Health Facility Assessment Sampling -

Appropriate Methods for Small N

24 February 2016
Geneva, Switzerland
Presented to LFAs
Principles of Sampling

• A subset of elements selected randomly from a population of elements can

tell us what we need to know about that population with a quantifiable level
of confidence that the results are correct, within a quantifiable margin of
error.
• The size of the subset required is largely determined by the desired level of
confidence and the acceptable margin of error.
• The sample needs to reflect the population from which it is derived in all
respects important to the information we wish to gather so that the
population parameters we are estimating are unbiased.
• Weighting of estimates can often be used post hoc to correct for intended
sources of bias in the sample.

1
Steps for deriving a sample of health facilities common to
all sampling methods
Identify survey design
• experimental
• non-experimental
Establish sampling frame
• Geographic coverage
• Where does the program offer services?
• Seasonal variation?
• Facility types
• Community, Primary, Secondary, Tertiary
• Management authority (e.g. public/private)
• NGO, donor-funded, disease specific, etc.
• Exclusion criteria
• safety, accessibility, political, programmatic
• decreasing the sampling frame must be done in a purposeful manner to avoid the appearance
of bias.
2
Survey of Sampling Methods Appropriate for Small N

• Purposive Sample
• Simple Random Sampling
• Cluster Sampling
• Lot Quality Assurance Sampling

3
Purposive Sampling

• In a purposive sample, health facilities are selected intentionally, i.e. for a

distinct purpose, such as investigating reporting anomalies or ensuring a
program is being implemented as designed.
• Purposive sampling (often called convenience sampling) is used when
resources are scarce and only a handful of health facilities can be selected.
• Purposive samples are also used when reported data indicate a problem of
program or data quality that requires immediate attention.
• While purposive samples are easy and inexpensive to conduct (no
complicated sampling methods) the findings are only applicable to the sites
selected, i.e. you can’t infer the findings to the larger population of health
facilities.

4
Purposive Sampling

Advantages:
• Simplicity of sampling and ease of research
• Helpful for pilot studies and for hypothesis generation
• Data collection can be facilitated in short duration of time
• Cost effectiveness

Disadvantages:
• Highly vulnerable to selection bias
• Generalizability unclear
• High level of sampling error

5
Simple Random Sampling (SRS) (1)

• SRS is the simplest form of probability sampling where elements are

selected from a larger population in which every element has the same
probability of being selected.
• Sample size requirements are determined by:
• the approximate prevalence of the parameter of interest in the population,
• the level of confidence desired in the estimate, and
• the minimum acceptable margin of error of the estimate.
• SRS has the disadvantage of often requiring large distances to cover to
reach sampled health facilities, which can be anywhere in the country.

6
Simple Random Sampling (2)

• Systemization can be introduced in the sampling of health facilities which

makes it easier to derive the sample.
• A master list of facilities can be used to select facilities systematically.
• Advantages of SRS
• Smaller samples compared to cluster sampling
• Methods are uncomplicated

• Disadvantages of SRS
• Often leads to high assessment implementation costs due to time and travel
required
• Requires a complete list of sample elements

7
Simple Random Sampling (3)

Stratified random sampling – a variation on SRS

• Sample frame is divided into strata based on meaningful criteria – e.g. facility
type, management authority, etc.
• A random sample is drawn from each strata
• Permits comparisons between strata
• Can increase the size of the overall sample since each stratum requires a
sample of elements large enough to infer results to the larger population

8
Simple Random Sampling (4)

Systematic sampling – a variation on SRS

• Every Nth unit on a list is selected based on dividing the size of the
population by the sample size.
• Convenient if facility list is available
• As good as random sampling if starting point is obtained randomly.
• Attention must be paid to avoiding recurring patterns within the sampling
frame.

9
Simple Random Sampling (5)

Relative variance:
• When the proportion of the parameter of interest in the population is small,
the margin of error is often too large for meaningful interpretation.
• Use relative variance (relative error vs. standard [or sampling] error) for
margin of error.
• Relative error also known as the coefficient of variation (cv).
• Statistically, the cv = SE(
• Results in larger samples, particularly for small values of p.

10
Simple Random Sampling (3)
The formula for calculating a sample size for SRS is:

�� ∗ � 1 − �
�=
��

n = required sample size

t = the value of the confidence level you have chosen (at 80% t = 1.28, 90% t = 1.64, 95 % t =
1.96)
p = estimated prevalence of the indicator. (The product of p and [1-p] is maximized when p =
0.5. Therefore, when prevalence is unknown, 0.5 should be used.)
m = margin of error

11
Simple Random Sampling (6)
The formula for calculating a sample size for SRS with relative error is:

�� ∗ � ∗ �
�=
�� ∗ �

n = required sample size

t = normal deviate needed to provide an estimate at the required confidence limit (e.g. 1.96 for 95%)
the 95 percent level of confidence. (at 80%, t = 1.28; 90%, t = 1.64, 95% t = 1.96)
p = estimated prevalence of the indicator.
q = 1 – p (the sample size is maximized when p and q (i.e.1-p) = 0.5. Therefore, when prevalence is
unknown, 0.5 should be used.)
f = design effect (deff)
V = relative variance (i.e. margin of error)
12
Simple Random Sampling (7)

Where there is a predetermined population (e.g., total number of facilities in the country),
the sample size needs to be multiplied by the Finite Population Correction (FPC) factor.
The formula can be expressed as:
n
New n =
n−1
1+
N
Where:
New n = the adjusted new sample size
N = the population size
n = the sample size obtained from the general formula
(the FPC should be applied if n/N ≥ 5%)
13
 Example: Stratified Random Sampling
Sample size formula for HF � � � �
]]*d
Sampling (with FPC & deff) :
In Excel: =(((D7*G7*H7)+F7)/(F7+D7*G7*H7/C7))*I7

square of
Ekiti State sample over
Strata Total HF normal ME ME^2 p q deff
Nigeria size sample
deviate
Ekiti State, Nigeria: By facility type and

Primary 294 3.8416 0.15 0.0225 0.5 0.5 1 38 38

Public
Primary 101 3.8416 0.15 0.0225 0.5 0.5 1 31 31
Private
managing authority

Secondary 18 3.8416 0.15 0.0225 0.5 0.5 1 13 18

Public
Secondary 44 3.8416 0.15 0.0225 0.5 0.5 1 22 30
Private
Tertiary 2 2 2
Public
Tertiary 0 0 0
Private
Total 459 106 119

14
Sample sizes for Conf = 95% Prevalence
Margin of error 50% 60% 70% 80% 90%
SRS 5% 384 369 323 246 138
– differing levels 10% 96 92 81 61 35
15% 43 41 36 27 15
of prevalence, 20% 24 23 20 15 9
confidence, and
margin of error Conf = 90% Prevalence
Margin of error 50% 60% 70% 80% 90%
5% 269 258 226 172 97
10% 67 65 56 43 24
15% 30 29 25 19 11
20% 17 16 14 11 6

Conf = 80% Prevalence

Margin of error 50% 60% 70% 80% 90%
5% 164 157 138 105 59
10% 41 39 34 26 15
15% 18 17 15 12 7
20% 10 10 9 7 4

15
Sample sizes for Conf = 95% Prevalence
Margin of error 50% 60% 70% 80% 90%
SRS with relative 5% 1537 1024 659 384 171
error 10% 384 256 165 96 43
15% 171 114 73 43 19
– differing levels 20% 96 64 41 24 11
of prevalence,
confidence, and Conf = 90% Prevalence
Margin of error 50% 60% 70% 80% 90%
margin of error 5% 1076 717 461 269 120
10% 269 179 115 67 30
15% 120 80 51 30 13
20% 67 45 29 17 7

Conf = 80% Prevalence

Margin of error 50% 60% 70% 80% 90%
5% 655 437 281 164 73
10% 164 109 70 41 18
15% 73 49 31 18 8
20% 41 27 18 10 5

16
Simple Random Sampling (8)
• Each estimate requires a different sample size – take the largest
• Adjust for non-response
• Adjust for equal reliability of domains (strata, e.g. urban/rural)
• If sampling rate in domain is ≥ 50% take a census
• If stratifying on size and don’t know the size, you can use a proxy measure:
• Client volume (e.g. indicator value), number of staff, number of beds
• Exact numbers can be rounded without loss of precision
• Large facilities should be sampled purposively to ensure inclusion
• Confidence level is more important than ME
• For our purposes, SRS probably only useful for larger values of p, and wide margin of
error

17
Cluster Sampling (1)

• Cluster sampling is a variation on SRS that allows for a smaller number of

sites to be sampled than with SRS
• Implemented in two stages:
• the population of interest is divided into geographically distinct groups or areas, and
a random sample of areas is obtained;
• facilities are sampled from within each area.
• The primary sampling areas are the clusters (e.g. provinces, districts,
enumeration areas or local government areas).
• Cluster sampling is primarily useful in reducing the total distance traveled to
assess facilities.
18
Cluster Sampling (2)

• Suitable for large countries or where travel is time-consuming and/or expensive

• If the number of facilities within clusters is known, PPS sampling can be used
(population proportionate to size).
• Clusters with more sites have a higher probability of selection, etc.

• Often results in increase in sample size relative to SRS due to fewer areas → ↓
representativeness → ↑ variance
• Design effect - the precision of the cluster sample compared to a simple random
sample of the same size
• Recommended to use design effect 1.2 for health facility sampling (with caveats)

19
Cluster Sampling (3)

Advantages of cluster sampling:

• Cost-effective due to less area to cover,
• Requires updated list of health facilities, but only for sampled clusters.
• Point estimate of accuracy of reporting can be obtained

Disadvantages of cluster sampling:

• Not intended for calculation of estimates from individual clusters,
• Less precise than a simple random sample,
• Increased sample size relative to SRS due to the design effect.

20
Cluster
Sampling (4)

Example

21
Cluster Conf = 95% Prevalence
Margin of error 50% 60% 70% 80% 90%
Sampling 5% 461 443 387 295 166
– Sample sizes 10% 115 111 97 74 41
15% 51 49 43 33 18
for differing 20% 29 28 24 18 10
levels of
prevalence, Conf = 90% Prevalence
Margin of error 50% 60% 70% 80% 90%
confidence, and 5% 323 310 271 207 116
margin of error 10% 81 77 68 52 29
15% 36 34 30 23 13
20% 20 19 17 13 7

Conf = 80% Prevalence

Margin of error 50% 60% 70% 80% 90%
5% 197 189 165 126 71
10% 49 47 41 31 18
15% 22 21 18 14 8
20% 12 12 10 8 4

22
Cluster Sampling

How many clusters? How many sites per cluster?

• Depends on calculated sample size
• Sample size is calculated like SRS + the adjustment for the design effect
• Can use the average HF / cluster, if known: if 100 HF in sample, and there
are on average 10 HF per district, take 10 districts (clusters)
• Precision increases with more clusters, but so does cost
• Often a trade-off between cost and precision
• Distribute HFs evenly among clusters if possible (unequal cluster sizes
reduces precision)

23
Cluster Sampling: Data Quality Example - Modified 2-stage
cluster sample

• Sampling method used for IDQA, DQA etc.

• Intended for use with the ratio statistic “Verification Factor”
• Extensively used to assess data quality, and evaluated
• Mathematical modeling conducted to determine precision levels for different
numbers of clusters and sites per cluster
• Modification is stratification within cluster on service volume (i.e. the indicator
value HFs for a representative period)
• HF are assigned to size strata, equal numbers of HF are randomly sampled
from within each strata
• Clusters are sampled systematically using PPS on the cluster value of
indicator of interest
24
Cluster Sampling: Data Quality Example - Modified 2-stage
cluster sample

Precision for varying number of clusters and sites per cluster - modified 2-stage cluster sample
Number of Districts (Clusters)
No. of HFs per
4 8 10 15 20 30
Stratum
1 0.472 0.249 0.212 0.164 0.138 0.110

2 0.446 0.238 0.202 0.156 0.131 0.104

6 0.469 0.245 0.209 0.159 0.134 0.106

25
Cluster Sampling: Data Quality Example - Modified 2-stage cluster sample
Procedure for Sampling Clusters
• Order districts and results alphabetically (or according to another method that will not introduce
periodicity or other forms of bias into the list of districts, e.g. serpentine method)
• Exclude districts for which it is not possible/advisable to conduct the assessment (e.g. due to
security concerns).
• Calculate the sampling interval - sum of indicator value over all clusters divided by the number of
clusters to sample.
• Select a starting point at random – number between 1 and sampling interval. This is your first
cluster.
• Calculate the running sum of cluster volume over all clusters.
• Add sampling interval to random starting point – find in running sum column the number for 2nd
cluster. If 1st cluster + sampling interval is below running sum value for the cluster then select.
• Proceed down the ordered list of clusters and select up to the desired number of clusters.
• Clusters with larger volume may be selected more than once. If a cluster is selected more than
once you will sample more health facilities per cluster in these clusters. (Number of times district is
selected as cluster x number of desired health facilities per cluster.)
• In the field, if a facility is deemed inaccessible, the next nearest non-sampled facility can used.
26
Cluster Sampling: Data Quality Example - Modified 2-stage
cluster sample
Example

Sampling Clusters with PPS

27
Cluster Sampling: Data Quality Example - Modified 2-stage cluster sample

Procedure for sampling health facilities within clusters

• From each selected district, stratify health facilities into large and small by
sorting largest to smallest on the indicator total. For an assessment of
quarterly reporting, aggregate monthly values up to the quarterly value, if
applicable.
• Exclude health facilities that are inaccessible for any reason.
• Select desired number of health facilities randomly from within each stratum.
• Select at least two back up facilities in case the originally selected sites
cannot be reached (e.g. take the next one down the list in terms of volume of
service).
• Try to ensure equal numbers of health facilities within clusters and strata as
this will improve the precision of the estimate of accuracy of reporting.

28
Cluster Sampling: Data Quality Example - Modified 2-stage cluster sample

District Health Facility Jan Feb Mar Total

Littoral HZ Suru-Léré 35 36 40 111

Littoral CNHU 36 42 23 101

Large Littoral CS Bethesda 14 18 22 54

Sites
Littoral ONG Arc en Ciel 14 22 9 45

CS Cotonou 1
Littoral 9 12 6 27
(DIST)

Littoral CNHPP Lazaret 12 7 8 27

Hôpital Camp
Littoral 4 19 3 26
Guézo

Small ONG Racines

Littoral 3 5 8 16
Sites (Cotonou)

Hôpital St Luc
Littoral 0 0 7 7
(Cotonou)

Littoral HOMEL 0

29
Cluster Sampling: Data Quality Example - Modified 2-stage cluster sample

Analysis:

Σ Recounted (HF in cluster) x Value for cluster at national level

Σ Reported (HF in cluster) Value for cluster at district level

• Calculate the weighted average of adjusted cluster-specific verification factors by

adding the products of the adjusted cluster-specific verification factors and the
cluster-specific volume of service (verified result for each cluster), and dividing by
the volume of service (verified results) over all clusters.

• Calculate a confidence interval for the point estimate:

Point estimate ± Standard Error (SE) * Z-score

30
Lot Quality Assurance Sample (LQAS)
• Often when small samples are necessary, classification is used to assess
elements of the population to make inferences to the larger population.
• LQAS is a method of classification which can also be seen as a stratified random
sampling design.
• LQAS consists of two stages:
• the first obtains random samples from, say, districts within a region in order to classify
each as belonging to one of two classes that is often labeled either “acceptable” or
“unacceptable”.
• The second combines information from such districts to obtain a measure of the region
as a whole (either as a stratified sample or a cluster sample)
• Regardless of the LQAS design selected for use, its primary purpose is the
classifications carried out at the local level.
• The second stage employs traditional sampling theory.

31
Lot Quality Assurance Sample (LQAS)
• A relatively rapid and inexpensive data collection approach that allows
program implementers to use small sample sizes and more frequent
sampling to categorize and prioritize areas by their performance on key
indicators.
• The lower expense is due to LQAS’s primary purpose of classification rather
than point estimate
• Since the 1980s it has been used to assess immunization coverage, post-
disaster assessment of health status, family planning and antenatal care,
growth and nutrition monitoring, diarrheal disease control, and quality
management, in urban zones, rural areas, or on a national scale in over 55
countries

32
Lot Quality Assurance Sample (LQAS)

• In LQAS, a program area is divided into supervision areas or “lots”, which

may consist of villages, urban zones, or health facility catchment areas.
• Lots are defined based on programmatic or administrative boundaries,
forming programmatically relevant clusters of sample elements.
• In the traditional LQAS method, simple random sampling is used to select a
small sample from each lot (often 19 for assessment of service provision
indicators)
• Sample sizes in these lots are too small to allow for precise estimation of
indicators, but lots can be classified as meeting or not meeting a
predetermined target level of performance and performance of different lots
can be compared.
33
Lot Quality Assurance Sample (LQAS)

• Threshold levels should be defined to correspond with performance levels to

be found in the best and worst performing lots
• The strength of LQAS is its focus on identifying the worst of the worst with
low levels of sampling error
• Program managers can make meaningful programmatic decisions (e.g.
where to target resources for improvement) using very small samples
• Results can be used to:
• prioritize programming in small, distinct geographic areas that may be
underperforming or
• to highlight areas that have been successful in achieving a target performance.

34
Lot Quality Assurance Sample (LQAS)

• Advantages of LQAS:
• Can use smaller samples since response for each lot is binary (yes/no)
• Simplified data analysis
• Good for routine monitoring of program and data quality
• Wide range of applications

• Disadvantages of LQAS:
• Requires updated list of health facilities
• Facilities need to be selected randomly within lots
• Initial conceptual leap
• Smaller, but growing body of research

35
Lot Quality Assurance Sample (LQAS)

Establish the sampling parameters:

• What is the health service being assessed?
• What constitutes a “lot” and the overall service area?
• What is the coverage target the program should reach (this is the standard
the sample units [i.e. HFs] should reach to be classified has having
acceptable performance)?
• What is the coverage which is unacceptably low and an indication of serious
problem of quality?
• What is the acceptable classification error (or provider and consumer risks)?

36
Lot Quality Assurance Sample (LQAS)
• Define lots: a ‘lot’ can be defined as a population unit assigned to health unit, a health
center, or even health records within a health center. An ideal lot is the smallest unit that
could provide meaningful information to a health planner when evaluating a health program
• PU = Bench mark for quality established equal to or above which service quality is deemed
acceptable
• PL = Bench mark for quality below which service quality is deemed very unacceptable
• n = sample size
• α error = the probability of misclassifying an area with unacceptable performance as
acceptable (1- α = specificity), also known as ‘consumer risk’
• β error = the probability of misclassifying an area with acceptable performance as
unacceptable (1 - β = sensitivity), also known as ‘provider risk’
• d = the decision rule – the number of sampled elements that must be deemed acceptable in
order for the entire ‘lot’ to be deemed acceptable. If the number of acceptable sampled
elements is not reached the ‘lot’ must be rejected as unacceptable. (The decision rule
depends on all the other sampling parameters).

37
LQAS
Provider vs. Consumer Risk
• The provider error measures the risk, or probability, that an acceptable lot will be
classified as unacceptable
• The consumer error measures risk that an unacceptable lot will be classified as
acceptable.
• In the health care setting, the provider is the organization providing the health
care interventions (most often the government), while the consumer is the
intended beneficiaries of the health care intervention (e.g. the general
population, or pregnant women, or children less than five years of age, etc.)
• Provider and consumer error are also known as sensitivity and specificity,
respectively. A low provider error typically results in a higher consumer error, and
vice versa. However, they should be kept as equal as possible.
• In general, a sample size of 19 permits both provider and consumer errors of ≤
10% which is typically sufficient confidence for both providers and consumers for
routine monitoring

38
LQAS – Sample size calculation
Sample sizes in LQAS are derived using the following binomial formula which calculates the
probability of finding a certain number of failures from a population for which, say, 80% are successes.

�!
�� = �� � � ���
�! (� − �)!
Where;
Pa = the probability of selecting “a” successes (e.g. rapid diagnostic tests) in a sample of “n” elements
p = the benchmark for quality (80% in this example)
q = the expected proportion of failures (q = 1-p)
n = the sample size
a = the exact number of ‘successes’ in the sample (i.e. the acceptable performance)
n-a = the number of ‘failures’ in the sample (i.e. the unacceptable performance; in LQAS this
expression is referred to as “d”, the “decision rule”)

39
LQAS – Table of cumulative probabilities, n=19
d 0.05 0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 0.90 0.95
0 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.001 0.014 0.135 0.377
1 0.000 0.000 0.000 0.000 0.000 0.000 0.001 0.010 0.083 0.420 0.755
2 0.000 0.000 0.000 0.000 0.000 0.000 0.005 0.046 0.237 0.705 0.933
3 0.000 0.000 0.000 0.000 0.000 0.002 0.023 0.133 0.455 0.885 0.987
4 0.000 0.000 0.000 0.000 0.001 0.010 0.070 0.282 0.673 0.965 0.998
5 0.000 0.000 0.000 0.000 0.003 0.032 0.163 0.474 0.837 0.991 1.000
6 0.000 0.000 0.000 0.001 0.012 0.084 0.308 0.666 0.932 0.998 1.000
7 0.000 0.000 0.000 0.003 0.035 0.180 0.488 0.818 0.977 1.000 1.000
8 0.000 0.000 0.000 0.011 0.088 0.324 0.667 0.916 0.993 1.000 1.000
9 0.000 0.000 0.002 0.033 0.186 0.500 0.814 0.967 0.998 1.000 1.000
10 0.000 0.000 0.007 0.084 0.333 0.676 0.912 0.989 1.000 1.000 1.000
11 0.000 0.000 0.023 0.182 0.512 0.820 0.965 0.997 1.000 1.000 1.000
12 0.000 0.002 0.068 0.334 0.692 0.916 0.988 0.999 1.000 1.000 1.000
13 0.000 0.009 0.163 0.526 0.837 0.968 0.997 1.000 1.000 1.000 1.000
14 0.002 0.035 0.327 0.718 0.930 0.990 0.999 1.000 1.000 1.000 1.000
15 0.013 0.115 0.545 0.867 0.977 0.998 1.000 1.000 1.000 1.000 1.000
16 0.067 0.295 0.763 0.954 0.995 1.000 1.000 1.000 1.000 1.000 1.000
17 0.245 0.580 0.917 0.990 0.999 1.000 1.000 1.000 1.000 1.000 1.000
18 0.623 0.865 0.986 0.999 1.000 1.000 1.000 1.000 1.000 1.000 1.000
19 1.000 1.000 1.000 1.000 1.000 1.000 1.000 1.000 1.000 1.000 1.000

40
LQAS – Operation Characteristic Curve (OC Curve) for
n=19, d = 13

41
 LQAS Table: Decision Rules for Sample Sizes of 12-30 and Coverage Targets / Average of 10% - 95%

LQAS Sample Average Coverage (Baselines) / Annual Coverage Target (Monitoring and Evaluation)
Size*
Table 12
10%
N/A
15%
N/A
20%
1
25%
1
30%
2
35%
2
40%
3
45%
4
50%
5
55%
5
60%
6
65%
7
70%
7
75%
8
80%
8
85%
9
90%
10
95%
11
13 N/A N/A 1 1 2 3 3 4 5 6 6 7 8 8 9 10 11 11
14 N/A N/A 1 1 2 3 4 4 5 6 7 8 8 9 10 11 11 12
15 N/A N/A 1 2 2 3 4 5 6 6 7 8 9 10 10 11 12 13
16 N/A N/A 1 2 2 3 4 5 6 7 8 9 9 10 11 12 13 14
17 N/A N/A 1 2 2 3 4 5 6 7 8 9 10 11 12 13 14 15
18 N/A N/A 1 2 2 3 5 6 7 8 9 10 11 11 12 13 14 16
19 N/A N/A 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
20 N/A N/A 1 2 3 4 5 6 7 8 9 11 12 13 14 15 16 17
21 N/A N/A 1 2 3 4 5 6 8 9 10 11 12 13 14 16 17 18
22 N/A N/A 1 2 3 4 5 7 8 9 10 12 13 14 15 16 18 19
23 N/A N/A 1 2 3 4 6 7 8 10 11 12 13 14 16 17 18 20
24 N/A N/A 1 2 3 4 6 7 9 10 11 13 14 15 16 18 19 21
25 N/A 1 2 2 4 5 6 8 9 10 12 13 14 16 17 18 20 21
26 N/A 1 2 3 4 5 6 8 9 11 12 14 15 16 18 19 21 22
27 N/A 1 2 3 4 5 7 8 10 11 13 14 15 17 18 20 21 23
28 N/A 1 2 3 4 5 7 8 10 12 13 15 16 18 19 21 22 24
29 N/A 1 2 3 4 5 7 9 10 12 13 15 17 18 20 21 23 25
30 N/A 1 2 3 4 5 7 9 11 12 14 16 17 19 20 22 24 26
N/A: Not applicable, meaning LQAS cannot be used in this assessment because the coverage is either too low or too high to assess an SA
Alpha or beta errors are ≥ 10%
42
Alpha or beta errors are > 15%
LQAS = Online sample calculator

43
 No. Health Facility Name Recounted Reported VF Acceptable (Y/N)

Example – 1 CS KAMOA 32 32 1.00 Y

LQAS and 2

3
CS KANZEZNZE

CS SAMBA
97

79
106

53
0.92

1.49
Y

data 4 CS LUKAMVUE 0 28 0.00 N

quality 5

6
CS LUMU

CS BURHIBA
265

127
271

127
0.98

1.00
Y

7 CS KASHEKE 89 98 0.91 Y

8 CS KAZIBA CENTRE 114 165 0.69 N

9 CS TSHIKAJI 40 37 1.08 Y

10 CS KALEMA MULUMBA 152 152 1.00 Y

11 CS LUBONDAIE 21 70 0.30 N

12 CS KATEBA 84 84 1.00 Y

13 CS TSHILOMBA 28 49 0.57 N

14 CS KALENDA GARE 27 100 0.27 N

15 CS TSHUMBE 1 37 43 0.86 Y

16 CS TSHUMBE 2 80 82 0.98 Y

17 CS CINDUNDU 25 41 0.61 N

18 CS KATABAYI 45 43 1.05 Y

19 CS BILOMBA 68 72 0.94 Y
44
Total Acceptable 12
 LQAS - Aggregating lots to derive a point estimate
• Lots can be aggregated to provide a point Sample sizes for LQAS with varying levels
estimate provided there are sufficient of precision (95% confidence)
sample elements to ensure adequate Precision Sample Size
precision (a total of 95 elements yields
±10% precision, at 95% confidence) 0.13 57
• Judgements about the health areas that
make up the total sample can also be 0.11 76
made (acceptable/ unacceptable).
• “Rolling sample”: Assessments of lots can 0.10 95
be staggered to yield the point estimate
over time (assess one lot per quarter, at 0.09 114
the end of the year derive point estimate)
0.08 133

45
 LQAS - Aggregating lots to derive a point estimate
Calculate the rate of Example–Weighted estimate of parameter of interest
retention on ART through a
Service wt. * n-
record review at a sample Area
n d n-d/n N wt.
d/n
of health facilities:
1 19 5 0.74 176 176/660 0.20

• n = LQAS sample size 2 19 4 0.79 150 150/660 0.18

• d = number of HIV
3 19 10 0.47 100 100/660 0.07
patients who were not
retained on ART after a 4 19 2 0.89 111 111/660 0.15
specified time period
5 19 9 0.53 123 123/660 0.10
• N = number of patients
alive and on ART at the Total 95 30 660 0.70
health facility
46
 LQAS – Targeted Spot Check of Programmatic Quality
Example–results of an assessment of safe management of insecticide for IRS

• It is believed that 90% of health Facility

Appropriate
storage
Appropriate
inventory control
Availability
of PPE
facilities manage insecticides safely.
1 Y Y Y
• System reform is necessary if only 50%
2 Y N N
of health facilities correctly manage
3 Y Y Y
insecticide.
4 Y N Y
• Competence is ‘bi-modal’ – i.e. they
know it and do it, or they don’t know it 5 Y N Y

and don’t do it. 6 N Y N

• sample size is 9 health facilities and the 7 Y Y Y

decision rule is 6 (actual α error = 8 Y N Y

0.045, actual β error = 0.080) 9 Y Y Y

• Wider performance range yields Number of failures 1 4 2

smaller samples Component

Yes No Yes
adequate

47
Sampling Health Facility Staff

48
Sampling Health Facility Clients

49