.42 8.057648 3,300 3,047,209.105
. matrix p_obs = 100 * r(_N)/e(N)

. matrix nsubp = r(_N_subp)
Now we call estat sd to report the stratum standard-deviation estimates and generate matrix
p neyman to contain the percent allocations according to (1). In the matrix expression, r(sd) is a
vector of the stratum standard deviations.
. estat sd
1: stratan = 1
2: stratan = 2
3: stratan = 3
4: stratan = 4
5: stratan = 5
6: stratan = 6
Over Mean Std. Dev.
birthwgt
1 1049.434 2305.931
2 2189.561 555.7971
3 3303.492 687.3575
4 1036.626 999.0867
5 2211.217 349.8068
6 3485.42 300.6945
. matrix p_neyman = 100 * hadamard(nsubp,r(sd))/el(nsubp*r(sd)’,1,1)

. matrix list p_obs, format(%4.1f)
p_obs[1,6]
birthwgt: birthwgt: birthwgt: birthwgt: birthwgt: birthwgt:
1 2 3 4 5 6
r1 8.5 8.1 36.0 7.1 7.2 33.2
. matrix list p_neyman, format(%4.1f)
p_neyman[1,6]
birthwgt: birthwgt: birthwgt: birthwgt: birthwgt: birthwgt:
1 2 3 4 5 6
r1 2.9 2.5 26.9 2.0 3.6 62.0
We can see that strata 3 and 6 each contain about one-third of the observed data, with the rest of
the observations spread out roughly equally to the remaining strata. However, plugging our sample
estimates into (1) indicates that stratum 6 should get 62% of the sampling units, stratum 3 should
get about 27%, and the remaining strata should get a roughly equal distribution of sampling units.
Example 6: Summarizing singleton and certainty strata

Use estat strata with svy estimation results to produce a table that reports the number of
singleton and certainty strata in each sampling stage. Here is an example using (fictional) data from
a complex survey with five sampling stages (the dataset is already svyset). If singleton strata are
present, estat strata will report their effect on the standard errors.
. use http://www.stata-press.com/data/r14/strata5
. svy: total y
(output omitted )
. estat strata
Singleton Certainty Total

Stage strata strata strata
1 0 1 4
2 1 0 10
3 0 3 29
4 2 0 110
5 204 311 865
Note: Missing standard error because of

stratum with single sampling unit.
estat strata also reports the scale factor used when the singleunit(scaled) option is
svyset. Of the 865 strata in the last stage, 204 are singleton strata and 311 are certainty strata. Thus
the scaling factor for the last stage is
865 − 311
≈ 1.58
865 − 311 − 204
. svyset, singleunit(scaled) noclear

(output omitted )
. svy: total y
(output omitted )
. estat strata
Singleton Certainty Total Scale

Stage strata strata strata factor
1 0 1 4 1
2 1 0 10 1.11
3 0 3 29 1
4 2 0 110 1.02
5 204 311 865 1.58
Note: Variances scaled within each stage to handle

strata with a single sampling unit.
The singleunit(scaled) option of svyset is one of three methods in which Stata’s svy commands
can automatically handle singleton strata when performing variance estimation; see [SVY] variance
estimation for a brief discussion of these methods.
Example 7: Goodness-of-fit test for svy: logistic

From example 2 in [SVY] svy estimation, we modeled the incidence of high blood pressure as a
function of height, weight, age, and sex (using the female indicator variable).
. use http://www.stata-press.com/data/r14/nhanes2d
. svyset
pweight: finalwgt
VCE: linearized
Strata 1: strata
SU 1: psu
FPC 1: <zero>
. svy: logistic highbp height weight age female
(running logistic on estimation sample)
Survey: Logistic regression
Design df = 31
F( 4, 28) = 368.33
Prob > F = 0.0000
Linearized
highbp Odds Ratio Std. Err. t P>|t| [95% Conf. Interval]
height .9657022 .0051511 -6.54 0.000 .9552534 .9762654

weight 1.053023 .0026902 20.22 0.000 1.047551 1.058524
age 1.050059 .0019761 25.96 0.000 1.046037 1.054097
female .6272129 .0368195 -7.95 0.000 .5564402 .706987
_cons .716868 .6106878 -0.39 0.699 .1261491 4.073749
We can use estat gof to perform a goodness-of-fit test for this model.
. estat gof
Logistic model for highbp, goodness-of-fit test
F(9,23) = 5.32
Prob > F = 0.0006
The F statistic is significant at the 5% level, indicating that the model is not a good fit for these data.
Stored results
estat svyset stores the following in r():
Scalars
r(stages) number of sampling stages
r(stages wt) last stage containing stage-level weights
Macros
r(wtype) weight type
r(wexp) weight expression
r(wvar) weight variable name
r(weight#) variable identifying weight for stage #
r(su#) variable identifying sampling units for stage #
r(strata#) variable identifying strata for stage #
r(fpc#) FPC for stage #
r(bsrweight) bsrweight() variable list
r(bsn) bootstrap mean-weight adjustment
r(brrweight) brrweight() variable list
r(fay) Fay’s adjustment
r(jkrweight) jkrweight() variable list
r(sdrweight) sdrweight() variable list
r(sdrfpc) fpc() value from within sdrweight()
r(vce) vcetype specified in vce()
r(dof) dof() value
r(mse) mse, if specified
r(poststrata) poststrata() variable
r(postweight) postweight() variable
r(settings) svyset arguments to reproduce the current settings
r(singleunit) singleunit() setting
estat strata stores the following in r():

Matrices
r( N strata single) number of strata with one sampling unit
r( N strata certain) number of certainty strata
r( N strata) number of strata
r(scale) variance scale factors used when singleunit(scaled) is svyset
estat effects stores the following in r():

Matrices
r(deff) vector of DEFF estimates
r(deft) vector of DEFT estimates
r(deffsub) vector of DEFF estimates for srssubpop
r(deftsub) vector of DEFT estimates for srssubpop
r(meff) vector of MEFF estimates
r(meft) vector of MEFT estimates
estat lceffects stores the following in r():

Scalars
r(estimate) point estimate
r(se) estimate of standard error
r(df) degrees of freedom
r(deff) DEFF estimate
r(deft) DEFT estimate
r(deffsub) DEFF estimate for srssubpop
r(deftsub) DEFT estimate for srssubpop
r(meff) MEFF estimate
r(meft) MEFT estimate
estat size stores the following in r():

Matrices
r( N) vector of numbers of nonmissing observations
r( N subp) vector of subpopulation size estimates
estat sd stores the following in r():

Macros
r(srssubpop) srssubpop, if specified
Matrices
r(mean) vector of subpopulation mean estimates
r(sd) vector of subpopulation standard-deviation estimates
r(variance) vector of subpopulation variance estimates
estat cv stores the following in r():

Matrices
r(b) estimates
r(se) standard errors of the estimates
r(cv) coefficients of variation of the estimates
estat gof stores the following in r():

Scalars
r(p) p-value associated with the test statistic
r(F) F statistic, if e(df r) was stored by estimation command
r(df1) numerator degrees of freedom for F statistic
r(df2) denominator degrees of freedom for F statistic
r(chi2) χ2 statistic, if e(df r) was not stored by estimation command
r(df) degrees of freedom for χ2 statistic
estat vce stores the following in r():

Matrices
r(V) VCE or correlation matrix

Methods and formulas are presented under the following headings:
Design effects
Linear combinations
Misspecification effects
Population and subpopulation standard deviations
Coefficient of variation
Goodness of fit for binary response models
Design effects
estat effects produces two estimators of design effect, DEFF and DEFT.
DEFF is estimated as described in Kish (1965) as
Vb (θ)
b
DEFF =
Vbsrswor (θesrs )
where Vb (θ)
b is the design-based estimate of variance for a parameter, θ, and Vbsrswor (θesrs ) is an
estimate of the variance for an estimator, θesrs , that would be obtained from a similar hypothetical
survey conducted using SRS without replacement (wor) and with the same number of sample elements,
m, as in the actual survey. For example, if θ is a total Y , then
m
M
c X 2
Vbsrswor (θesrs ) = (1 − f ) wj yj − Yb (1)
m − 1 j=1
where Yb = Yb /M c. The factor (1 − f ) is a finite population correction. If the user sets an FPC for
the first stage, f = m/Mc is used; otherwise, f = 0.
DEFT is estimated as described in Kish (1987, 41) as
s
Vb (θ)
b
DEFT =
Vbsrswr (θesrs )
where Vbsrswr (θesrs ) is an estimate of the variance for an estimator, θesrs , obtained from a similar survey
conducted using SRS with replacement (wr). Vbsrswr (θesrs ) is computed using (1) with f = 0.
When computing estimates for a subpopulation, S , and the srssubpop option is not specified
(that is, the default), (1) is used with wSj = IS (j) wj in place of wj , where
(
1, if j ∈ S
IS (j) =
0, otherwise
The sums in (1) are still calculated over all elements in the sample, regardless of whether they belong
to the subpopulation: by default, the SRS is assumed to be done across the full population.
When the srssubpop option is specified, the SRS is carried out within subpopulation S . Here
(1) is used with the sums restricted to those elements belonging to the subpopulation; m is replaced
with mS , the number of sample elements from the subpopulation; M c is replaced with M
cS , the sum
of the weights from the subpopulation; and Y is replaced with Y = Y /M , the weighted mean
b b b S
c S S
across the subpopulation.
Linear combinations
estat lceffects estimates η = Cθ, where θ is a q × 1 vector of parameters (for example,
population means or population regression coefficients) and C is any 1 × q vector of constants. The
estimate of η is ηb = C θb, and its variance estimate is
Vb (b b 0
η ) = C Vb (θ)C
Similarly, the SRS without replacement (srswor) variance estimator used in the computation of DEFF
is
ηsrs ) = C Vbsrswor (θbsrs )C 0

Vbsrswor (e
and the SRS with replacement (srswr) variance estimator used in the computation of DEFT is
ηsrs ) = C Vbsrswr (θbsrs )C 0

Vbsrswr (e
The variance estimator used in computing MEFF and MEFT is
ηmsp ) = C Vbmsp (θbmsp )C 0

Vbmsp (e
estat lceffects was originally developed under a different command name; see Eltinge and
Sribney (1996b).
Misspecification effects
estat effects produces two estimators of misspecification effect, MEFF and MEFT.
Vb (θ)
b
MEFF =
Vbmsp (θbmsp )
√
MEFT = MEFF
where Vb (θ)
b is the design-based estimate of variance for a parameter, θ, and Vbmsp (θbmsp ) is the variance
estimate for θbmsp . These estimators, θbmsp and Vbmsp (θbmsp ), are based on the incorrect assumption
that the observations were obtained through SRS with replacement: they are the estimators obtained
by simply ignoring weights, stratification, and clustering. When θ is a total Y , the estimator and its
variance estimate are computed using the standard formulas for an unweighted total:
m
cy = M
cX
Ybmsp = M yj
m j=1
c2 m
M X 2
Vbmsp (Ybmsp ) = yj − y
m(m − 1) j=1
When computing MEFF and MEFT for a subpopulation, sums are restricted to those elements
belonging to the subpopulation, and mS and M
cS are used in place of m and M
c.
Population and subpopulation standard deviations

For srswr designs, the variance of the mean estimator is
Vsrswr (y) = σ 2 /n
where n is the sample size and σ is the population standard deviation. estat sd uses this formula
and the results from mean and svy: mean to estimate the population standard deviation via
q
σ
b = n Vbsrswr (y)
Subpopulation standard deviations are computed similarly, using the corresponding variance estimate
and sample size.
Coefficient of variation
The coefficient of variation (CV) for estimate θb is
q
Vb (θ)
b
CV(θ)
b = × 100%
|θ|
b
A missing value is reported when θb is zero.
Goodness of fit for binary response models

Let yj be the j th observed value of the dependent variable, pbj be the predicted probability of a
positive outcome, and rbj = yj − pbj . Let g be the requested number of groups from the group()
option; then the rbj are placed in g quantile groups as described in Methods and formulas for the
xtile command in [D] pctile. Let r = (r1 , . . . , rg ), where ri is the subpopulation mean of the rbj
for the ith quantile group. The standard Wald statistic for testing H0 : r = 0 is
b 2 = r{Vb (r)}−1 r0
X
b 2 is approximately distributed as a
where Vb (r) is the design-based variance estimate for r. Here X
χ2 with g − 1 degrees of freedom. This Wald statistic is one of the three goodness-of-fit statistics
discussed in Graubard, Korn, and Midthune (1997). estat gof reports this statistic when the design
degrees of freedom is missing, such as with svy bootstrap results.
According to Archer and Lemeshow (2006), the F -adjusted mean residual test is given by
b 2 (d − g + 2)/(dg)
Fb = X
where d is the design degrees of freedom. Here Fb is approximately distributed as an F with g − 1

numerator and d − g + 2 denominator degrees of freedom.
With the total option, estat gof uses the subpopulation total estimator instead of the subpopulation
mean estimator.
References
Archer, K. J., and S. A. Lemeshow. 2006. Goodness-of-fit test for a logistic regression model fitted using survey
sample data. Stata Journal 6: 97–105.
Eltinge, J. L., and W. M. Sribney. 1996a. Accounting for point-estimation bias in assessment of misspecification
effects, confidence-set coverage rates and test sizes. Unpublished manuscript, Department of Statistics, Texas A&M
University.
. 1996b. svy5: Estimates of linear combinations and hypothesis tests for survey data. Stata Technical Bulletin
31: 31–42. Reprinted in Stata Technical Bulletin Reprints, vol. 6, pp. 246–259. College Station, TX: Stata Press.
Gonzalez, J. F., Jr., N. Krauss, and C. Scott. 1992. Estimation in the 1988 National Maternal and Infant Health
Survey. Proceedings of the Section on Statistics Education, American Statistical Association 343–348.
Graubard, B. I., E. L. Korn, and D. Midthune. 1997. Testing goodness-of-fit for logistic regression with survey data.
In Proceedings of the Section on Survey Research Methods, Joint Statistical Meetings, 170–174. Alexandria, VA:
American Statistical Association.
Kish, L. 1965. Survey Sampling. New York: Wiley.
. 1987. Statistical Design for Research. New York: Wiley.

Scott, A. J., and D. Holt. 1982. The effect of two-stage sampling on ordinary least squares methods. Journal of the
American Statistical Association 77: 848–854.
Skinner, C. J. 1989. Introduction to part A. In Analysis of Complex Surveys, ed. C. J. Skinner, D. Holt, and
T. M. F. Smith, 23–58. New York: Wiley.
West, B. T., and S. E. McCabe. 2012. Incorporating complex sample design effects when only final survey weights
are available. Stata Journal 12: 718–725.
Also see
[SVY] svy postestimation — Postestimation tools for svy
[SVY] subpopulation estimation — Subpopulation estimation for survey data
[SVY] variance estimation — Variance estimation for survey data
Title
jackknife options — More options for jackknife variance estimation
Description
svy accepts more options when performing jackknife variance estimation.
Syntax
jackknife options Description
SE
keep keep pseudovalues
trace trace command
saving(), keep, verbose, noisily, trace, title(), nodrop, and reject() are not shown in the dialog boxes
for estimation commands.
Options

SE
value of the statistic based on the entire dataset. By default, svy computes the variance by using
deviations of the pseudovalues from their mean.
successful replication. A red ‘x’ is displayed if command returns with an error, ‘e’ is displayed
if at least one of the values in the exp list is missing, ‘n’ is displayed if the sample size is not
correct, and a yellow ‘s’ is displayed if the dropped sampling unit is outside the subpopulation
sample.
saving(), keep, verbose, noisily, trace, title(), nodrop, reject(); see [SVY] svy jackknife.
Also see
[SVY] svy jackknife — Jackknife estimation for survey data
51
Title
ml for svy — Maximum pseudolikelihood estimation for survey data
Remarks and examples Reference Also see

Stata’s ml command can fit maximum likelihood–based models for survey data. Many ml-based
estimators can now be modified to handle one or more stages of clustering, stratification, sampling
weights, finite population correction, poststratification, and subpopulation estimation. See [R] ml for
details.
See [P] program properties for a discussion of the programming requirements for an estimation
command to work with the svy prefix. See Gould, Pitblado, and Poi (2010) for examples of user-written
estimation commands that support the svy prefix.
Example 1: User-written survey regression

The ml command requires a program that computes likelihood values to perform maximum
likelihood. Here is a likelihood evaluator used in Gould, Pitblado, and Poi (2010) to fit linear
regression models using likelihood from the normal distribution.
program mynormal_lf
version 14
args lnf mu lnsigma
quietly replace ‘lnf’ = ln(normalden($ML_y1,‘mu’,exp(‘lnsigma’)))
end
Here we fit a survey regression model using a multistage survey dataset with ml and the above
likelihood evaluator.
. use http://www.stata-press.com/data/r14/multistage
. svyset county [pw=sampwgt], strata(state) fpc(ncounties) || school,
> fpc(nschools)
pweight: sampwgt
VCE: linearized
Strata 1: state
SU 1: county
FPC 1: ncounties
Strata 2: <one>
SU 2: school
FPC 2: nschools
. ml model lf mynormal_lf (mu: weight = height) /lnsigma, svy
52
ml for svy — Maximum pseudolikelihood estimation for survey data 53
. ml max
initial: log pseudolikelihood = -<inf> (could not be evaluated)
feasible: log pseudolikelihood = -7.301e+08
rescale: log pseudolikelihood = -51944380
rescale eq: log pseudolikelihood = -47565331
Iteration 6: log pseudolikelihood = -39379181 (backed up)
Design df = 50
F( 1, 50) = 593.99
Prob > F = 0.0000
Linearized
mu
height .716311 .0293908 24.37 0.000 .6572778 .7753442
_cons -149.6181 12.57266 -11.90 0.000 -174.871 -124.3652
lnsigma
_cons 3.372153 .0180777 186.54 0.000 3.335843 3.408464
Reference
Gould, W. W., J. S. Pitblado, and B. P. Poi. 2010. Maximum Likelihood Estimation with Stata. 4th ed. College
Also see
[P] program properties — Properties of user-defined programs
[R] maximize — Details of iterative maximization
[R] ml — Maximum likelihood estimation
Title
poststratification — Poststratification for survey data

References Also see
Description
Poststratification is a method for adjusting the sampling weights, usually to account for underrep-
resented groups in the population.
See [SVY] direct standardization for a similar method of adjustment that allows the comparison
of rates that come from different frequency distributions.

Overview
Video example
Overview
Poststratification involves adjusting the sampling weights so that they sum to the population
sizes within each poststratum. This usually results in decreasing bias because of nonresponse and
underrepresented groups in the population. Poststratification also tends to result in smaller variance
estimates.
The svyset command has options to set variables for applying poststratification adjustments to the
sampling weights. The poststrata() option takes a variable that contains poststratum identifiers,
and the postweight() option takes a variable that contains the poststratum population sizes.
In the following example, we use an example from Levy and Lemeshow (2008) to show how
poststratification affects the point estimates and their variance.
Example 1: Poststratified mean

Levy and Lemeshow (2008, sec. 6.6) give an example of poststratification by using simple survey
data from a veterinarian’s client list. The data in poststrata.dta were collected using simple
random sampling without replacement. The totexp variable contains the total expenses to the client,
type identifies the cats and dogs, postwgt contains the poststratum sizes (450 for cats and 850 for
dogs), and fpc contains the total number of clients (850 + 450 = 1300).
. use http://www.stata-press.com/data/r14/poststrata
. svyset, poststrata(type) postweight(postwgt) fpc(fpc)
pweight: <none>
VCE: linearized
Poststrata: type
Postweight: postwgt
Strata 1: <one>
FPC 1: fpc
54
poststratification — Poststratification for survey data 55
. svy: mean totexp

Number of PSUs = 50 Population size = 1,300
N. of poststrata = 2 Design df = 49
Linearized
totexp 40.11513 1.163498 37.77699 42.45327
The mean total expenses is $40.12 with a standard error of $1.16. In the following, we omit the
poststratification information from svyset, resulting in mean total expenses of $39.73 with standard
error $2.22. The difference between the mean estimates is explained by the facts that expenses tend
to be larger for dogs than for cats and that the dogs were slightly underrepresented in the sample
(850/1,300 ≈ 0.65 for the population; 32/50 = 0.64 for the sample). This reasoning also explains why
the variance estimate from the poststratified mean is smaller than the one that was not poststratified.
. svyset, fpc(fpc)
pweight: <none>
VCE: linearized
Strata 1: <one>
FPC 1: fpc
. svy: mean totexp
Design df = 49
Linearized
totexp 39.7254 2.221747 35.26063 44.19017
Video example
Specifying the poststratification of survey data to Stata

Suppose that you used a complex survey design to sample m individuals from a population of
size M . Let Pk be the set of individuals in the sample that belong to poststratum k , and let IPk (j)
indicate if the j th individual is in poststratum k , where
(
1, if j ∈ Pk
IPk (j) =
0, otherwise
Also let LP be the number of poststrata and Mk be the population size for poststratum k .
56 poststratification — Poststratification for survey data
If wj is the unadjusted sampling weight for the j th sampled individual, the poststratification
adjusted sampling weight is
LP
X Mk
wj∗ = IPk (j) w
ck j
M
k=1
where M
ck is
m
X
M
ck = IPk (j)wj
j=1
The point estimates are computed using these adjusted weights. For example, the poststratified total
estimator is
Xm
Yb P = wj∗ yj
j=1
where yj is an item from the j th sampled individual.

For replication-based variance estimation, the BRR and jackknife replicate-weight variables are
similarly adjusted to produce the replicate values used in the respective variance formulas.
The score variable for the linearized variance estimator of a poststratified total is
LP
!
bP
X Mk Ybk
zj (Y ) = IPk (j) yj − (1)
k=1 M
ck M
ck
where Ybk is the total estimator for the k th poststratum,

m
X
Ybk = IPk (j)wj yj
j=1
For the poststratified ratio estimator, the score variable is
bP bP bP bP
bP ) = X zj (Y ) − Y zj (X )
zj ( R (2)
b P )2
(X
where Xb P is the poststratified total estimator for item xj . For regression models, the equation-level
scores are adjusted as in (1). These score variables were derived using the method described in
[SVY] variance estimation for the ratio estimator and are a direct result of the methods described in
Deville (1999), Demnati and Rao (2004), and Shah (2004).
References
17–26.
Levy, P. S., and S. A. Lemeshow. 2008. Sampling of Populations: Methods and Applications. 4th ed. Hoboken, NJ:
Wiley.
poststratification — Poststratification for survey data 57
Also see
Title
sdr options — More options for SDR variance estimation
Description
svy accepts more options when performing successive difference replication (SDR) variance
estimation. See [SVY] svy sdr for a complete discussion.
Syntax
sdr options Description
SE
trace trace command
saving(), verbose, noisily, trace, title(), nodrop, and reject() are not shown in the dialog boxes for
estimation commands.
Options

SE
saving(), verbose, noisily, trace, title(), nodrop, reject(); see [SVY] svy sdr.
Also see
[SVY] svy sdr — Successive difference replication for survey data
58
Title
subpopulation estimation — Subpopulation estimation for survey data

References Also see
Description
Subpopulation estimation focuses on part of the population. This entry discusses subpopulation
estimation and explains why you should use the subpop() option instead of if and in for your
survey data analysis.

Subpopulation estimation involves computing point and variance estimates for part of the population.
This is not the same as restricting the estimation sample to the collection of observations within
the subpopulation because variance estimation for survey data measures sample-to-sample variability,
assuming that the same survey design is used to collect the data; see Methods and formulas
for a detailed explanation. West, Berglund, and Heeringa (2008) provides further information on
subpopulation analysis.
The svy prefix command’s subpop() option performs subpopulation estimation. The svy: mean,
svy: proportion, svy: ratio, and svy: total commands also have the over() option to perform
estimation for multiple subpopulations.
The following examples illustrate how to use the subpop() and over() options.
Example 1
Suppose that we are interested in estimating the proportion of women in our population who have
had a heart attack. In our NHANES II dataset (McDowell et al. 1981), the female participants can
be identified using the female variable, and the heartatk variable indicates whether an individual
has ever had a heart attack. Below we use svy: mean with the heartatk variable to estimate the
proportion of individuals who have had a heart attack, and we use subpop(female) to identify our
subpopulation of interest.
. svy, subpop(female): mean heartatk
Design df = 31
Linearized
heartatk .0193276 .0017021 .0158562 .0227991
59
60 subpopulation estimation — Subpopulation estimation for survey data
The subpop(varname) option takes a 0/1 variable, and the subpopulation of interest is defined by
varname = 1. All other members of the sample not in the subpopulation are indicated by varname = 0.
If a person’s subpopulation status is unknown, varname should be set to missing (.), so those
observations will be omitted from the analysis. For instance, in the preceding analysis, if a person’s
sex was not recorded, female should be coded as missing rather than as male (female = 0).
Technical note
Actually, the subpop(varname) option takes a zero/nonzero variable, and the subpopulation is
defined by varname 6= 0 and not missing. All other members of the sample not in the subpopulation
are indicated by varname = 0, but 0, 1, and missing are typically the only values used for the
subpop() variable.
Furthermore, you can specify an if qualifier within subpop() to identify a subpopulation. The
result is the same as generating a variable equal to the conditional expression and supplying it as the
subpop() variable. If a varname and an if qualifier are specified within the subpop() option, the
subpopulation is identified by their logical conjunction (logical and ), and observations with missing
values in either are dropped from the estimation sample.
Example 2: Multiple subpopulation estimation

Means, proportions, ratios, and totals for multiple subpopulations can be estimated using the
over() option with svy: mean, svy: proportion, svy: ratio, and svy: total, respectively.
Here is an example using the NMIHS data (Gonzalez, Krauss, and Scott 1992), estimating mean
birthweight over the categories of the race variable.
. svy: mean birthwgt, over(race)
Number of PSUs = 9,946 Population size = 3,895,562
Design df = 9,940
black: race = black
Linearized
birthwgt
nonblack 3402.32 7.609532 3387.404 3417.236
black 3127.834 6.529814 3115.035 3140.634
subpopulation estimation — Subpopulation estimation for survey data 61
More than one variable can be used in the over() option.

. svy: mean birthwgt, over(race marital)
Design df = 9,940
Over: race marital
_subpop_1: nonblack single
_subpop_2: nonblack married
_subpop_3: black single
_subpop_4: black married
Linearized
birthwgt
_subpop_1 3291.045 20.18795 3251.472 3330.617
_subpop_2 3426.407 8.379497 3409.982 3442.833
_subpop_3 3073.122 8.752553 3055.965 3090.279
_subpop_4 3221.616 12.42687 3197.257 3245.975
Here the race and marital variables have value labels. race has the value 0 labeled “nonblack”
(that is, white and other) and 1 labeled “black”; marital has the value 0 labeled “single” and 1
labeled “married”. Value labels on the over() variables make for a more informative legend above
the table of point estimates. See [U] 12.6.3 Value labels for information on creating value labels.
We can also combine the subpop() option with the over() option.
. generate nonblack = (race == 0) if !missing(race)
. svy, subpop(nonblack): mean birthwgt, over(marital age20)
Design df = 4,721
Over: marital age20
_subpop_1: single age20+
_subpop_2: single age<20
_subpop_3: married age20+
_subpop_4: married age<20
Linearized
birthwgt
_subpop_1 3312.012 24.2869 3264.398 3359.625
_subpop_2 3244.709 36.85934 3172.448 3316.971
_subpop_3 3434.923 8.674633 3417.916 3451.929
_subpop_4 3287.301 34.15988 3220.332 3354.271
Note: 3 strata omitted because they contain no subpopulation

members.
This time, we estimated means for the marital status and age (<20 or ≥20) subpopulations for race
== 0 (nonblack) only. We carefully define nonblack so that it is missing when race is missing.
If we omitted the if !missing(race) in our generate statement, then nonblack would be 0

when race was missing. This would improperly assume that all individuals with a missing value
for race were black and could cause our results to have incorrect standard errors. The standard
errors could be incorrect because those observations for which race is missing would be counted as
part of the estimation sample, potentially inflating the number of PSUs used in the formula for the
variance estimator. For this reason, observations with missing values for any of the over() variables
are omitted from the analysis.

Cochran (1977, sec. 2.13) discusses a method by which you can derive estimates for subpopulation
totals. This section uses this method to derive the formulas for a subpopulation total from a simple
random sample (without replacement) to explain how the subpop() option works, shows why this
method will often produce different results from those produced using an equivalent if (or in)
qualifier (outside the subpop() option), and discusses how this method applies to subpopulation
means, proportions, ratios, and regression models.
Subpopulation totals
Subpopulation estimates other than the total
Subpopulation with replication methods
Subpopulation totals
Let Yj be a survey item for individual j in the population, where j = 1, . . . , N and N is the
population size. Let S be a subset of individuals in the population and IS (j) indicate if the j th
individual is in S , where (
1, if j ∈ S
IS (j) =
0, otherwise
The subpopulation total is
N
X
YS = IS (j)Yj
j=1
and the subpopulation size is
N
X
NS = IS (j)
j=1
Let yj be the items for those individuals selected in the sample, where j = 1, . . . , n and n is the
sample size. The number of individuals sampled from the subpopulation is
X n
nS = IS (j)
j=1
The estimator for the subpopulation total is
n
X
YbS = IS (j)wj yj (1)
j=1
subpopulation estimation — Subpopulation estimation for survey data 63
where wj = N/n is the unadjusted sampling weight for this design. The estimator for NS is
n
X
NbS = IS (j)wj
j=1
The replicate values for the BRR and jackknife variance estimators are computed using the same
method.
The linearized variance estimator for YbS is
n 2
n n X 1
Vb (YbS ) = 1 − IS (j)wj yj − YbS (2)
N n−1 j=1
n
The covariance estimator for the subpopulation totals YbS and X

bS (notation for XS is defined similarly
to that of YS ) is
n
n n X 1b 1 b
Cov(YS , XS ) = 1 −
d b b IS (j)wj yj − YS IS (j)wj xj − XS (3)
N n−1 j=1
n n
Equation (2) is not the same formula that results from restricting the estimation sample to
the observations within S . The formula using this restricted sample (assuming a svyset with the
corresponding FPC) is
n 2
nS nS X 1 b
V (YS ) = 1 −
e b IS (j) wj yj − YS (4)
NbS nS − 1 nS
j=1
These variance estimators, (2) and (4), assume two different survey designs. In (2), n individuals are
sampled without replacement from the population comprising the NS values from the subpopulation
with N − NS additional zeros. In (4), nS individuals are sampled without replacement from the
subpopulation of NS values. We discourage using (4) by warning against using the if and in
qualifiers for subpopulation estimation because this variance estimator does not accurately measure
the sample-to-sample variability of the subpopulation estimates for the survey design that was used
to collect the data.
For survey data, there are only a few circumstances that require using the if qualifier. For example,
if you suspected laboratory error for a certain set of measurements, then using the if qualifier to
omit these observations from the analysis might be proper.
Subpopulation estimates other than the total

To generalize the above results, note that the other point estimators—such as means, proportions,
ratios, and regression coefficients—yield a linearized variance estimator based on one or more (equation
level) score variables. For example, the weighted sample estimation equations of a regression model
for a given subpopulation (see (3) from [SVY] variance estimation) is
X n
G(
b βS ) = IS (j)wj S(βS ; yj , xj ) = 0 (5)
j=1
You can write G(

b βS ) as
n
X
G(
b βS ) = IS (j)wj dj
j=1
which is an estimator for the subpopulation total G(βS ), so its variance estimator can be computed
using the design-based variance estimator for a subpopulation total.
Subpopulation with replication methods

The above comparison between the variance estimator from the subpop() option and the variance
estimator from the if and in qualifiers is also true for the replication methods.
For the BRR method, the same number of replicates is produced with or without the subpop()
option. The difference is how the replicate values are computed. Using the if and in qualifiers may
cause an error because svy brr checks that there are two PSUs in every stratum within the restricted
sample.
For the jackknife method, every PSU produces a replicate, even if it does not contain an observation
within the subpopulation specified using the subpop() option. When the if and in qualifiers are
used, only the PSUs that have at least 1 observation within the restricted sample will produce a
replicate.
For methods using replicate weight variables, every weight variable produces a replicate, even if
it does not contain an observation within the subpopulation specified using the subpop() option.
When the if and in qualifiers are used, only the PSUs that have at least 1 observation within the
restricted sample will produce a replicate.
References
West, B. T., P. A. Berglund, and S. G. Heeringa. 2008. A closer examination of subpopulation analysis of complex-sample
survey data. Stata Journal 8: 520–531.
Also see
Title
svy — The survey prefix command
Description Quick start Syntax Options

Remarks and examples Stored results Methods and formulas References
Also see
Description
svy fits statistical models for complex survey data by adjusting the results of a command for
survey settings identified by svyset. Any Stata estimation command listed in [SVY] svy estimation
may be used with svy. User-written commands that meet the requirements in [P] program properties
may also be used.
Quick start
Data for a two-stage design with sampling weight wvar1, strata defined by levels of svar, sampling
units are identified by su1, and second-stage clustering is defined by su2
svyset su1 [pweight=wvar1], strata(svar) || su2
Adjust linear regression for complex survey design settings specified in svyset
svy: regress . . .
As above, but restrict estimation to the subpopulation where group equals 4
svy, subpop(if group==4): regress . . .
Same as above, but use new binary variable insample to indicate the subpopulation
generate insample = (group==4)
svy, subpop(insample): regress . . .
Specify that the design degrees of freedom is 135 instead of the difference between the number of
unique values of su1 and the number of levels of svar
svy, dof(135): regress . . .
Note: Any estimation command meeting the requirements specified in the Description may be
substituted for regress in the examples above.
65
66 svy — The survey prefix command
Syntax

svy vcetype , svy options eform option : command
vcetype Description
SE
linearized Taylor-linearized variance estimation
bootstrap bootstrap variance estimation; see [SVY] svy bootstrap
brr BRR variance estimation; see [SVY] svy brr
jackknife jackknife variance estimation; see [SVY] svy jackknife
sdr SDR variance estimation; see [SVY] svy sdr
Specifying a vcetype overrides the default from svyset.
svy options Description

if/in

subpop( varname if ) identify a subpopulation
SE
dof(#) design degrees of freedom
bootstrap options more options allowed with bootstrap variance estimation;
see [SVY] bootstrap options
brr options more options allowed with BRR variance estimation;
see [SVY] brr options
jackknife options more options allowed with jackknife variance estimation;
see [SVY] jackknife options
sdr options more options allowed with SDR variance estimation;
see [SVY] sdr options
Reporting
level(#) set confidence level; default is level(95)
nocnsreport do not display constraints
display options control columns and column formats, row spacing, line width,
display of omitted variables and base and empty cells, and
factor-variable labeling
noheader suppress table header
nolegend suppress table legend
noadjust do not adjust model Wald statistic
trace trace command
coeflegend display legend instead of statistics
svy — The survey prefix command 67
svy requires that the survey design variables be identified using svyset; see [SVY] svyset.
command defines the estimation command to be executed. The by prefix cannot be part of command.
mi estimate may be used with svy linearized if the estimation command allows mi estimate; it may not be
used with svy bootstrap, svy brr, svy jackknife, or svy sdr.
noheader, nolegend, noadjust, noisily, trace, and coeflegend are not shown in the dialog boxes for estimation
commands.
Warning: Using if or in restrictions will often not produce correct variance estimates for subpopulations. To compute
estimates for subpopulations, use the subpop() option.
See [U] 20 Estimation and postestimation commands for more capabilities of estimation commands.
Options

if/in
subpop(subpop) specifies that estimates be computed for the single subpopulation identified by
subpop, which is

varname if
Thus the subpopulation is defined by the observations for which varname 6= 0 that also meet
the if conditions. Typically, varname = 1 defines the subpopulation, and varname = 0 indicates
observations not belonging to the subpopulation. For observations whose subpopulation status is
uncertain, varname should be set to a missing value; such observations are dropped from the
estimation sample.
See [SVY] subpopulation estimation and [SVY] estat.

SE
dof(#) specifies the design degrees of freedom, overriding the default calculation, df = Npsu −Nstrata .
bootstrap options are other options that are allowed with bootstrap variance estimation specified by svy
bootstrap or specified as svyset using the vce(bootstrap) option; see [SVY] bootstrap options.
brr options are other options that are allowed with BRR variance estimation specified by svy brr or
specified as svyset using the vce(brr) option; see [SVY] brr options.
jackknife options are other options that are allowed with jackknife variance estimation specified by svy
jackknife or specified as svyset using the vce(jackknife) option; see [SVY] jackknife options.
sdr options are other options that are allowed with SDR variance estimation specified by svy sdr or
specified as svyset using the vce(sdr) option; see [SVY] sdr options.

Reporting
level(#) specifies the confidence level, as a percentage, for confidence intervals. The default is
level(95) or as set by set level; see [U] 20.7 Specifying the width of confidence intervals.
nocnsreport; see [R] estimation options.
display options: noci, nopvalues, noomitted, vsquish, noemptycells, baselevels,
allbaselevels, nofvlabel, fvwrap(#), fvwrapon(style), cformat(% fmt), pformat(% fmt),
sformat(% fmt), and nolstretch; see [R] estimation options.
The following options are available with svy but are not shown in the dialog boxes:
noheader prevents the table header from being displayed. This option implies nolegend.
nolegend prevents the table legend identifying the subpopulations from being displayed.
noadjust specifies that the model Wald test be carried out as W/k ∼ F (k, d), where W is the Wald
test statistic, k is the number of terms in the model excluding the constant term, d is the total
number of sampled PSUs minus the total number of strata, and F (k, d) is an F distribution with
k numerator degrees of freedom and d denominator degrees of freedom. By default, an adjusted
Wald test is conducted: (d − k + 1)W/(kd) ∼ F (k, d − k + 1).
See Korn and Graubard (1990) for a discussion of the Wald test and the adjustments thereof. Using
the noadjust option is not recommended.
noisily requests that any output from command be displayed.
trace causes a trace of the execution of command to be displayed.
coeflegend; see [R] estimation options.
The following option is usually available with svy at the time of estimation or on replay but is not
shown in all dialog boxes:
eform option; see [R] eform option.

The svy prefix is designed for use with complex survey data. Typical survey design characteristics
include sampling weights, one or more stages of clustered sampling, and stratification. For a general
discussion of various aspects of survey designs, including multistage designs, see [SVY] svyset.
Below we present an example of the effects of weights, clustering, and stratification. This is a
typical case, but drawing general rules from any one example is still dangerous. You could find
particular analyses from other surveys that are counterexamples for each of the trends for standard
errors exhibited here.
Example 1: The effects of weights, clustering, and stratification

We use data from the Second National Health and Nutrition Examination Survey (NHANES II)
(McDowell et al. 1981) as our example. This is a national survey, and the dataset has sampling
weights, strata, and clustering. In this example, we will consider the estimation of the mean serum
zinc level of all adults in the United States.
First, consider a proper design-based analysis, which accounts for weighting, clustering, and
stratification. Before we issue our svy estimation command, we set the weight, strata, and PSU
identifier variables:
. use http://www.stata-press.com/data/r14/nhanes2f
. svyset psuid [pweight=finalwgt], strata(stratid)
pweight: finalwgt
VCE: linearized
Strata 1: stratid
SU 1: psuid
FPC 1: <zero>
We now estimate the mean by using the proper design-based analysis:

. svy: mean zinc
Design df = 31
Linearized
zinc 87.18207 .4944827 86.17356 88.19057
If we ignore the survey design and use mean to estimate the mean, we get
. mean zinc
Mean estimation Number of obs = 9,189
zinc 86.51518 .1510744 86.21904 86.81132
The point estimate from the unweighted analysis is smaller by more than one standard error than
the proper design-based estimate. Also, design-based analysis produced a standard error that is 3.27
times larger than the standard error produced by our incorrect analysis.
Example 2: Halfway is not enough—the importance of stratification and clustering

When some people analyze survey data, they say, “I know I have to use my survey weights, but
I will just ignore the stratification and clustering information.” If we follow this strategy, we will
obtain the proper design-based point estimates, but our standard errors, confidence intervals, and test
statistics will usually be wrong.
To illustrate this effect, suppose that we used the svy: mean procedure with pweights only.
. svyset [pweight=finalwgt]
pweight: finalwgt
VCE: linearized
Strata 1: <one>
FPC 1: <zero>
. svy: mean zinc
Design df = 9,188
Linearized
zinc 87.18207 .1828747 86.82359 87.54054

This approach gives us the same point estimate as our design-based analysis, but the reported
standard error is less than one-half the design-based standard error. If we accounted only for clustering
and weights and ignored stratification in NHANES II, we would obtain the following analysis:
. svyset psuid [pweight=finalwgt]
pweight: finalwgt
VCE: linearized
Strata 1: <one>
SU 1: psuid
FPC 1: <zero>
. svy: mean zinc
Design df = 1
Linearized
zinc 87.18207 .7426221 77.74616 96.61798
Here our standard error is about 50% larger than what we obtained in our proper design-based analysis.
Example 3
Let’s look at a regression. We model zinc on the basis of age, weight, sex, race, and rural or urban
residence. We compare a proper design-based analysis with an ordinary regression (which assumes
independent and identically distributed error).
Here is our design-based analysis:

pweight: finalwgt
VCE: linearized
Strata 1: stratid
SU 1: psuid
FPC 1: <zero>
. svy: regress zinc age c.age#c.age weight female black orace rural
Design df = 31
F( 7, 25) = 62.50
Prob > F = 0.0000
R-squared = 0.0698
Linearized
zinc Coef. Std. Err. t P>|t| [95% Conf. Interval]
age -.1701161 .0844192 -2.02 0.053 -.3422901 .002058
c.age#c.age .0008744 .0008655 1.01 0.320 -.0008907 .0026396
weight .0535225 .0139115 3.85 0.001 .0251499 .0818951

female -6.134161 .4403625 -13.93 0.000 -7.032286 -5.236035
black -2.881813 1.075958 -2.68 0.012 -5.076244 -.687381
orace -4.118051 1.621121 -2.54 0.016 -7.424349 -.8117528
rural -.5386327 .6171836 -0.87 0.390 -1.797387 .7201216
_cons 92.47495 2.228263 41.50 0.000 87.93038 97.01952
If we had improperly ignored our survey weights, stratification, and clustering (that is, if we had
used the usual Stata regress command), we would have obtained the following results:
. regress zinc age c.age#c.age weight female black orace rural
Source SS df MS Number of obs = 9,189
F(7, 9181) = 79.72
Model 110417.827 7 15773.9753 Prob > F = 0.0000
Residual 1816535.3 9,181 197.85811 R-squared = 0.0573
Adj R-squared = 0.0566
Total 1926953.13 9,188 209.724982 Root MSE = 14.066
age -.090298 .0638452 -1.41 0.157 -.2154488 .0348528
c.age#c.age -.0000324 .0006788 -0.05 0.962 -.0013631 .0012983
weight .0606481 .0105986 5.72 0.000 .0398725 .0814237

female -5.021949 .3194705 -15.72 0.000 -5.648182 -4.395716
black -2.311753 .5073536 -4.56 0.000 -3.306279 -1.317227
orace -3.390879 1.060981 -3.20 0.001 -5.470637 -1.311121
rural -.0966462 .3098948 -0.31 0.755 -.7041089 .5108166
_cons 89.49465 1.477528 60.57 0.000 86.59836 92.39093
The point estimates differ by 3% – 100%, and the standard errors for the proper designed-based analysis
are 30% – 110% larger. The differences are not as dramatic as we saw with the estimation of the
mean, but they are still substantial.
Stored results
svy stores the following in e():
Scalars
e(N) number of observations
e(N sub) subpopulation observations
e(N strata) number of strata
e(N strata omit) number of strata omitted
e(singleton) 1 if singleton strata, 0 otherwise
e(census) 1 if census data, 0 otherwise
e(F) model F statistic
e(df m) model degrees of freedom
e(df r) variance degrees of freedom
e(N pop) estimate of population size
e(N subpop) estimate of subpopulation size
e(N psu) number of sampled PSUs
e(stages) number of sampling stages
e(k eq) number of equations in e(b)
e(k aux) number of ancillary parameters
e(p) p-value
e(rank) rank of e(V)
Macros
e(prefix) svy
e(cmdname) command name from command
e(cmd) same as e(cmdname) or e(vce)
e(command) command
e(cmdline) command as typed
e(wtype) weight type
e(wexp) weight expression
e(weight#) variable identifying weight for stage #
e(wvar) weight variable name
e(singleunit) singleunit() setting
e(strata) strata() variable
e(strata#) variable identifying strata for stage #
e(psu) psu() variable
e(su#) variable identifying sampling units for stage #
e(fpc) fpc() variable
e(fpc#) FPC for stage #
e(title) title in estimation output
e(poststrata) poststrata() variable
e(postweight) postweight() variable
e(vce) vcetype specified in vce()
e(vcetype) title used to label Std. Err.
e(mse) mse, if specified
e(subpop) subpop from subpop()
e(adjust) noadjust, if specified
e(properties) b V
e(estat cmd) program used to implement estat
e(predict) program used to implement predict
e(marginsnotok) predictions disallowed by margins
e(marginswtype) weight type for margins
Matrices
e(b) estimates
e(V) design-based variance
e(V srs) simple-random-sampling-without-replacement variance, V
bsrswor
e(V srssub) subpopulation simple-random-sampling-without-replacement variance, V

bsrswor
(created only when subpop() is specified)
e(V srswr) simple-random-sampling-with-replacement variance, V
bsrswr
(created only when fpc() option is svyset)
e(V srssubwr) subpopulation simple-random-sampling-with-replacement variance, V
bsrswr
(created only when subpop() is specified)
e(V modelbased) model-based variance
e(V msp) variance from misspecified model fit, V
bmsp
e( N strata single) number of strata with one sampling unit
e( N strata certain) number of certainty strata
e( N strata) number of strata
Functions
e(sample) marks estimation sample
svy also carries forward most of the results already in e() from command.

See [SVY] variance estimation for all the details behind the point estimate and variance calculations
made by svy.
References
Korn, E. L., and B. I. Graubard. 1990. Simultaneous testing of regression coefficients with complex survey data: Use
of Bonferroni t statistics. American Statistician 44: 270–276.
Also see
[U] 20 Estimation and postestimation commands
Title
svy bootstrap — Bootstrap for survey data

References Also see
Description
svy bootstrap performs nonparametric bootstrap estimation of specified statistics (or expressions)
for a Stata command or a user-written command. The command is executed once for each replicate
using sampling weights that are adjusted according to the bootstrap methodology. Any Stata estimation
command listed in [SVY] svy estimation may be used with svy bootstrap. User-written commands
that meet the requirements in [P] program properties may also be used.
Quick start
Estimate population mean of v1 using bootstrap standard-error estimates and variables with prefix
rwvar as the bootstrap replicate weights
svyset, bsrweight(rwvar*)
svy bootstrap _b: mean v1
Same as above
svyset, bsrweight(rwvar*) vce(bootstrap)
svy: mean v1
As above, and specify that 3 replicates were used to calculate each bootstrap replicate weight
svy, bsn(3): mean v1
Bootstrap standard error of the difference between the means of v2 and v3 using either svyset
command above
svy bootstrap (_b[v2]-_b[v3]): mean v2 v3
As above, but name the result diff and save results from each replication to mydata.dta
svy bootstrap diff=(_b[v2]-_b[v3]), saving(mydata): mean v2 v3
substituted for mean in the examples above.
Menu
Statistics > Survey data analysis > Resampling > Bootstrap estimation
74
svy bootstrap — Bootstrap for survey data 75
Syntax

svy bootstrap exp list , svy options bootstrap options eform option : command
if/in

Reporting
coeflegend is not shown in the dialog boxes for estimation commands.
bootstrap options Description

Main
Options

saving( filename , . . . ) save results to filename; save statistics in double precision;
save results to filename every # replications
Reporting
trace trace command
title(text) use text as title for bootstrap results
Advanced
command defines the statistical command to be executed. The by prefix cannot be part of command.
svy bootstrap requires that the bootstrap replicate weights be identified using svyset.
76 svy bootstrap — Bootstrap for survey data
exp list specifies the statistics to be collected from the execution of command. exp list is required
unless command has the svyb program property, in which case exp list defaults to b; see [P] program
properties. The expressions in exp list are assumed to conform to the following:
exp list contains (name: elist)
elist
eexp
elist contains newvarname = (exp)
(exp)
eexp is specname
[eqno]specname
specname is b
b[]
se
se[]
eqno is ##
name
exp is a standard Stata expression; see [U] 13 Functions and expressions.

Distinguish between [ ], which are to be typed, and , which indicate optional arguments.
Options
svy options; see [SVY] svy.

Main
mean-weight variable specified in the bsrweight() option of svyset. The default is bsn(1).
The bsn() option of svy bootstrap overrides the bsn() option of svyset; see [SVY] svyset.

Options

saving( filename , suboptions ) creates a Stata data file (.dta file) consisting of (for each statistic
in exp list) a variable containing the replicates.
double specifies that the results for each replication be saved as doubles, meaning 8-byte reals.
By default, they are saved as floats, meaning 4-byte reals. This option may be used without
the saving() option to compute the variance estimates by using double precision.
every(#) specifies that results be written to disk every #th replication. every() should be specified
in conjunction with saving() only when command takes a long time for each replication.
This will allow recovery of partial results should some other software crash your computer.
See [P] postfile.
replace indicates that filename be overwritten if it exists. This option does not appear in the
dialog box.
mse specifies that svy bootstrap compute the variance by using deviations of the replicates from the
observed value of the statistics based on the entire dataset. By default, svy bootstrap computes
the variance by using deviations of the replicates from their mean.

Reporting
verbose requests that the full table legend be displayed.
successful replication. A red ‘x’ is printed if command returns with an error, and ‘e’ is printed if
one of the values in exp list is missing.
noisily requests that any output from command be displayed. This option implies the nodots
option.
trace causes a trace of the execution of command to be displayed. This option implies the noisily
option.
title(text) specifies a title to be displayed above the table of bootstrap results; the default title is
“Bootstrap results”.
eform option; see [R] eform option. This option is ignored if exp list is not b.

Advanced
nodrop prevents observations outside e(sample) and the if and in qualifiers from being dropped
before the data are resampled.
reject(exp) identifies an expression that indicates when results should be rejected. When exp is
true, the resulting values are reset to missing values.

The bootstrap methods for survey data used in recent years are largely due to McCarthy and
Snowden (1985), Rao and Wu (1988), and Rao, Wu, and Yue (1992). For example, Yeo, Mantel, and
Liu (1999) cites Rao, Wu, and Yue (1992) as the method for variance estimation used in the National
Population Health Survey conducted by Statistics Canada.
In the survey bootstrap, the model is fit multiple times, once for each of a set of adjusted sampling
weights. The variance is estimated using the resulting replicated point estimates.
Example 1
Suppose that we need to estimate the average birthweight for the population represented by the
National Maternal and Infant Health Survey (NMIHS) (Gonzalez, Krauss, and Scott 1992).
In [SVY] svy estimation, the dataset nmihs.dta contained the following design information:
• Primary sampling units are mothers; that is, PSUs are individual observations — there is no
separate PSU variable.
• The finalwgt variable contains the sampling weights.
• The stratan variable identifies strata.
• There is no variable for the finite population correction.
nmihs bs.dta is equivalent to nmihs.dta except that the stratum identifier variable stratan is
replaced by bootstrap replicate-weight variables. The replicate-weight variables are already svyset,
and the default method for variance estimation is vce(bootstrap).
. use http://www.stata-press.com/data/r14/nmihs_bs
. svyset
pweight: finwgt
VCE: bootstrap
MSE: off
bsrweight: bsrw1 bsrw2 bsrw3 bsrw4 bsrw5 bsrw6 bsrw7 bsrw8 bsrw9 bsrw10
bsrw11 bsrw12 bsrw13 bsrw14 bsrw15 bsrw16 bsrw17 bsrw18 bsrw19
(output omitted )
bsrw989 bsrw990 bsrw991 bsrw992 bsrw993 bsrw994 bsrw995
bsrw996 bsrw997 bsrw998 bsrw999 bsrw1000
Strata 1: <one>
FPC 1: <zero>
Now we can use svy: mean to estimate the average birthweight for our population, and the standard
errors will be estimated using the survey bootstrap.
. svy, nodots: mean birthwgt
Survey: Mean estimation Number of obs = 9,946
Replications = 1,000
Observed Bootstrap Normal-based

birthwgt 3355.452 6.520637 3342.672 3368.233
From these results, we are 95% confident that the mean birthweight for our population is between
3,343 and 3,368 grams.
To accommodate privacy concerns, many public-use datasets contain replicate-weight variables

derived from the “mean bootstrap” described by Yung (1997). In the mean bootstrap, each adjusted
weight is derived from more than one bootstrap sample. When replicate-weight variables for the
mean bootstrap are svyset, the bsn() option identifying the number of bootstrap samples used to
generate the adjusted-weight variables should also be specified. This number is used in the variance
calculation; see [SVY] variance estimation.
Example 2
nmihs mbs.dta is equivalent to nmihs.dta except that the strata identifier variable stratan is
replaced by mean bootstrap replicate-weight variables. The replicate-weight variables and variance
adjustment are already svyset, and the default method for variance estimation is vce(bootstrap).
. use http://www.stata-press.com/data/r14/nmihs_mbs
. svyset
pweight: finwgt
VCE: bootstrap
MSE: off
bsrweight: mbsrw1 mbsrw2 mbsrw3 mbsrw4 mbsrw5 mbsrw6 mbsrw7 mbsrw8 mbsrw9
mbsrw10 mbsrw11 mbsrw12 mbsrw13 mbsrw14 mbsrw15 mbsrw16
(output omitted )
mbsrw192 mbsrw193 mbsrw194 mbsrw195 mbsrw196 mbsrw197 mbsrw198
mbsrw199 mbsrw200
bsn: 5
Strata 1: <one>
FPC 1: <zero>
Notice that the 200 mean bootstrap replicate-weight variables were generated from 5 bootstrap samples;
in fact, the mean bootstrap weight variables in nmihs mbs.dta were generated from the bootstrap
weight variables in nmihs bs.dta.
Here we use svy: mean to estimate the average birthweight for our population.
. svy, nodots: mean birthwgt
Replications = 200
Observed Bootstrap Normal-based

birthwgt 3355.452 5.712574 3344.256 3366.649
The standard error and confidence limits differ from the previous example. This merely illustrates
that the mean bootstrap is not numerically equivalent to the standard bootstrap, even when the
replicate-weight variables are generated from the same resampled datasets.
Stored results
In addition to the results documented in [SVY] svy, svy bootstrap stores the following in e():
Scalars
e(N reps) number of replications
e(N misreps) number of replications with missing values
e(k exp) number of standard expressions
e(k eexp) number of b/ se expressions
e(k extra) number of extra estimates added to b
e(bsn) bootstrap mean-weight adjustment
Macros
e(cmd) same as e(cmdname) or bootstrap
e(vce) bootstrap
e(exp#) #th expression
e(bsrweight) bsrweight() variable list
Matrices
e(b bs) bootstrap means
e(V) bootstrap variance estimates
When exp list is b, svy bootstrap will also carry forward most of the results already in e() from
command.

See [SVY] variance estimation for details regarding bootstrap variance estimation.
References
Kolenikov, S. 2010. Resampling variance estimation for complex survey data. Stata Journal 10: 165–199.
McCarthy, P. J., and C. B. Snowden. 1985. The bootstrap and finite population sampling. In Vital and Health Statistics,
1–23. Washington, DC: U.S. Government Printing Office.
Rao, J. N. K., and C. F. J. Wu. 1988. Resampling inference with complex survey data. Journal of the American
Statistical Association 83: 231–241.
Rao, J. N. K., C. F. J. Wu, and K. Yue. 1992. Some recent work on resampling methods for complex surveys. Survey
Methodology 18: 209–217.
Yeo, D., H. Mantel, and T.-P. Liu. 1999. Bootstrap variance estimation for the National Population Health Survey.
In Proceedings of the Survey Research Methods Section, 778–785. American Statistical Association.
Yung, W. 1997. Variance estimation for public use files under confidentiality constraints. In Proceedings of the Survey
Research Methods Section, 434–439. American Statistical Association.
Also see
[R] bootstrap — Bootstrap sampling and estimation
Title
svy brr — Balanced repeated replication for survey data

References Also see
Description
svy brr performs balanced repeated replication (BRR) estimation of specified statistics (or ex-
pressions) for a Stata command or a user-written command. The command is executed once for
each replicate using sampling weights that are adjusted according to the BRR methodology. Any
Stata estimation command listed in [SVY] svy estimation may be used with svy brr. User-written
commands that meet the requirements in [P] program properties may also be used.
Quick start
Estimate population mean of v1 using BRR standard-error estimates with sampling weight wvar1 and
replicate weights in variables with prefix rwvar
svyset [pweight = wvar1], brrweight(rwvar*)
svy brr _b: mean v1
BRR estimate of the standard error of the difference between the means of v2 and v3
svy brr (_b[v2]-_b[v3]): mean v2 v3
svy brr diff=(_b[v2]-_b[v3]), saving(mydata): mean v2 v3
Same as above
brr diff=(_b[v2]-_b[v3]), saving(mydata): mean v2 v3
Menu
Statistics > Survey data analysis > Resampling > Balanced repeated replications estimation
82
svy brr — Balanced repeated replication for survey data 83
Syntax

svy brr exp list , svy options brr options eform option : command
if/in

Reporting
brr options Description

Main
hadamard(matrix) Hadamard matrix
Options

Reporting
trace trace command
title(text) use text as title for BRR results
Advanced
84 svy brr — Balanced repeated replication for survey data
elist
eexp
(exp)
eexp is specname
[eqno]specname
specname is b
b[]
se
se[]
eqno is ##
name

Options

Main
hadamard(matrix) specifies the Hadamard matrix to be used to determine which PSUs are chosen
for each replicate.
fay(#) specifies Fay’s adjustment (Judkins 1990), where 0 ≤ # ≤ 2, but excluding 1. This option
overrides the fay(#) option of svyset; see [SVY] svyset.

Options

See [P] postfile.
dialog box.
mse specifies that svy brr compute the variance by using deviations of the replicates from the
observed value of the statistics based on the entire dataset. By default, svy brr computes the
variance by using deviations of the replicates from their mean.

Reporting
option.
option.
title(text) specifies a title to be displayed above the table of BRR results; the default title is “BRR
results”.

Advanced

BRR was first introduced by McCarthy (1966, 1969a, 1969b) as a method of variance estimation
for designs with two PSUs in every stratum. The BRR variance estimator tends to give more reasonable
variance estimates for this design than the linearized variance estimator, which can result in large
values and undesirably wide confidence intervals.
In BRR, the model is fit multiple times, once for each of a balanced set of combinations where
one PSU is dropped from each stratum. The variance is estimated using the resulting replicated point
estimates. Although the BRR method has since been generalized to include other designs, Stata’s
implementation of BRR requires two PSUs per stratum.
To protect the privacy of survey participants, public survey datasets may contain replicate-weight
variables instead of variables that identify the PSUs and strata. These replicate-weight variables are
adjusted copies of the sampling weights. For BRR, the sampling weights are adjusted for dropping
one PSU from each stratum; see [SVY] variance estimation for more details.
Example 1: BRR replicate-weight variables

The survey design for the NHANES II data (McDowell et al. 1981) is specifically suited to BRR;
there are two PSUs in every stratum.
. svydescribe
pweight: finalwgt
VCE: linearized
Strata 1: strata
SU 1: psu
FPC 1: <zero>
#Obs per Unit
1 2 380 165 190.0 215

2 2 185 67 92.5 118
3 2 348 149 174.0 199
4 2 460 229 230.0 231
5 2 252 105 126.0 147
(output omitted )
29 2 503 215 251.5 288
30 2 365 166 182.5 199
31 2 308 143 154.0 165
32 2 450 211 225.0 239
31 62 10,351 67 167.0 288
Here is a privacy-conscious dataset equivalent to the one above; all the variables and values remain,
except strata and psu are replaced with BRR replicate-weight variables. The BRR replicate-weight
variables are already svyset, and the default method for variance estimation is vce(brr).
. use http://www.stata-press.com/data/r14/nhanes2brr
. svyset
pweight: finalwgt
VCE: brr
MSE: off
brrweight: brr_1 brr_2 brr_3 brr_4 brr_5 brr_6 brr_7 brr_8 brr_9 brr_10
brr_29 brr_30 brr_31 brr_32
Strata 1: <one>
FPC 1: <zero>
Suppose that we were interested in the population ratio of weight to height. Here we use total
to estimate the population totals of weight and height and the svy brr prefix to estimate their
ratio and variance; we use total instead of ratio (which is otherwise preferable here) to illustrate
how to specify an exp list.
. svy brr WtoH = (_b[weight]/_b[height]): total weight height

1 2 3 4 5
................................
Replications = 32
Design df = 31
BRR
WtoH .4268116 .0008904 479.36 0.000 .4249957 .4286276
The mse option causes svy brr to use the MSE form of the BRR variance estimator. This variance
estimator will tend to be larger than the previous because of the addition of the familiar squared
bias term in the MSE; see [SVY] variance estimation for more details. The header for the column of
standard errors in the table of results is BRR * for the BRR variance estimator using the MSE formula.
. svy brr WtoH = (_b[weight]/_b[height]), mse: total weight height
1 2 3 4 5
................................
Replications = 32
Design df = 31
BRR *
WtoH .4268116 .0008904 479.36 0.000 .4249957 .4286276
The bias term here is too small to see any difference in the standard errors.
Example 2: Survey data without replicate-weight variables

For survey data with the PSU and strata variables but no replication weights, svy brr can compute
adjusted sampling weights within its replication loop. Here the hadamard() option must be supplied
with the name of a Stata matrix that is a Hadamard matrix of appropriate order for the number of
strata in your dataset (see the following technical note for a quick introduction to Hadamard matrices).
There are 31 strata in nhanes2.dta, so we need a Hadamard matrix of order 32 (or more) to
use svy brr with this dataset. Here we use h32 (from the following technical note) to estimate the
population ratio of weight to height by using the BRR variance estimator.
. svy brr, hadamard(h32): ratio (WtoH: weight/height)
1 2 3 4 5
................................
Replications = 32
Design df = 31
WtoH: weight/height
BRR
WtoH .4268116 .0008904 .4249957 .4286276
Technical note
A Hadamard matrix is a square matrix with r rows and columns that has the property
Hr0 Hr = rIr
where Ir is the identity matrix of order r. Generating a Hadamard matrix with order r = 2p is
easily accomplished. Start with a Hadamard matrix of order 2 (H2 ), and build your Hr by repeatedly
applying Kronecker products with H2 . Here is the Stata code to generate the Hadamard matrix for
the previous example.
matrix h2 = (-1, 1 \ 1, 1)
matrix h32 = h2
forvalues i = 1/4 {
matrix h32 = h2 # h32
}
svy brr consumes Hadamard matrices from left to right, so it is best to make sure that r is greater
than the number of strata and that the last column is the one consisting of all 1s. This will ensure
full orthogonal balance according to Wolter (2007).
Stored results
In addition to the results documented in [SVY] svy, svy brr stores the following in e():
Scalars
e(fay) Fay’s adjustment
Macros
e(cmd) same as e(cmdname) or brr
e(vce) brr
e(brrweight) brrweight() variable list
Matrices
e(b brr) BRR means
e(V) BRR variance estimates
When exp list is b, svy brr will also carry forward most of the results already in e() from
command.

See [SVY] variance estimation for details regarding BRR variance estimation.
References
Judkins, D. R. 1990. Fay’s method for variance estimation. Journal of Official Statistics 6: 223–239.
Also see
Title
svy estimation — Estimation commands for survey data
Description Menu Remarks and examples References Also see
Description
Survey data analysis in Stata is essentially the same as standard data analysis. The standard syntax
applies; you just need to also remember the following:
• Use svyset to identify the survey design characteristics.
• Prefix the estimation commands with “svy:”.
For example,
. use http://www.stata-press.com/data/r14/nhanes2f
. svy: regress zinc age c.age#c.age weight female black orace rural
See [SVY] svyset and [SVY] svy.
The following estimation commands support the svy prefix:
Descriptive statistics
mean [R] mean — Estimate means
proportion [R] proportion — Estimate proportions
ratio [R] ratio — Estimate ratios
total [R] total — Estimate totals
Linear regression models

churdle [R] churdle — Cragg hurdle regression
cnsreg [R] cnsreg — Constrained linear regression
etregress [TE] etregress — Linear regression with endogenous treatment effects
glm [R] glm — Generalized linear models
intreg [R] intreg — Interval regression
nl [R] nl — Nonlinear least-squares estimation
regress [R] regress — Linear regression
tobit [R] tobit — Tobit regression
truncreg [R] truncreg — Truncated regression
Structural equation models

sem [SEM] sem — Structural equation model estimation command
gsem [SEM] gsem — Generalized structural equation model estimation command
Survival-data regression models

stcox [ST] stcox — Cox proportional hazards model
streg [ST] streg — Parametric survival models
90
svy estimation — Estimation commands for survey data 91
Binary-response regression models

biprobit [R] biprobit — Bivariate probit regression
cloglog [R] cloglog — Complementary log-log regression
hetprobit [R] hetprobit — Heteroskedastic probit model
logistic [R] logistic — Logistic regression, reporting odds ratios
logit [R] logit — Logistic regression, reporting coefficients
probit [R] probit — Probit regression
scobit [R] scobit — Skewed logistic regression
Discrete-response regression models

clogit [R] clogit — Conditional (fixed-effects) logistic regression
mlogit [R] mlogit — Multinomial (polytomous) logistic regression
mprobit [R] mprobit — Multinomial probit regression
ologit [R] ologit — Ordered logistic regression
oprobit [R] oprobit — Ordered probit regression
slogit [R] slogit — Stereotype logistic regression
Fractional-response regression models

betareg [R] betareg — Beta regression
fracreg [R] fracreg — Fractional response regression
Poisson regression models

cpoisson [R] cpoisson — Censored Poisson regression
etpoisson [TE] etpoisson — Poisson regression with endogenous treatment effects
gnbreg Generalized negative binomial regression in [R] nbreg
nbreg [R] nbreg — Negative binomial regression
poisson [R] poisson — Poisson regression
tnbreg [R] tnbreg — Truncated negative binomial regression
tpoisson [R] tpoisson — Truncated Poisson regression
zinb [R] zinb — Zero-inflated negative binomial regression
zip [R] zip — Zero-inflated Poisson regression
Instrumental-variables regression models

ivprobit [R] ivprobit — Probit model with continuous endogenous covariates
ivregress [R] ivregress — Single-equation instrumental-variables regression
ivtobit [R] ivtobit — Tobit model with continuous endogenous covariates
Regression models with selection

heckman [R] heckman — Heckman selection model
heckoprobit [R] heckoprobit — Ordered probit model with sample selection
heckprobit [R] heckprobit — Probit model with sample selection
92 svy estimation — Estimation commands for survey data
Multilevel mixed-effects models

mecloglog [ME] mecloglog — Multilevel mixed-effects complementary log-log regression
meglm [ME] meglm — Multilevel mixed-effects generalized linear model
melogit [ME] melogit — Multilevel mixed-effects logistic regression
menbreg [ME] menbreg — Multilevel mixed-effects negative binomial regression
meologit [ME] meologit — Multilevel mixed-effects ordered logistic regression
meoprobit [ME] meoprobit — Multilevel mixed-effects ordered probit regression
mepoisson [ME] mepoisson — Multilevel mixed-effects Poisson regression
meprobit [ME] meprobit — Multilevel mixed-effects probit regression
mestreg [ME] mestreg — Multilevel mixed-effects parametric survival models
Item response theory

irt 1pl [IRT] irt 1pl — One-parameter logistic model
irt 2pl [IRT] irt 2pl — Two-parameter logistic model
irt 3pl [IRT] irt 3pl — Three-parameter logistic model
irt grm [IRT] irt grm — Graded response model
irt nrm [IRT] irt nrm — Nominal response model
irt pcm [IRT] irt pcm — Partial credit model
irt rsm [IRT] irt rsm — Rating scale model
irt hybrid [IRT] irt hybrid — Hybrid IRT models
Menu
Statistics > Survey data analysis > ...
Dialog boxes for all statistical estimators that support svy can be found on the above menu path.
In addition, you can access survey data estimation from standard dialog boxes on the SE/Robust or
SE/Cluster tab.

Overview of survey analysis in Stata
Regression models
Health surveys
Overview of survey analysis in Stata

Many Stata commands estimate the parameters of a process or population by using sample data.
For example, mean estimates means, ratio estimates ratios, regress fits linear regression models,
poisson fits Poisson regression models, and logistic fits logistic regression models. Some of these
estimation commands support the svy prefix, that is, they may be prefixed by svy: to produce results
appropriate for complex survey data. Whereas poisson is used with standard, nonsurvey data, svy:
poisson is used with survey data. In what follows, we refer to any estimation command not prefixed
by svy: as the standard command. A standard command prefixed by svy: is referred to as a svy
command.
Most standard commands (and all standard commands supported by svy) allow pweights and
the vce(cluster clustvar) option, where clustvar corresponds to the PSU variable that you svyset.
If your survey data exhibit only sampling weights or first-stage clusters (or both), you can get by
with using the standard command with pweights, vce(cluster clustvar), or both. Your parameter
estimates will always be identical to those you would have obtained from the svy command, and the
standard command uses the same robust (linearization) variance estimator as the svy command with
a similarly svyset design.
Most standard commands are also fit using maximum likelihood. When used with independently
distributed, nonweighted data, the likelihood to be maximized reflects the joint probability distribution
of the data given the chosen model. With complex survey data, however, this interpretation of the
likelihood is no longer valid, because survey data are weighted, not independently distributed, or
both. Yet for survey data, (valid) parameter estimates for a given model can be obtained using the
associated likelihood function with appropriate weighting. Because the probabilistic interpretation no
longer holds, the likelihood here is instead called a pseudolikelihood, but likelihood-ratio tests are no
longer valid. See Skinner (1989, sec. 3.4.4) for a discussion of maximum pseudolikelihood estimators.
Here we highlight the other features of svy commands:

• svy commands handle stratified sampling, but none of the standard commands do. Because
stratification usually makes standard errors smaller, ignoring stratification is usually conservative.
So not using svy with stratified sample data is not a terrible thing to do. However, to get the
smallest possible “honest” standard-error estimates for stratified sampling, use svy.
• svy commands use t statistics with n − L degrees of freedom to test the significance of
coefficients, where n is the total number of sampled PSUs (clusters) and L is the number of
strata in the first stage. Some of the standard commands use t statistics, but most use z statistics.
If the standard command uses z statistics for its standard variance estimator, then it also uses
z statistics with the robust (linearization) variance estimator. Strictly speaking, t statistics are
appropriate with the robust (linearization) variance estimator; see [P] robust for the theoretical
rationale. But, using z rather than t statistics yields a nontrivial difference only when there is
a small number of clusters (< 50). If a regression model command uses t statistics and the
vce(cluster clustvar) option is specified, then the degrees of freedom used is the same as
that of the svy command (in the absence of stratification).
• svy commands produce an adjusted Wald test for the model test, and test can be used to
produce adjusted Wald tests for other hypotheses after svy commands. Only unadjusted Wald
tests are available if the svy prefix is not used. The adjustment can be important when the
degrees of freedom, n − L, is small relative to the dimension of the test. (If the dimension is
one, then the adjusted and unadjusted Wald tests are identical.) This fact along with the point
made in the second bullet make using the svy command important if the number of sampled
PSUs (clusters) is small (< 50).
• svy: regress differs slightly from regress and svy: ivregress differs slightly from
ivregress in that they use different multipliers for the variance estimator. regress and
ivregress (when the small option is specified) use a multiplier of {(N − 1)/(N −k)}{n/(n−
1)}, where N is the number of observations, n is the number of clusters (PSUs), and k is
the number of regressors including the constant. svy: regress and svy: ivregress use
n/(n − 1) instead. Thus they produce slightly different standard errors. The (N − 1)/(N − k)
is ad hoc and has no rigorous theoretical justification; hence, the purist svy commands do not
use it. The svy commands tacitly assume that N k . If (N − 1)/(N − k) is not close to 1,
you may be well advised to use regress or ivregress so that some punishment is inflicted on
your variance estimates. Maximum likelihood estimators in Stata (for example, logit) do no
such adjustment but rely on the sensibilities of the analyst to ensure that N is reasonably larger
than k . Thus the maximum pseudolikelihood estimators (for example, svy: logit) produce
the same standard errors as the corresponding maximum likelihood commands (for example,
logit), but p-values are slightly different because of the point made in the second bullet.
• svy commands can produce proper estimates for subpopulations by using the subpop() option.
Using an if restriction with svy or standard commands can yield incorrect standard-error esti-
mates for subpopulations. Often an if restriction will yield the same standard error as subpop();
most other times, the two standard errors will be slightly different; but sometimes — usually for
thinly sampled subpopulations — the standard errors can be appreciably different. Hence, the
svy command with the subpop() option should be used to obtain estimates for thinly sampled
subpopulations. See [SVY] subpopulation estimation for more information.
• svy commands handle zero sampling weights properly. Standard commands ignore any obser-
vation with a weight of zero. Usually, this will yield the same standard errors, but sometimes
they will differ. Sampling weights of zero can arise from various postsampling adjustment
procedures. If the sum of weights for one or more PSUs is zero, svy and standard commands
will produce different standard errors, but usually this difference is very small.
• You can svyset iweights and let these weights be negative. Negative sampling weights can
arise from various postsampling adjustment procedures. If you want to use negative sampling
weights, then you must svyset iweights instead of pweights; no standard command will
allow negative sampling weights.
• The svy commands compute finite population corrections (FPCs).
• After a svy command, estat effects will compute the design effects DEFF and DEFT and
the misspecification effects MEFF and MEFT.
• svy commands can perform variance estimation that accounts for multiple stages of clustered
sampling.
• svy commands can perform variance estimation that accounts for poststratification adjustments
to the sampling weights.
• Some standard options are not allowed with the svy prefix. For example, vce() and weights
cannot be specified when using the svy prefix because svy is already using the variance
estimation and sampling weights identified by svyset. Some options are not allowed with
survey data because they would be statistically invalid, such as noskip for producing optional
likelihood-ratio tests. Other options are not allowed because they change how estimation results
are reported (for example, nodisplay, first, plus) or are not compatible with svy’s variance
estimation methods (for example, irls, mse1, hc2, hc3).
• Estimation results are presented in the standard way, except that svy has its own table header:
In addition to the sample size, model test, and R2 (if present in the output from the standard
command), svy will also report the following information in the header:
a. number of strata and PSUs
b. number of poststrata, if specified to svyset
c. population size estimate
d. subpopulation sizes, if the subpop() option was specified
e. design degrees of freedom
Use svy: mean, svy: ratio, svy: proportion, and svy: total to estimate finite population and
subpopulation means, ratios, proportions, and totals, respectively. You can also estimate standardized
means, ratios, and proportions for survey data; see [SVY] direct standardization. Estimates for multiple
subpopulations can be obtained using the over() option; see [SVY] subpopulation estimation.
Example 1
Suppose that we need to estimate the average birthweight for the population represented by the
National Maternal and Infant Health Survey (NMIHS) (Gonzalez, Krauss, and Scott 1992).
First, we gather the survey design information.
• Primary sampling units are mothers; that is, PSUs are individual observations — there is no
separate PSU variable.
• The finalwgt variable contains the sampling weights.
• The stratan variable identifies strata.
• There is no variable for the finite population correction.
Then we use svyset to identify the variables for sampling weights and stratification.
. svyset [pweight=finwgt], strata(stratan)
pweight: finwgt
VCE: linearized
Strata 1: stratan
FPC 1: <zero>
Now we can use svy: mean to estimate the average birthweight for our population.
. svy: mean birthwgt
Design df = 9,940
Linearized
birthwgt 3355.452 6.402741 3342.902 3368.003
From these results, we are 95% confident that the mean birthweight for our population is between
3,343 and 3,368 grams.
Regression models
As exhibited in the table at the beginning of this manual entry, many of Stata’s regression model
commands support the svy prefix. If you know how to use one of these commands with standard
data, then you can also use the corresponding svy command with your survey data.
Example 2
Let’s model the incidence of high blood pressure with a dataset from the Second National Health and
Nutrition Examination Survey (NHANES II) (McDowell et al. 1981). The survey design characteristics
are already svyset, so we will just replay them.
. svyset
pweight: finalwgt
VCE: linearized
Strata 1: strata
SU 1: psu
FPC 1: <zero>
Now we can use svy: logistic to model the incidence of high blood pressure as a function of
height, weight, age, and sex (using the female indicator variable).
. svy: logistic highbp height weight age female
Design df = 31
F( 4, 28) = 368.33
Prob > F = 0.0000
Linearized
height .9657022 .0051511 -6.54 0.000 .9552534 .9762654

weight 1.053023 .0026902 20.22 0.000 1.047551 1.058524
age 1.050059 .0019761 25.96 0.000 1.046037 1.054097
female .6272129 .0368195 -7.95 0.000 .5564402 .706987
_cons .716868 .6106878 -0.39 0.699 .1261491 4.073749
The odds ratio for the female predictor is 0.63 (rounded to two decimal places) and is significantly
less than 1. This finding implies that females have a lower incidence of high blood pressure than do
males.
Here we use the subpop() option to model the incidence of high blood pressure in the subpopulation
identified by the female variable.
. svy, subpop(female): logistic highbp height weight age

Design df = 31
F( 3, 29) = 227.53
Prob > F = 0.0000
Linearized
height .9630557 .0074892 -4.84 0.000 .9479018 .9784518

weight 1.053197 .003579 15.25 0.000 1.045923 1.060522
age 1.066112 .0034457 19.81 0.000 1.059107 1.073163
_cons .3372393 .4045108 -0.91 0.372 .029208 3.893807
Because the odds ratio for the age predictor is significantly greater than 1, we can conclude that
older females are more likely to have high blood pressure than are younger females.
Health surveys
There are many sources of bias when modeling the association between a disease and its risk
factors (Korn, Graubard, and Midthune 1997; Korn and Graubard 1999, sec. 3.7). In cross-sectional
health surveys, inference is typically restricted to the target population as it stood when the data were
collected. This type of survey cannot capture the fact that participants may change their habits over
time. Some health surveys collect data retrospectively, relying on the participants to recall the status
of risk factors as they stood in the past. This type of survey is vulnerable to recall bias.
Longitudinal surveys collect data over time, monitoring the survey participants over several years.
Although the above biases are minimized, analysts are still faced with some tough choices/situations
when modeling time-to-event data. For example:
• Time scale. When studying cancer, should we measure the time scale by using the participant’s
age or the initial date from which data were collected?
• Time-varying covariates. Were all relevant risk factors sampled over time, or do we have only
the baseline measurement?
• Competing risks. When studying mortality, do we have the data specific to cause of death?
Binder (1983) provides the foundation for fitting most of the common parametric models by using
survey data. Similarly, Lin and Wei (1989) provide the foundational theory for robust inference by
using the proportional hazards model. Binder (1992) describes how to estimate standard errors for
the proportional hazards model from survey data, and Lin (2000) provides a rigorous justification for
Binder’s method. Korn and Graubard (1999) discuss many aspects of model fitting by using data from
health surveys. O’Donnell et al. (2008, chap. 10) use Stata survey commands to perform multivariate
analysis using health survey data.
Example 3: Cox’s proportional hazards model

Suppose that we want to model the incidence of lung cancer by using three risk factors: smoking
status, sex, and place of residence. Our dataset comes from a longitudinal health survey: the First
National Health and Nutrition Examination Survey (NHANES I) (Miller 1973; Engel et al. 1978) and its
1992 Epidemiologic Follow-up Study (NHEFS) (Cox et al. 1997); see the National Center for Health
Statistics website at http://www.cdc.gov/nchs/. We will be using data from the samples identified by
NHANES I examination locations 1–65 and 66–100; thus we will svyset the revised pseudo-PSU and
strata variables associated with these locations. Similarly, our pweight variable was generated using
the sampling weights for the nutrition and detailed samples for locations 1–65 and the weights for
the detailed sample for locations 66–100.
. use http://www.stata-press.com/data/r14/nhefs
. svyset psu2 [pw=swgt2], strata(strata2)
pweight: swgt2
VCE: linearized
Strata 1: strata2
SU 1: psu2
FPC 1: <zero>
The lung cancer information was taken from the 1992 NHEFS interview data. We use the participants’
ages for the time scale. Participants who never had lung cancer and were alive for the 1992 interview
were considered censored. Participants who never had lung cancer and died before the 1992 interview
were also considered censored at their age of death.
. stset age_lung_cancer [pw=swgt2], fail(lung_cancer)
failure event: lung_cancer != 0 & lung_cancer < .
obs. time interval: (0, age_lung_cancer]
exit on or before: failure
14407 total observations

5126 event time missing (age_lung_cancer>=.) PROBABLE ERROR
9281 observations remaining, representing

83 failures in single-record/single-failure data
599691 total analysis time at risk and under observation
at risk from t = 0
earliest observed entry t = 0
last observed exit t = 97
Although stset warns us that it is a “probable error” to have 5,126 observations with missing event
times, we can verify from the 1992 NHEFS documentation that there were indeed 9,281 participants
with complete information.
For our proportional hazards model, we pulled the risk factor information from the NHANES I and
1992 NHEFS datasets. Smoking status was taken from the 1992 NHEFS interview data, but we filled
in all but 132 missing values by using the general medical history supplement data in NHANES I.
Smoking status is represented by separate indicator variables for former smokers and current smokers;
the base comparison group is nonsmokers. Sex was determined using the 1992 NHEFS vitality data
and is represented by an indicator variable for males. Place-of-residence information was taken from
the medical history questionnaire in NHANES I and is represented by separate indicator variables for
rural and heavily populated (more than 1 million people) urban residences; the base comparison group
is urban residences with populations of fewer than 1 million people.
. svy: stcox former_smoker smoker male urban1 rural

Design df = 70
F( 5, 66) = 14.07
Prob > F = 0.0000
Linearized
former_smoker 2.788113 .6205102 4.61 0.000 1.788705 4.345923

smoker 7.849483 2.593249 6.24 0.000 4.061457 15.17051
male 1.187611 .3445315 0.59 0.555 .6658757 2.118142
urban1 .8035074 .3285144 -0.54 0.594 .3555123 1.816039
rural 1.581674 .5281859 1.37 0.174 .8125799 3.078702
From the above results, we can see that both former and current smokers have a significantly
higher risk for developing lung cancer than do nonsmokers.
Technical note
In the previous example, we specified a sampling weight variable in the calls to both svyset
and stset. When the svy prefix is used with stcox and streg, it identifies the sampling weight
variable by using the data characteristics from both svyset and stset. svy will report an error if
the svyset pweight variable is different from the stset pweight variable. The svy prefix will use
the specified pweight variable, even if it is svyset but not stset. If a pweight variable is stset
but not svyset, svy will note that it will be using the stset pweight variable and then svyset it.
The standard st commands will not use the svyset pweight variable if it is not also stset.
Example 4: Multiple baseline hazards

We can assess the proportional-hazards assumption across the observed race categories for the
model fit in the previous example. The race information in our 1992 NHEFS dataset is contained in
the revised race variable. We will use stphplot to produce a log-log plot for each category of
revised race. As described in [ST] stcox PH-assumption tests, if the plotted lines are reasonably
parallel, the proportional-hazards assumption has not been violated. We will use the zero option to
reset the risk factors to their base comparison group.
. stphplot, strata(revised_race) adjust(former_smoker smoker male urban1 rural)

> zero legend(col(1))
failure _d: lung_cancer
analysis time _t: age_lung_cancer
10
−ln[−ln(Survival Probability)]
4 6 2 8
3.6 3.8 4 4.2 4.4 4.6

ln(analysis time)
revised_race = Aleut, Eskimo or American Indian

revised_race = Asian/Pacific Islander
revised_race = Black
revised_race = White
revised_race = Other
As we can see from the graph produced above, the lines for the black and white race categories
intersect. This indicates a violation of the proportional-hazards assumption, so we should consider
using separate baseline hazard functions for each race category in our model fit. We do this next, by
specifying strata(revised race) in our call to svy: stcox.
. svy: stcox former_smoker smoker male urban1 rural, strata(revised_race)
Design df = 70
F( 5, 66) = 13.95
Prob > F = 0.0000
Linearized
former_smoker 2.801797 .6280352 4.60 0.000 1.791761 4.381201

smoker 7.954921 2.640022 6.25 0.000 4.103709 15.42038
male 1.165724 .3390339 0.53 0.600 .6526527 2.082139
urban1 .784031 .3120525 -0.61 0.543 .3544764 1.73412
rural 1.490269 .5048569 1.18 0.243 .7582848 2.928851
Stratified by revised_race
References
Binder, D. A. 1983. On the variances of asymptotically normal estimators from complex surveys. International
Statistical Review 51: 279–292.
. 1992. Fitting Cox’s proportional hazards models for survey data. Biometrika 79: 139–147.
Cox, C. S., M. E. Mussolino, S. T. Rothwell, M. A. Lane, C. D. Golden, J. H. Madans, and J. J. Feldman. 1997.
Plan and operation of the NHANES I Epidemiologic Followup Study, 1992. In Vital and Health Statistics, series 1,
no. 35. Hyattsville, MD: National Center for Health Statistics.
Engel, A., R. S. Murphy, K. Maurer, and E. Collins. 1978. Plan and operation of the HANES I augmentation survey
of adults 25–74 years: United States 1974–75. In Vital and Health Statistics, series 1, no. 14. Hyattsville, MD:
National Center for Health Statistics.
Korn, E. L., B. I. Graubard, and D. Midthune. 1997. Time-to-event analysis of longitudinal follow-up of a survey:
Choice of time-scale. American Journal of Epidemiology 145: 72–80.
Lin, D. Y. 2000. On fitting Cox’s proportional hazards models to survey data. Biometrika 87: 37–47.
Lin, D. Y., and L. J. Wei. 1989. The robust inference for the Cox proportional hazards model. Journal of the American
Miller, H. W. 1973. Plan and operation of the Health and Nutrition Examination Survey: United States 1971–1973.
Hyattsville, MD: National Center for Health Statistics.
O’Donnell, O., E. van Doorslaer, A. Wagstaff, and M. Lindelow. 2008. Analyzing Health Equity Using Household
Survey Data: A Guide to Techniques and Their Implementation. Washington, DC: The World Bank.
Also see
[SVY] estat — Postestimation statistics for survey data
[SVY] direct standardization — Direct standardization of means, proportions, and ratios
Title
svy jackknife — Jackknife estimation for survey data

References Also see
Description
svy jackknife performs jackknife estimation of specified statistics (or expressions) for a Stata
command or a user-written command. The command is executed once for each replicate using sampling
weights that are adjusted according to the jackknife methodology. Any Stata estimation command
listed in [SVY] svy estimation may be used with svy jackknife. User-written commands that meet
the requirements in [P] program properties may also be used.
Quick start
Estimate population mean of v1 using jackknife standard-error estimates with sampling weight wvar1
and sampling units identified by su1
svyset su1 [pweight = wvar1]
svy jackknife _b: mean v1
Same as above, but with jackknife replication weights in variables with prefix rwvar
svyset [pweight = wvar1], vce(jackknife) jkrweight(rwvar*)
svy: mean v1
Jackknife standard error of the difference between the means of v2 and v3 using either svyset
command above
svy jackknife (_b[v2]-_b[v3]): mean v2 v3
svy jackknife diff=(_b[v2]-_b[v3]), saving(mydata): mean v2 v3
Menu
Statistics > Survey data analysis > Resampling > Jackknife estimation
102
svy jackknife — Jackknife estimation for survey data 103
Syntax

svy jackknife exp list , svy options jackknife options eform option : command

if/in

Reporting
jackknife options Description

Main
eclass number of observations is in e(N)
rclass number of observations is in r(N)
n(exp) specify exp that evaluates to number of observations used
Options

keep keep pseudovalues
Reporting
trace trace command
title(text) use text as title for jackknife results
Advanced
104 svy jackknife — Jackknife estimation for survey data
unless command has the svyj program property, in which case exp list defaults to b; see [P] program
elist
eexp
(exp)
eexp is specname
[eqno]specname
specname is b
b[]
se
se[]
eqno is ##
name

Options

Main
eclass, rclass, and n(exp) specify where command stores the number of observations on which
it based the calculated results. We strongly advise you to specify one of these options.
eclass specifies that command store the number of observations in e(N).
rclass specifies that command store the number of observations in r(N).
n(exp) allows you to specify an expression that evaluates to the number of observations used.
Specifying n(r(N)) is equivalent to specifying the rclass option. Specifying n(e(N)) is equiv-
alent to specifying the eclass option. If command stores the number of observations in r(N1),
specify n(r(N1)).
If you specify none of these options, svy jackknife will assume eclass or rclass depending
upon which of e(N) and r(N) is not missing (in that order). If both e(N) and r(N) are missing,
svy jackknife assumes that all observations in the dataset contribute to the calculated result. If
that assumption is incorrect, then the reported standard errors will be incorrect. For instance, say
that you specify
. svy jackknife coef=_b[x2]: myreg y x1 x2 x3
where myreg uses e(n) instead of e(N) to identify the number of observations used in calculations.
Further assume that observation 42 in the dataset has x3 equal to missing. The 42nd observation
plays no role in obtaining the estimates, but svy jackknife has no way of knowing that and will
use the wrong N . If, on the other hand, you specify
. svy jackknife coef=_b[x2], n(e(n)): myreg y x1 x2 x3
Then svy jackknife will notice that observation 42 plays no role. The n(e(n)) option is
specified because myreg is an estimation command, but it stores the number of observations
used in e(n) (instead of the standard e(N)). When svy jackknife runs the regression omitting
the 42nd observation, svy jackknife will observe that e(n) has the same value as when svy
jackknife previously ran the regression by using all the observations. Thus svy jackknife will
know that myreg did not use the observation.

Options

See [P] postfile.
dialog box.
keep specifies that new variables be added to the dataset containing the pseudovalues of the requested
statistics. For instance, if you typed
. svy jackknife coef=_b[x2], eclass keep: regress y x1 x2 x3
Then the new variable coef would be added to the dataset containing the pseudovalues for b[x2].
Let b be defined as the value of b[x2] when all observations are used to fit the model, and let
b(j) be the value when the j th observation is omitted. The pseudovalues are defined as
pseudovaluej = N × {b − b(j)} + b(j)
where N is the number of observations used to produce b.

keep implies the nodrop option.
mse specifies that svy jackknife compute the variance by using deviations of the replicates from the
observed value of the statistics based on the entire dataset. By default, svy jackknife computes
the variance by using deviations of the pseudovalues from their mean.

Reporting
successful replication. A red ‘x’ is printed if command returns with an error, ‘e’ is printed if one
of the values in exp list is missing, ‘n’ is printed if the sample size is not correct, and a yellow
‘s’ is printed if the dropped sampling unit is outside the subpopulation sample.
option.
option.
title(text) specifies a title to be displayed above the table of jackknife results; the default title is
“Jackknife results”.

Advanced

The jackknife is
• an alternative, first-order unbiased estimator for a statistic;
• a data-dependent way to calculate the standard error of the statistic and to obtain significance
levels and confidence intervals; and
• a way of producing measures called pseudovalues for each observation, reflecting the observation’s
influence on the overall statistic.
The idea behind the simplest form of the jackknife—the one implemented in [R] jackknife —is to
repeatedly calculate the statistic in question, each time omitting just one of the dataset’s observations.
Assume that our statistic of interest is the sample mean. Let yj be the j th observation of our data
on some measurement y , where j = 1, . . . , N and N is the sample size. If y is the sample mean of
y using the entire dataset and y (j) is the mean when the j th observation is omitted, then
(N − 1) y (j) + yj
y=
N
Solving for yj , we obtain
yj = N y − (N − 1) y (j)
These are the pseudovalues that svy: jackknife calculates. To move this discussion beyond the
sample mean, let θb be the value of our statistic (not necessarily the sample mean) using the entire
dataset, and let θb(j) be the computed value of our statistic with the j th observation omitted. The
pseudovalue for the j th observation is
θbj∗ = N θb − (N − 1) θb(j)
The mean of the pseudovalues is the alternative, first-order unbiased estimator mentioned above,
and the standard error of the mean of the pseudovalues is an estimator for the standard error of θb
(Tukey 1958, Shao and Tu 1995).
When the jackknife is applied to survey data, primary sampling units (PSUs) are omitted instead
of observations, N is the number of PSUs instead of the sample size, and the sampling weights are
adjusted owing to omitting PSUs; see [SVY] variance estimation for more details.
Because of privacy concerns, many public survey datasets contain jackknife replication-weight
variables instead of variables containing information on the PSUs and strata. These replication-weight
variables are the adjusted sampling weights, and there is one replication-weight variable for each
omitted PSU.
Example 1: Jackknife with information on PSUs and strata

Suppose that we were interested in a measure of association between the weight and height of
individuals in the population represented by the NHANES II data (McDowell et al. 1981). To measure
the association, we will use the slope estimate from a linear regression of weight on height. We
also use svy jackknife to estimate the variance of the slope.
. svyset
pweight: finalwgt
VCE: linearized
Strata 1: strata
SU 1: psu
FPC 1: <zero>
. svy jackknife slope = _b[height]: regress weight height
Jackknife replications (62)
1 2 3 4 5
.................................................. 50
............
Linear regression
Replications = 62
Design df = 31
command: regress weight height
slope: _b[height]
n(): e(N)
Jackknife
slope .8014753 .0160281 50.00 0.000 .7687858 .8341648
Example 2: Jackknife replicate-weight variables

nhanes2jknife.dta is a privacy-conscious dataset equivalent to nhanes2.dta; all the variables
and values remain, except that strata and psu are replaced with jackknife replicate-weight variables.
The replicate-weight variables are already svyset, and the default method for variance estimation is
vce(jackknife).
. use http://www.stata-press.com/data/r14/nhanes2jknife
. svyset
pweight: finalwgt
VCE: jackknife
MSE: off
jkrweight: jkw_1 jkw_2 jkw_3 jkw_4 jkw_5 jkw_6 jkw_7 jkw_8 jkw_9 jkw_10
jkw_11 jkw_12 jkw_13 jkw_14 jkw_15 jkw_16 jkw_17 jkw_18 jkw_19
jkw_56 jkw_57 jkw_58 jkw_59 jkw_60 jkw_61 jkw_62
Strata 1: <one>
FPC 1: <zero>
Here we perform the same analysis as in the previous example, using jackknife replication weights.
. svy jackknife slope = _b[height], nodots: regress weight height
Linear regression
Replications = 62
Design df = 31
slope: _b[height]
Jackknife
slope .8014753 .0160281 50.00 0.000 .7687858 .8341648
The mse option causes svy jackknife to use the MSE form of the jackknife variance estimator.
This variance estimator will tend to be larger than the previous because of the addition of the familiar
squared bias term in the MSE; see [SVY] variance estimation for more details. The header for the
column of standard errors in the table of results is Jknife * for the jackknife variance estimator,
which uses the MSE formula.
. svy jackknife slope = _b[height], mse nodots: regress weight height
Linear regression
Replications = 62
Design df = 31
slope: _b[height]
Jknife *
slope .8014753 .0160284 50.00 0.000 .7687852 .8341654

Stored results
In addition to the results documented in [SVY] svy, svy jackknife stores the following in e():
Scalars
Macros
e(cmd) same as e(cmdname) or jackknife
e(vce) jackknife
e(jkrweight) jkrweight() variable list
Matrices
e(b jk) jackknife means
e(V) jackknife variance estimates
When exp list is b, svy jackknife will also carry forward most of the results already in e() from
command.

See [SVY] variance estimation for details regarding jackknife variance estimation.
References
Tukey, J. W. 1958. Bias and confidence in not-quite large samples. Abstract in Annals of Mathematical Statistics 29:
614.
Also see
[R] jackknife — Jackknife estimation
Title
svy postestimation — Postestimation tools for svy
Postestimation commands predict margins

Remarks and examples References Also see
Postestimation commands
The following postestimation commands are available after svy:
command Description
contrast contrasts and ANOVA-style joint tests of estimates
estat (svy) postestimation statistics for survey data
estimates cataloging estimation results
lincom point estimates, standard errors, testing, and inference for linear
combinations of coefficients
margins marginal means, predictive margins, marginal effects, and average marginal
effects
marginsplot graph the results from margins (profile plots, interaction plots, etc.)
nlcom point estimates, standard errors, testing, and inference for nonlinear
combinations of coefficients
predict predictions, residuals, influence statistics, and other diagnostic measures
predictnl point estimates, standard errors, testing, and inference for generalized
predictions
pwcompare pairwise comparisons of estimates
suest seemingly unrelated estimation
test Wald tests of simple and composite linear hypotheses
testnl Wald tests of nonlinear hypotheses
See [SVY] estat.
predict
The syntax of predict (and even if predict is allowed) after svy depends on the command used
with svy. Specifically, predict is not allowed after svy: mean, svy: proportion, svy: ratio,
svy: tabulate, or svy: total.
margins
The syntax of margins (and even if margins is allowed) after svy depends on the command used
with svy. Specifically, margins is not allowed after svy: mean, svy: proportion, svy: ratio,
svy: tabulate, or svy: total.
110
svy postestimation — Postestimation tools for svy 111

What follows are some examples of applications of postestimation commands using survey data.
The examples are meant only to introduce the commands in a survey context and explore a few of
the possibilities for postestimation analysis. See the individual entries for each command in the Base
Reference Manual for complete syntax and many more examples.
Example 1: Linear and nonlinear combinations

lincom will display an estimate of a linear combination of parameters, along with its standard
error, a confidence interval, and a test that the linear combination is zero. nlcom will do likewise for
nonlinear combinations of parameters.
lincom is commonly used to compute the differences of two subpopulation means. For example,
suppose that we wish to estimate the difference of zinc levels in white males versus black males
in the population represented by the NHANES II data (McDowell et al. 1981). Because the survey
design characteristics are already svyset in nhanes2.dta, we only need to generate a variable for
identifying the male subpopulation before using svy: mean.
. generate male = (sex == 1)
. svy, subpop(male): mean zinc, over(race)
Design df = 31
White: race = White
Black: race = Black
Other: race = Other
Linearized
zinc
White 91.15725 .541625 90.0526 92.2619
Black 88.269 1.208336 85.80458 90.73342
Other 85.54716 2.608974 80.22612 90.8682
Then we run lincom to estimate the difference of zinc levels between the two subpopulations.
. lincom [zinc]White - [zinc]Black
( 1) [zinc]White - [zinc]Black = 0
Mean Coef. Std. Err. t P>|t| [95% Conf. Interval]
(1) 2.888249 1.103999 2.62 0.014 .6366288 5.139868
The t statistic and its p-value give a survey analysis equivalent of a two-sample t test.
lincom and nlcom can be used after any of the estimation commands described in [SVY] svy
estimation. lincom can, for example, display results as odds ratios after svy: logit and can be
used to compute odds ratios for one covariate group relative to another. nlcom can display odds
112 svy postestimation — Postestimation tools for svy
ratios, as well, and allows more general nonlinear combinations of the parameters. See [R] lincom
and [R] nlcom for full details. Also see Eltinge and Sribney (1996) for an earlier implementation of
lincom for survey data.
Finally, lincom and nlcom operate on the estimated parameters only. To obtain estimates and
inference for functions of the parameters and of the data, such as for an exponentiated linear predictor
or a predicted probability of success from a logit model, use predictnl; see [R] predictnl.
Example 2: Quadratic terms

From example 2 in [SVY] svy estimation, we modeled the incidence of high blood pressure as a
function of height, weight, age, and sex (using the female indicator variable). Here we also include
c.age#c.age, a squared term for age.
. use http://www.stata-press.com/data/r14/nhanes2d, clear
. svy: logistic highbp height weight age c.age#c.age female
Design df = 31
F( 5, 27) = 284.33
Prob > F = 0.0000
Linearized
height .9656421 .005163 -6.54 0.000 .9551693 .9762298

weight 1.052911 .0026433 20.54 0.000 1.047534 1.058316
age 1.055829 .0133575 4.29 0.000 1.028935 1.083426
c.age#c.age .9999399 .0001379 -0.44 0.666 .9996588 1.000221
female .6260988 .0364945 -8.03 0.000 .5559217 .7051347

_cons .653647 .5744564 -0.48 0.632 .108869 3.924483
Because our model includes a quadratic in the age variable, the peak incidence of high blood
pressure with respect to age will occur at - b[age]/(2* b[c.age#c.age]), which we can estimate,
along with its standard error, using nlcom.
. nlcom peak: -_b[age]/(2*_b[c.age#c.age])
peak: -_b[age]/(2*_b[c.age#c.age])
highbp Coef. Std. Err. z P>|z| [95% Conf. Interval]
peak 452.0979 933.512 0.48 0.628 -1377.552 2281.748
Or we can use testnl to test that the peak incidence of high blood pressure in the population is
70 years.
. testnl -_b[age]/(2*_b[c.age#c.age]) = 70
(1) -_b[age]/(2*_b[c.age#c.age]) = 70
chi2(1) = 0.17
Prob > chi2 = 0.6823
These data do not reject our theory. testnl allows multiple hypotheses to be tested jointly and
applies the degrees-of-freedom adjustment for survey results; see [R] testnl.
Example 3: Predictive margins

Changing our logistic regression for high blood pressure slightly, we add a factor variable for the
levels of race. Level 1 of race represents whites, level 2 represents blacks, and level 3 represents
others. We also specify that female is a factor variable, which does not change its coefficient but
does increase its functionality with some postestimation commands.
. svy: logistic highbp height weight age c.age#c.age i.female i.race, baselevels
Design df = 31
F( 7, 25) = 230.16
Prob > F = 0.0000
Linearized
height .9675961 .0052361 -6.09 0.000 .9569758 .9783343

weight 1.052683 .0026091 20.72 0.000 1.047376 1.058018
age 1.056628 .0134451 4.33 0.000 1.029559 1.084408
c.age#c.age .9999402 .0001382 -0.43 0.668 .9996585 1.000222
female
0 1 (base)
1 .6382331 .0377648 -7.59 0.000 .5656774 .720095
race
White 1 (base)
Black 1.422003 .1556023 3.22 0.003 1.137569 1.777557
Other 1.63456 .2929919 2.74 0.010 1.13405 2.355971
_cons .4312846 .378572 -0.96 0.345 .0719901 2.583777
Our point estimates indicate that the odds of females having high blood pressure is about 64% of
the odds for men and that the odds of blacks having high blood pressure is about 1.4 times that of
whites. The odds ratios give us the relative effects of their covariates, but they do not give us any
sense of the absolute size of the effects. The odds ratio comparing blacks with whites is clearly large
and statistically significant, but does it represent a sizable change? One way to answer that question
is to explore the probabilities of high blood pressure from our fitted model. Let’s first look at the
predictive margins of the probability of high blood pressure for the three levels of race.
. margins race, vce(unconditional)

Predictive margins Number of obs = 10,351
Expression : Pr(highbp), predict()
Linearized
Margin Std. Err. t P>|t| [95% Conf. Interval]
race
White .3600722 .0150121 23.99 0.000 .3294548 .3906895
Black .4256413 .0211311 20.14 0.000 .3825441 .4687385
Other .4523404 .0311137 14.54 0.000 .3888836 .5157972
Because our response is a probability, these margins are sometimes called predicted marginal
proportions or model-adjusted risks. They let us compare the effect of our three racial groups while
controlling for the distribution of other covariates in the groups. Computationally, these predictive
margins are the weighted average of the predicted probabilities for each observation in the estimation
sample. The marginal probability for whites is the average probability, assuming that everyone in the
sample is white; the margin for blacks assumes that everyone is black; and the margin for others
assumes that everyone is something other than black or white.
There is a sizable difference in blood pressure between whites and blacks, with the marginal
probability of high blood pressure for whites being about 36% and that for blacks being almost 43%.
These are the adjusted probability levels. A more direct answer to our question about whether the odds
ratios represent a substantial effect requires looking at the differences of these marginal probabilities.
Researchers in the health-related sciences call such differences risk differences, whereas researchers
in the social sciences usually call them average marginal effects or average partial effects.
Regardless of terminology, we are interested in the difference in the probability of blacks having
high blood pressure as compared with whites, while adjusting for all other covariates in the model.
We request risk differences by specifying the variables of interest in a dydx() option.
. margins, vce(unconditional) dydx(race)
Average marginal effects Number of obs = 10,351
dy/dx w.r.t. : 2.race 3.race
Linearized
dy/dx Std. Err. t P>|t| [95% Conf. Interval]
race
Black .0655691 .0204063 3.21 0.003 .0239501 .1071881
Other .0922682 .0343809 2.68 0.012 .0221478 .1623886
Note: dy/dx for factor levels is the discrete change from the base level.
Looking in the column labeled dy/dx, we see that the risk difference between blacks and whites
is about 6.6% (0.6557). That is a sizable as well as significant difference.
Because they are population-weighted averages over the whole sample, these margins are estimates
of the population average risk differences. And because we specified the vce(unconditional)
option, their standard errors and confidence intervals can be used to make inferences about the
population average risk differences. See Methods and formulas in [R] margins for details.
We can also compute margins or risk differences for subpopulations. To compute risk differences
for the four subpopulations that are the regions of the United States—Northeast, Midwest, South,
and West—we add the over(region) option.
. margins, vce(unconditional) dydx(race) by(region)

Average marginal effects Number of obs = 10,351
dy/dx w.r.t. : 2.race 3.race
over : region
Linearized
dy/dx Std. Err. t P>|t| [95% Conf. Interval]
2.race
region
NE .0662951 .0207354 3.20 0.003 .0240051 .1085852
MW .065088 .0204357 3.19 0.003 .0234091 .106767
S .0663448 .0202173 3.28 0.003 .0251112 .1075783
W .0647221 .0203523 3.18 0.003 .0232134 .1062308
3.race
region
NE .093168 .0343919 2.71 0.011 .0230253 .1633106
MW .091685 .034247 2.68 0.012 .0218379 .1615322
S .0932303 .0345933 2.70 0.011 .0226769 .1637837
W .0912062 .034322 2.66 0.012 .021206 .1612063
Note: dy/dx for factor levels is the discrete change from the base level.
The differences in the covariate distributions across the regions have little effect on the risk
differences between blacks and whites, or between other races and whites.
Rather than explore the probabilities after logistic regression, we might have explored the hazards
or mean survival times after fitting a survival model. See [R] margins for many more applications of
margins.
Example 4: Predictive means with replication-based variance estimators

When performing estimations with linearized standard errors, we use the vce(unconditional)
option to compute marginal effects so that we can use the results to make inferences on the population.
margins with vce(unconditional) uses linearization to compute the unconditional variance of the
marginal means.
The vce(unconditional) option, therefore, cannot be used when a different variance estimation
method has been specified for the model. If you are using a replication-based method to estimate
the variance in your model, you may want to use this method to perform the variance estimation for
your margins as well. To do that, you can write a program that performs both your main estimation
and the computation of your margins and use the replication method with your program.
Continuing with the logistic example, we will see how to estimate the marginal means for race
by using the jackknife variance estimator. The program below accepts an argument that contains the
estimation command line. Notice that the program should accept the if qualifier and also weights.
In addition, the set buildfvinfo on command is included so that margins checks for estimability.
buildfvinfo is usually set on, but is set off because it increases the computation time when you
use replication methods; thus you need to set it on. The option post of margins posts the results
to e(b), so they can be used by svy jackknife.
program mymargins, eclass

version 14
syntax anything [if] [iw pw]
if "‘weight’" != "" {
local wgtexp "[‘weight’ ‘exp’]"
}
set buildfvinfo on
‘anything’ ‘if’ ‘wgtexp’
margins race, post
end
We can now type

. local mycmdline logistic highbp height weight age c.age#c.age i.race i.female
. quietly mymargins ‘mycmdline’
. svy jackknife _b: mymargins ‘mycmdline’
(running mymargins on estimation sample)
1 2 3 4 5
.................................................. 50
............
Predictive margins Number of obs = 10,351
Replications = 62
Jackknife
race
White .3600722 .0150128 23.98 0.000 .3294534 .390691
Black .4256413 .0211504 20.12 0.000 .3825048 .4687778
Other .4523404 .0322488 14.03 0.000 .3865684 .5181124
You can see that now the jackknife standard errors are being reported.
Example 5: Nonlinear predictions and their standard errors

Continuing with the NHANES II data, we fit a linear regression of log of blood lead level on age,
age-squared, gender, race, and region.
. svy: regress loglead age c.age#c.age i.female i.race i.region
Design df = 31
F( 8, 24) = 156.24
Prob > F = 0.0000
R-squared = 0.2379
Linearized
loglead Coef. Std. Err. t P>|t| [95% Conf. Interval]
age .0158388 .0027352 5.79 0.000 .0102603 .0214173
c.age#c.age -.0001464 .0000295 -4.96 0.000 -.0002066 -.0000862
1.female -.3655338 .0116157 -31.47 0.000 -.3892242 -.3418434
race
Black .178402 .0314173 5.68 0.000 .114326 .242478
Other -.0516952 .0402381 -1.28 0.208 -.1337614 .030371
region
MW -.02283 .0389823 -0.59 0.562 -.1023349 .0566749
S -.1685453 .056004 -3.01 0.005 -.2827662 -.0543244
W -.0362295 .0387508 -0.93 0.357 -.1152623 .0428032
_cons 2.440671 .0627987 38.86 0.000 2.312592 2.568749
Given that we modeled the natural log of the lead measurement, we can use predictnl to compute
the exponentiated linear prediction (in the original units of the lead variable), along with its standard
error.
. predictnl leadhat = exp(xb()) if e(sample), se(leadhat_se)
(5,403 missing values generated)
. sort lead leadhat
. generate showobs = inrange(_n,1,5) + inrange(_n,2501,2505) +
> inrange(_n,4945,4948)
. list lead leadhat leadhat_se age c.age#c.age if showobs, abbrev(10)
c.age#
lead leadhat leadhat_se age c.age
1. 2 9.419804 .5433255 29 841

2. 3 8.966098 .5301117 23 529
3. 3 9.046788 .5298448 24 576
4. 3 9.046788 .5298448 24 576
5. 3 9.27693 .5347956 27 729
2501. 13 16.88317 .7728783 37 1369

2502. 13 16.90057 2.296082 71 5041
2503. 13 16.90057 2.296082 71 5041
2504. 13 16.90237 1.501056 48 2304
2505. 13 16.90852 2.018708 60 3600
4945. 61 17.18581 2.052034 58 3364

4946. 64 15.08437 .647629 24 576
4947. 66 17.78698 1.641349 56 3136
4948. 80 16.85864 1.333927 42 1764
Example 6: Multiple-hypothesis testing

Joint-hypothesis tests can be performed after svy commands with the test command. Using the
results from the regression model fit in the previous example, we can use test to test the joint
significance of 2.region, 3.region, and 4.region. (1.region is the Northeast, 2.region is the
Midwest, 3.region is the South, and 4.region is the West.) We test the hypothesis that 2.region
= 0, 3.region = 0, and 4.region = 0.
. test 2.region 3.region 4.region
Adjusted Wald test
( 1) 2.region = 0
( 2) 3.region = 0
( 3) 4.region = 0
F( 3, 29) = 2.96
Prob > F = 0.0486
The nosvyadjust option on test produces an unadjusted Wald test.

. test 2.region 3.region 4.region, nosvyadjust
Unadjusted Wald test
( 1) 2.region = 0
( 2) 3.region = 0
( 3) 4.region = 0
F( 3, 31) = 3.17
Prob > F = 0.0382
For one-dimensional tests, the adjusted and unadjusted F statistics are identical, but they differ for
higher-dimensional tests. Using the nosvyadjust option is not recommended because the unadjusted
F statistic can produce extremely anticonservative p-values (that is, p-values that are too small) when
the variance degrees of freedom (equal to the number of sampled PSUs minus the number of strata)
is not large relative to the dimension of the test.
Bonferroni-adjusted p-values can also be computed:

. test 2.region 3.region 4.region, mtest(bonferroni)
Adjusted Wald test
( 1) 2.region = 0
( 2) 3.region = 0
( 3) 4.region = 0
F(df,29) df p
(1) 0.34 1 1.0000 #

(2) 9.06 1 0.0155 #
(3) 0.87 1 1.0000 #
all 2.96 3 0.0486
# Bonferroni-adjusted p-values
See Korn and Graubard (1990) for a discussion of these three different procedures for conducting
joint-hypothesis tests. See Eltinge and Sribney (1996) for an earlier implementation of test for
survey data.
Example 7: Contrasts
After svy commands, we can estimate contrasts and make pairwise comparisons with the contrast
and pwcompare commands. First, we will fit a regression of serum zinc levels on health status:
. use http://www.stata-press.com/data/r14/nhanes2f, clear
. label list hlthgrp
hlthgrp:
1 poor
2 fair
3 average
4 good
5 excellent
. svy: regress zinc i.health
Design df = 31
F( 4, 28) = 15.61
Prob > F = 0.0000
R-squared = 0.0098
Linearized
health
fair .9272308 .7690396 1.21 0.237 -.6412357 2.495697
average 2.444004 .6407097 3.81 0.001 1.137268 3.75074
good 4.038285 .6830349 5.91 0.000 2.645226 5.431344
excellent 4.770911 .7151641 6.67 0.000 3.312324 6.229498
_cons 83.94729 .8523379 98.49 0.000 82.20893 85.68564

Higher levels of zinc are associated with better health. We can use reverse adjacent contrasts to
compare each health status with the preceding status.
. contrast ar.health
Contrasts of marginal linear predictions
Design df = 31
Margins : asbalanced
df F P>F
health
(fair vs poor) 1 1.45 0.2371
(average vs fair) 1 5.49 0.0257
(good vs average) 1 10.92 0.0024
(excellent vs good) 1 1.93 0.1744
Joint 4 15.61 0.0000
Design 31
Note: F statistics are adjusted for the survey design.
Contrast Std. Err. [95% Conf. Interval]
health
(fair vs poor) .9272308 .7690396 -.6412357 2.495697
(average vs fair) 1.516773 .6474771 .1962347 2.837311
(good vs average) 1.594281 .4824634 .6102904 2.578271
(excellent vs good) .7326264 .5270869 -.3423744 1.807627
The first table reports significance tests for each contrast, along with a joint test of all the contrasts.
The row labeled (fair vs poor), for example, tests the null hypothesis that the first two health
statuses have the same mean zinc level. The test statistics are automatically adjusted for the survey
design.
The second table reports estimates, standard errors, and confidence limits for each contrast. The
row labeled (good vs average), for example, shows that those in good health have a mean zinc
level about 1.6 units higher than those of average health. The standard errors and confidence intervals
also account for the survey design.
If we would like to go further and make all possible pairwise comparisons of the health groups,
we can use the pwcompare command. We will specify the mcompare(sidak) option to account for
multiple comparisons and the cformat(%3.1f) option to reduce the number of decimal places in
the output:
. pwcompare health, mcompare(sidak) cformat(%3.1f)

Pairwise comparisons of marginal linear predictions
Design df = 31
Margins : asbalanced
Number of
Comparisons
health 10
Sidak
Contrast Std. Err. [95% Conf. Interval]
health
fair vs poor 0.9 0.8 -1.4 3.2
average vs poor 2.4 0.6 0.5 4.4
good vs poor 4.0 0.7 2.0 6.1
excellent vs poor 4.8 0.7 2.6 6.9
average vs fair 1.5 0.6 -0.4 3.5
good vs fair 3.1 0.5 1.5 4.8
excellent vs fair 3.8 0.7 1.7 6.0
good vs average 1.6 0.5 0.1 3.0
excellent vs average 2.3 0.7 0.3 4.4
excellent vs good 0.7 0.5 -0.9 2.3
Seven of the ten Šidák intervals exclude the null value of zero. See [R] pwcompare for more
information on pairwise comparisons and multiple-comparison adjustments.
Example 8: Using suest with survey data, the svy prefix

suest can be used to obtain the variance estimates for a series of estimators that used the svy
prefix. To use suest for this purpose, perform the following steps:
1. Be sure to set the survey design characteristics correctly by using svyset. Do not use the
vce() option to change the default variance estimator from the linearized variance estimator.
vce(brr) and vce(jackknife) are not supported by suest.
2. Fit the model or models by using the svy prefix command, optionally including subpopulation
estimation with the subpop() option.
3. Store the estimation results with estimates store name.
In the following, we illustrate how to use suest to compare the parameter estimates between two
ordered logistic regression models.
In the NHANES II dataset, we have the variable health containing self-reported health status,
which takes on the values 1–5, with 1 being “poor” and 5 being “excellent”. Because this is an
ordered categorical variable, it makes sense to model it by using svy: ologit. We use some basic
demographic variables as predictors: female (an indicator of female individuals), black (an indicator
for black individuals), age in years, and c.age#c.age (age squared).
. svyset psuid [pw=finalwgt], strata(stratid)
pweight: finalwgt
VCE: linearized
Strata 1: stratid
SU 1: psuid
FPC 1: <zero>
. svy: ologit health female black age c.age#c.age
(running ologit on estimation sample)
Survey: Ordered logistic regression
Design df = 31
F( 4, 28) = 223.27
Prob > F = 0.0000
Linearized
health Coef. Std. Err. t P>|t| [95% Conf. Interval]
female -.1615219 .0523678 -3.08 0.004 -.2683267 -.054717

black -.986568 .0790277 -12.48 0.000 -1.147746 -.8253899
age -.0119491 .0082974 -1.44 0.160 -.0288717 .0049736
c.age#c.age -.0003234 .000091 -3.55 0.001 -.000509 -.0001377
/cut1 -4.566229 .1632561 -27.97 0.000 -4.899192 -4.233266

/cut2 -3.057415 .1699944 -17.99 0.000 -3.404121 -2.710709
/cut3 -1.520596 .1714342 -8.87 0.000 -1.870239 -1.170954
/cut4 -.242785 .1703965 -1.42 0.164 -.590311 .104741
The self-reported health variable takes five categories. Categories 1 and 2 denote negative
categories, whereas categories 4 and 5 denote positive categories. We wonder whether the distinctions
between the two positive categories and between the two negative categories are produced in accordance
with one latent dimension, which is an assumption of the ordered logistic model. To test one-
dimensionality, we will collapse the five-point health measure into a three-point measure, refit the
ordered logistic model, and compare the regression coefficients and cutpoints between the two analyses.
If the single latent variable assumption is valid, the coefficients and cutpoints should match. This can
be seen as a Hausman-style specification test. Estimation of the ordered logistic model parameters
for survey data is by maximum pseudolikelihood. Neither estimator is fully efficient, and thus the
assumptions for the classic Hausman test and for the hausman command are not satisfied. With
suest, we can obtain an appropriate Hausman test for survey data.
To perform the Hausman test, we are already almost halfway there by following steps 1 and 2 for
one of the models. We just need to store the current estimation results before moving on to the next
model. Here we store the results with estimates store under the name H5, indicating that in this
analysis, the dependent variable health has five categories.
. estimates store H5
We proceed by generating a new dependent variable health3, which maps values 1 and 2 into
2, 3 into 3, and 4 and 5 into 4. This transformation is conveniently accomplished with the clip()
function. We then fit an ologit model with this new dependent variable and store the estimation
results under the name H3.
. generate health3 = clip(health, 2, 4)
(2 missing values generated)
. svy: ologit health3 female black age c.age#c.age
(running ologit on estimation sample)
Survey: Ordered logistic regression
Design df = 31
F( 4, 28) = 197.08
Prob > F = 0.0000
Linearized
health3 Coef. Std. Err. t P>|t| [95% Conf. Interval]
female -.1551238 .0563809 -2.75 0.010 -.2701133 -.0401342

black -1.046316 .0728274 -14.37 0.000 -1.194849 -.8977836
age -.0365408 .0073653 -4.96 0.000 -.0515624 -.0215192
c.age#c.age -.00009 .0000791 -1.14 0.264 -.0002512 .0000713
/cut1 -3.655498 .1610211 -22.70 0.000 -3.983903 -3.327093

/cut2 -2.109584 .1597057 -13.21 0.000 -2.435306 -1.783862
We can now obtain the combined estimation results of the two models stored under H5 and H3
with design-based standard errors.
. suest H5 H3
Simultaneous survey results for H5, H3
Design df = 31
Linearized
H5_health
female -.1615219 .0523678 -3.08 0.004 -.2683267 -.054717
black -.986568 .0790277 -12.48 0.000 -1.147746 -.8253899
age -.0119491 .0082974 -1.44 0.160 -.0288717 .0049736
c.age#c.age -.0003234 .000091 -3.55 0.001 -.000509 -.0001377
H5_cut1
_cons -4.566229 .1632561 -27.97 0.000 -4.899192 -4.233266
H5_cut2
_cons -3.057415 .1699944 -17.99 0.000 -3.404121 -2.710709
H5_cut3
_cons -1.520596 .1714342 -8.87 0.000 -1.870239 -1.170954
H5_cut4
_cons -.242785 .1703965 -1.42 0.164 -.590311 .104741
H3_health3
female -.1551238 .0563809 -2.75 0.010 -.2701133 -.0401342
black -1.046316 .0728274 -14.37 0.000 -1.194849 -.8977836
age -.0365408 .0073653 -4.96 0.000 -.0515624 -.0215192
c.age#c.age -.00009 .0000791 -1.14 0.264 -.0002512 .0000713
H3_cut1
_cons -3.655498 .1610211 -22.70 0.000 -3.983903 -3.327093
H3_cut2
_cons -2.109584 .1597057 -13.21 0.000 -2.435306 -1.783862
The coefficients of H3 and H5 look rather similar. We now use test to perform a formal Hausman-
type test for the hypothesis that the regression coefficients are indeed the same, as we would expect
if there is indeed a one-dimensional latent dimension for health. Thus we test that the coefficients in
the equation H5 health are equal to those in H3 health3.
. test [H5_health=H3_health3]
Adjusted Wald test
( 1) [H5_health]female - [H3_health3]female = 0
( 2) [H5_health]black - [H3_health3]black = 0
( 3) [H5_health]age - [H3_health3]age = 0
( 4) [H5_health]c.age#c.age - [H3_health3]c.age#c.age = 0
F( 4, 28) = 17.13
Prob > F = 0.0000
We can reject the null hypothesis, which indicates that the ordered logistic regression model is
indeed misspecified. Another specification test can be conducted with respect to the cutpoints. Variable
health3 was constructed from health by collapsing the two worst categories into value 2 and the
two best categories into value 4. This action effectively has removed two cutpoints, but if the model
fits the data, it should not affect the other two cutpoints. The comparison is hampered by a difference
in the names of the cutpoints between the models, as illustrated in the figure below:
cut1 cut2 cut3 cut4

H5 latent ———x———x———x———x———
observed 1 2 3 4 5
cut1 cut2
H3 latent ——————–x———x——————-
observed 2 3 4
Cutpoint /cut2 of model H5 should be compared with cutpoint /cut1 of H3, and similarly, /cut3
of H5 with /cut2 of H3.
. test ([H5_cut2]_cons=[H3_cut1]_cons) ([H5_cut3]_cons=[H3_cut2]_cons)
Adjusted Wald test
( 1) [H5_cut2]_cons - [H3_cut1]_cons = 0
( 2) [H5_cut3]_cons - [H3_cut2]_cons = 0
F( 2, 30) = 33.49
Prob > F = 0.0000
We conclude that the invariance of the cutpoints under the collapse of categories is not supported
by the data, again providing evidence against the reduced specification of the ordered logistic model
in this case.
Example 9: Using suest with survey data, the svy option

Not all estimation commands support the svy prefix, but you can use the svy option with suest
to get survey estimation results. If you can use suest after a command, you can use suest, svy.
Here are the corresponding Stata commands to perform the analysis in the previous example, using
the svy option instead of the svy prefix.
. svyset psuid [pw=finalwgt], strata(stratid)
. ologit health female black age c.age#c.age [iw=finalwgt]
. generate health3 = clip(health,2,4)
. ologit health3 female black age c.age#c.age [iw=finalwgt]
. suest H5 H3, svy
. test [H5_health=H3_health3]
. test ([H5_cut2]_cons=[H3_cut1]_cons) ([H5_cut3]_cons=[H3_cut2]_cons)
The calls to ologit now use iweights instead of the svy prefix, and the svy option was added
to suest. No other changes are required.
References
Eltinge, J. L., and W. M. Sribney. 1996. svy5: Estimates of linear combinations and hypothesis tests for survey data.
Stata Technical Bulletin 31: 31–42. Reprinted in Stata Technical Bulletin Reprints, vol. 6, pp. 246–259. College
Graubard, B. I., and E. L. Korn. 2004. Predictive margins with survey data. Biometrics 55: 652–659.
Also see
[SVY] estat — Postestimation statistics for survey data
[U] 13.5 Accessing coefficients and standard errors
Title
svy sdr — Successive difference replication for survey data

Reference Also see
Description
svy sdr performs successive difference replication (SDR) estimation of specified statistics (or
expressions) for a Stata command or a user-written command. The command is executed once for
each replicate using sampling weights that are adjusted according to the SDR methodology. Any
Stata estimation command listed in [SVY] svy estimation may be used with svy sdr. User-written
commands that meet the requirements in [P] program properties may also be used.
Quick start
Estimate population mean of v1 using SDR standard-error estimates with sampling weight wvar1 and
replicate weights in variables with prefix rwvar
svyset [pweight=wvar1], sdrweight(rwvar*)
svy sdr _b: mean v1
Same as above
svyset [pweight=wvar1], sdrweight(rwvar*) vce(sdr)
svy: mean v1
SDR estimate of the standard error of the difference between the means of v2 and v3 using either
svyset command above
svy sdr (_b[v2]-_b[v3]): mean v2 v3
svy sdr diff=(_b[v2]-_b[v3]), saving(mydata): mean v2 v3
Same as above
sdr diff=(_b[v2]-_b[v3]), saving(mydata): mean v2 v3
Menu
Statistics > Survey data analysis > Resampling > Successive difference replications estimation
127
128 svy sdr — Successive difference replication for survey data
Syntax

svy sdr exp list , svy options sdr options eform option : command

if/in

Reporting
sdr options Description

Options

Reporting
trace trace command
title(text) use text as title for SDR results
Advanced
svy sdr requires that the successive difference replicate weights be identified using svyset.
svy sdr — Successive difference replication for survey data 129
elist
eexp
(exp)
eexp is specname
[eqno]specname
specname is b
b[]
se
se[]
eqno is ##
name

Options

Options

double specifies that the results for each replication be stored as doubles, meaning 8-byte reals.
By default, they are stored as floats, meaning 4-byte reals. This option may be used without
See [P] postfile.
dialog box.
mse specifies that svy sdr compute the variance by using deviations of the replicates from the
observed value of the statistics based on the entire dataset. By default, svy sdr computes the
variance by using deviations of the replicates from their mean.

Reporting
option.
option.
title(text) specifies a title to be displayed above the table of SDR results; the default title is “SDR
results”.

Advanced

SDR was first introduced by Fay and Train (1995) as a method of variance estimation for annual
demographic supplements to the Current Population Survey (CPS). In SDR, the model is fit multiple
times, once for each of a set of adjusted sampling weights. The variance is estimated using the
resulting replicated point estimates.
Example 1
The U.S. Census Bureau publishes public-use data from several of its surveys. This data can
be downloaded from http://factfinder.census.gov. We downloaded the American Community Survey
(ACS) Public Use Microdata Sample (PUMS) data collected in 2007. We extracted data for the state of
Texas and kept the variables containing age, sex, and sampling weight for each person in the dataset.
This sample dataset also contains 80 SDR weight variables.
. use http://www.stata-press.com/data/r14/ss07ptx
. svyset
pweight: pwgtp
VCE: sdr
MSE: off
sdrweight: pwgtp1 pwgtp2 pwgtp3 pwgtp4 pwgtp5 pwgtp6 pwgtp7 pwgtp8 pwgtp9
pwgtp10 pwgtp11 pwgtp12 pwgtp13 pwgtp14 pwgtp15 pwgtp16
(output omitted )
pwgtp73 pwgtp74 pwgtp75 pwgtp76 pwgtp77 pwgtp78 pwgtp79
pwgtp80
Strata 1: <one>
FPC 1: <zero>
This dataset was already svyset as

. svyset [pw=pwgtp], sdrweight(pwgtp1-pwgtp80) vce(sdr)
Here we estimate the average age of the males and of the females for our Texas subpopulation.
The standard errors are estimated using SDR.
svy sdr — Successive difference replication for survey data 131
. svy: mean agep, over(sex)

SDR replications (80)
1 2 3 4 5
.................................................. 50
..............................
Replications = 80
Male: sex = Male
SDR
agep
Male 33.24486 .0470986 33.15255 33.33717
Female 35.23908 .0386393 35.16335 35.31481
Stored results
In addition to the results documented in [SVY] svy, svy sdr stores the following in e():
Scalars
Macros
e(cmd) same as e(cmdname) or sdr
e(vce) sdr
e(sdrweight) sdrweight() variable list
Matrices
e(b sdr) SDR means
e(V) SDR variance estimates
When exp list is b, svy sdr will also carry forward most of the results already in e() from
command.

See [SVY] variance estimation for details regarding SDR variance estimation.
Reference
Also see
Title
svy: tabulate oneway — One-way tables for survey data
Description Quick start Menu

Syntax Options Remarks and examples
Stored results Methods and formulas Reference
Also see
Description
svy: tabulate produces one-way tabulations for complex survey data. See [SVY] svy: tabulate
twoway for two-way tabulations for complex survey data.
Quick start
One-way table showing weighted proportions for categories of v1 using svyset data
svy: tabulate v1
Add 95% confidence intervals and weighted counts
svy: tabulate v1, ci count
As above, and display large counts in a more readable format
svy: tabulate v1 ci count format(%11.3g)
Unweighted numbers of observations and weighted proportions for categories of v2
svy: tabulate v2, obs
Weighted proportions and CIs for categories of v3 in the subpopulation defined by v4 > 40
svy, subpop(if v4>40): tabulate v3, ci
Menu
Statistics > Survey data analysis > Tables > One-way tables
133
134 svy: tabulate oneway — One-way tables for survey data
Syntax
Basic syntax
svy: tabulate varname
Full syntax

svy vcetype , svy options : tabulate varname if in

, tabulate options display items display options
Syntax to report results

svy , display items display options
vcetype Description
SE

if/in

SE
svy: tabulate oneway — One-way tables for survey data 135
tabulate options Description

Model
stdize(varname) variable identifying strata for standardization
stdweight(varname) weight variable for standardization
tab(varname) variable for which to compute cell totals/proportions
missing treat missing values like other values
display items Description

Table items
cell cell proportions
count weighted cell counts
se standard errors
ci confidence intervals
deff display the DEFF design effects
deft display the DEFT design effects
cv display the coefficient of variation
srssubpop report design effects assuming SRS within subpopulation
obs cell observations
When any of se, ci, deff, deft, cv, or srssubpop is specified, only one of cell or count can be specified. If
none of se, ci, deff, deft, cv, or srssubpop is specified, both cell and count can be specified.
display options Description

Reporting
proportion display proportions; the default
percent display percentages instead of proportions
nomarginal suppress column marginal
nolabel suppress displaying value labels
cellwidth(#) cell width
csepwidth(#) column-separation width
stubwidth(#) stub width
format(% fmt) cell format; default is format(%6.0g)
proportion is not shown in the dialog box.
Options

Model
stdize(varname) specifies that the point estimates be adjusted by direct standardization across the
strata identified by varname. This option requires the stdweight() option.
stdweight(varname) specifies the weight variable associated with the standard strata identified in
the stdize() option. The standardization weights must be constant within the standard strata.
tab(varname) specifies that counts be cell totals of this variable and that proportions (or percentages)
be relative to (that is, weighted by) this variable. For example, if this variable denotes income, then
the cell “counts” are instead totals of income for each cell, and the cell proportions are proportions
of income for each cell.
missing specifies that missing values in varname be treated as another row category rather than be
omitted from the analysis (the default).

Table items
cell requests that cell proportions (or percentages) be displayed. This is the default if count is not
specified.
count requests that weighted cell counts be displayed.
se requests that the standard errors of cell proportions (the default) or weighted counts be displayed.
When se (or ci, deff, deft, or cv) is specified, only one of cell or count can be selected.
The standard error computed is the standard error of the one selected.
ci requests confidence intervals for cell proportions or weighted counts.
deff and deft request that the design-effect measures DEFF and DEFT be displayed for each cell
proportion or weighted count. See [SVY] estat for details.
or poststratification.
cv requests that the coefficient of variation be displayed for each cell proportion, count, or row or
column proportion. See [SVY] estat for details.
srssubpop requests that DEFF and DEFT be computed using an estimate of SRS (simple random
sampling) variance for sampling within a subpopulation. By default, DEFF and DEFT are computed
using an estimate of the SRS variance for sampling from the entire population. Typically, srssubpop
would be given when computing subpopulation estimates by strata or by groups of strata.
obs requests that the number of observations for each cell be displayed.

Reporting
proportion, the default, requests that proportions be displayed.
percent requests that percentages be displayed instead of proportions.
nomarginal requests that the column marginal not be displayed.
nolabel requests that variable labels and value labels be ignored.
cellwidth(#), csepwidth(#), and stubwidth(#) specify widths of table elements in the output;
see [P] tabdisp. Acceptable values for the stubwidth() option range from 4 to 32.
format(% fmt) specifies a format for the items in the table. The default is format(%6.0g). See
[U] 12.5 Formats: Controlling how data are displayed.
svy: tabulate uses the tabdisp command (see [P] tabdisp) to produce the table. Only five items
can be displayed in the table at one time. The ci option implies two items. If too many items are
selected, a warning will appear immediately. To view more items, redisplay the table while specifying
different options.

Despite the long list of options for svy: tabulate, it is a simple command to use. Using the
svy: tabulate command is just like using tabulate to produce one-way tables for ordinary data.
The main difference is that svy: tabulate computes standard errors appropriate for complex survey
data.
Standard errors and confidence intervals can optionally be displayed for weighted counts or cell
proportions. The confidence intervals for proportions are constructed using a logit transform so that
their endpoints always lie between 0 and 1; see [SVY] svy: tabulate twoway. Associated design
effects (DEFF and DEFT) can be viewed for the variance estimates.
Example 1
Here we use svy: tabulate to estimate the distribution of the race category variable from our
NHANES II dataset (McDowell et al. 1981). Before calling svy: tabulate, we use svyset to declare
the survey structure of the data.
pweight: finalwgt
VCE: linearized
Strata 1: stratid
SU 1: psuid
FPC 1: <zero>
. svy: tabulate race
Design df = 31
1=white,
2=black,
3=other proportion
White .8792
Black .0955
Other .0253
Total 1
Key: proportion = cell proportion
Here we display weighted counts for each category of race along with the 95% confidence bounds,
as well as the design effects DEFF and DEFT. We also use the format() option to improve the look
of the table.
. svy: tabulate race, format(%11.3g) count ci deff deft

Design df = 31
1=white,
2=black,
3=other count lb ub deff deft
White 102999549 97060400 108938698 60.2 7.76

Black 11189236 8213964 14164508 18.6 4.31
Other 2968728 414930 5522526 47.9 6.92
Total 117157513
Key: count = weighted count

lb = lower 95% confidence bound for weighted count
ub = upper 95% confidence bound for weighted count
deff = deff for variance of weighted count
deft = deft for variance of weighted count
From the above results, we can conclude with 95% confidence that the number of people in the
population that fall within the White category is between 97,060,400 and 108,938,698.
Stored results
In addition to the results documented in [SVY] svy, svy: tabulate stores the following in e():
Scalars
e(r) number of rows e(total) weighted sum of tab() variable
Macros
e(cmd) tabulate e(rowvlab) row variable label
e(tab) tab() variable e(rowvar) varname, the row variable
e(rowlab) label or empty e(setype) cell or count
Matrices
e(Prop) matrix of cell proportions e(V row) variance for row totals
e(Obs) matrix of observation counts e(V srs row) Vsrs for row totals
e(Deff) DEFF vector for e(setype) items e(Deff row) DEFF for row totals
e(Deft) DEFT vector for e(setype) items e(Deft row) DEFT for row totals
e(Row) values for row variable

See Methods and formulas in [SVY] svy: tabulate twoway for a discussion of how table items
and confidence intervals are computed. A one-way table is really just a two-way table that has one
row or column.

Margaret E. Martin (1912–2012) is best known for her work developing the U.S. Current Population
Survey (CPS). Martin was born in New York City and had an early love for mathematics. She
received a bachelor’s degree in economics from Barnard College and went on to earn an MA
and a PhD in economics from Columbia University. Martin began her career in the midst of the
Great Depression, working for a New Deal agency in New York to classify employers covered
by the unemployment insurance system. Despite having the third highest score on the qualifying
civil service exam, she almost did not take the job because she “had been trained by economists
primarily, and they had a very low opinion of government work”.
Her work in New York allowed her to later move to the U.S. Bureau of Budget (now the Office
of Management and Budget), where she joined the team that developed the CPS. The majority
of Martin’s work focused on the CPS, a survey of employment and demographics among U.S.
households. She worked to explain differences in previous unemployment survey results derived
from sampling businesses. She also oversaw an effort to improve the reliability of information
from the CPS by adding questions that addressed labor-force participation and the use of paid
and unpaid leave. Today, the CPS is a continuous monthly survey and the primary source of
information about characteristics of the U.S. labor force.
In 1973, Martin became the first executive director of the National Academy of Sciences’
Committee on National Statistics. She was elected president of the American Statistical Association
(ASA) in 1980 and was the first recipient of the ASA’s Founders Award.

Reference
Also see
[SVY] svydescribe — Describe survey data
[R] tabulate oneway — One-way table of frequencies
[SVY] svy: tabulate twoway — Two-way tables for survey data
Title
svy: tabulate twoway — Two-way tables for survey data
Description Quick start Menu

Syntax Options Remarks and examples
Stored results Methods and formulas References
Also see
Description
svy: tabulate produces two-way tabulations with tests of independence for complex survey data.
See [SVY] svy: tabulate oneway for one-way tabulations for complex survey data.
Quick start
Two-way table of weighted cell proportions for v1 and v2 using svyset data
svy: tabulate v1 v2
As above, but with a test of independence using Pearson’s χ2 statistic with and without correction
for the complex design
svy: tabulate v1 v2, pearson
Within-row and within-column proportions
svy: tabulate v1 v2, row column
95% confidence intervals for within-column proportions
svy: tabulate v1 v2, column ci
Unweighted numbers of observations and weighted counts
svy: tabulate v1 v2, obs count
As above, but display large counts in a more readable format
svy: tabulate v1 v2, obs count format(%11.0fc)
Weighted counts in the subpopulation defined by v3 > 0
svy, subpop(v3): tabulate v1 v2, count
Menu
Statistics > Survey data analysis > Tables > Two-way tables
140
svy: tabulate twoway — Two-way tables for survey data 141
Syntax
Basic syntax
svy: tabulate varname1 varname2
Full syntax

svy vcetype , svy options : tabulate varname1 varname2 if in

, tabulate options display items display options statistic options
Syntax to report results

svy , display items display options statistic options
vcetype Description
SE

if/in

SE
142 svy: tabulate twoway — Two-way tables for survey data
tabulate options Description

Model
stdize(varname) variable identifying strata for standardization
stdweight(varname) weight variable for standardization
tab(varname) variable for which to compute cell totals/proportions
missing treat missing values like other values
display items Description

Table items
cell cell proportions
count weighted cell counts
column within-column proportions
row within-row proportions
se standard errors
ci confidence intervals
deff display the DEFF design effects
deft display the DEFT design effects
cv display the coefficient of variation
srssubpop report design effects assuming SRS within subpopulation
obs cell observations
When any of se, ci, deff, deft, cv, or srssubpop is specified, only one of cell, count, column, or row can
be specified. If none of se, ci, deff, deft, cv, or srssubpop is specified, any of or all cell, count, column,
and row can be specified.
display options Description

Reporting
proportion display proportions; the default
percent display percentages instead of proportions
vertical stack confidence interval endpoints vertically
nomarginals suppress row and column marginals
nolabel suppress displaying value labels
notable suppress displaying the table
cellwidth(#) cell width
csepwidth(#) column-separation width
stubwidth(#) stub width
format(% fmt) cell format; default is format(%6.0g)
proportion and notable are not shown in the dialog box.
statistic options Description

Test statistics
pearson Pearson’s chi-squared
lr likelihood ratio
null display null-based statistics
wald adjusted Wald
llwald adjusted log-linear Wald
noadjust report unadjusted Wald statistics
Options

Model
stdize(varname) specifies that the point estimates be adjusted by direct standardization across the
strata identified by varname. This option requires the stdweight() option.
stdweight(varname) specifies the weight variable associated with the standard strata identified in
the stdize() option. The standardization weights must be constant within the standard strata.
tab(varname) specifies that counts be cell totals of this variable and that proportions (or percentages)
be relative to (that is, weighted by) this variable. For example, if this variable denotes income, the
cell “counts” are instead totals of income for each cell, and the cell proportions are proportions
of income for each cell.
missing specifies that missing values in varname1 and varname2 be treated as another row or column
category rather than be omitted from the analysis (the default).

Table items
cell requests that cell proportions (or percentages) be displayed. This is the default if none of count,
row, or column is specified.
count requests that weighted cell counts be displayed.
column or row requests that column or row proportions (or percentages) be displayed.
se requests that the standard errors of cell proportions (the default), weighted counts, or row or
column proportions be displayed. When se (or ci, deff, deft, or cv) is specified, only one of
cell, count, row, or column can be selected. The standard error computed is the standard error
of the one selected.
ci requests confidence intervals for cell proportions, weighted counts, or row or column proportions.
The confidence intervals are constructed using a logit transform so that their endpoints always lie
between 0 and 1.
deff and deft request that the design-effect measures DEFF and DEFT be displayed for each cell
proportion, count, or row or column proportion. See [SVY] estat for details. The mean generalized
DEFF is also displayed when deff, deft, or subpop is requested; see Methods and formulas for
an explanation.
or poststratification.
cv requests that the coefficient of variation be displayed for each cell proportion, count, or row or
column proportion. See [SVY] estat for details.
srssubpop requests that DEFF and DEFT be computed using an estimate of SRS (simple random
sampling) variance for sampling within a subpopulation. By default, DEFF and DEFT are computed
using an estimate of the SRS variance for sampling from the entire population. Typically, srssubpop
would be given when computing subpopulation estimates by strata or by groups of strata.
obs requests that the number of observations for each cell be displayed.

Reporting
proportion, the default, requests that proportions be displayed.
percent requests that percentages be displayed instead of proportions.
vertical requests that the endpoints of confidence intervals be stacked vertically on display.
nomarginals requests that row and column marginals not be displayed.
nolabel requests that variable labels and value labels be ignored.
notable prevents the header and table from being displayed in the output. When specified, only the
results of the requested test statistics are displayed. This option may not be specified with any
other option in display options except the level() option.
cellwidth(#), csepwidth(#), and stubwidth(#) specify widths of table elements in the output;
see [P] tabdisp. Acceptable values for the stubwidth() option range from 4 to 32.
format(% fmt) specifies a format for the items in the table. The default is format(%6.0g). See
[U] 12.5 Formats: Controlling how data are displayed.

Test statistics
pearson requests that the Pearson χ2 statistic be computed. By default, this is the test of independence
that is displayed. The Pearson χ2 statistic is corrected for the survey design with the second-order
correction of Rao and Scott (1984) and is converted into an F statistic. One term in the correction
formula can be calculated using either observed cell proportions or proportions under the null
hypothesis (that is, the product of the marginals). By default, observed cell proportions are used.
If the null option is selected, then a statistic corrected using proportions under the null hypothesis
is displayed as well.
lr requests that the likelihood-ratio test statistic for proportions be computed. This statistic is not
defined when there are one or more zero cells in the table. The statistic is corrected for the survey
design by using the same correction procedure that is used with the pearson statistic. Again either
observed cell proportions or proportions under the null hypothesis can be used in the correction
formula. By default, the former is used; specifying the null option gives both the former and the
latter. Neither variant of this statistic is recommended for sparse tables. For nonsparse tables, the
lr statistics are similar to the corresponding pearson statistics.
null modifies the pearson and lr options only. If null is specified, two corrected statistics are
displayed. The statistic labeled “D-B (null)” (“D-B” stands for design-based) uses proportions
under the null hypothesis (that is, the product of the marginals) in the Rao and Scott (1984)
correction. The statistic labeled merely “Design-based” uses observed cell proportions. If null is
not specified, only the correction that uses observed proportions is displayed.
wald requests a Wald test of whether observed weighted counts equal the product of the marginals
(Koch, Freeman, and Freeman 1975). By default, an adjusted F statistic is produced; an unadjusted
statistic can be produced by specifying noadjust. The unadjusted F statistic can yield extremely
anticonservative p-values (that is, p-values that are too small) when the degrees of freedom of the
variance estimates (the number of sampled PSUs minus the number of strata) are small relative
to the (R − 1)(C − 1) degrees of freedom of the table (where R is the number of rows and C
is the number of columns). Hence, the statistic produced by wald and noadjust should not be
used for inference unless it is essentially identical to the adjusted statistic.
This option must be specified at run time in order to be used on subsequent calls to svy to report
results.
llwald requests a Wald test of the log-linear model of independence (Koch, Freeman, and Free-
man 1975). The statistic is not defined when there are one or more zero cells in the table. The
adjusted statistic (the default) can produce anticonservative p-values, especially for sparse tables,
when the degrees of freedom of the variance estimates are small relative to the degrees of freedom of
the table. Specifying noadjust yields a statistic with more severe problems. Neither the adjusted
nor the unadjusted statistic is recommended for inference; the statistics are made available only
for pedagogical purposes.
noadjust modifies the wald and llwald options only. It requests that an unadjusted F statistic be
displayed in addition to the adjusted statistic.
svy: tabulate uses the tabdisp command (see [P] tabdisp) to produce the table. Only five items
can be displayed in the table at one time. The ci option implies two items. If too many items are
selected, a warning will appear immediately. To view more items, redisplay the table while specifying
different options.

Introduction
The Rao and Scott correction
Wald statistics
Properties of the statistics
Introduction
Despite the long list of options for svy: tabulate, it is a simple command to use. Using the
svy: tabulate command is just like using tabulate to produce two-way tables for ordinary data.
The main difference is that svy: tabulate computes a test of independence that is appropriate for
complex survey data.
The test of independence that is displayed by default is based on the usual Pearson χ2 statistic
for two-way tables. To account for the survey design, the statistic is turned into an F statistic with
noninteger degrees of freedom by using a second-order Rao and Scott (1981, 1984) correction.
Although the theory behind the Rao and Scott correction is complicated, the p-value for the corrected
F statistic can be interpreted in the same way as a p-value for the Pearson χ2 statistic for “ordinary”
data (that is, data that are assumed independent and identically distributed [i.i.d.]).
svy: tabulate, in fact, computes four statistics for the test of independence with two variants
of each, for a total of eight statistics. The option combination for each of the eight statistics are the
following:
1. pearson (the default)
2. pearson null
3. lr
4. lr null
5. wald
6. wald noadjust
7. llwald
8. llwald noadjust
The wald and llwald options with noadjust yield the statistics developed by Koch, Freeman, and
Freeman (1975), which have been implemented in the CROSSTAB procedure of the SUDAAN software
(Research Triangle Institute 1997, release 7.5).
These eight statistics, along with other variants, have been evaluated in simulations (Sribney 1998).
On the basis of these simulations, we advise researchers to use the default statistic (the pearson
option) in all situations. We recommend that the other statistics be used only for comparative or
pedagogical purposes. Sribney (1998) gives a detailed comparison of the statistics; a summary of his
conclusions is provided later in this entry.
Other than the test-statistic options (statistic options) and the survey design options (svy options),
most of the other options of svy: tabulate simply relate to different choices for what can be
displayed in the body of the table. By default, cell proportions are displayed, but viewing either row
or column proportions or weighted counts usually makes more sense.
Standard errors and confidence intervals can optionally be displayed for weighted counts or cell,
row, or column proportions. The confidence intervals for proportions are constructed using a logit
transform so that their endpoints always lie between 0 and 1. Associated design effects (DEFF and
DEFT) can be viewed for the variance estimates. The mean generalized DEFF (Rao and Scott 1984)
is also displayed when option deff, deft, or srssubpop is specified. The mean generalized DEFF
is essentially a design effect for the asymptotic distribution of the test statistic; see the Methods and
formulas section at the end of this entry.
Example 1
Using data from the Second National Health and Nutrition Examination Survey (NHANES II)
(McDowell et al. 1981), we identify the survey design characteristics with svyset and then produce
a two-way table of cell proportions with svy: tabulate.
pweight: finalwgt
VCE: linearized
Strata 1: stratid
SU 1: psuid
FPC 1: <zero>
. svy: tabulate race diabetes
Design df = 31
1=white,
3=other 0 1 Total
White .851 .0281 .8791

Black .0899 .0056 .0955
Other .0248 5.2e-04 .0253
Total .9658 .0342 1
Key: cell proportion

Pearson:
Design-based F(1.52, 47.26) = 15.0056 P = 0.0000
The default table displays only cell proportions, and this makes it difficult to compare the incidence
of diabetes in white, black, and “other” racial groups. It would be better to look at row proportions.
This can be done by redisplaying the results (that is, reissuing the command without specifying any
variables) with the row option.
. svy: tabulate, row
Design df = 31
1=white, diabetes, 1=yes,

2=black, 0=no
3=other 0 1 Total
White .968 .032 1

Black .941 .059 1
Other .9797 .0203 1
Total .9658 .0342 1
Key: row proportion

Pearson:
Design-based F(1.52, 47.26) = 15.0056 P = 0.0000
This table is much easier to interpret. A larger proportion of blacks have diabetes than do whites
or persons in the “other” racial category. The test of independence for a two-way contingency table
is equivalent to the test of homogeneity of row (or column) proportions. Hence, we can conclude
that there is a highly significant difference between the incidence of diabetes among the three racial
groups.
We may now wish to compute confidence intervals for the row proportions. If we try to redisplay,
specifying ci along with row, we get the following result:
. svy: tabulate, row ci
confidence intervals are only available for cells to compute row confidence
intervals, rerun command with row and ci options
r(111);
There are limits to what svy: tabulate can redisplay. Basically, any of the options relating to
variance estimation (that is, se, ci, deff, and deft) must be specified at run time along with the
single item (that is, count, cell, row, or column) for which you want standard errors, confidence
intervals, DEFF, or DEFT. So to get confidence intervals for row proportions, we must rerun the
command. We do so below, requesting not only ci but also se.
. svy: tabulate race diabetes, row se ci format(%7.4f)
Design df = 31
1=white,
3=other 0 1 Total
White 0.9680 0.0320 1.0000

(0.0020) (0.0020)
[0.9638,0.9718] [0.0282,0.0362]
Black 0.9410 0.0590 1.0000

(0.0061) (0.0061)
[0.9271,0.9523] [0.0477,0.0729]
Other 0.9797 0.0203 1.0000

(0.0076) (0.0076)
[0.9566,0.9906] [0.0094,0.0434]
Total 0.9658 0.0342 1.0000

(0.0018) (0.0018)
[0.9619,0.9693] [0.0307,0.0381]
Key: row proportion

(linearized standard error of row proportion)
[95% confidence interval for row proportion]
Pearson:
Design-based F(1.52, 47.26) = 15.0056 P = 0.0000
In the above table, we specified a %7.4f format rather than using the default %6.0g format.
The single format applies to every item in the table. We can omit the marginal totals by specifying
nomarginals. If the above style for displaying the confidence intervals is obtrusive—and it can be
in a wider table—we can use the vertical option to stack the endpoints of the confidence interval,
one over the other, and omit the brackets (the parentheses around the standard errors are also omitted
when vertical is specified). To express results as percentages, as with the tabulate command (see
[R] tabulate twoway), we can use the percent option. Or we can play around with these display
options until we get a table that we are satisfied with, first making changes to the options on redisplay
(that is, omitting the cross-tabulated variables when we issue the command).
Technical note
The standard errors computed by svy: tabulate are the same as those produced by svy: mean,
svy: proportion, and svy: ratio. Indeed, svy: tabulate uses these commands as subroutines
to produce its table.
In the previous example, the estimate of the proportion of African Americans with diabetes (the
second proportion in the second row of the preceding table) is simply a ratio estimate; hence, we can
also obtain the same estimates by using svy: ratio:
. drop black
. generate black = (race==2) if !missing(race)
. generate diablk = diabetes*black
(2 missing values generated)
. svy: ratio diablk/black
Design df = 31
_ratio_1: diablk/black
Linearized
_ratio_1 .0590349 .0061443 .0465035 .0715662
Although the standard errors are the same, the confidence intervals are slightly different. The
svy: tabulate command produced the confidence interval [ 0.0477, 0.0729 ], and svy: ratio
gave [ 0.0465, 0.0716 ]. The difference is because svy: tabulate uses a logit transform to produce
confidence intervals whose endpoints are always between 0 and 1. This transformation also shifts the
confidence intervals slightly toward 0.5, which is beneficial because the untransformed confidence
intervals tend to be, on average, biased away from 0.5. See Methods and formulas for details.
Example 2: The tab() option

The tab() option allows us to compute proportions relative to a certain variable. Suppose that
we wish to compare the proportion of total income among different racial groups in males with that
of females. We do so below with fictitious data:
. use http://www.stata-press.com/data/r14/svy_tabopt, clear

. svy: tabulate gender race, tab(income) row
Design df = 31
Race
Gender White Black Other Total
Male .8857 .0875 .0268 1

Female .884 .094 .022 1
Total .8848 .0909 .0243 1
Tabulated variable: income

Key: row proportion
Pearson:
Design-based F(1.91, 59.12) = 0.8626 P = 0.4227
The Rao and Scott correction

svy: tabulate can produce eight different statistics for the test of independence. By default,
svy: tabulate displays the Pearson χ2 statistic with the Rao and Scott (1981, 1984) second-order
correction. On the basis of simulations Sribney (1998), we recommend that you use this statistic
in all situations. The statistical literature, however, contains several alternatives, along with other
possibilities for implementing the Rao and Scott correction. Hence, for comparative or pedagogical
purposes, you may want to view some of the other statistics computed by svy: tabulate. This
section briefly describes the differences among these statistics; for a more detailed discussion, see
Sribney (1998).
Two statistics commonly used for i.i.d. data for the test of independence of R × C tables (R rows
and C columns) are the Pearson χ2 statistic
R X
X C
2
XP2 = m prc − pb0rc ) /b
(b p0rc
r=1 c=1
and the likelihood-ratio χ2 statistic

R X
X C
2
XLR = 2m pbrc ln (b
prc /b
p0rc )
r=1 c=1
where m is the total number of sampled individuals, pbrc is the estimated proportion for the cell in the
rth row and cth column of the table, and pb0rc is the estimated proportion under the null hypothesis of
PC
independence; that is, pb0rc = pbr· pb·c , the product of the row and column marginals: pbr· = c=1 pbrc
PR
and pb·c = r=1 pbrc .
For i.i.d. data, both these statistics are distributed asymptotically as χ2(R−1)(C−1) . The likelihood-
ratio statistic is not defined when one or more of the cells in the table are empty. The Pearson statistic,
however, can be calculated when one or more cells in the table are empty—the statistic may not have
good properties in this case, but the statistic still has a computable value.
For survey data, XP2 and XLR

2
can be computed using weighted estimates of pbrc and pb0rc . However,
for a complex sampling design, one can no longer claim that they are distributed as χ2(R−1)(C−1) , but
you can estimate the variance of pbrc under the sampling design. For instance, in Stata, this variance
can be estimated via linearization methods by using svy: mean or svy: ratio.
Rao and Scott (1981, 1984) derived the asymptotic distribution of XP2 and XLR 2
in terms of the
variance of pbrc . Unfortunately, the result (see (1) in Methods and formulas) is not computationally
feasible, but it can be approximated using correction formulas. svy: tabulate uses the second-order
correction developed by Rao and Scott (1984). By default, or when the pearson option is specified,
svy: tabulate displays the second-order correction of the Pearson statistic. The lr option gives the
second-order correction of the likelihood-ratio statistic. Because it is the default of svy: tabulate,
the correction computed with pbrc is referred to as the default correction.
The Rao and Scott papers, however, left some details outstanding about the computation of the
correction. One term in the correction formula can be computed using either pbrc or pb0rc . Because
under the null hypothesis both are asymptotically equivalent, theory offers no guidance about which
is best. By default, svy: tabulate uses pbrc for the corrections of the Pearson and likelihood-ratio
statistics. If the null option is specified, the correction is computed using pb0rc . For nonsparse tables,
these two correction methods yield almost identical results. However, in simulations of sparse tables,
Sribney (1998) found that the null-corrected statistics were extremely anticonservative for 2 × 2 tables
(that is, under the null, “significance” was declared too often) and were too conservative for other
tables. The default correction, however, had better properties. Hence, we do not recommend using
null.
For the computational details of the Rao and Scott–corrected statistics, see Methods and formulas.
Wald statistics
Prior to the work by Rao and Scott (1981, 1984), Wald tests for the test of independence for
two-way tables were developed by Koch, Freeman, and Freeman (1975). Two Wald statistics have
been proposed. The first, similar to the Pearson statistic, is based on
brc − N
Ybrc = N br· N
b·c /N
b··
where N brc is the estimated weighted count for the r, cth cell. The delta method can be used to
approximate the variance of Ybrc , and a Wald statistic can be calculated as usual. A second Wald
statistic can be constructed based on a log-linear model for the table. Like the likelihood-ratio statistic,
this statistic is undefined when there is a zero proportion in the table.
These Wald statistics are initially χ2 statistics, but they have better properties when converted
into F statistics with denominator degrees of freedom that account for the degrees of freedom of the
variance estimator. They can be converted to F statistics in two ways.
One method is the standard manner: divide by the χ2 degrees of freedom d0 = (R − 1)(C − 1)
to get an F statistic with d0 numerator degrees of freedom and ν = n − L denominator degrees of
freedom. This is the form of the F statistic suggested by Koch, Freeman, and Freeman (1975) and
implemented in the CROSSTAB procedure of the SUDAAN software (Research Triangle Institute 1997,
release 7.5), and it is the method used by svy: tabulate when the noadjust option is specified
with wald or llwald.
Another technique is to adjust the F statistic by using
Fadj = (ν − d0 + 1)W/(νd0 ) with Fadj ∼ F (d0 , ν − d0 + 1)
This is the default adjustment for svy: tabulate. test and the other svy estimation com-
mands produce adjusted F statistics by default, using the same adjustment procedure. See Korn
and Graubard (1990) for a justification of the procedure.
The adjusted F statistic is identical to the unadjusted F statistic when d0 = 1, that is, for 2 × 2
tables.
As Thomas and Rao (1987) point out (also see Korn and Graubard [1990]), the unadjusted
F statistics can become extremely anticonservative as d0 increases when ν is small or moderate;
that is, under the null, the statistics are “significant” far more often than they should be. Because
the unadjusted statistics behave so poorly for larger tables when ν is not large, their use can be
justified only for small tables or when ν is large. But when the table is small or when ν is large,
the unadjusted statistic is essentially identical to the adjusted statistic. Hence, for statistical inference,
looking at the unadjusted statistics has no point.
The adjusted “Pearson” Wald F statistic usually behaves reasonably under the null. However, even
the adjusted F statistic for the log-linear Wald test tends to be moderately anticonservative when ν
is not large (Thomas and Rao 1987; Sribney 1998).
Example 3
With the NHANES II data, we tabulate, for the male subpopulation, high blood pressure (highbp)
versus a variable (sizplace) that indicates the degree of urbanity/ruralness. We request that all eight
statistics for the test of independence be displayed.
. generate male = (sex==1) if !missing(sex)
. svy, subpop(male): tabulate highbp sizplace, col obs pearson lr null wald
> llwald noadj
Design df = 31
1 if BP >
140/90, 0 1=urban,..., 8=rural
otherwise 1 2 3 4 5 6 7 8 Total
0 .8489 .8929 .9213 .8509 .8413 .9242 .8707 .8674 .8764

431 527 558 371 186 210 314 1619 4216
1 .1511 .1071 .0787 .1491 .1587 .0758 .1293 .1326 .1236

95 80 64 74 36 20 57 273 699
Total 1 1 1 1 1 1 1 1 1
526 607 622 445 222 230 371 1892 4915
Key: column proportion

number of observations
Pearson:
D-B (null) F(5.30, 164.45) = 2.2078 P = 0.0522
Design-based F(5.54, 171.87) = 2.6863 P = 0.0189
Likelihood ratio:
D-B (null) F(5.30, 164.45) = 2.3372 P = 0.0408
Design-based F(5.54, 171.87) = 2.8437 P = 0.0138
Wald (Pearson):
Unadjusted chi2(7) = 21.2704
Unadjusted F(7, 31) = 3.0386 P = 0.0149
Adjusted F(7, 25) = 2.4505 P = 0.0465
Wald (log-linear):
Unadjusted chi2(7) = 25.7644
Unadjusted F(7, 31) = 3.6806 P = 0.0052
Adjusted F(7, 25) = 2.9683 P = 0.0208
The p-values from the null-corrected Pearson and likelihood-ratio statistics (lines labeled “D-B
(null)”; “D-B” stands for “design-based”) are bigger than the corresponding default-corrected statistics
(lines labeled “Design-based”). Simulations (Sribney 1998) show that the null-corrected statistics are
overly conservative for many sparse tables (except 2 × 2 tables); this appears to be the case here,
although this table is hardly sparse. The default-corrected Pearson statistic has good properties under
the null for both sparse and nonsparse tables; hence, the smaller p-value for it should be considered
reliable.
The default-corrected likelihood-ratio statistic is usually similar to the default-corrected Pearson
statistic except for sparse tables, when it tends to be anticonservative. This example follows this
pattern, with its p-value being slightly smaller than that of the default-corrected Pearson statistic.
For tables of these dimensions (2 × 8), the unadjusted “Pearson” Wald and log-linear Wald
F statistics are extremely anticonservative under the null when the variance degrees of freedom is
small. Here the variance degrees of freedom is only 31 (62 PSUs minus 31 strata), so we expect that
the unadjusted Wald F statistics yield smaller p-values than the adjusted F statistics. Because of
their poor behavior under the null for small variance degrees of freedom, they cannot be trusted here.
Simulations show that although the adjusted “Pearson” Wald F statistic has good properties under
the null, it is often less powerful than the default Rao and Scott–corrected statistics. That is probably
the explanation for the larger p-value for the adjusted “Pearson” Wald F statistic than that for the
default-corrected Pearson and likelihood-ratio statistics.
The p-value for the adjusted log-linear Wald F statistic is about the same as that for the trustworthy
default-corrected Pearson statistic. However, that is probably because of the anticonservatism of the
log-linear Wald under the null balancing out its lower power under alternative hypotheses.
The “uncorrected” χ2 Pearson and likelihood-ratio statistics displayed in the table are misspecified
statistics; that is, they are based on an i.i.d. assumption, which is not valid for complex survey data.
Hence, they are not correct, even asymptotically. The “unadjusted” Wald χ2 statistics, on the other
hand, are completely different. They are valid asymptotically as the variance degrees of freedom
becomes large.
Properties of the statistics

This section briefly summarizes the properties of the eight statistics computed by svy: tabulate.
For details, see Sribney (1998), Rao and Thomas (1989), Thomas and Rao (1987), and Korn and
Graubard (1990).
pearson is the Rao and Scott (1984) second-order corrected Pearson statistic, computed using pbrc
in the correction (default correction). It is displayed by default. Simulations show it to have good
properties under the null for both sparse and nonsparse tables. Its power is similar to that of the
lr statistic in most situations. It often appears to be more powerful than the adjusted “Pearson”
Wald F statistic (wald option), especially for larger tables. We recommend using this statistic in
all situations.
pearson null is the Rao and Scott second-order corrected Pearson statistic, computed using pb0rc in
the correction. It is numerically similar to the pearson statistic for nonsparse tables. For sparse
tables, it can be erratic. Under the null, it can be anticonservative for sparse 2 × 2 tables but
conservative for larger sparse tables.
lr is the Rao and Scott second-order corrected likelihood-ratio statistic, computed using pbrc in the
correction (default correction). The correction is identical to that for pearson. It is numerically
similar to the pearson statistic for nonsparse tables. It can be anticonservative (p-values too small)
in sparse tables. If there is a zero cell, it cannot be computed.
lr null is the Rao and Scott second-order corrected likelihood-ratio statistic, computed using pb0rc
in the correction. The correction is identical to that for pearson null. It is numerically similar
to the lr statistic for nonsparse tables. For sparse tables, it can be overly conservative. If there is
a zero cell, it cannot be computed.
wald statistic is the adjusted “Pearson” Wald F statistic. It has good properties under the null for
nonsparse tables. It can be erratic for sparse 2 × 2 tables and some sparse large tables. The pearson
statistic often appears to be more powerful.
wald noadjust is the unadjusted “Pearson” Wald F statistic. It can be extremely anticonservative
under the null when the table degrees of freedom (number of rows minus one times the number of
columns minus one) approaches the variance degrees of freedom (number of sampled PSUs minus
the number of strata). It is the same as the adjusted wald statistic for 2 × 2 tables. It is similar
to the adjusted wald statistic for small tables, large variance degrees of freedom, or both.
llwald statistic is the adjusted log-linear Wald F statistic. It can be anticonservative for both sparse
and nonsparse tables. If there is a zero cell, it cannot be computed.
llwald noadjust statistic is the unadjusted log-linear Wald F statistic. Like wald noadjust, it
can be extremely anticonservative under the null when the table degrees of freedom approaches
the variance degrees of freedom. It also suffers from the same general anticonservatism of the
llwald statistic. If there is a zero cell, it cannot be computed.
Stored results
In addition to the results documented in [SVY] svy, svy: tabulate stores the following in e():
Scalars
e(r) number of rows e(c) number of columns
e(cvgdeff) c.v. of generalized DEFF eigenval- e(mgdeff) mean generalized DEFF
ues
e(total) weighted sum of tab() variable
e(F Pear) default-corrected Pearson F e(F Penl) null-corrected Pearson F
e(df1 Pear) numerator d.f. for e(F Pear) e(df1 Penl) numerator d.f. for e(F Penl)
e(df2 Pear) denominator d.f. for e(F Pear) e(df2 Penl) denominator d.f. for e(F Penl)
e(p Pear) p-value for e(F Pear) e(p Penl) p-value for e(F Penl)
e(cun Pear) uncorrected Pearson χ2 e(cun Penl) null variant uncorrected Pearson χ2
e(F LR) default-corrected likelihood-ratio F e(F LRnl) null-corrected likelihood-ratio F
e(df1 LR) numerator d.f. for e(F LR) e(df1 LRnl) numerator d.f. for e(F LRnl)
e(df2 LR) denominator d.f. for e(F LR) e(df2 LRnl) denominator d.f. for e(F LRnl)
e(p LR) p-value for e(F LR) e(p LRnl) p-value for e(F LRnl)
e(cun LR) uncorrected likelihood-ratio χ2 e(cun LRln) null variant uncorrected
likelihood-ratio χ2
e(F Wald) adjusted “Pearson” Wald F e(F LLW) adjusted log-linear Wald F
e(p Wald) p-value for e(F Wald) e(p LLW) p-value for e(F LLW)
e(Fun Wald) unadjusted “Pearson” Wald F e(Fun LLW) unadjusted log-linear Wald F
e(pun Wald) p-value for e(Fun Wald) e(pun LLW) p-value for e(Fun LLW)
e(cun Wald) unadjusted “Pearson” Wald χ2 e(cun LLW) unadjusted log-linear Wald χ2
Macros
e(cmd) tabulate e(colvlab) column variable label
e(tab) tab() variable e(rowvar) varname1 , the row variable
e(rowlab) label or empty e(colvar) varname2 , the column variable
e(collab) label or empty e(setype) cell, count, column, or row
e(rowvlab) row variable label
Matrices
e(Prop) matrix of cell proportions e(V col) variance for column totals
e(Obs) matrix of observation counts e(V srs row) Vsrs for row totals
e(Deff) DEFF vector for e(setype) items e(V srs col) Vsrs for column totals
e(Deft) DEFT vector for e(setype) items e(Deff row) DEFF for row totals
e(Row) values for row variable e(Deff col) DEFF for column totals
e(Col) values for column variable e(Deft row) DEFT for row totals
e(V row) variance for row totals e(Deft col) DEFT for column totals

The table items
Confidence intervals
The test statistics
See Coefficient of variation under Methods and formulas of [SVY] estat for information on the
coefficient of variation (the cv option).
The table items

For a table of R rows by C columns with cells indexed by r, c, let
1 if the j th observation of the data is in the r, cth cell
n
y(rc)j =
0 otherwise
where j = 1, . . . , m indexes individuals in the sample. Weighted cell counts (count option) are
m
X
N
brc = wj y(rc)j
j=1
where wj is a sampling weight. If a variable, xj , is specified with the tab() option, N

brc becomes
m
X
N
brc = wj xj y(rc)j
j=1
Let
C
X R
X R X
X C
N
br· = N
brc , N
b·c = N
brc , and N
b·· = N
brc
c=1 r=1 r=1 c=1
Estimated cell proportions are pbrc = Nbrc /N

b·· ; estimated row proportions (row option) are pbrow rc =
Nrc /Nr· ; estimated column proportions (column option) are pbcol rc = N
b b brc /Nb·c ; estimated row
marginals are pbr· = Nr· /N·· ; and estimated column marginals are pb·c = N·c /N·· .
b b b b
N
brc is a total, the proportion estimators are ratios, and their variances can be estimated using
linearization methods as outlined in [SVY] variance estimation. svy: tabulate computes the variance
estimates by using svy: mean, svy: ratio, and svy: total.
Confidence intervals for proportions are calculated using a logit transform so that the endpoints
lie between 0 and 1. Let pb be an estimated proportion and sb be an estimate of its standard error. Let

pb
f (b
p) = ln
1 − pb
be the logit transform of the proportion. In this metric, an estimate of the standard error is
sb
SE p)} = f 0 (b
c {f (b p)b
s =
pb(1 − pb)
Thus a 100(1 − α)% confidence interval in this metric is

pb t1−α/2,ν sb
ln ±
1 − pb pb(1 − pb)
where t1−α/2,ν is the (1 − α/2)th quantile of Student’s t distribution with ν degrees of freedom.
The endpoints of this confidence interval are transformed back to the proportion metric by using the
inverse of the logit transform
ey
f −1 (y) =
1 + ey
Hence, the displayed confidence intervals for proportions are

−1 pb t1−α/2,ν sb
f ln ±
1 − pb pb(1 − pb)
Confidence intervals for weighted counts are untransformed and are identical to the intervals produced
by svy: total.
The test statistics

The uncorrected Pearson χ2 statistic is
R X
X C
2
XP2 = m prc − pb0rc ) /b
(b p0rc
r=1 c=1
and the uncorrected likelihood-ratio χ2 statistic is
R X
X C
2
XLR = 2m pbrc ln (b
prc /b
p0rc )
r=1 c=1
where m is the total number of sampled individuals, pbrc is the estimated proportion for the cell in the
rth row and cth column of the table as defined earlier, and pb0rc is the estimated proportion under the
null hypothesis of independence; that is, pb0rc = pbr· pb·c , the product of the row and column marginals.
Rao and Scott (1981, 1984) show that, asymptotically, XP2 and XLR
2
are distributed as
(R−1)(C−1)
X
X2 ∼ δ k Wk (1)
k=1
where the Wk are independent χ21 variables and the δk are the eigenvalues of
e 0 Vsrs X
∆ = (X e 2 )−1 (X
e 0 VX
e 2) (2)
2 2
where V is the variance of the pbrc under the survey design and Vsrs is the variance of the pbrc that
you would have if the design were simple random sampling; namely, Vsrs has diagonal elements
prc (1 − prc )/m and off-diagonal elements −prc pst /m.
X
e 2 is calculated as follows. Rao and Scott do their development in a log-linear modeling context,
so consider [ 1 | X1 | X2 ] as predictors for the cell counts of the R × C table in a log-linear model.
The X1 matrix of dimension RC × (R + C − 2) contains the R − 1 “main effects” for the rows
and the C − 1 “main effects” for the columns. The X2 matrix of dimension RC × (R − 1)(C − 1)
contains the row and column “interactions”. Hence, fitting [ 1 | X1 | X2 ] gives the fully saturated
model (that is, fits the observed values perfectly) and [ 1 | X1 ] gives the independence model. The
X
e 2 matrix is the projection of X2 onto the orthogonal complement of the space spanned by the
columns of X1 , where the orthogonality is defined with respect to Vsrs ; that is, Xe 0 Vsrs X1 = 0.
2
See Rao and Scott (1984) for the proof justifying (1) and (2). However, even without a full
understanding, you can get a feeling for ∆. It is like a ratio (although remember that it is a matrix) of
two variances. The variance in the numerator involves the variance under the true survey design, and
the variance in the denominator involves the variance assuming that the design was simple random
sampling. The design effect DEFF for an estimated proportion (see [SVY] estat) is defined as
Vb (b
prc )
DEFF =
Vsrsor (e
e prc )
Hence, ∆ can be regarded as a design-effects matrix, and Rao and Scott call its eigenvalues, the δk s,
the “generalized design effects”.
Computing an estimate for ∆ by using estimates for V and Vsrs is easy. Rao and Scott (1984)
derive a simpler formula for ∆
b:
b = C0 D−1 Vbsrs D−1 C −1 C0 D−1 Vb D−1 C

∆
p
b p
b p
b p
b
Here C is a contrast matrix that is any RC × (R − 1)(C − 1) full-rank matrix orthogonal to [ 1 | X1 ];

that is, C0 1 = 0 and C0 X1 = 0. Db p
is a diagonal matrix with the estimated proportions pbrc on the
diagonal. When one of the pbrc is zero, the corresponding variance estimate is also zero; hence, the
corresponding element for D−1 is immaterial for computing ∆ b.
p
b
Unfortunately, (1) is not practical for computing a p-value. However, you can compute simple
first-order and second-order corrections based on it. A first-order correction is based on downweighting
the i.i.d. statistics by the average eigenvalue of ∆b ; namely, you compute
XP2 (δb· ) = XP2 /δb· and 2

XLR 2
(δb· ) = XLR /δb·
where δb· is the mean-generalized DEFF

(R−1)(C−1)
1 X
δb· = δk
(R − 1)(C − 1)
k=1
These corrected statistics are asymptotically distributed as χ2(R−1)(C−1) . Thus, to first-order, you can
view the i.i.d. statistics XP2 and XLR
2
as being “too big” by a factor of δb· for true survey design.
A better second-order correction can be obtained by using the Satterthwaite approximation to the
distribution of a weighted sum of χ21 variables. Here the Pearson statistic becomes
XP2
XP2 (δb· , b
a) = (3)
a2 + 1)
δb· (b
a is the coefficient of variation of the eigenvalues:
where b
P b2
δk
a2 =
b −1
(R − 1)(C − 1)δb·2
Because
Pb
δk = tr ∆ b 2 , (3) can be written in an easily computable form as
b and P δb2 = tr ∆
k
tr ∆
b
XP2 (δb· , b
a) = X2
b2 P
tr ∆
These corrected statistics are asymptotically distributed as χ2d , with
(R − 1)(C − 1) b 2
(tr ∆)
d= 2
=
a +1
b b2
tr ∆
that is, a χ2 with, in general, noninteger degrees of freedom. The likelihood-ratio statistic XLR
2
can
also be given this second-order correction in an identical manner.
Two issues remain. First, there are two possible ways to compute the variance estimate Vbsrs ,
which is used to compute ∆ b . Vsrs has diagonal elements prc (1 − prc )/m and off-diagonal elements
−prc pst /m, but here prc is the true, not estimated, proportion. Hence, the question is what to use
to estimate prc : the observed proportions, pbrc , or the proportions estimated under the null hypothesis
of independence, pb0rc = pbr· pb·c ? Rao and Scott (1984, 53) leave this as an open question.
Because of the question of using pbrc or pb0rc to compute Vbsrs , svy: tabulate can compute both
corrections. By default, when the null option is not specified, only the correction based on pbrc is
displayed. If null is specified, two corrected statistics and corresponding p-values are displayed, one
computed using pbrc and the other using pb0rc .
The second outstanding issue concerns the degrees of freedom resulting from the variance estimate,
Vb , of the cell proportions under the survey design. The customary degrees of freedom for t statistics
resulting from this variance estimate is ν = n − L, where n is the number of PSUs in the sample
and L is the number of strata.
Rao and Thomas (1989) suggest turning the corrected χ2 statistic into an F statistic by dividing
it by its degrees of freedom, d0 = (R − 1)(C − 1). The F statistic is then taken to have numerator
degrees of freedom equal to d0 and denominator degrees of freedom equal to νd0 . Hence, the corrected
Pearson F statistic is
XP2 b 2
(tr ∆)
FP = with FP ∼ F (d, νd) where d = and ν = n − L (4)
tr ∆
b b2
tr ∆
This is the corrected statistic that svy: tabulate displays by default or when the pearson option
is specified. When the lr option is specified, an identical correction is produced for the likelihood-ratio
2
statistic XLR . When null is specified, (4) is also used. For the statistic labeled “D-B (null)”, ∆ b is
computed using pb0rc . For the statistic labeled “Design-based”, ∆ is computed using pbrc .
b
The Wald statistics computed by svy: tabulate with the wald and llwald options were developed
by Koch, Freeman, and Freeman (1975). The statistic given by the wald option is similar to the
Pearson statistic because it is based on
brc − N
Ybrc = N br· N
b·c /N
b··
where r = 1, . . . , R − 1 and c = 1, . . . , C − 1. The delta method can be used to estimate the

variance of Y
b (which is Ybrc stacked into a vector), and a Wald statistic can be constructed in the
usual manner:
b N 0 −1 Y
b 0 JN Vb (N)J b0

W =Y b where JN = ∂ Y/∂
b N
The statistic given by the llwald option is based on the log-linear model with predictors [1|X1 |X2 ]
that was mentioned earlier. This Wald statistic is
0 −1 0
WLL = X02 lnp b X02 Jp Vb (b p)Jp 0 X2

X2 lnp
b
where Jp is the matrix of first derivatives of lnp

b with respect to p
b , which is, of course, just a matrix
with pb−1
rc on the diagonal and zero elsewhere. This log-linear Wald statistic is undefined when there
is a zero cell in the table.
Unadjusted F statistics (noadjust option) are produced using
Funadj = W/d0 with Funadj ∼ F (d0 , ν)
Adjusted F statistics are produced using
Fadj = (ν − d0 + 1)W/(νd0 ) with Fadj ∼ F (d0 , ν − d0 + 1)
The other svy estimators also use this adjustment procedure for F statistics. See Korn and
Graubard (1990) for a justification of the procedure.
References
Fuller, W. A., W. J. Kennedy, Jr., D. Schnell, G. Sullivan, and H. J. Park. 1986. PC CARP. Software package. Ames,
IA: Statistical Laboratory, Iowa State University.
Jann, B. 2008. Multinomial goodness-of-fit: Large-sample tests with survey design correction and exact tests for small
samples. Stata Journal 8: 147–169.
Koch, G. G., D. H. Freeman, Jr., and J. L. Freeman. 1975. Strategies in the multivariate analysis of data from
complex surveys. International Statistical Review 43: 59–78.
Rao, J. N. K., and A. J. Scott. 1981. The analysis of categorical data from complex sample surveys: Chi-squared
tests for goodness of fit and independence in two-way tables. Journal of the American Statistical Association 76:
221–230.
. 1984. On chi-squared tests for multiway contingency tables with cell proportions estimated from survey data.
Annals of Statistics 12: 46–60.
Rao, J. N. K., and D. R. Thomas. 1989. Chi-squared tests for contingency tables. In Analysis of Complex Surveys,
ed. C. J. Skinner, D. Holt, and T. M. F. Smith, 89–114. New York: Wiley.
Research Triangle Institute. 1997. SUDAAN User’s Manual, Release 7.5. Research Triangle Park, NC: Research
Triangle Institute.
Sribney, W. M. 1998. svy7: Two-way contingency tables for survey or clustered data. Stata Technical Bulletin 45:
33–49. Reprinted in Stata Technical Bulletin Reprints, vol. 8, pp. 297–322. College Station, TX: Stata Press.
Thomas, D. R., and J. N. K. Rao. 1987. Small-sample comparisons of level and power for simple goodness-of-fit
statistics under cluster sampling. Journal of the American Statistical Association 82: 630–636.
Also see
[R] tabulate twoway — Two-way table of frequencies
[R] test — Test linear hypotheses after estimation
[SVY] svy: tabulate oneway — One-way tables for survey data
Title
svydescribe — Describe survey data
Description Menu Syntax Options Remarks and examples

Methods and formulas References Also see
Description
svydescribe displays a table that describes the strata and the sampling units for a given sampling
stage in a survey dataset.
Menu
Statistics > Survey data analysis > Setup and utilities > Describe survey data
Syntax

svydescribe varlist if in , options
options Description
Main
stage(#) sampling stage to describe; default is stage(1)
finalstage display information per sampling unit in the final stage
single display only the strata with one sampling unit
generate(newvar) generate a variable identifying strata with one sampling unit
svydescribe requires that the survey design variables be identified using svyset; see [SVY] svyset.
Options

Main
stage(#) specifies the sampling stage to describe. The default is stage(1).
finalstage specifies that results be displayed for each sampling unit in the final sampling stage;
that is, a separate line of output is produced for every sampling unit in the final sampling stage.
This option is not allowed with stage(), single, or generate().
single specifies that only the strata containing one sampling unit be displayed in the table.
generate(newvar) stores a variable that identifies strata containing one sampling unit for a given
sampling stage.
161
162 svydescribe — Describe survey data

Survey datasets are typically the result of a stratified survey design with cluster sampling in one
or more stages. Within a stratum for a given sampling stage, there are sampling units, which may be
either clusters of observations or individual observations.
svydescribe displays a table that describes the strata and sampling units for a given sampling
stage. One row of the table is produced for each stratum. Each row contains the number of sampling
units, the range and mean of the number of observations per sampling unit, and the total number
of observations. If the finalstage option is specified, one row of the table is produced for each
sampling unit of the final stage. Here each row contains the number of observations for the respective
sampling unit.
If a varlist is specified, svydescribe reports the number of sampling units that contain at least
one observation with complete data (that is, no missing values) for all variables in varlist. These are
the sampling units that would be used to compute point estimates by using the variables in varlist
with a given svy estimation command.
Example 1: Strata with one sampling unit

We use data from the Second National Health and Nutrition Examination Survey (NHANES II)
(McDowell et al. 1981) as our example. First, we set the PSU, pweight, and strata variables.
pweight: finalwgt
VCE: linearized
Strata 1: stratid
SU 1: psuid
FPC 1: <zero>
svydescribe will display the strata and PSU arrangement of the dataset.
. svydescribe
pweight: finalwgt
VCE: linearized
Strata 1: stratid
SU 1: psuid
FPC 1: <zero>
#Obs per Unit
1 2 380 165 190.0 215

2 2 185 67 92.5 118
3 2 348 149 174.0 199
(output omitted )
17 2 393 180 196.5 213
18 2 359 144 179.5 215
20 2 285 125 142.5 160
21 2 214 102 107.0 112
(output omitted )
31 2 308 143 154.0 165
32 2 450 211 225.0 239
31 62 10,351 67 167.0 288

svydescribe — Describe survey data 163
Our NHANES II dataset has 31 strata (stratum 19 is missing) and two PSUs per stratum.
The hdresult variable contains serum levels of high-density lipoprotein (HDL). If we try to
estimate the mean of hdresult, we get a missing value for the standard-error estimate and a note
explaining why.
. svy: mean hdresult
Design df = 29
Linearized
hdresult 49.67141 . . .
Note: Missing standard error because of stratum with single

sampling unit.
Running svydescribe with hdresult and the single option will show which strata have only one
PSU.
. svydescribe hdresult, single
Survey: Describing strata with a single sampling unit in stage 1
pweight: finalwgt
VCE: linearized
Strata 1: stratid
SU 1: psuid
FPC 1: <zero>
#Obs with #Obs with #Obs per included Unit
#Units #Units complete missing
Stratum included omitted data data min mean max
1 1* 1 114 266 114 114.0 114

2 1* 1 98 87 98 98.0 98
Both stratid = 1 and stratid = 2 have only one PSU with nonmissing values of hdresult.
Because this dataset has only 62 PSUs, the finalstage option produces a manageable amount of
output:
. svydescribe hdresult, finalstage

Survey: Describing final stage sampling units
pweight: finalwgt
VCE: linearized
Strata 1: stratid
SU 1: psuid
FPC 1: <zero>
#Obs with #Obs with
complete missing
Stratum Unit data data
1 1 0 215
1 2 114 51
2 1 98 20
2 2 0 67
(output omitted )
32 2 203 8
31 62 8,720 1,631
10,351
It is rather striking that there are two PSUs with no values for hdresult. All other PSUs have only
a moderate number of missing values. Obviously, here a data analyst should first try to ascertain why
these data are missing. The answer here (C. L. Johnson, 1995, pers. comm.) is that HDL measurements
could not be collected until the third survey location. Thus there are no hdresult data for the first
two locations: stratid = 1, psuid = 1 and stratid = 2, psuid = 2.
Assuming that we wish to go ahead and analyze the hdresult data, we must collapse strata—that
is, merge them—so that every stratum has at least two PSUs with some nonmissing values. We can
accomplish this by collapsing stratid = 1 into stratid = 2. To perform the stratum collapse, we
create a new strata identifier, newstr, and a new PSU identifier, newpsu.
. generate newstr = stratid
. generate newpsu = psuid
. replace newpsu = psuid + 2 if stratid == 1
(380 real changes made)
. replace newstr = 2 if stratid == 1
(380 real changes made)
svyset the new PSU and strata variables.

. svyset newpsu [pweight=finalwgt], strata(newstr)
pweight: finalwgt
VCE: linearized
Strata 1: newstr
SU 1: newpsu
FPC 1: <zero>
Then use svydescribe to check what we have done.

svydescribe — Describe survey data 165
. svydescribe hdresult, finalstage

Survey: Describing final stage sampling units
pweight: finalwgt
VCE: linearized
Strata 1: newstr
SU 1: newpsu
FPC 1: <zero>
#Obs with #Obs with
complete missing
Stratum Unit data data
2 1 98 20
2 2 0 67
2 3 0 215
2 4 114 51
3 1 161 38
3 2 116 33
(output omitted )
32 1 180 59
32 2 203 8
30 62 8,720 1,631
10,351
The new stratum, newstr = 2, has four PSUs, two of which contain some nonmissing values of
hdresult. This is sufficient to allow us to estimate the mean of hdresult and get a nonmissing
standard-error estimate.
Design df = 30
Linearized
hdresult 49.67141 .3830147 48.88919 50.45364
Example 2: Using e(sample) to find strata with one sampling unit

Some estimation commands drop observations from the estimation sample when they encounter
collinear predictors or perfect predictors. Ascertaining which strata contain one sampling unit is
therefore difficult. We can then use if e(sample) instead of varlist when faced with the problem
of strata with one sampling unit. We revisit the previous analysis to illustrate.
. use http://www.stata-press.com/data/r14/nhanes2b, clear

Design df = 29
Linearized
hdresult 49.67141 . . .
Note: Missing standard error because of stratum with single

sampling unit.
. svydescribe if e(sample), single
Survey: Describing strata with a single sampling unit in stage 1
pweight: finalwgt
VCE: linearized
Strata 1: stratid
SU 1: psuid
FPC 1: <zero>
#Obs per Unit
1 1* 114 114 114.0 114

2 1* 98 98 98.0 98

See Eltinge and Sribney (1996) for an earlier implementation of svydescribe.
References
Eltinge, J. L., and W. M. Sribney. 1996. svy3: Describing survey data: Sampling design and missing data. Stata
Technical Bulletin 31: 23–26. Reprinted in Stata Technical Bulletin Reprints, vol. 6, pp. 235–239. College Station,
TX: Stata Press.
Also see
Title
svymarkout — Mark observations for exclusion on the basis of survey characteristics
Description Syntax Remarks and examples Stored results Also see
Description
svymarkout is a programmer’s command that resets the values of markvar to contain 0 wherever
any of the survey-characteristic variables (previously set by svyset) contain missing values.
Syntax

svymarkout markvar

svymarkout assumes that markvar was created by marksample or mark; see [P] mark. This
command is most helpful for developing estimation commands that use ml to fit models using
maximum pseudolikelihood directly, instead of relying on the svy prefix; see [P] program properties
for a discussion of how to write programs to be used with the svy prefix.
Example 1
program mysvyprogram, ...
...
syntax ...
marksample touse
svymarkout ‘touse’
...
end
Stored results
svymarkout stores the following in s():
Macros
s(weight) weight variable set by svyset
Also see
[P] mark — Mark observations for inclusion
167
Title
svyset — Declare survey design for dataset

Options Remarks and examples Stored results References
Also see
Description
svyset manages the survey analysis settings of a dataset. You use svyset to designate variables
that contain information about the survey design, such as the sampling units and weights. svyset
is also used to specify other design characteristics, such as the number of sampling stages and the
sampling method, and analysis defaults, such as the method for variance estimation. You must svyset
your data before using any svy command; see [SVY] svy estimation.
svyset without arguments reports the current settings. svyset, clear removes the current survey
settings.
Quick start
One-stage design with sampling weight wvar1, strata defined by levels of svar, and sampling units
identified by su1
svyset su1 [pweight=wvar1], strata(svar)
Two-stage design with finite population correction fpc and n indicating second-stage sampling units
are the sampled individuals
svyset su1 [pweight=wvar1], strata(svar) fpc(fpc) || _n
Two-stage design with second-stage clustering defined by su2
svyset su1 [pweight=wvar1], strata(svar) fpc(fpc) || su2
Jackknife variance estimation as the default for svy commands, with sampling weight wvar2, and
replicate-weight variables rwvar*
svyset [pweight=wvar2], vce(jackknife) jkrweight(rwvar*)
As above, but use the MSE formula
svyset [pweight=wvar2], vce(jackknife) jkrweight(rwvar*) mse
Display current survey settings
svyset
Clear current survey settings
svyset, clear
Menu
Statistics > Survey data analysis > Setup and utilities > Declare survey design for dataset
168
svyset — Declare survey design for dataset 169
Syntax
Single-stage design

svyset psu weight , design options options
Multiple-stage design

svyset psu weight , design options || ssu, design options . . . options
Clear the current settings

svyset, clear
Report the current settings

svyset
psu identifies the primary sampling units and may be n or varname. In the single-stage syntax, psu
is optional and defaults to n.
n indicates that individuals were randomly sampled if the design does not involve clustered
sampling.
varname contains identifiers for the clusters in a clustered sampling design.
ssu is n or varname containing identifiers for sampling units (clusters) in subsequent stages of the
survey design.
n indicates that individuals were randomly sampled within the last sampling stage.
design options Description

Main
strata(varname) variable identifying strata
fpc(varname) finite population correction
weight(varname) stage-level sampling weight
170 svyset — Declare survey design for dataset
options Description
Weights
brrweight(varlist) balanced repeated replicate (BRR) weights
bsrweight(varlist) bootstrap replicate weights
jkrweight(varlist, . . . ) jackknife replicate weights
sdrweight(varlist, . . . ) successive difference replicate (SDR) weights
SE
vce(linearized) Taylor linearized variance estimation
vce(bootstrap) bootstrap variance estimation
vce(brr) BRR variance estimation
vce(jackknife) jackknife variance estimation
vce(sdr) SDR variance estimation
mse use the MSE formula with vce(bootstrap), vce(brr),
vce(jackknife), or vce(sdr)
singleunit(method) strata with a single sampling unit; method may be missing,
certainty, scaled, or centered
Poststratification
poststrata(varname) variable identifying poststrata
postweight(varname) poststratum population sizes
clear clear all settings from the data
noclear change some of the settings without clearing the others
clear(opnames) clear specified settings without clearing all others; opnames may be
one or more of weight, vce, dof, mse, bsrweight, brrweight,
jkrweight, sdrweight, or poststrata
pweights and iweights are allowed; see [U] 11.1.6 weight.
The full specification for jkrweight() is

jkrweight(varlist , stratum(# # . . . ) fpc(# # . . . ) multiplier(# # . . . ) reset )
The full specification for sdrweight() is

sdrweight(varlist , fpc(#) )
clear, noclear, and clear() are not shown in the dialog box.
Options

Main
strata(varname) specifies the name of a variable (numeric or string) that contains stratum identifiers.
fpc(varname) requests a finite population correction for the variance estimates. If varname has
values less than or equal to 1, it is interpreted as a stratum sampling rate fh = nh /Nh , where
nh = number of units sampled from stratum h and Nh = total number of units in the population
belonging to stratum h. If varname has values greater than or equal to nh , it is interpreted as
containing Nh . It is an error for varname to have values between 1 and nh or to have a mixture
of sampling rates and stratum sizes.
weight(varname) specifies a stage-level sampling weight variable. For most models, stage-level
sampling weights are multiplied together to create a single observation-level sampling weight
variable used for weighted estimation. For commands such as gsem and meglm, each stage-level
weight variable is assumed to correspond with a hierarchical group level in the model and is used
to compute the pseudolikelihood at that associated group level. Stage-level sampling weights are
required to be constant within their corresponding group level. For examples of fitting a multilevel
model with stage-level sampling weights, see example 5 and example 6 in [ME] meglm.

Weights
brrweight(varlist) specifies the replicate-weight variables to be used with vce(brr) or with svy
brr.
fay(#) specifies Fay’s adjustment (Judkins 1990). The value specified in fay(#) is used to adjust
the BRR weights and is present in the BRR variance formulas.
The sampling weight of the selected PSUs for a given replicate is multiplied by 2-#, where the
sampling weight for the unselected PSUs is multiplied by #. When brrweight(varlist) is specified,
the replicate-weight variables in varlist are assumed to be adjusted using #.
fay(0) is the default and is equivalent to the original BRR method. # must be between 0 and 2,
inclusive, and excluding 1. fay(1) is not allowed because this results in unadjusted weights.
bsrweight(varlist) specifies the replicate-weight variables to be used with vce(bootstrap) or
with svy bootstrap.
mean-weight variable specified in the bsrweight() option. The default is bsn(1). The value
specified in bsn(#) is used to adjust the variance estimate to account for mean bootstrap weights.
jkrweight(varlist, . . . ) specifies the replicate-weight variables to be used with vce(jackknife)
or with svy jackknife.
The following options set characteristics on the jackknife replicate-weight variables. If one value
is specified, all the specified jackknife replicate-weight variables will be supplied with the same
characteristic. If multiple values are specified, each replicate-weight variable will be supplied with
the corresponding value according to the order specified. These options are not shown in the dialog
box.

stratum(# # . . . ) specifies an identifier for the stratum in which the sampling weights have
been adjusted.

fpc(# # . . . ) specifies the FPC value to be added as a characteristic of the jackknife
replicate-weight variables. The values set by this suboption have the same interpretation as
the fpc(varname) option.

multiplier(# # . . . ) specifies the value of a jackknife multiplier to be added as a characteristic
of the jackknife replicate-weight variables.
reset indicates that the characteristics for the replicate-weight variables may be overwritten or
reset to the default, if they exist.
sdrweight(varlist, . . . ) specifies the replicate-weight variables to be used with vce(sdr) or with
svy sdr.
fpc(#) specifies the FPC value associated with the SDR weights. The value set by this suboption
has the same interpretation as the fpc(varname) option. This option is not shown in the dialog
box.

SE
vce(vcetype) specifies the default method for variance estimation; see [SVY] variance estimation.
vce(linearized) sets the default to Taylor linearization.
vce(bootstrap) sets the default to the bootstrap; also see [SVY] svy bootstrap.
vce(brr) sets the default to BRR; also see [SVY] svy brr.
vce(jackknife) sets the default to the jackknife; also see [SVY] svy jackknife.
vce(sdr) sets the default to the SDR; also see [SVY] svy sdr.
mse specifies that the MSE formula be used when vce(bootstrap), vce(brr), vce(jackknife),
or vce(sdr) is specified. This option requires vce(bootstrap), vce(brr), vce(jackknife),
or vce(sdr).
singleunit(method) specifies how to handle strata with one sampling unit.
singleunit(missing) results in missing values for the standard errors and is the default.
singleunit(certainty) causes strata with single sampling units to be treated as certainty units.
Certainty units contribute nothing to the standard error.
singleunit(scaled) results in a scaled version of singleunit(certainty). The scaling factor
comes from using the average of the variances from the strata with multiple sampling units for
each stratum with one sampling unit.
singleunit(centered) specifies that strata with one sampling unit are centered at the grand
mean instead of the stratum mean.

Poststratification
poststrata(varname) specifies the name of the variable (numeric or string) that contains poststratum
identifiers. See [SVY] poststratification for more information.
postweight(varname) specifies the name of the numeric variable that contains poststratum population
totals (or sizes), that is, the number of elementary sampling units in the population within each
poststratum. See [SVY] poststratification for more information.
The following options are available with svyset but are not shown in the dialog box:
clear clears all the settings from the data. Typing
. svyset, clear
clears the survey design characteristics from the data in memory. Although this option may be
specified with some of the other svyset options, it is redundant because svyset automatically
clears the previous settings before setting new survey design characteristics.
noclear allows some of the options in options to be changed without clearing all the other settings.
This option is not allowed with psu, ssu, design options, or clear.
clear(opnames) allows some of the options in options to be cleared without clearing all the other
settings. opnames refers to an option name and may be one or more of the following:
weight vce dof mse brrweight bsrweight jkrweight sdrweight
poststrata
This option implies the noclear option.

Introduction to survey design characteristics
Finite population correction (FPC)
Multiple-stage designs and with-replacement sampling
Replication-weight variables
Combining datasets from multiple surveys
Video example
Introduction to survey design characteristics

Stata’s suite of commands for survey data analysis relies on properly identified survey design
characteristics for point estimation, model fitting, and variance estimation. In fact, the svy prefix will
report an error if no survey design characteristics have been identified using svyset. Settings made
by svyset are saved with a dataset. So, if a dataset is saved after it has been svyset, it does not
have to be set again.
Typical survey design characteristics include sampling weights, one or more stages of clustered
sampling, and stratification. O’Donnell et al. (2008, 26–27) show four survey sample designs with
the corresponding svyset specification. Use svyset to declare your dataset to be complex survey
data by specifying the survey design variables. We will use the following contrived dataset for the
examples in this section.
. use http://www.stata-press.com/data/r14/stage5a
Example 1: Simple random sampling with replacement

Use n for psu to specify that the primary sampling units (PSUs) are the sampled individuals.
. svyset _n
pweight: <none>
VCE: linearized
Strata 1: <one>
FPC 1: <zero>
The output from svyset states that there are no sampling weights (each observation is given a
sampling weight of 1), there is only one stratum (which is the same as no stratification), and the
PSUs are the observed individuals.
Example 2: One-stage clustered design with stratification

The most commonly specified design, one-stage clustered design with stratification, can be used to
approximate multiple-stage designs when only the first-stage information is available. In this design,
the population is partitioned into strata and the PSUs are sampled independently within each stratum.
A dataset from this design will have a variable that identifies the strata, another variable that identifies
the PSUs, and a variable containing the sampling weights. Let’s assume that these variables are,
respectively, strata, su1, and pw.
. svyset su1 [pweight=pw], strata(strata)

pweight: pw
VCE: linearized
Strata 1: strata
SU 1: su1
FPC 1: <zero>
Example 3: Two-stage designs

In two-stage designs, the PSUs are sampled without replacement and then collections of individuals
are sampled within the selected PSUs. svyset uses || (double “or” bars) to separate the stage-
specific design specifications. The first-stage information is specified before ||, and the second-stage
information is specified afterward. We will assume that the variables containing the finite population
correction (FPC) information for the two stages are named fpc1 and fpc2; see Finite population
correction (FPC) for a discussion about the FPC.
Use n for ssu to specify that the second-stage sampling units are the sampled individuals.
. svyset su1 [pweight=pw], fpc(fpc1) || _n, fpc(fpc2)
pweight: pw
VCE: linearized
Strata 1: <one>
SU 1: su1
FPC 1: fpc1
Strata 2: <one>
FPC 2: fpc2
Suppose that su2 identifies the clusters of individuals sampled in the second stage.
. svyset su1 [pweight=pw], fpc(fpc1) || su2, fpc(fpc2)
pweight: pw
VCE: linearized
Strata 1: <one>
SU 1: su1
FPC 1: fpc1
Strata 2: <one>
SU 2: su2
FPC 2: fpc2
Stratification can take place in one or both of the sampling stages. Suppose that strata identifies
the second-stage strata and the first stage was not stratified.
. svyset su1 [pweight=pw], fpc(fpc1) || su2, fpc(fpc2) strata(strata)
pweight: pw
VCE: linearized
Strata 1: <one>
SU 1: su1
FPC 1: fpc1
Strata 2: strata
SU 2: su2
FPC 2: fpc2
Example 4: Multiple-stage designs

Specifying designs with three or more stages is not much more difficult than specifying two-stage
designs. Each stage will have its own variables for identifying strata, sampling units, and the FPC.
Not all stages will be stratified and some will be sampled with replacement; thus some stages may
not have a variable for identifying strata or the FPC.
Suppose that we have a three-stage design with variables su# and fpc# for the sampling unit and
FPC information in stage #. Also assume that the design called for stratification in the first stage only.
. svyset su1 [pweight=pw], fpc(fpc1) strata(strata)
> || su2, fpc(fpc2)
> || su3, fpc(fpc3)
pweight: pw
VCE: linearized
Strata 1: strata
SU 1: su1
FPC 1: fpc1
Strata 2: <one>
SU 2: su2
FPC 2: fpc2
Strata 3: <one>
SU 3: su3
FPC 3: fpc3
Use n for ssu in the last stage if the individuals are sampled within the third stage of clustered
sampling.
. svyset su1 [pweight=pw], fpc(fpc1) strata(strata)
> || su2, fpc(fpc2)
> || su3, fpc(fpc3)
> || _n
pweight: pw
VCE: linearized
Strata 1: strata
SU 1: su1
FPC 1: fpc1
Strata 2: <one>
SU 2: su2
FPC 2: fpc2
Strata 3: <one>
SU 3: su3
FPC 3: fpc3
Strata 4: <one>
FPC 4: <zero>
Finite population correction (FPC)

An FPC accounts for the reduction in variance that occurs when sampling without replacement from
a finite population compared to sampling with replacement from the same population. Specifying
an FPC variable for stage i indicates that the sampling units in that stage were sampled without
replacement. See Cochran (1977) for an introduction to variance estimation and sampling without
replacement.
Example 5
Consider the following dataset:
. use http://www.stata-press.com/data/r14/fpc
. list
stratid psuid weight nh Nh x
1. 1 1 3 5 15 2.8
2. 1 2 3 5 15 4.1
3. 1 3 3 5 15 6.8
4. 1 4 3 5 15 6.8
5. 1 5 3 5 15 9.2
6. 2 1 4 3 12 3.7
7. 2 2 4 3 12 6.6
8. 2 3 4 3 12 4.2
Here the variable nh is the number of PSUs per stratum that were sampled, Nh is the total number of
PSUs per stratum in the sampling frame (that is, the population), and x is our survey item of interest.
If we wish to use a finite population correction in our computations, we must svyset an FPC
variable when we specify the variables for sampling weights, PSUs, and strata. The FPC variable
typically contains the number of sampling units per stratum in the population; Nh is our FPC variable.
Here we estimate the population mean of x assuming sampling without replacement.
. svyset psuid [pweight=weight], strata(stratid) fpc(Nh)
pweight: weight
VCE: linearized
Strata 1: stratid
SU 1: psuid
FPC 1: Nh
. svy: mean x
Design df = 6
Linearized
x 5.448148 .6160407 3.940751 6.955545
We must respecify the survey design before estimating the population mean of x assuming sampling
with replacement.
. svyset psuid [pweight=weight], strata(stratid)
pweight: weight
VCE: linearized
Strata 1: stratid
SU 1: psuid
FPC 1: <zero>
. svy: mean x
Design df = 6
Linearized
x 5.448148 .7412683 3.63433 7.261966
Including an FPC always reduces the variance estimate. However, the reduction in the variance estimates
will be small when the Nh are large relative to the nh .
Rather than having a variable that represents the total number of PSUs per stratum in the sampling
frame, we sometimes have a variable that represents a sampling rate fh = nh /Nh . The syntax for
svyset is the same whether the FPC variable contains Nh or fh . The survey variance-estimation
routines in Stata are smart enough to identify what type of FPC information has been specified. If
the FPC variable is less than or equal to 1, it is interpreted as a sampling rate; if it is greater than
or equal to nh , it is interpreted as containing Nh . It is an error for the FPC variable to have values
between 1 and nh or to have a mixture of sampling rates and stratum sizes.
Multiple-stage designs and with-replacement sampling

Although survey data are seldom collected using with-replacement sampling, dropping the FPC
information when the sampling fractions are small is common. In either case, svyset ignores the
design variables specified in later sampling stages because this information is not necessary for
variance estimation. In the following, we describe why this is true.
Example 6
Consider the two-stage design where PSUs are sampled with replacement and individuals are
sampled without replacement within the selected PSUs. Sampling the individuals with replacement
would change some of the details in the following discussion, but the result would be the same.
Our population contains 100 PSUs, with five individuals in each, so our population size is 500.
We will sample 10 PSUs with replacement and then sample two individuals without replacement from
within each selected PSU. This results in a dataset with 10 PSUs, each with 2 observations, for a total
of 20 observations. If our dataset contained the PSU information in variable su1 and the second-stage
FPC information in variable fpc2, our svyset command would be as follows.
. use http://www.stata-press.com/data/r14/svyset_wr
. svyset su1 || _n, fpc(fpc2)
Note: Stage 1 is sampled with replacement; further stages will be ignored for
variance estimation.
pweight: <none>
VCE: linearized
Strata 1: <one>
SU 1: su1
FPC 1: <zero>
As expected, svyset tells us that it is ignoring the second-stage information because the first-stage
units were sampled with replacement. Because we do not have an FPC variable for the first stage, we
can regard the sampling of PSUs as a series of independently and identically distributed draws. The
second-sampled PSU is drawn independently from the first and has the same sampling distribution
because the first-sampled PSU is eligible to be sampled again.
Consider the following alternative scenario. Because there are 10 ways to pick two people of five,
let’s expand the 100 PSUs to form 100 × 10 = 1,000 “new PSUs” (NPSUs), each of size 2, representing
all possible two-person groups that can be sampled from the original 100 groups of five people. We
now have a population of 1,000 × 2 = 2,000 “new people”; each original person was replicated four
times. We can select 10 NPSUs with replacement to end up with a dataset consisting of 10 groups of
two to form samples of 20 people. If our “new” dataset contained the PSU information in variable
nsu1, our svyset command would be as follows:
. svyset nsu1
pweight: <none>
VCE: linearized
Strata 1: <one>
SU 1: nsu1
FPC 1: <zero>
There is nothing from a sampling standpoint to distinguish between our two scenarios. The
information contained in the variables su1 and nsu1 is equivalent; thus svyset can behave as if our
dataset came from the second scenario.
The following questions may spring to mind after reading the above:
• The population in the first scenario has 500 people; the second has 2,000. Does that not
invalidate the comparison between the two scenarios?
Although the populations are different, the sampling schemes described for each scenario result
in the same sampling space. By construction, each possible sample from the first scenario is
also a possible sample from the second scenario. For the first scenario, the number of possible
samples of 10 of 100 PSUs sampled with replacement, where two of five individuals are sampled
without replacement, is
10
10 5
100 × = 1030
2
For the second scenario, the number of possible samples of 10 of 1,000 NPSUs sampled with
replacement, where each NPSU is sampled as a whole, is
1,00010 = 1030
• Does the probability of being in the sample not depend on what happens in the first sampling
stage?
Not when the first stage is sampled with replacement. Sampling with replacement means that
all PSUs have the same chance of being selected even after one of the PSUs has been selected.
Thus each of the two-person groups that can possibly be sampled has the same chance of being
sampled even after a specific two-person group has been selected.
• Is it valid to have replicated people in the population like the one in the second scenario?
Yes, because each person in the population can be sampled more than once. Sampling with
replacement allows us to construct the replicated people.
Replication-weight variables
Many groups that collect survey data for public use have taken steps to protect the privacy of
the survey participants. This may result in datasets that have replicate-weight variables instead of
variables that identify the strata and sampling units from the sampling stages. These datasets require
replication methods for variance estimation.
The brrweight(), jkrweight(), bsrweight(), and sdrweight() options allow svyset to
identify the set of replication weights for use with BRR, jackknife, bootstrap, and SDR variance
estimation (svy brr, svy jackknife, svy bootstrap, and svy sdr), respectively. In addition to
the weight variables, svyset also allows you to change the default variance estimation method from
linearization to BRR, jackknife, bootstrap, or SDR.
Example 7
Here are two simple examples using jackknife replication weights.
1. Data containing only sampling weights and jackknife replication weights, and we set the default
variance estimator to the jackknife:
. use http://www.stata-press.com/data/r14/stage5a_jkw
. svyset [pweight=pw], jkrweight(jkw_*) vce(jackknife)
pweight: pw
VCE: jackknife
MSE: off
jkrweight: jkw_1 jkw_2 jkw_3 jkw_4 jkw_5 jkw_6 jkw_7 jkw_8 jkw_9
Strata 1: <one>
FPC 1: <zero>
2. Data containing only sampling weights and jackknife replication weights, and we set the default
variance estimator to the jackknife by using the MSE formula:
. svyset [pweight=pw], jkrweight(jkw_*) vce(jackknife) mse
pweight: pw
VCE: jackknife
MSE: on
jkrweight: jkw_1 jkw_2 jkw_3 jkw_4 jkw_5 jkw_6 jkw_7 jkw_8 jkw_9
Strata 1: <one>
FPC 1: <zero>
Example 8: Characteristics for jackknife replicate-weight variables

The jkrweight() option has suboptions that allow you to identify certain characteristics of the
jackknife replicate-weight variables. These characteristics include the following:
• An identifier for the stratum in which the sampling weights have been adjusted because one
of its PSUs was dropped. We use the stratum() suboption to set these values. The default is
one stratum for all the replicate-weight variables.
• The FPC value. We use the fpc() suboption to set these values. The default value is zero.
This characteristic is ignored when the mse option is supplied to svy jackknife.
• A jackknife multiplier used in the formula for variance estimation. The multiplier for the
standard leave-one-out jackknife method is
nh − 1
nh
where nh is the number of PSUs sampled from stratum h. We use the multiplier() suboption
to set these values. The default is derived from the above formula, assuming that nh is equal
to the number of replicate-weight variables for stratum h.
Because of privacy concerns, public survey datasets may not contain stratum-specific information.
However, the population size and an overall jackknife multiplier will probably be provided. You must
then supply this information to svyset for the jackknife replicate-weight variables. We will use the
1999–2000 NHANES data to illustrate how to set these characteristics.
The NHANES datasets for years 1999–2000 are available for download from the Centers for Disease
Control and Prevention (CDC) website, http://www.cdc.gov. This particular release of the NHANES data
contains jackknife replication weights in addition to the usual PSU and stratum information. These
variables are contained in the demographic dataset. In our web browser, we saved the demographic data
from the CDC website ftp://ftp.cdc.gov/pub/Health Statistics/NCHS/nhanes/1999-2000/DEMO.xpt. We
suggest that you rename the data to demo.xpt.
The 1999–2000 NHANES datasets are distributed in SAS Transport format, so we use Stata’s
import sasxport command to read the data into memory. Because of the nature of the survey
design, the demographic dataset demo.xpt has two sampling-weight variables. wtint2yr contains
the sampling weights appropriate for the interview data, and wtmec2yr contains the sampling weights
appropriate for the Mobile Examination Center (MEC) exam data. Consequently, there are two sets of
jackknife replicate-weight variables. The jackknife replicate-weight variables for the interview data
are named wtirep01, wtirep02, . . . , wtirep52. The jackknife replicate-weight variables for the
MEC exam data are named wtmrep01, wtmrep02, . . . , wtmrep52. The documentation published with
the NHANES data gives guidance on which weight variables to use.
. import sasxport demo.xpt
. describe wtint2yr wtmec2yr wtirep01 wtmrep01
storage display value
variable name type format label variable label
wtint2yr double %10.0g Full Sample 2 Year Interview

Weight
wtmec2yr double %10.0g Full Sample 2 Year MEC Exam Weight
wtirep01 double %10.0g Interview Weight Jack Knife
Replicate 01
wtmrep01 double %10.0g MEC Exam Weight Jack Knife
Replicate 01
The number of PSUs in the NHANES population is not apparent, so we will not set an FPC value, but
we can set the standard jackknife multiplier for the 52 replicate-weight variables and save the results
as a Stata dataset for future use. Also the NHANES datasets all contain a variable called seqn. This
variable has a respondent sequence number that allows the dataset users to merge the demographic
dataset with other 1999–2000 NHANES datasets, so we sort on seqn before saving demo99 00.dta.
. local mult = 51/52

. svyset, jkrweight(wtmrep*, multiplier(‘mult’))
(output omitted )
. svyset, jkrweight(wtirep*, multiplier(‘mult’))
(output omitted )
. svyset, clear
. sort seqn
. save demo99_00
file demo99_00.dta saved
To complete this example, we will perform a simple analysis using the blood pressure data;
however, before we can perform any analysis, we have to merge the blood pressure dataset, bpx.xpt,
with our demographic dataset, demo99 00.dta. In our web browser, we saved the blood pressure
data from the CDC website ftp://ftp.cdc.gov/pub/Health Statistics/NCHS/nhanes/1999-2000/BPX.xpt.
We suggest that you rename the data to bpx.xpt.
We can then use import sasxport to read in the blood pressure data, sort on seqn, and save
the resulting dataset to bpx99 00.dta. We read in our copy of the demographic data, drop the
irrelevant weight variables, and merge in the blood pressure data from bpx99 00.dta. A quick call
to tabulate on the merge variable generated by merge indicates that 683 observations in the
demographic data are not present in the blood pressure data. We do not drop these observations;
otherwise, the estimate of the population size will be incorrect. Finally, we set the appropriate sampling
and replicate-weight variables with svyset before replacing bpx99 00.dta with a more complete
copy of the blood pressure data.
. import sasxport bpx.xpt
. sort seqn
. save bpx99_00
file bpx99_00.dta saved
. use demo99_00
. drop wtint?yr wtirep*
. merge 1:1 seqn using bpx99_00
Result # of obs.
not matched 683

from master 683 (_merge==1)
from using 0 (_merge==2)
matched 9,282 (_merge==3)
. drop _merge
. svyset [pw=wtmec2yr], jkrweight(wtmrep*) vce(jackknife)
(output omitted )
. save bpx99_00, replace
Having saved our merged dataset (with svysettings), we estimate the mean systolic blood pressure
for the population, using the MEC exam replication weights for jackknife variance estimation.
. svy: mean bpxsar

1 2 3 4 5
.................................................. 50
..
Replications = 52
Design df = 51
Jackknife
bpxsar 119.7056 .5109122 118.6799 120.7313
Combining datasets from multiple surveys

The 2001–2002 NHANES datasets are also available from the CDC website, http://www.cdc.gov.
The guidelines that are published with these datasets recommend that the 1999–2000 and 2001–2002
NHANES datasets be combined to increase the accuracy of results. Combining datasets from multiple
surveys is a complicated process, and Stata has no specific tools for this task. However, the distributors
of the NHANES datasets provide sampling-weight variables for the 1999–2002 combined data in the
respective demographic datasets. They also provide some simple instructions on how to combine the
datasets from these two surveys.
In the previous example, we worked with the 1999–2000 NHANES data. The 2001–2002 NHANES
demographics data are contained in demo b.xpt, and the blood pressure data are contained in
bpx b.xpt. We follow the same steps as in the previous example to merge the blood pressure data
with the demographic data for 2001–2002.
Visit the following CDC websites and save the data:
ftp://ftp.cdc.gov/pub/Health Statistics/NCHS/nhanes/2001-2002/BPX B.xpt
ftp://ftp.cdc.gov/pub/Health Statistics/NCHS/nhanes/2001-2002/DEMO B.xpt
We suggest that you rename the data to bpx b.xpt and demo b.xpt. We can then continue with
our example:
. import sasxport bpx_b.xpt
. sort seqn
. save bpx01_02
. import sasxport demo_b.xpt
. drop wtint?yr
. sort seqn
. merge 1:1 seqn using bpx01_02

Result # of obs.
not matched 562

from master 562 (_merge==1)
from using 0 (_merge==2)
matched 10,477 (_merge==3)
. drop _merge
. svyset sdmvpsu [pw=wtmec2yr], strata(sdmvstra)
pweight: wtmec2yr
VCE: linearized
Strata 1: sdmvstra
SU 1: sdmvpsu
FPC 1: <zero>
. save bpx01_02, replace
The demographic dataset for 2001–2002 does not contain replicate-weight variables, but there are
variables that provide information on PSUs and strata for variance estimation. The PSU information
is contained in sdmvpsu, and the stratum information is in sdmvstra. See the documentation that
comes with the NHANES datasets for the details regarding these variables.
This new blood pressure dataset (bpx01 02.dta) is all we need if we are interested in analyzing
blood pressure data only for 2001–2002. However, we want to use the 1999–2002 combined data,
so we will follow the advice in the guidelines and just combine the datasets from the two surveys.
For those concerned about overlapping stratum identifiers between the two survey datasets, it is a
simple exercise to check that sdmvstra ranges from 1 to 13 for 1999–2000 but ranges from 14 to
28 for 2001–2002. Thus the stratum identifiers do not overlap, so we can simply append the data.
The 2001–2002 NHANES demographic dataset has no jackknife replicate-weight variables, so
we drop the replicate-weight variables from the 1999–2000 dataset. The sampling-weight variable
wtmec2yr is no longer appropriate for use with the combined data because its values are based on
the survey designs individually, so we drop it from the combined dataset. Finally, we use svyset
to identify the design variables for the combined surveys. wtmec4yr is the sampling-weight variable
for the MEC exam data developed by the data producers for the combined 1999–2002 NHANES data.
. use bpx99_00
. drop wt?rep*
. append using bpx01_02
. drop wtmec2yr
. svyset sdmvpsu [pw=wtmec4yr], strata(sdmvstra)
pweight: wtmec4yr
VCE: linearized
Strata 1: sdmvstra
SU 1: sdmvpsu
FPC 1: <zero>
. save bpx99_02
Now we can estimate the mean systolic blood pressure for our population by using the combined
surveys and jackknife variance estimation.
. svy jackknife: mean bpxsar

1 2 3 4 5
.................................................. 50
.......
Replications = 57
Design df = 29
Jackknife
bpxsar 119.8914 .3828434 119.1084 120.6744
Video example
Specifying the design of your survey data to Stata
Stored results
svyset stores the following in r():
Scalars
r(stages) number of sampling stages
r(stages wt) last stage containing stage-level weights
Macros
r(wtype) weight type
r(wexp) weight expression
r(wvar) weight variable name
r(weight#) variable identifying weight for stage #
r(su#) variable identifying sampling units for stage #
r(strata#) variable identifying strata for stage #
r(fpc#) FPC for stage #
r(bsrweight) bsrweight() variable list
r(bsn) bootstrap mean-weight adjustment
r(brrweight) brrweight() variable list
r(fay) Fay’s adjustment
r(jkrweight) jkrweight() variable list
r(sdrweight) sdrweight() variable list
r(sdrfpc) fpc() value from within sdrweight()
r(vce) vcetype specified in vce()
r(dof) dof() value
r(mse) mse, if specified
r(poststrata) poststrata() variable
r(postweight) postweight() variable
r(settings) svyset arguments to reproduce the current settings
r(singleunit) singleunit() setting
References
O’Donnell, O., E. van Doorslaer, A. Wagstaff, and M. Lindelow. 2008. Analyzing Health Equity Using Household
Survey Data: A Guide to Techniques and Their Implementation. Washington, DC: The World Bank.
Also see
Title
variance estimation — Variance estimation for survey data
Description Remarks and examples References Also see
Description
Stata’s suite of estimation commands for survey data use the most commonly used variance estima-
tion techniques: bootstrap, balanced repeated replication, jackknife, successive difference replication,
and linearization. The bootstrap, balanced repeated replication, jackknife, and successive difference
replication techniques are known as replication methods in the survey literature. We stick with that
nomenclature here, but note that these techniques are also known as resampling methods. This entry
discusses the details of these variance estimation techniques.
Also see Cochran (1977), Wolter (2007), and Shao and Tu (1995) for some background on these
variance estimators.

Variance of the total
Stratified single-stage design
Stratified two-stage design
Variance for census data
Certainty sampling units
Strata with one sampling unit
Ratios and other functions of survey data
Revisiting the total estimator
The ratio estimator
A note about score variables
Linearized/robust variance estimation
The bootstrap
BRR
The jackknife
The delete-one jackknife
The delete-k jackknife
Successive difference replication
Variance of the total

This section describes the methods and formulas for svy: total. The variance estimators not
using replication methods use the variance of a total as an important ingredient; this section therefore
also introduces variance estimation for survey data.
We will discuss the variance estimators for two complex survey designs:
1. The stratified single-stage design is the simplest design that has the elements present in most
complex survey designs.
2. Adding a second stage of clustering to the previous design results in a variance estimator for
designs with multiple stages of clustered sampling.
186
variance estimation — Variance estimation for survey data 187
Stratified single-stage design

The population is partitioned into groups called strata. Clusters of observations are randomly
sampled—with or without replacement—from within each stratum. These clusters are called primary
sampling units (PSUs). In single-stage designs, data are collected from every member of the sampled
PSUs. When the observed data are analyzed, sampling weights are used to account for the survey
design. If the PSUs were sampled without replacement, a finite population correction (FPC) is applied
to the variance estimator.
The svyset syntax to specify this design is
svyset psu [pweight=weight], strata(strata) fpc(fpc)
The stratum identifiers are contained in the variable named strata, PSU identifiers are contained in
variable psu, the sampling weights are contained in variable weight, and the values for the FPC are
contained in variable fpc.
Let h = 1, . . . , L count the strata and (h, i) denote the ith PSU in stratum h, where i = 1, . . . , Nh
and Nh is the number of PSUs in stratum h. Let (h, i, j) denote the j th individual from PSU (h, i)
and Mhi be the number of individuals in PSU (h, i); then
Nh
L X
X
M= Mhi
h=1 i=1
is the number of individuals in the population. Let Yhij be a survey item for individual (h, i, j); for
example, Yhij might be income for adult j living in block i of county h. The associated population
total is
XL XNh MX hi
Y = Yhij
h=1 i=1 j=1
Let yhij denote the items for individuals who are members of the sampled PSUs; here h = 1,
. . . , L; i = 1, . . . , nh ; and j = 1, . . . , mhi . The number of individuals in the sample (number of
observations) is
XL Xnh
m= mhi
h=1 i=1
The estimator for Y is

nh X
L X
X mhi
Yb = whij yhij
h=1 i=1 j=1
where whij is a sampling weight, and its unadjusted value for this design is whij = Nh /nh . The
estimator for the number of individuals in the population (population size) is
nh X
L X
X mhi
M
c= whij
h=1 i=1 j=1
188 variance estimation — Variance estimation for survey data
The estimator for the variance of Yb is

L h n
X nh X
Vb (Yb ) = (1 − fh ) (yhi − y h )2 (1)
nh − 1 i=1
h=1
where yhi is the weighted total for PSU (h, i),

mhi
X
yhi = whij yhij
j=1
and y h is the mean of the PSU totals for stratum h:

nh
1 X
yh = yhi
nh i=1
The factor (1 − fh ) is the FPC for stratum h, and fh is the sampling rate for stratum h. The sampling
rate fh is derived from the variable specified in the fpc() option of svyset. If an FPC variable is
not svyset, then fh = 0. If an FPC variable is set and its values are greater than or equal to nh ,
then the variable is assumed to contain the values of Nh , and fh is given by fh = nh /Nh . If its
values are less than or equal to 1, then the variable is assumed to contain the sampling rates fh .
If multiple variables are supplied to svy: total, covariances are also computed. The estimator
for the covariance between Yb and X b (notation for X is defined similarly to that of Y ) is
L hn
X nh X
Cov(
d Yb , X)
b = (1 − fh ) (yhi − y h )(xhi − xh )
nh − 1 i=1
h=1
Stratified two-stage design

The population is partitioned into strata. PSUs are randomly sampled without replacement from within
each stratum. Clusters of observations are then randomly sampled—with or without replacement—
from within the sampled PSUs. These clusters are called secondary sampling units (SSUs). Data are then
collected from every member of the sampled SSUs. When the observed data are analyzed, sampling
weights are used to account for the survey design. Each sampling stage provides a component to the
variance estimator and has its own FPC.
The svyset syntax to specify this design is
svyset psu [pweight=weight], strata(strata) fpc(fpc1 ) || ssu, fpc(fpc2 )
The stratum identifiers are contained in the variable named strata, PSU identifiers are contained in
variable psu, the sampling weights are contained in variable weight, the values for the FPC for the
first sampling stage are contained in variable fpc1 , SSU identifiers are contained in variable ssu, and
the values for the FPC for the second sampling stage are contained in variable fpc2 .
The notation for this design is based on the previous notation. There still are L strata, and (h, i)
identifies the ith PSU in stratum h. Let Mhi be the number of SSUs in PSU (h, i), Mhij be the number
of individuals in SSU (h, i, j), and
Nh M
L X
X Xhi
M= Mhij
h=1 i=1 j=1
be the population size. Let Yhijk be a survey item for individual (h, i, j, k); for example, Yhijk might
be income for adult k living in block j of county i of state h. The associated population total is
X Nh M
L X Xhi M
Xhij
Y = Yhijk
h=1 i=1 j=1 k=1
Let yhijk denote the items for individuals who are members of the sampled SSUs; here h = 1,
. . . , L; i = 1, . . . , nh ; j = 1, . . . , mhi ; and k = 1, . . . , mhij . The number of observations is
nh X
L X
X mhi
m= mhij
h=1 i=1 j=1
The estimator for Y is

X mhi m
nh X
L X X hij
Yb = whijk yhijk
h=1 i=1 j=1 k=1
where whijk is a sampling weight, and its unadjusted value for this design is

Nh Mhi
whijk =
nh mhi
The estimator for the population size is
X mhi m
nh X
L X X hij
M
c= whijk
h=1 i=1 j=1 k=1
The estimator for the variance of Yb is

L hn
X nh X
Vb (Yb ) = (1 − fh ) (yhi − y h )2
nh − 1 i=1
h=1
(2)
L nh hi m
X X mhi X
+ fh (1 − fhi ) (yhij − y hi )2
i=1
mhi − 1 j=1
h=1
where yhi is the weighted total for PSU (h, i); y h is the mean of the PSU totals for stratum h; yhij
is the weighted total for SSU (h, i, j),
mhij
X
yhij = whijk yhijk
k=1
and y hi is the mean of the SSU totals for PSU (h, i),
mhi
1 X
y hi = yhij
mhi j=1
Equation (2) is equivalent to (1) with an added term representing the increase in variability because
of the second stage of sampling. The factor (1 − fh ) is the FPC, and fh is the sampling rate for the
first stage of sampling. The factor (1 − fhi ) is the FPC, and fhi is the sampling rate for PSU (h, i).
The sampling rate fhi is derived in the same manner as fh .
If multiple variables are supplied to svy: total, covariances are also computed. For estimated
totals Yb and X
b (notation for X is defined similarly to that of Y ), the covariance estimator is
L n
h
X nh X
Cov(
d Yb , X)
b = (1 − fh ) (yhi − y h )(xhi − xh )
nh − 1 i=1
h=1
L nh mhi
X X mhi X
+ fh (1 − fhi ) (yhij − y hi )(xhij − xhi )
i=1
mhi − 1 j=1
h=1
On the basis of the formulas (1) and (2), writing down the variance estimator for a survey design
with three or more stages is a matter of deriving the variance component for each sampling stage.
The sampling units from a given stage pose as strata for the next sampling stage.
All but the last stage must be sampled without replacement to get nonzero variance components
from each stage of clustered sampling. For example, if fh = 0 in (2), the second stage contributes
nothing to the variance estimator.
Variance for census data

The point estimates that result from the analysis of census data, in which the entire population
was sampled without replacement, are the population’s parameters instead of random variables. As
such, there is no sample-to-sample variation if we consider the population fixed. Here the sampling
fraction is one; thus, if the FPC variable you svyset for the first sampling stage is one, Stata will
report a standard error of zero.
Certainty sampling units

Stata’s svy commands identify strata with an FPC equal to one as units sampling with certainty.
To properly determine the design degrees of freedom, certainty sampling units should be contained
within their own strata, one for each certainty unit, in each sampling stage. Although the observations
contained in certainty units from a given sampling stage play a role in parameter estimation, they
contribute nothing to the variance for that stage.
Strata with one sampling unit

By default, Stata’s svy commands report missing standard errors when they encounter a stratum
with one sampling unit. Although the best way to solve this problem is to reassign the sampling unit
to another appropriately chosen stratum, there are three automatic alternatives that you can choose
from, in the singleunit() option, when you svyset your data.
singleunit(certainty) treats the strata with single sampling units as certainty units.
singleunit(scaled) treats the strata with single sampling units as certainty units but multiplies
the variance components from each stage by a scaling factor. For a given sampling stage, suppose
that L is the total number of strata, Lc is the number of certainty strata, and Ls is the number
of strata with one sampling unit, and then the scaling factor is (L − Lc )/(L − Lc − Ls ). Using
this scaling factor is the same as using the average of the variances from the strata with multiple
sampling units for each stratum with one sampling unit.
singleunit(centered) specifies that strata with one sampling unit are centered at the population
mean instead of the stratum mean. The quotient nh /(nh − 1) in the variance formula is also taken
to be 1 if nh = 1.
Ratios and other functions of survey data

Shah (2004) points out a simple procedure for deriving the linearized variance for functions of
survey data that are continuous functions of the sampling weights. Let θ be a (possibly vector-valued)
function of the population data and θb be its associated estimator based on survey data.
1. Define the j th observation of the score variable by
∂ θb
zj =
∂wj
If θb is implicitly defined through estimating equations, zj can be computed by taking the partial
derivative of the estimating equations with respect to wj .
2. Define the weighted total of the score variable by
m
X
Zb = wj zj
j=1
3. Estimate the variance V (Z)

b by using the design-based variance estimator for the total Zb. This
variance estimator is an approximation of V (θ)
b.
Revisiting the total estimator

As a first example, we derive the variance of the total from a stratified single-stage design. Here
you have θb = Yb , and deriving the score variable for Yb results in the original values of the variable
of interest.
b = zj (Yb ) = ∂ Y = yj
b
zj (θ)
∂wj
Thus you trivially recover the variance of the total given in (1) and (2).
The ratio estimator

The estimator for the population ratio is
b= Y
b
R
X
b
and its score variable is
∂Rb yj − Rb xj
zj (R)
b = =
∂wj X
b
Plugging this into (1) or (2) results in a variance estimator that is algebraically equivalent to the
variance estimator derived from directly applying the delta method (a first-order Taylor expansion
with respect to y and x)
b = 1 Vb (Yb ) − 2R b2 Vb (X)

Vb (R) b Cov(
d Yb , X)b +R b
X
b 2
A note about score variables

The functional form of the score variable for each estimation command is detailed in the Methods
and formulas section of its manual entry; see [R] total, [R] ratio, and [R] mean.
Although Deville (1999) and Demnati and Rao (2004) refer to zj as the linearized variable, here
it is referred to as the score variable to tie it more closely to the model-based estimators discussed
in the following section.
Linearized/robust variance estimation

The regression models for survey data that allow the vce(linearized) option use linearization-
based variance estimators that are natural extensions of the variance estimator for totals. For gen-
eral background on regression and generalized linear model analysis of complex survey data, see
Binder (1983); Cochran (1977); Fuller (1975); Godambe (1991); Kish and Frankel (1974); Särndal,
Swensson, and Wretman (1992); and Skinner (1989).
Suppose that you observed (Yj , xj ) for the entire population and are interested in modeling the
relationship between Yj and xj by the vector of parameters β that solve the following estimating
equations:
XM
G(β) = S(β; Yj , xj ) = 0
j=1
For ordinary least squares, G(β) is the normal equations
G(β) = X 0 Y − X 0 X β = 0
where Y is the vector of outcomes for the full population and X is the matrix of explanatory variables
for the full population. For a pseudolikelihood model—such as logistic regression— G(β) is the first
derivative of the log-pseudolikelihood function with respect to β. Estimate β by solving for βb from
the weighted sample estimating equations
m
X
G(
b β) = wj S(β; yj , xj ) = 0 (3)
j=1
The associated estimation command with iweights will produce point estimates β
b equal to the
solution of (3).
A first-order matrix Taylor-series expansion yields
( )−1
b−β ≈ − ∂ G(
b β)
β G(
b β)
∂β
with the following variance estimator for β:

b
( )−1 ( )−T 
b ) =  ∂ G(β) ∂ G(
b b β)
Vb (β Vb {G(
b β)} = DVb {G(
b β)} D0

∂β ∂β b b
β=β β=β
where D is (Xs0 W Xs )−1 for linear regression (where W is a diagonal matrix of the sampling weights
and Xs is the matrix of sampled explanatory variables) or the inverse of the negative Hessian matrix
from the pseudolikelihood model. Write G( b β) as
m
X
G(
b β) = wj dj
j=1
where dj = sj xj and sj is a residual for linear regression or an equation-level score from the
pseudolikelihood model. The term equation-level score means the derivative of the log pseudolikelihood
with respect to xj β. In either case, G(
b βb ) is an estimator for the total G(β), and the variance estimator
V {G(β)}| b is computed using the design-based variance estimator for a total.
b b
β=β
The above result is easily extended to models with ancillary parameters, multiple regression
equations, or both.
The bootstrap
The bootstrap methods for survey data used in recent years are largely due to McCarthy and
Snowden (1985), Rao and Wu (1988), and Rao, Wu, and Yue (1992). For example, Yeo, Mantel,
and Liu (1999) cite Rao, Wu, and Yue (1992) with the method for variance estimation used in the
National Population Health Survey conducted by Statistics Canada.
In the survey bootstrap, the model is fit multiple times, once for each of a set of adjusted sampling
weights that mimic bootstrap resampling. The variance is estimated using the resulting replicated
point estimates.
Let bθ be the vector of point estimates computed using the sampling weights for a given survey
θ could be a vector of means, ratios, or regression coefficients). Each bootstrap
dataset (for example, b
replicate is produced by fitting the model with adjusted sampling weights. The adjusted sampling
weights are derived from the method used to resample the original survey data.
According to Yeo, Mantel, and Liu (1999), if nh is the number of observed PSUs in stratum h,
then nh − 1 PSUs are sampled with replacement from within stratum h. This sampling is performed
independently across the strata to produce one bootstrap sample of the survey data. Let r be the
number of bootstrap samples. Suppose that we are about to generate the adjusted-weight variable for
the ith bootstrap replication and whij is the sampling weight attached to the j th observation in the
ith PSU of stratum h. The adjusted weight is
∗ nh
whij = m∗ whij
nh − 1 hi
where m∗hi is the number of times the ith cluster in stratum h was resampled.
To accommodate privacy concerns, many public-use datasets contain replicate-weight variables
derived from the “mean bootstrap” described by Yung (1997). In the mean bootstrap, each adjusted
weight is derived from b bootstrap samples instead of one. The adjusted weight is
∗ nh
whij = m∗ whij
nh − 1 hi
where
b
1X ∗
m∗hi = mhik
b
k=1
is the average of the number of times the ith cluster in stratum h was resampled among the b bootstrap
samples.
Each replicate is produced using an adjusted-weight variable with the estimation command that
computed bθ. The adjusted-weight variables must be supplied to svyset with the bsrweight() option.
For the mean bootstrap, b must also be supplied to svyset with the bsn() option; otherwise, bsn(1)
is assumed. We call the variables supplied to the bsrweight() option bootstrap replicate-weight
variables when b = 1 and mean bootstrap replicate-weight variables when b > 1.
θ(i) be the vector of point estimates from the ith replication. When the mse option is specified,
Let b
the variance estimator is
r
b X b
V (θ) =
b b {θ(i) − b
θ}{bθ(i) − bθ}0
r i=1
Otherwise, the variance estimator is

r
b X b
Vb (b
θ) = θ(i) − θ(.) }0
{θ(i) − θ(.) }{b
r i=1
where θ(.) is the bootstrap mean,

r
1 Xb
θ(.) = θ(i)
r i=1
BRR
BRR was first introduced by McCarthy (1966, 1969a, and 1969b) as a method of variance estimation
variance estimates for this design than the linearized variance estimator, which can result in large
values and undesirably wide confidence intervals.
The model is fit multiple times, once for each of a balanced set of combinations where one PSU is
dropped (or downweighted) from each stratum. The variance is estimated using the resulting replicated
point estimates (replicates). Although the BRR method has since been generalized to include other
designs, Stata’s implementation of BRR requires two PSUs per stratum.
Let bθ be the vector of point estimates computed using the sampling weights for a given stratified
θ could be a vector of means, ratios, or regression coefficients). Each BRR
survey design (for example, b
replicate is produced by dropping (or downweighting) a PSU from every stratum. This could result
in as many as 2L replicates for a dataset with L strata; however, the BRR method uses Hadamard
matrices to identify a balanced subset of the combinations from which to produce the replicates.
A Hadamard matrix is a square matrix, Hr (with r rows and columns), such that Hr0 Hr = rI ,
where I is the identity matrix. The elements of Hr are +1 and −1; −1 causes the first PSU to be
downweighted and +1 causes the second PSU to be downweighted. Thus r must be greater than or
equal to the number of strata.
Suppose that we are about to generate the adjusted-weight variable for the ith replication and wj
is the sampling weight attached to the j th observation, which happens to be in the first PSU of stratum
h. The adjusted weight is
(
∗
f wj , if Hr [i, h] = −1
wj =
(2 − f )wj , if Hr [i, h] = +1
where f is Fay’s adjustment (Judkins 1990). By default, f = 0.
computed b θ. The adjusted-weight variables can be generated by Stata or supplied to svyset with the
brrweight() option. We call the variables supplied to the brrweight() option BRR replicate-weight
variables.
θ(i) be the vector of point estimates from the ith replication. When the mse option is specified,
Let b
r
1 X
Vb (b
θ) = {b
θ(i) − b
θ}{b θ}0
θ(i) − b
r(1 − f )2 i=1

r
1 X
Vb (b
θ) = θ(i) − θ(.) }{b
{b θ(i) − θ(.) }0
r(1 − f )2 i=1
where θ(.) is the BRR mean,

r
1 Xb
θ(.) = θ(i)
r i=1
The jackknife
The jackknife method for variance estimation is appropriate for many models and survey designs.
The model is fit multiple times, and each time one or more PSUs are dropped from the estimation
sample. The variance is estimated using the resulting replicates (replicated point estimates).
θ be the vector of point estimates computed using the sampling weights for a given survey
Let b
θ could be a vector of means, ratios, or regression coefficients). The dataset
design (for example, b
is resampled by dropping one or more PSUs from one stratum and adjusting the sampling weights
θ.
before recomputing a replicate for b
Let whij be the sampling weight for the j th individual from PSU i in stratum h. Suppose that
you are about to generate the adjusted weights for the replicate resulting from dropping k PSUs from
stratum h. The adjusted weight is

 0, if a = h and b is dropped

∗
 nh
wabj = wabj , if a = h and b is not dropped
 nh − k


wabj , otherwise
Each replicate is produced by using the adjusted-weight variable with the estimation command
that produced bθ. For the delete-one jackknife (where one PSU is dropped for each replicate), adjusted
weights can be generated by Stata or supplied to svyset with the jkrweight() option. For the delete-
k jackknife (where k > 1 PSUs are dropped for each replicate), the adjusted-weight variables must
be supplied to svyset using the jkrweight() option. The variables supplied to the jkrweight()
option are called jackknife replicate-weight variables.
The delete-one jackknife

θ(h,i) be the point estimates (replicate) from leaving out the ith PSU from stratum h. The
Let b
pseudovalue for replicate (h, i) is
∗
θ(h,i) + nh {b
θh,i = b
b θ−b
θ(h,i) }
When the mse option is specified, the variance estimator is

L
X nh
X
Vb (b
θ) = (1 − fh ) mh {b
θ(h,i) − b
θ}{b θ}0
θ(h,i) − b
h=1 i=1
and the jackknife mean is
L nh
1 XX
θ(.) = θ(h,i)
b
n i=1
h=1
where fh is the sampling rate and mh is the jackknife multiplier associated with stratum h. Otherwise,
L nh nh
X X 1 X
Vb (b
θ) = (1 − fh ) mh {b θ(h,i) − θh }0 ,
θ(h,i) − θh }{b θh = θ(h,i)
b
i=1
nh i=1
h=1
L nh
∗ 1 XX ∗
θ = θh,i
b
n i=1
h=1
The multiplier for the delete-one jackknife is
nh − 1
mh =
nh
The delete-k jackknife

θ(h,d) be one of the point estimates that resulted from leaving out k PSUs from stratum h. Let
Let e
ch be the number of such combinations that were used to generate a replicate for stratum h; then
d = 1, . . . , ch . If all combinations were used, then
nh !
ch =
(nh − k)!k!
The pseudovalue for replicate (h, d) is
∗
θ(h,d) + ch {b
θh,d = e
e θ−e
θ(h,d) }
When the mse option is specified, the variance estimator is

L
X ch
X
Vb (b
θ) = (1 − fh ) mh {e
θ(h,d) − b
θ}{e θ}0
θ(h,d) − b
h=1 d=1
L ch L
1 XX X
θ(.) = θ(h,d) ,
e C= ch
C
h=1 d=1 h=1

L ch ch
X X 1 X
Vb (b
θ) = (1 − fh ) mh θ(h,d) − θh }0 ,
θ(h,d) − θh }{e
{e θh = θ(h,d)
e
ch
h=1 d=1 d=1

L ch
∗ 1 XX ∗
θ = θh,d
e
C
h=1 d=1
The multiplier for the delete-k jackknife is

nh − k
mh =
ch k
Variables containing the values for the stratum identifier h, the sampling rate fh , and the jackknife
multiplier mh can be svyset using the respective suboptions of the jkrweight() option: stratum(),
fpc(), and multiplier().
Successive difference replication

Successive difference replication (SDR) was first introduced by Fay and Train (1995) as a method
of variance estimation for annual demographic supplements to the Current Population Survey. This
method is typically applied to systematic samples, where the observed sampling units are somehow
ordered.
In SDR, the model is fit multiple times, once for each of a set of adjusted sampling weights. The
variance is estimated using the resulting replicated point estimates.
Let bθ be the vector of point estimates computed using the sampling weights for a given survey
dataset (for example, bθ could be a vector of means, ratios, or regression coefficients). Each SDR
replicate is produced by fitting the model with adjusted sampling weights. The SDR method uses
Hadamard matrices to generate these adjustments.
A Hadamard matrix is a square matrix, Hr (with r rows and columns), such that Hr0 Hr = rI ,
where I is the identity matrix. Let hij be an element of Hr ; then hij = 1 or hij = −1. In SDR, if
n is the number of PSUs, then we must find Hr with r ≥ n + 2.
Without loss of generality, we will assume the ordered PSUs are individuals instead of clusters.
Suppose that we are about to generate the adjusted-weight variable for the ith replication and that wj
is the sampling weight attached to the j th observation. The adjusted weight is wj∗ = fji wj , where
fji is
1
fji = 1 + √ (hj+1,i − hj+2,i )
2 2
Here we assume that the elements of the first row of Hr are all 1.
computed bθ. The adjusted-weight variables must be supplied to svyset with the sdrweight()
option. We call the variables supplied to the sdrweight() option SDR replicate-weight variables.
θ(i) be the vector of point estimates from the ith replication, and let f be the sampling fraction
Let b
computed using the FPC information svyset in the fpc() suboption of the sdrweight() option,
where f = 0 when fpc() is not specified. When the mse option is specified, the variance estimator
is
r
4 X
θ) = (1 − f )
Vb (b {b
θ(i) − b
θ}{b θ}0
θ(i) − b
r i=1

r
4 X
θ) = (1 − f )
Vb (b {b θ(i) − θ(.) }0
θ(i) − θ(.) }{b
r i=1
where θ(.) is the SDR mean,

r
1 Xb
θ(.) = θ(i)
r i=1
In survey data analysis, the customary number of degrees of freedom attributed to a test statistic is
d = n − L, where n is the number of PSUs and L is the number of strata. Under regularity conditions,
an approximate 100(1 − α)% confidence interval for a parameter θ (for example, θ could be a total,
ratio, or regression coefficient) is
b 1/2
θb ± t1−α/2,d {Vb (θ)}
Cochran (1977, sec. 2.8) and Korn and Graubard (1990) give some theoretical justification for
using d = n − L to compute univariate confidence intervals and p-values. However, for some cases,
inferences based on the customary n − L degrees-of-freedom calculation may be excessively liberal;
the resulting confidence intervals may have coverage rates substantially less than the nominal 1 − α.
This problem generally is of the greatest practical concern when the population of interest has a
skewed or heavy-tailed distribution or is concentrated in a few PSUs. In some of these cases, the user
may want to consider constructing confidence intervals based on alternative degrees-of-freedom terms,
based on the Satterthwaite (1941, 1946) approximation and modifications thereof; see, for example,
Cochran (1977, sec. 5.4) and Eltinge and Jang (1996).
Sometimes there is no information on n or L for datasets that contain replicate-weight variables
but no PSU or strata variables. Each of svy’s replication commands has its own default behavior
when the design degrees of freedom are not svyset or specified using the dof() option. svy brr:
and svy jackknife: use d = r − 1, where r is the number of replications. svy bootstrap: and
svy sdr: use z1−α/2 for the critical value instead of t1−α/2,d .
References
Binder, D. A. 1983. On the variances of asymptotically normal estimators from complex surveys. International
Statistical Review 51: 279–292.
17–26.
Eltinge, J. L., and D. S. Jang. 1996. Stability measures for variance component estimators under a stratified multistage
design. Survey Methodology 22: 157–165.
Fuller, W. A. 1975. Regression analysis for sample survey. Sankhyā, Series C 37: 117–132.
Godambe, V. P., ed. 1991. Estimating Functions. Oxford: Oxford University Press.
Kish, L., and M. R. Frankel. 1974. Inference from complex samples. Journal of the Royal Statistical Society, Series
B 36: 1–37.
Kolenikov, S. 2010. Resampling variance estimation for complex survey data. Stata Journal 10: 165–199.
McCarthy, P. J., and C. B. Snowden. 1985. The bootstrap and finite population sampling. In Vital and Health Statistics,
1–23. Washington, DC: U.S. Government Printing Office.
Rao, J. N. K., and C. F. J. Wu. 1988. Resampling inference with complex survey data. Journal of the American
Rao, J. N. K., C. F. J. Wu, and K. Yue. 1992. Some recent work on resampling methods for complex surveys. Survey
Methodology 18: 209–217.
Särndal, C.-E., B. Swensson, and J. Wretman. 1992. Model Assisted Survey Sampling. New York: Springer.
Satterthwaite, F. E. 1941. Synthesis of variance. Psychometrika 6: 309–316.
. 1946. An approximate distribution of estimates of variance components. Biometrics Bulletin 2: 110–114.
Yeo, D., H. Mantel, and T.-P. Liu. 1999. Bootstrap variance estimation for the National Population Health Survey.
In Proceedings of the Survey Research Methods Section, 778–785. American Statistical Association.
Yung, W. 1997. Variance estimation for public use files under confidentiality constraints. In Proceedings of the Survey
Research Methods Section, 434–439. American Statistical Association.
Also see
Glossary
100% sample. See census.

balanced repeated replication. Balanced repeated replication (BRR) is a method of variance estimation
variance estimates for this design than does the linearized variance estimator, which can result in
large values and undesirably wide confidence intervals. The BRR variance estimator is described
in [SVY] variance estimation.
bootstrap. The bootstrap is a method of variance estimation. The bootstrap variance estimator for
survey data is described in [SVY] variance estimation.
BRR. See balanced repeated replication.
census. When a census of the population is conducted, every individual in the population participates
in the survey. Because of the time, cost, and other constraints, the data collected in a census are
typically limited to items that can be quickly and easily determined, usually through a questionnaire.
cluster. A cluster is a collection of individuals that are sampled as a group. Although the cost in time
and money can be greatly decreased, cluster sampling usually results in larger variance estimates
when compared with designs in which individuals are sampled independently.
DEFF and DEFT. DEFF and DEFT are design effects. Design effects compare the sample-to-sample
variability from a given survey dataset with a hypothetical SRS design with the same number of
individuals sampled from the population.
DEFF is the ratio of two variance estimates. The design-based variance is in the numerator; the
hypothetical SRS variance is in the denominator.
DEFT is the ratio of two standard-error estimates. The design-based standard error is in the
numerator; the hypothetical SRS with-replacement standard error is in the denominator. If the given
survey design is sampled with replacement, DEFT is the square root of DEFF.
delta method. See linearization.
design effects. See DEFF and DEFT.
direct standardization. Direct standardization is an estimation method that allows comparing rates
that come from different frequency distributions.
Estimated rates (means, proportions, and ratios) are adjusted according to the frequency distribution
from a standard population. The standard population is partitioned into categories called standard
strata. The stratum frequencies for the standard population are called standard weights. The
standardizing frequency distribution typically comes from census data, and the standard strata are
most commonly identified by demographic information such as age, sex, and ethnicity.
finite population correction. Finite population correction (FPC) is an adjustment applied to the variance
of a point estimator because of sampling without replacement, resulting in variance estimates that
are smaller than the variance estimates from comparable with-replacement sampling designs.
FPC. See finite population correction.
Hadamard matrix. A Hadamard matrix is a square matrix with r rows and columns that has the
property
Hr0 Hr = rIr
201
202 Glossary
where Ir is the identity matrix of order r. Generating a Hadamard matrix with order r = 2p
is easily accomplished. Start with a Hadamard matrix of order 2 (H2 ), and build your Hr by
repeatedly applying Kronecker products with H2 .
jackknife. The jackknife is a data-dependent way to estimate the variance of a statistic, such as a
mean, ratio, or regression coefficient. Unlike BRR, the jackknife can be applied to practically any
survey design. The jackknife variance estimator is described in [SVY] variance estimation.
linearization. Linearization is short for Taylor linearization. Also known as the delta method or
the Huber/White/robust sandwich variance estimator, linearization is a method for deriving an
approximation to the variance of a point estimator, such as a ratio or regression coefficient. The
linearized variance estimator is described in [SVY] variance estimation.
MEFF and MEFT. MEFF and MEFT are misspecification effects. Misspecification effects compare
the variance estimate from a given survey dataset with the variance from a misspecified model. In
Stata, the misspecified model is fit without weighting, clustering, or stratification.
MEFF is the ratio of two variance estimates. The design-based variance is in the numerator; the
misspecified variance is in the denominator.
MEFT is the ratio of two standard-error estimates. The design-based standard error is in the
numerator; the misspecified standard error is in the denominator. MEFT is the square root of MEFF.
misspecification effects. See MEFF and MEFT.
point estimate. A point estimate is another name for a statistic, such as a mean or regression
coefficient.
poststratification. Poststratification is a method for adjusting sampling weights, usually to account
for underrepresented groups in the population. This usually results in decreased bias because of
nonresponse and underrepresented groups in the population. Poststratification also tends to result
in smaller variance estimates.
The population is partitioned into categories, called poststrata. The sampling weights are adjusted
so that the sum of the weights within each poststratum is equal to the respective poststratum size.
The poststratum size is the number of individuals in the population that are in the poststratum.
The frequency distribution of the poststrata typically comes from census data, and the poststrata
are most commonly identified by demographic information such as age, sex, and ethnicity.
predictive margins. Predictive margins provide a way of exploring the response surface of a fitted
model in any response metric of interest—means, linear predictions, probabilities, marginal effects,
risk differences, and so on. Predictive margins are estimates of responses (or outcomes) for
the groups represented by the levels of a factor variable, controlling for the differing covariate
distributions across the groups. They are the survey-data and nonlinear response analogue to what
are often called estimated marginal means or least-squares means for linear models.
Because these margins are population-weighted averages over the estimation sample or subsamples,
and because they take account of the sampling distribution of the covariates, they can be used to
make inferences about treatment effects for the population.
primary sampling unit. Primary sampling unit (PSU) is a cluster that was sampled in the first sampling
stage; see cluster.
probability weight. Probability weight is another term for sampling weight.
pseudolikelihood. A pseudolikelihood is a weighted likelihood that is used for point estimation.
Pseudolikelihoods are not true likelihoods because they do not represent the distribution function
for the sample data from a survey. The sampling distribution is instead determined by the survey
design.
Glossary 203
PSU. See primary sampling unit.

replicate-weight variable. A replicate-weight variable contains sampling weight values that were
adjusted for resampling the data; see [SVY] variance estimation for more details.
resampling. Resampling refers to the process of sampling from the dataset. In the delete-one jackknife,
the dataset is resampled by dropping one PSU and producing a replicate of the point estimates. In
the BRR method, the dataset is resampled by dropping combinations of one PSU from each stratum.
The resulting replicates of the point estimates are used to estimate their variances and covariances.
sample. A sample is the collection of individuals in the population that were chosen as part of the
survey. Sample is also used to refer to the data, typically in the form of answered questions,
collected from the sampled individuals.
sampling stage. Complex survey data are typically collected using multiple stages of clustered
sampling. In the first stage, the PSUs are independently selected within each stratum. In the second
stage, smaller sampling units are selected within the PSUs. In later stages, smaller and smaller
sampling units are selected within the clusters from the previous stage.
sampling unit. A sampling unit is an individual or collection of individuals from the population that
can be selected in a specific stage of a given survey design. Examples of sampling units include
city blocks, high schools, hospitals, and houses.
sampling weight. Given a survey design, the sampling weight for an individual is the reciprocal of
the probability of being sampled. The probability for being sampled is derived from stratification
and clustering in the survey design. A sampling weight is typically considered to be the number
of individuals in the population represented by the sampled individual.
sampling with and without replacement. Sampling units may be chosen more than once in designs
that use sampling with replacement. Sampling units may be chosen at most once in designs that
use sampling without replacement. Variance estimates from with-replacement designs tend to be
larger than those from corresponding without-replacement designs.
SDR. See successive difference replication.
secondary sampling unit. Secondary sampling unit (SSU) is a cluster that was sampled from within a
PSU in the second sampling stage. SSU is also used as a generic term unit to indicate any sampling
unit that is not from the first sampling stage.
simple random sample. In a simple random sample (SRS), individuals are independently sampled—
each with the same probability of being chosen.
SRS. See simple random sample.
SSU. See secondary sampling unit.
standard strata. See direct standardization.
standard weights. See direct standardization.
stratification. The population is partitioned into well-defined groups of individuals, called strata.
In the first sampling stage, PSUs are independently sampled from within each stratum. In later
sampling stages, SSUs are independently sampled from within each stratum for that stage.
Survey designs that use stratification typically result in smaller variance estimates than do similar
designs that do not use stratification. Stratification is most effective in decreasing variability when
sampling units are more similar within the strata than between them.
subpopulation estimation. Subpopulation estimation focuses on computing point and variance esti-
mates for part of the population. The variance estimates measure the sample-to-sample variability,
assuming that the same survey design is used to select individuals for observation from the
204 Glossary
population. This approach results in a different variance than measuring the sample-to-sample vari-
ability by restricting the samples to individuals within the subpopulation; see [SVY] subpopulation
estimation.
successive difference replication. Successive difference replication (SDR) is a method of variance
typically applied to systematic samples, where the observed sampling units are somehow ordered.
The SDR variance estimator is described in [SVY] variance estimation.
survey data. Survey data consist of information about individuals that were sampled from a population
according to a survey design. Survey data distinguishes itself from other forms of data by the
complex nature under which individuals are selected from the population.
In survey data analysis, the sample is used to draw inferences about the population. Furthermore,
the variance estimates measure the sample-to-sample variability that results from the survey design
applied to the fixed population. This approach differs from standard statistical analysis, in which the
sample is used to draw inferences about a physical process and the variance measures the sample-
to-sample variability that results from independently collecting the same number of observations
from the same process.
survey design. A survey design describes how to sample individuals from the population. Survey
designs typically include stratification and cluster sampling at one or more stages.
Taylor linearization. See linearization.
variance estimation. Variance estimation refers to the collection of methods used to measure the
amount of sample-to-sample variation of point estimates; see [SVY] variance estimation.
Subject and author index
See the combined subject index and the combined author index in the Glossary and Index.
205

Stata Survey Data Reference Manual: Release 14

Uploaded by

Copyright:

Stata Survey Data Reference Manual: Release 14

Uploaded by

Document Information

Original Title

Copyright

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Stata Survey Data Reference Manual: Release 14

Uploaded by

Copyright:

STATA SURVEY DATA REFERENCE

A Stata Press Publication

The suggested citation for this software is

intro . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Introduction to survey data manual 1

bootstrap options . . . . . . . . . . . . . . . . . . . . . More options for bootstrap variance estimation 23

direct standardization . . . . . . . . . . . Direct standardization of means, proportions, and ratios 25

estat . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Postestimation statistics for survey data 29

jackknife options . . . . . . . . . . . . . . . . . . . . . . More options for jackknife variance estimation 51

ml for svy . . . . . . . . . . . . . . . . . . . . . Maximum pseudolikelihood estimation for survey data 52

poststratification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Poststratification for survey data 54

sdr options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . More options for SDR variance estimation 58

variance estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Variance estimation for survey data 186

Subject and author index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205

[GSM] Getting Started with Stata for Mac

[M] Mata Reference Manual

Description Also see

[SVY] survey Introduction to survey commands

Description Remarks and examples Acknowledgments References Also see

Survey design tools

Survey data analysis tools

Tools for programmers of new survey commands

Remarks and examples

Survey design tools

Example 1: Survey data from a one-stage design

Example 2: Multistage survey data

Example 3: Survey describe

50 100 0 4,071 0 14 40.7 81

Survey data analysis tools

Example 4: Estimating a population mean

weight 160.2863 .7412512 158.7974 161.7751

Example 5: Survey regression

height .7163115 .0293908 24.37 0.000 .6572784 .7753447

Example 6: Cox’s proportional hazards model

14407 total observations

9281 observations remaining, representing

. svy: stcox former_smoker smoker male urban1 rural

former_smoker 2.788113 .6205102 4.61 0.000 1.788705 4.345923

Example 7: Two-way tables for survey data

White 0.9680 0.0320 1.0000

Black 0.9410 0.0590 1.0000

Other 0.9797 0.0203 1.0000

Total 0.9658 0.0342 1.0000

Key: row proportion

Example 8: Comparing means

Example 9: Design effects

Note: Weights must represent population totals for deff to

Mean Coef. Std. Err. DEFF DEFT

(1) 29.27691 1.515201 2.42759 1.55768

Note: Weights must represent population totals for deff to

Example 10: BRR and replicate-weight variables

Stratum #Units #Obs min mean max

1 2 380 165 190.0 215

31 62 10,351 67 167.0 288

WtoH .4268116 .0008904 479.36 0.000 .4249957 .4286276

Survey data concepts

Example 11: Subpopulation estimation

sex byte %9.0g sex 1=male, 2=female

height .7632911 .0508432 15.01 0.000 .6611696 .8654127

Example 12: Standardized rates

age bgliving bgdeaths hsliving hsdeaths

0-5 10739 850 5738 463

Sum 74088 1764 66453 1418

Note: Zero standard errors because of 100% sampling rate