See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/24096700

Confidence Intervals for Rank Statistics: Somers’ D and Extensions

Article  in  Stata Journal · September 2006

DOI: 10.1177/1536867X0600600302 · Source: RePEc

CITATIONS READS

150 1,624

1 author:

Roger Newson
Imperial College London
191 PUBLICATIONS   6,158 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

GA2LEN Survey View project

All content following this page was uploaded by Roger Newson on 02 May 2018.

The user has requested enhancement of the downloaded file.

 snp15.y Confidence intervals for rank order statistics and their differences
Author: Roger Newson, Imperial College London, UK. Email: [email protected] Date: 26 February 2006.
Abstract: Rank order or so–called “non–parametric” methods are in fact based on population parameters, which
are zero under the null hypothesis. Two of these parameters are Kendall’s τa and Somers’ D, the parameter
tested by a Wilcoxon rank–sum test. Confidence limits for these parameters are more informative than P –values
alone, for three reasons. First, confidence intervals show that a high P –value does not prove a null hypothesis.
Second, for continuous data, Kendall’s τa can often be used to define robust confidence limits for Pearson’s
correlation by Greiner’s relation. Third, we can define confidence limits for differences between two Kendall’s
τa s or Somers’ Ds, and these are informative, because a larger Kendall’s τa or Somers’ D cannot be secondary
to a smaller one. The program somersd calculates confidence intervals for Somers’ D or Kendall’s τa , using
jackknife variances. There is a choice of transformations, including Fisher’s z, Daniels’ arcsine, Greiner’s ρ,
the z–transform of Greiner’s ρ, and Harrell’s c. A cluster() option is available for clustered samples. The
funtype() option allows the user to estimate between–cluster, within–cluster or von Mises versions of Somers’ D
or Kendall’s τa . The wstrata() option allows the user to estimate stratified versions of Somers’ D or Kendall’s
τa , which can be used to measure associations between a predictor and an outcome within strata specified by
the values of a confounder. The cenind() option allows the user to specify left or right censorship indicators
for the X and/or Y variables. The estimation results are saved as for a model fit, so that differences can be
estimated using lincom.
Keywords: Somers’ D; Kendall’s tau; Harrell’s c; ROC area; Gini index; rank correlation; rank–sum test; Wilcoxon
test; sign test; confidence intervals; non–parametric methods.

1 Syntax
£ ¤£ ¤£ ¤£ ¤£
somersd varlist weight if exp in range , taua tdist transf(transformation name)
cenind(cenind list) cluster(varname) cfweight(expression) funtype(functional type)
¤
wstrata(varlist) bstrata(varlist | n) notree level(#) cimatrix(new matrix)
where transformation name is one of
iden | z | asin | rho | zrho | c
and functional type is one of
wcluster | bcluster | vonmises
and cenind list is a list of variable names and/or zeros.
fweights, iweights and pweights are allowed; see help for weight. They are treated as described in Interpretation
of weights and Methods and formulas below.
bootstrap, by, jackknife, statsby, and svy jackknife are allowed; see help for prefix.

1.1 Description
somersd calculates the rank order statistics Somers’ D (corresponding to rank–sum tests) and Kendall’s τa ,
with confidence limits. Somers’ D or Kendall’s τa is calculated for the first variable of varlist as a predictor of each
of the other variables in varlist, with estimates and jackknife variances and confidence intervals output and saved in
e() as if for the parameters of a model fit. It is possible to use lincom to output confidence limits for differences
between the population Somers’ D or Kendall’s τa values.

1.2 Options
taua causes somersd to calculate Kendall’s τa . If taua is absent, then somersd calculates Somers’ D.
tdist specifies that the estimates are assumed to have a t–distribution with N − 1 degrees of freedom, where N is
the number of clusters if cluster() is specified, or the number of observations if cluster() is not specified.
transf(transformation name) specifies that the estimates are to be transformed, defining estimates for the trans-
formed population value. iden (identity or untransformed) is the default. z specifies Fisher’s z (the hyperbolic
arctangent), asin specifies Daniels’ arcsine, rho specifies Greiner’s ρ (Pearson correlation estimated using
Greiner’s relation), zrho specifies the z–transform of Greiner’s ρ, and c specifies Harrell’s c. If the first variable
of varlist is a binary indicator of a disease and the other variables are quantitative predictors for that disease,
then Harrell’s c is the area under the receiver operating characteristic (ROC) curve. somersd recognises the
transformation names arctanh and atanh as synonyms for z, arcsin and arsin as synonyms for asin, sinph
2 Post-publication update RBN-1

as a synonym for rho, zsinph as a synonym for zrho, and roc and auroc as synonyms for c. It also recognizess
unambiguous abbreviations for transformation names, such as id for iden, or aur for auroc.
cenind(cenind list) specifies a list of left– or right–censorship indicators, corresponding to the variables mentioned in
the varlist. Each censorship indicator is either a variable name or a zero. If the censorship indicator corresponding
to a variable is the name of a second variable, then this second variable is used to indicate the censorship status
of the first variable, which is assumed to be left–censored (at or below its stated value) in observations in which
the second variable is negative, right–censored (at or above its stated value) in observations in which the second
variable is positive, and uncensored (equal to its stated value) in observations in which the second variable
is zero. If the censorship indicator corresponding to a variable is a zero, then the variable is assumed to be
uncensored. If cenind() is unspecified, then all variables in the varlist are assumed to be uncensored. If the list
of censorship indicators specified by cenind() is shorter than the list of variables specified in the varlist, then
the list of censorship indicators is completed with the required number of zeros on the right.
cluster(varname) specifies the variable which defines sampling clusters. If cluster() is defined, then the variances
and confidence limits are calculated assuming that the data represent a sample of clusters from a population of
clusters, rather than a sample of observations from a population of observations.
cfweight(expression) specifies an expression giving the cluster frequency weights. These cluster frequency weights
must have the same value for all observations in a cluster. If cfweight() and cluster() are both specified, then
each cluster in the dataset is assumed to represent a number of identical clusters equal to the cluster frequency
weight for that cluster. If cfweight() is specified and cluster() is unspecified, then each observation in the
dataset is treated as a cluster, and assumed to represent a number of identical one–observation clusters equal to
the cluster frequency weight. For more details on the interpretation of weights, see Interpretation of weights
below.
funtype(functional type) specifies whether the Somers’ D or Kendall’s τa functionals estimated are ratios of
between–cluster, within–cluster or von Mises functionals. These three functional types are specified by the
options funtype(bcluster), funtype(wcluster) or funtype(vonmises), respectively. If funtype() is not
specified, then funtype(bcluster) is assumed, and between–cluster functionals are estimated. The within–
cluster Somers’ D is a generalization of the confidence interval corresponding to the sign test (see [R] signrank).
The Gini coefficient is a special case of the clustered von Mises Somers’ D. For further details, see Methods
and Formulas.
wstrata(varlist) specifies a list of variables whose value combinations are the W –strata. If wstrata() is specified,
then somersd estimates stratified Somers’ D or Kendall’s τa parameters, applying only to pairs of observations
within the same W –stratum. These parameters can be used to measure associations within strata, such as
associations between an outcome and an exposure within groups defined by values of a confounder, or by values
of a propensity score based on multiple confounders.
bstrata(varlist | n) specifies the B–strata. If bstrata() is specified, then somersd estimates Somers’ D or Kendall’s
τa parameters specific to pairs of observations from different B–strata. These B–strata are either combinations
of values of a list of variables (if varlist is specified) or the individual observations (if n is specified). B–strata
will not often be required. However, if we are estimating the within–cluster Kendall’s τa (using the options taua
funtype(wcluster)), then the additional option bstrata( n) will ensure that the within–cluster Kendall’s τa
can take the whole range of values from −1 (in the case of complete discordance within clusters) to +1 (in the
case of complete concordance within clusters).
notree specifies that somersd does not use the default search tree algorithm based on Newson (2006), but instead
uses a trivial algorithm, which compares every pair of observations and requires much more time with large
datasets. This option is rarely used except to compare performance.
level(#) specifies the confidence level, in percent, for confidence intervals of the estimates; see [R] level.
cimatrix(new matrix) specifies an output matrix to be created, containing estimates and confidence limits for the
untransformed Somers’ D, Kendall’s τa or Greiner’s ρ parameters. If transf() is specified, then the confidence
limits will be asymmetric and based on symmetric confidence limits for the transformed parameters. This option
(like level()) may be used in replay mode as well as in non–replay mode.
If a varlist is supplied, then all options are allowed. If not, then somersd replays the previous somersd estimation
(if available), and the only options allowed are level() and cimatrix().

1.3 Interpretation of weights

 Post-publication update 3

somersd inputs up to two weight expressions, which are the ordinary Stata weights given by the weight and the
cluster frequency weights given by the cfweight() option. Internally, somersd defines and uses three distinct sets
of weights, which are the cluster frequency weights, the observation frequency weights, and the importance weights.
The cluster frequency weights must be the same for different observations in a cluster, and imply that each
cluster in the input dataset represents a number of identical clusters equal to the cluster frequency weight in that
cluster. If cluster() is not specified, then the individual observations are clusters, and the cluster frequency weight
implies that each one–observation cluster represents a number of identical one–observation clusters equal to the
cluster frequency weight. The cluster frequency weights are given by cfweight() if that option is specified, are set
to 1 if cfweight() is unspecified and cluster() is specified, are equal to the ordinary Stata weights if neither
cluster() nor cfweight() is specified and the ordinary Stata weights are fweights, and are equal to 1 otherwise.
The observation frequency weights are summed over all observations in the input dataset to produce the number
of observations reported by somersd and returned in the estimation result e(N), and are not used in any other way.
They are set by cfweight() if that option is specified and the ordinary Stata weights are not fweights, are equal
to the ordinary Stata weights if cfweight() is unspecified and the ordinary Stata weights are fweights, are equal
to the product of the cfweight() expression and the ordinary Stata weights if cfweight() is specified and the
ordinary Stata weights are fweights, and are equal to 1 otherwise.
The importance weights are used as described in Methods and Formulas below. They are equal to the ordinary
Stata weights if these are specified and either cluster() or cfweight() is specified, are equal to the ordinary Stata
weights if neither of these two options is specified and the ordinary Stata weights are specified as pweights or
iweights, and are equal to 1 otherwise.

1.4 Saved results

somersd saves in e():
Scalars
e(N) number of observations e(df r) residual degrees of freedom
e(N clust) number of clusters
Macros
e(cmd) somersd e(param) parameter (somersd or taua)
e(parmlab) parameter label in output e(tdist) tdist if specified
e(depvar) name of X–variable e(clustvar) name of cluster variable
e(vcetype) title used to label standard error e(wtype) weight type
e(wexp) weight expression e(cfweight) cfweight() expression
e(funtype) funtype() option e(wstrata) wstrata() option
e(bstrata) bstrata() option e(predict) program called by predict (somers p)
e(transf) transformation specified by transf() e(tranlab) transformation label in output
e(properties) "b V"
Matrices
e(b) coefficient vector e(vV) variance–covariance matrix of the estimators
Functions
e(sample) marks estimation sample

Note that (confusingly) e(depvar) is the X–variable, or predictor variable, in the conventional terminology for
defining Somers’ D. somersd is also different from most estimation commands in that its results are not designed to
be used by predict. If the user tries to do so, then the program somers p is called, and tells the user that predict
should not be used after somersd.

2. Methods and formulas

The population value of Kendall’s τa (Kendall and Gibbons, 1990) is traditionally defined as

τXY = E [sign(X1 − X2 ) sign(Y1 − Y2 )] , (1)

where (X1 , Y1 ) and (X2 , Y2 ) are bivariate random variables sampled independently from the same population, and E[·]
denotes expectation. This definition can be generalized to possibly left– or right–censored and/or stratified and/or
clustered and/or weighted data as follows. Suppose that 4–variate observations (Xi , Ri , Yi , Si ) are sampled from an
arbitrary population, using an arbitrary sampling scheme. The Ri are censorship indicators for the corresponding
Xi , and the Si are censorship indicators for the corresponding Yi . These censorship indicators are negative in the
case of left–censorship (where the “true” value of the indicated variable is known to be equal to or less than its
recorded value), positive in the case of right–censorship (in which the “true” value of the indicated variable is known
4 Post-publication update RBN-1

to be equal to or greater than its recorded value), and zero in the case of non–censorship (in which the “true” value
is known to be equal to the recorded value). We define a “censored sign difference” for two values u and v, with
respective censorship indicators p and q, as
(
1, u > v and p ≥ 0 ≥ q
csign(u, p, v, q) = −1, u < v and p ≤ 0 ≤ q (2)
0, otherwise.

Given two observations (Xi , Ri , Yi , Si ) and (Xj , Rj , Yj , Sj ), we will denote the product of csign(Xi , Ri , Xj , Rj )
and csign(Yi , Si , Yj , Sj ) as the concordance–discordance difference for the two observations, and say that the two
observations are concordant if this product is 1, discordant if the product is −1, and neither concordant nor discordant
if the product is zero. We can now redefine Kendall’s τa as

τXY = E [csign(Xi , Ri , Xj , Rj ) csign(Yi , Si , Yj , Sj )] , (3)

or (in words) as the mean concordance–discordance difference. This expectation can be defined using weights specific
to the observations and/or restrictions to subsets of pairs of observations, defined in terms of the sampling scheme.
The population value of Somers’ D (Somers, 1962) is defined as
τXY
DY X = . (4)
τXX

Therefore, τXY is the difference between two probabilities, namely the probability that the larger of the two X–values
is associated with the larger of the two Y –values and the probability that the larger X–value is associated with the
smaller Y –value. DY X is the difference between the two corresponding conditional probabilities, given that the two X–
values are known to be unequal. Somers’ D is related to Harrell’s c index by the formula D = 2c − 1 (see Harrell et al.,
1982 and Harrell et al., 1996). Kendall’s τa is the covariance between csign(Xi , Ri , Xj , Yj ) and csign(Yi , Si , Yj , Sj ),
whereas Somers’ D is the regression coefficient of csign(Yi , Si , Yj , Sj ) with respect to csign(Xi , Ri , Xj , Rj ). The
correlation coefficient between csign(Xi , Ri , Xj , Rj ) and csign(Yi , Si , Yj , Sj ) is known as Kendall’s τb , and is equal
to the geometric mean of the absolute values of DY X and DXY multiplied by their common sign.
Given a sample of observations (Xi , Ri , Yi , Si ), we may estimate and test the population values of Kendall’s τa and
Somers’ D by the corresponding sample statistics τ̂XY and D̂Y X . These are commonly known as “non–parametric”
statistics, even though τXY and DY X are parameters. The two Wilcoxon rank–sum tests (see [R] ranksum and
[R] signrank) both test hypotheses predicting DY X = 0. The two–sample rank–sum test represents the case where
X is a binary variable indicating membership of one of two sub–populations. If the binary X–variable indicates that
a patient has a disease, and the Y –variable is a continuous diagnostic test indicator with high values indicating a
high probability that the patient has the disease, then the area A under the receiver operating characteristic (ROC)
curve, or sensitivity–specificity curve, is linked to Somers’ D by the relation DY X = 2A − 1. (See [R] roc or Hanley
and McNeil, 1982.) The matched–pairs rank–sum test represents the case where there are paired data (Wi1 , Wi2 ),
such that Xi = sign(Wi1 − Wi2 ), and Yi = |Wi1 − Wi2 |. Kendall’s τa is usually tested on “continuous” data, using
ktau (see [R] spearman).

2.1 Motivation for confidence intervals

There are several reasons for preferring confidence intervals to P –values alone:
1. Non–statisticians often quote a “non–significant” result for a “non–parametric” test and argue as if they have
“proved” a null hypothesis, when a confidence interval would show a wide range of other hypotheses which also
fit the data.
2. In the case of continuous bivariate data, there is a correspondence between Kendall’s τa and the better–known
Pearson’s correlation coefficient ρ, known as Greiner’s relation (Kendall and Gibbons, 1990). This states that
³π ´
ρ = sin τa , (5)
2
and holds if the joint distribution of X and Y is bivariate normal. Under this relation, Kendall’s τa –values
of 0, ± 31 , ± 12 and ±1 correspond to Pearson’s correlations of 0, ± 12 , ± √12 and ±1, respectively. A similar
correspondence is likely to hold in a wider range of continuous bivariate distributions (Kendall, 1949; Newson,
2002).
 Post-publication update 5

3. Kendall’s τa has the desirable property that a larger τa cannot be secondary to a smaller τa . That is to say, if
a positive τXY is caused entirely by a monotonic positive relationship of both variables with a third variable
W , then τW X and τW Y must both be greater than τXY . If we can show that τXY − τW Y > 0 (or, equivalently,
that DXY − DW Y > 0), then this implies that the correlation between X and Y is not caused entirely by the
influence of W .
To understand the third point, assume that data points (Wi , Xi , Yi , Si ) are sampled independently from a
common population, with discrete probability mass function fW,X,Y,S (·, ·, ·, ·) and marginal probability mass function
fW,X (·, ·). We assume that the Si are censorship indicators for the corresponding outcome variables Yi , and that the
Xi and Wi are alternative uncensored predictors. Define the conditional expectation
Z(w1 , x1 , w2 , x2 ) = E [ csign(Y2 , S2 , Y1 , S1 ) | W1 = w1 , X1 = x1 , W2 = w2 , X2 = x2 ] (6)
for any w1 and w2 in the range of W –values and any x1 and x2 in the range of X–values. If we state that the positive
relationship between Xi and Yi is caused entirely by a monotonic positive relationship between both variables and
Wi , then that is equivalent to stating that
Z(w1 , x1 , w2 , x2 ) ≥ 0 (7)
whenever w1 ≤ w2 and x2 ≤ x1 . However, the difference between the two τa coefficients is
X X
τW Y − τXY =2 fW,X (w, x1 ) fW,X (w, x2 ) Z(w, x1 , w, x2 )
w x2 <x1
X X
+2 fW,X (w1 , x) fW,X (w2 , x) Z(w1 , x, w2 , x)
x w1 <w2
X X
+4 fW,X (w1 , x1 ) fW,X (w2 , x2 ) Z(w1 , x1 , w2 , x2 ). (8)
w1 <w2 x2 <x1

This difference must be non–negative whenever the inequality (7) applies. In particular, if the distribution of the Wi
and Xi is nearly continuous, then the difference (8) will be dominated by the third term, representing discordant
(Wi , Xi )–pairs.

2.2 Estimation formulas for Somers’ D and Kendall’s τa

Somers’ D and Kendall’s τa , in their various forms, can be expressed as ratios of sample means, Hoeffding
U –statistics or von Mises V –statistics, depending on the functional type specified by the funtype() option. somersd
works by jackknifing the original means, U –statistics and V –statistics (Arvesen, 1969), and by using Taylor polynomials
to derive variances for the ratios. Normalizing and/or variance–stabilizing transformations may then be applied.
We assume the general case where the observations are clustered, which becomes the familiar unclustered case
when there is one observation per cluster, and that there are N clusters in the sample, sampled from a common
population. We assume that there are one or more indexed W –strata (defaulting to one all–inclusive W –stratum if
wstrata() is not specified). Two slightly different versions of the notation will be used, depending on whether or
not the user has specified B–strata using the bstrata() option.
If there are no B–strata, then we define wf gi , Xf gi , Yf gi , Rf gi and Sf gi to be the importance weight, X–value,
Y –value, X–censorship indicator and Y –censorship indicator, respectively, for the ith observation belonging to the
gth W –stratum in the f th cluster. Not every possible index combination f gi will correspond to an observation, so
all summation over index combinations will be over index combinations corresponding to an observation. For index
combinations f gi and jkm corresponding to observations, we can define
vf gi,jkm = wf gi wjkm ,
(XY )
tf gi,jkm = vf gi,jkm csign(Xf gi , Rf gi , Xjkm , Rjkm ) csign(Yf gi , Sf gi , Yjkm , Sjkm ). (9)

We will use the usual dot–substitution notation to define (for instance)

X (XY )
X (XY ) X (XY )
X (XY )
vf gi,jk. = vf gi,jkm , tf gi,jk. = tf gi,jkm , vf gi,j.. = vf gi,jk. , tf gi,j.. = tf gi,jk. , (10)
m m k k

and any other sums over any other indices. For clusters f and j, we define
(V )
X (XY )
X (XY )
φf j = vf g.,jg. , φf j = tf g.,jg. . (11)
g g
6 Post-publication update RBN-1

(V ) (XY )
In other words, φf j is the sum of pairwise importance weights, and φf j is the sum of pairwise importance–weighted
concordance–discordance differences, belonging to pairs of observations, in the same W –stratum, of which the first
(V ) (XY )
observation is in cluster f and the second observation is in cluster j. The quantities φf j and φf j are known as
kernels in the terminology of Chapter 5 of Serfling (1980), and are defined for any pair of clusters.
(V ) (XY )
If the user has defined B–strata, then we define the kernels φf j and φf j by a slightly different formula.
We define wf ghi , Xf ghi , Yf ghi , Rf ghi and Sf ghi to be the importance weight, X–value, Y –value, X–censorship
indicator and Y –censorship indicator, respectively, for the ith observation belonging to cluster f , W –stratum g and
B–stratum h. For index combinations f ghi and jklm corresponding to observations, we define

vf ghi,jklm = wf ghi wjklm ,

(XY )
tf ghi,jklm = vf ghi,jklm csign(Xf ghi , Rf ghi , Xjklm , Rjklm ) csign(Yf ghi , Sf ghi , Yjklm , Sjklm ), (12)

and for clusters f and j we define

(V )
X XX (XY )
X (XY )
XX (XY )
φf j = vf g..,jg.. − vf gh.,jgh. , φf j = tf g..,jg.. − tf gh.,jgh. . (13)
g g h g g h

(V ) (XY )
This time, φf j is the sum of products of importance weights, and φf j is the sum of importance–weighted
concordance–discordance differences, belonging to pairs of observations, in the same W –stratum and different B–
strata, of which the first observation is in cluster f and the second observation is in cluster j. Note that, if the user
(V )
has specified bstrata( n), then every observation is in its own B–stratum, and the second terms in the φf j and
(XY )
φf j of (13) will then contain only pairs in which an observation is paired with itself.
(V ) (XY )
The kernels φf j and φf j of (11) or (13) can be “averaged” over their indices to produce parameters denoted
as V and TXY , respectively. Kendall’s τa and Somers’ D are defined as ratios of these “averages” by

τXY = TXY /V, DY X = TXY /TXX = τXY /τXX . (14)

The way in which the kernels are averaged depends on the funtype() option. If the user specifies funtype(wcluster),
then V and TXY are “within–cluster averages”. If the user specifies funtype(bcluster) (the default), then V and
TXY are “between–cluster averages”. If the user specifies funtype(vonmises), then V and TXY are “overall averages”.
In all cases, we estimate the population parameters V and TXY using sample statistics V̂ and T̂XY as point estimates,
and estimate the sampling variances of these point estimates using a jackknife method, with pseudovalues denoted
(V ) (XY )
ψj and ψj for the jth cluster.
If the user specifies funtype(wcluster), then somersd estimates the parameters
h i h i
(V ) (XY )
V = E φjj , TXY = E φjj . (15)

These functionals are population means of within–cluster kernels, and their point estimates are the corresponding
sample means
N
X XN
(V ) (XY )
V̂ = N −1 φjj , T̂XY = N −1 φjj , (16)
j=1 j=1

and the jackknife pseudovalues for the jth cluster are given by

(V ) (V ) (XY ) (XY )
ψj = φjj , ψj = φjj . (17)

If the user has specified funtype(bcluster) (the default) or funtype(vonmises), then somersd estimates the
parameters h i h i
(V ) (XY )
V = E φf j , TXY = E φf j , (18)

for f 6= j. These parameters are known as Hoeffding functionals if clusters f and j are assumed to be sampled
without replacement, and as von Mises functionals if clusters f and j are assumed to be sampled with replacement.
 Post-publication update 7

(If the population from which the clusters are sampled is infinite, then the population Hoeffding functional is equal
to the corresponding population von Mises functional.)
If the user specifies funtype(bcluster), or does not specify a funtype() option, then the point estimates
of the population Hoeffding functionals are the corresponding sample Hoeffding functionals, or U –statistics in the
terminology of Hoeffding (1948), defined as V̂ = T̂XY = 0 if N = 1, and otherwise as
PN (V ) PN (XY )
φ(V
..
)
− j=1 φjj φ(XY
..
)
− j=1 φjj
V̂ = , T̂XY = . (19)
N (N − 1) N (N − 1)

(V ) (XY )
The jackknife pseudovalues for the jth cluster are given by ψj = ψj = 0 if N = 1, by

(V ) (V ) (V ) (XY ) (XY ) (XY )

ψj = φj. − φjj , ψj = φj. − φjj (20)

if N = 2, and otherwise as
à ! " #
N
X N
X ³ ´
(V ) −1 (V ) (V ) (V ) (V )
ψj = (N − 1) φ..(V ) − φkk − (N − 2) −1
φ(V
..
)
− φkk −2 φj. − φjj ,
k=1 k=1
à ! " #
N
X N
X ³ ´
(XY ) −1 (XY ) (XY ) (XY ) (XY )
ψj = (N − 1) φ(XY
..
)
− φkk − (N − 2) −1
φ(XY
..
)
− φkk −2 φj. − φjj .
k=1 k=1
(21)

If the user specifies funtype(vonmises), then the point estimates of the population von Mises functionals are
the corresponding sample von Mises functionals, or V –statistics in the terminology of Chapter 5 of Serfling (1980).
These are defined as
V̂ = N −2 φ(V )
.. , T̂XY = N −2 φ(XY
..
)
, (22)
and the jackknife pseudovalues for the jth cluster are given by

(V ) (V ) (XY ) (XY )
ψj = φjj , ψj = φjj (23)

if N = 1, and otherwise by
³ ´
(V ) (V ) (V )
ψj = N −1 φ(V
..
)
− (N − 1)−1 φ(V
..
)
− 2φj. + φ jj ,
³ ´
(XY ) (XY ) (XY )
ψj = N −1 φ(XY
..
)
− (N − 1)−1 φ(XY..
)
− 2φj. + φjj . (24)

Note that the estimates and jackknife pseudovalues of formulas (15) to (24) can all be expressed in terms of
(V ) (V ) (XY ) (XY )
the φjj , φj. , φjj and φj. . Newson (2006) introduced an algorithm to calculate these quantities, using binary
search trees, which requires an amount of computation time of order Nobs log Nobs , where Nobs is the total number
of observations. A version of this algorithm is used by somersd, unless the user specifies the notree option, in which
case somersd uses a trivial algorithm, which compares all pairs of observations and requires an amount of time
quadratic in Nobs . The difference in performance can be spectacular in large datasets (Nobs > 1000).
The parameters we really want to estimate are Kendall’s τa and/or Somers’ D, defined by (14). These formulas
are equal to the familiar formulas (3) and (4) if each cluster contains one observation with an importance weight of
one. To estimate them, we use the jackknife method on V and TXY , and use appropriate Taylor polynomials. somersd
calculates correlation measures for a single variable X with a set of Y –variates (Y (1) , . . . , Y (p) ). (The X–variate may
have a censorship indicator R, and the Y –variates may have censorship indicators (S (1) , . . . , S (p) ).) It calculates, in
the first instance, the covariance matrix for V̂ , T̂XX , and T̂XY (i) for 1 ≤ i ≤ p. This is done using the jackknife
influence matrix Υ, which has N rows labelled by the cluster subscripts, and p + 2 columns labelled (in Stata fashion)
by the names V , X, and Y (i) for 1 ≤ i ≤ p. It is defined by
h i
(V ) (XX) (XY (i) ) (i)
Υ [j, V ] = ψj − ψ̄ (V ) , Υ [j, X] = ψj − ψ̄ (XX) , Υ j, Y (i) = ψj − ψ̄ (XY ) , (25)
8 Post-publication update RBN-1

where the quantities

N
X N
X N
X
(V ) (XX) (i) (XY (i) )
ψ̄ (V ) = N −1 ψk , ψ̄ (XX) = N −1 ψk , ψ̄ (XY )
= N −1 ψk (26)
k=1 k=1 k=1

are the mean pseudovalues. (These mean pseudovalues are equal to the corresponding point estimates unless
funtype(vonmises) is specified, in which case the mean pseudovalue is equal to the corresponding Hoeffding
U –statistic.) The jackknife covariance matrix is equal to
−1
Ĉ = [N (N − 1)] Υ0 Υ. (27)
The estimates for Kendall’s τa and Somers’ D, for variables Y and X, are defined by
τ̂XY = T̂XY /V̂ , D̂Y X = T̂XY /T̂XX , (28)
unless the denominators of these expressions are zero, in which case the numerators must also be zero, and somersd
therefore sets the estimates and their covariances to zero. If the denominator is nonzero, then the covariance matrix
is defined using Taylor polynomials. In the case of Somers’ D, we define the p × (p + 2) matrix of estimated derivatives
Γ̂(D) , whose rows are labelled by the names Y (1) , . . . , Y (p) , and whose columns are labelled by V, X, Y (1) , . . . , Y (p) .
This matrix is defined by
h i ∂ D̂Y (i) X T̂ (i)
Γ̂(D) Y (i) , X = = − XY 2
,
∂ T̂XX T̂XX
h i ∂ D̂Y (i) X 1
Γ̂(D) Y (i) , Y (i) = = , (29)
∂ T̂XY (i) T̂XX
all other entries being 0. In the case of Kendall’s τa , we define a (p + 1) × (p + 2) matrix of estimated derivatives Γ̂(τ ) ,
whose rows are labelled by X, Y (1) , . . . , Y (p) , and whose columns are labelled by V, X, Y (1) , . . . , Y (p) . This matrix is
defined by
∂ τ̂XX T̂XX
Γ̂(τ ) [X, V ] = = − ,
∂ V̂ V̂ 2
∂ τ̂XX 1
Γ̂(τ ) [X, X] = = ,
∂ T̂XX V̂
h i ∂ τ̂XY (i) T̂ (i)
Γ̂(τ ) Y (i) , V = = − XY ,
∂ V̂ V̂ 2
h i ∂ τ̂ (i) 1
Γ̂(τ ) Y (i) , Y (i) = XY = , (30)
∂ T̂XY (i) V̂
all other entries again being 0. The estimated dispersion matrices of the Somers’ D and τa estimates are therefore
Ĉ (D) and Ĉ (τ ) , respectively, defined by
Ĉ (D) = Γ̂(D) Ĉ Γ̂(D) 0 , Ĉ (τ ) = Γ̂(τ ) Ĉ Γ̂(τ ) 0 . (31)

2.3 Transformations
The transf() option offers a choice of transformations. Since these are available both for Somers’ D and for
Kendall’s τa , we will denote the original estimate as θ (which can stand for D or τ ) and the transformed estimate
as ζ. They are summarized below, together with their derivatives dζ/dθ and their inverses θ(ζ).
transf() Transform name ζ(θ) dζ/dθ θ(ζ)

iden Untransformed θ 1 ζ
¡ ¢−1
z Fisher’s z arctanh(θ) = 1 − θ2 tanh(ζ) =
1
2 log[(1 + θ)/(1 − θ)] [exp(2ζ) − 1]/[exp(2ζ) + 1]
¡ ¢−1/2
asin Daniels’ arcsine arcsin(θ) 1 − θ2 sin(ζ)
rho Greiner’s ρ sin( π2 θ) π π
2 cos( 2 θ) (2/π) arcsin(ζ)
π π π π 2 −1
zrho Greiner’s ρ arctanh[sin( 2 θ)] 2 cos( 2 θ)[1 − sin( 2 θ) ] (2/π) arcsin[tanh(ζ)]
(z-transformed)
c Harrell’s c (θ + 1)/2 1/2 2ζ − 1
 Post-publication update 9

(Note that all of these expressions are defined for θ = 0, but some are undefined for θ = 1 or θ = −1, and, in
those cases, somersd enters a substitute θ–argument very close to 1 or −1.) If transf() is specified, then somersd
displays and saves the transformed estimates and their estimated covariance, instead of the untransformed versions.
If Ĉ (θ) is the covariance matrix for the untransformed estimates given by (31), and Γ̂(ζ) is the diagonal matrix whose
diagonal entries are the dζ/dθ estimates specified in the table, then the transformed parameter and its covariance
matrix are
ζ̂ = ζ(θ̂), Ĉ (ζ) = Γ̂(ζ) Ĉ (θ) Γ̂(ζ) 0 . (32)
Fisher’s z–transform was originally recommended for the Pearson correlation coefficient by Fisher (1921) (see
also Gayen (1951)), but Edwardes (1995) recommended it specifically for Somers’ D on the basis of simulation
studies. Daniels’ arcsine was suggested as a normalizing transform in Daniels and Kendall (1947). If transf(z)
or transf(asin) is specified, then somersd prints asymmetric confidence intervals for the untransformed D or
τa parameters, calculated from symmetric confidence intervals for the transformed parameters using the inverse
function θ(ζ). (This feature corresponds to the eform option of other estimation commands.) Greiner’s ρ (Kendall
and Gibbons, 1990) is based on the relation (5), and is designed to estimate the Pearson correlation coefficient
corresponding to the measured τa . If transf(zrho) is specified, then somersd prints asymmetric confidence intervals
for the untransformed Greiner’s ρ, using the inverse z–transform on symmetric confidence intervals for the z–
transformed Greiner’s ρ. Harrell’s c is usually a reparameterization of Somers’ D, and is recommended in Harrell et
al. (1982) and Harrell et al. (1996) as a general measure of the predictive power of a prognostic score arising from a
medical test.

3 Examples
We present 4 examples, using datasets provided by Stata Press at http://www.stata-press.com/data/.

3.1 Somers’ D in the auto data

In the auto data, we compare US cars with foreign cars regarding weight and fuel efficiency. First, we use
ranksum to give significance tests without confidence intervals:
. ranksum mpg,by(foreign)
Two-sample Wilcoxon rank-sum (Mann-Whitney) test
foreign | obs rank sum expected
-------------+---------------------------------
Domestic | 52 1688.5 1950
Foreign | 22 1086.5 825
-------------+---------------------------------
combined | 74 2775 2775
unadjusted variance 7150.00
adjustment for ties -36.95
----------
adjusted variance 7113.05
Ho: mpg(foreign==Domestic) = mpg(foreign==Foreign)
z = -3.101
Prob > |z| = 0.0019
. ranksum weight,by(foreign)
Two-sample Wilcoxon rank-sum (Mann-Whitney) test
foreign | obs rank sum expected
-------------+---------------------------------
Domestic | 52 2379.5 1950
Foreign | 22 395.5 825
-------------+---------------------------------
combined | 74 2775 2775
unadjusted variance 7150.00
adjustment for ties -1.06
----------
adjusted variance 7148.94
Ho: weight(foreign==Domestic) = weight(foreign==Foreign)
z = 5.080
Prob > |z| = 0.0000

We note that US cars are typically heavier and travel fewer miles per gallon than foreign cars. For confidence
intervals, we use somersd:
. somersd foreign mpg weight
Somers’ D with variable: foreign
Transformation: Untransformed
10 Post-publication update RBN-1

Valid observations: 74
Symmetric 95% CI
------------------------------------------------------------------------------
| Jackknife
foreign | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
mpg | .4571678 .135146 3.38 0.001 .1922866 .7220491
weight | -.7508741 .0832485 -9.02 0.000 -.9140383 -.58771
------------------------------------------------------------------------------

We see that, given a randomly–chosen foreign car and a randomly–chosen US car, the foreign car is 46% more
likely to travel more miles per gallon than the US car than vice versa, with confidence limits from 19% to 72% more
likely. However, being foreign seems to be more reliable as a negative predictor of weight than as a positive predictor
of “fuel efficiency”. We can use lincom to define confidence limits for the difference:
. lincom -weight-mpg
( 1) - mpg - weight = 0
------------------------------------------------------------------------------
foreign | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
(1) | .2937063 .0884397 3.32 0.001 .1203677 .4670449
------------------------------------------------------------------------------

The difference between Somers’ D–values is positive. This indicates that, if there are two cars, one heavier and
consuming fewer gallons per mile, the other lighter and consuming more gallons per mile, then the second is more
likely to be foreign. So maybe 1970s US cars were not as wasteful as some people think, and were, if anything, more
fuel–efficient for their weight than non–US cars at the time. Figure 1 illustrates this graphically. Data points are
domestic cars (“D”) and foreign cars (“F”). A regression analysis could show the same thing, but Somers’ D shows
it in stronger terms, without contentious assumptions such as linearity. (On the other hand, a regression model is
more informative if its assumptions are true, so the two methods are mutually complementary.)

45
F
40

FF
35 D

F
F D
30 D
FD D
Mileage (mpg)

F D D
FF DF F
25 F DDD
F F F
DD D DD
F D F D D
D DD
20 D
D DD DDD
D
F F DD D D DD
D
F F D D
D D D D
D D
15 F DD DD D
D D

10

0
1,000 2,000 3,000 4,000 5,000
Weight (lbs.)
Figure 1. Mileage and weight in US cars (D) and non–US cars (F)

The confidence intervals for such high values of Somers’ D would probably be more reliable if we used the
z–transform, recommended by Edwardes (1995). The results of this are as follows:
 Post-publication update 11

. somersd foreign mpg weight,tran(z)

Somers’ D with variable: foreign
Transformation: Fisher’s z
Valid observations: 74
Symmetric 95% CI for transformed Somers’ D
------------------------------------------------------------------------------
| Jackknife
foreign | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
mpg | .4937249 .1708551 2.89 0.004 .1588551 .8285947
weight | -.9749561 .1908547 -5.11 0.000 -1.349024 -.6008878
------------------------------------------------------------------------------
Asymmetric 95% CI for untransformed Somers’ D
Somers_D Minimum Maximum
mpg .45716783 .15753219 .67972072
weight -.75087413 -.87382282 -.53768098
. lincom -weight-mpg
( 1) - mpg - weight = 0
------------------------------------------------------------------------------
foreign | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
(1) | .4812312 .1235452 3.90 0.000 .2390871 .7233753
------------------------------------------------------------------------------

Note that somersd gives not only symmetric confidence limits for the z–transformed Somers’ D estimates, but
also the more informative asymmetric confidence limits for the untransformed Somers’ D estimates (corresponding
to the eform option). The asymmetric confidence limits for the untransformed estimates are closer to zero than
the symmetric confidence limits for the untransformed estimates in the previous output, and are probably more
realistic. The output to lincom gives confidence limits for the difference between z–transformed Somers’ D values.
This difference is expressed in z–units, but must, of course, be in the same direction as the difference between
untransformed Somers’ D values. The conclusions are similar.

3.2 Kendall’s τa in the auto data

In this example, we demonstrate Kendall’s τa by comparing weight (pounds) and displacement (cubic inches)
as predictors of fuel efficiency (miles per gallon). We first use ktau to carry out significance tests with no confidence
limits:
. ktau mpg mpg
Number of obs = 74
Kendall’s tau-a = 0.9471
Kendall’s tau-b = 1.0000
Kendall’s score = 2558
SE of score = 212.989 (corrected for ties)
Test of Ho: mpg and mpg are independent
Prob > |z| = 0.0000 (continuity corrected)
. ktau mpg weight
Number of obs = 74
Kendall’s tau-a = -0.6857
Kendall’s tau-b = -0.7059
Kendall’s score = -1852
SE of score = 213.605 (corrected for ties)
Test of Ho: mpg and weight are independent
Prob > |z| = 0.0000 (continuity corrected)
. ktau mpg displacement
Number of obs = 74
Kendall’s tau-a = -0.5942
Kendall’s tau-b = -0.6257
Kendall’s score = -1605
SE of score = 212.850 (corrected for ties)
Test of Ho: mpg and displacement are independent
Prob > |z| = 0.0000 (continuity corrected)

We then use somersd (with the taua option and the z–transform) to compute the same statistics with confidence
limits. Note that somersd also outputs the τa of mpg with mpg, which is simply the probability that two independently
sampled mpg–values are not equal.

. somersd mpg weight displacement,taua tr(z)

Kendall’s tau-a with variable: mpg
12 Post-publication update RBN-1

Transformation: Fisher’s z
Valid observations: 74
Symmetric 95% CI for transformed Kendall’s tau-a
------------------------------------------------------------------------------
| Jackknife
mpg | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
mpg | 1.802426 .0748368 24.08 0.000 1.655748 1.949103
weight | -.8397412 .084022 -9.99 0.000 -1.004421 -.6750612
displacement | -.6841711 .093055 -7.35 0.000 -.8665556 -.5017866
------------------------------------------------------------------------------
Asymmetric 95% CI for untransformed Kendall’s tau-a
Tau_a Minimum Maximum
mpg .94705665 .92964223 .96024957
weight -.68567197 -.76344472 -.58829928
displacement -.59422436 -.69961991 -.46352103

We can use lincom to compare the two predictors and test whether smaller and heavier cars travel fewer miles
per gallon than larger and lighter cars. This seems to be the case, as weight is a more negative predictor of mpg
than displacement:
. lincom weight-displacement
( 1) weight - displacement = 0
------------------------------------------------------------------------------
mpg | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
(1) | -.1555701 .0742717 -2.09 0.036 -.3011399 -.0100003
------------------------------------------------------------------------------

We demonstrate the cluster option using the variable manuf, equal to the first word of make, to denote
manufacturer. This analysis assumes that we are sampling from the population of car manufacturers, rather than
from the population of car models. The results are as follows:
. somersd mpg weight displacement,taua tr(z) cluster(manuf)
Kendall’s tau-a with variable: mpg
Transformation: Fisher’s z
Valid observations: 74
Number of clusters: 23
Symmetric 95% CI for transformed Kendall’s tau-a
(Std. Err. adjusted for 23 clusters in manuf)
------------------------------------------------------------------------------
| Jackknife
mpg | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
mpg | 1.83398 .0821029 22.34 0.000 1.673061 1.994898
weight | -.8391083 .0917593 -9.14 0.000 -1.018953 -.6592633
displacement | -.694607 .0976751 -7.11 0.000 -.8860467 -.5031674
------------------------------------------------------------------------------
Asymmetric 95% CI for untransformed Kendall’s tau-a
Tau_a Minimum Maximum
mpg .95021392 .93195521 .96366535
weight -.68533644 -.76943983 -.57787293
displacement -.60093349 -.70943563 -.46460448
. lincom weight-displacement
( 1) weight - displacement = 0
------------------------------------------------------------------------------
mpg | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
(1) | -.1445012 .0801437 -1.80 0.071 -.30158 .0125775
------------------------------------------------------------------------------

Note that, in contrast to the case of most estimation commands, the cluster option affects the estimates as well
as their confidence limits. This is because, in default, the clustered estimates are calculated only from between–cluster
comparisons, in this case pairs of car models from different manufacturers. However, we can alter this by specifying
funtype(vonmises) bstrata( n), in which case the estimates will be as before, but the confidence limits will be
different:
. somersd mpg weight displacement,taua tr(z) cluster(manuf) ///
> funtype(vonmises) bstrata(_n)
Von Mises Kendall’s tau-a with variable: mpg
 Post-publication update 13

Transformation: Fisher’s z
Between strata defined by: _n
Valid observations: 74
Number of clusters: 23
Symmetric 95% CI for transformed Kendall’s tau-a
(Std. Err. adjusted for 23 clusters in manuf)
------------------------------------------------------------------------------
| Jackknife
mpg | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
mpg | 1.802426 .0813496 22.16 0.000 1.642983 1.961868
weight | -.8397412 .0905411 -9.27 0.000 -1.017199 -.6622839
displacement | -.6841711 .0940933 -7.27 0.000 -.8685905 -.4997517
------------------------------------------------------------------------------
Asymmetric 95% CI for untransformed Kendall’s tau-a
Tau_a Minimum Maximum
mpg .94705665 .92788848 .96123213
weight -.68567197 -.76872301 -.57988136
displacement -.59422436 -.70065729 -.46192188
. lincom weight-displacement
( 1) weight - displacement = 0
------------------------------------------------------------------------------
mpg | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
(1) | -.1555701 .0817275 -1.90 0.057 -.315753 .0046128
------------------------------------------------------------------------------

Alternatively, we might want to compare the power of weight and displacement to predict mileage when comparing
cars made by the same manufacturer. To do this, we use the options funtype(wcluster) bstrata( n):
. somersd mpg weight displacement,taua tr(z) cluster(manuf) ///
> funtype(wcluster) bstrata(_n)
Within-cluster Kendall’s tau-a with variable: mpg
Transformation: Fisher’s z
Between strata defined by: _n
Valid observations: 74
Number of clusters: 23
Symmetric 95% CI for transformed Kendall’s tau-a
(Std. Err. adjusted for 23 clusters in manuf)
------------------------------------------------------------------------------
| Jackknife
mpg | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
mpg | 1.396604 .1289473 10.83 0.000 1.143872 1.649336
weight | -.852374 .1701861 -5.01 0.000 -1.185933 -.5188155
displacement | -.4992644 .070216 -7.11 0.000 -.6368852 -.3616436
------------------------------------------------------------------------------
Asymmetric 95% CI for untransformed Kendall’s tau-a
Tau_a Minimum Maximum
mpg .88461538 .81571376 .92876647
weight -.69230769 -.82931315 -.47678526
displacement -.46153846 -.56277501 -.34666094
. lincom weight-displacement
( 1) weight - displacement = 0
------------------------------------------------------------------------------
mpg | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
(1) | -.3531096 .1551786 -2.28 0.023 -.6572541 -.0489652
------------------------------------------------------------------------------

We see that, once again, both weight and displacement are negative predictors of mileage, and that weight is a
more negative predictor than displacement.
Suppose that we are writing for an audience more familiar with Pearson’s correlation than with Kendall’s τa .
To estimate the Pearson correlations corresponding to our τa coefficients, we use the zrho transform. The results
are as follows:
. somersd mpg weight displacement,taua tr(zrho)
Kendall’s tau-a with variable: mpg
Transformation: z-transform of Greiner’s rho
Valid observations: 74
14 Post-publication update RBN-1

Symmetric 95% CI for transformed Greiner’s rho

------------------------------------------------------------------------------
| Jackknife
mpg | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
mpg | 3.179521 .1458796 21.80 0.000 2.893602 3.465439
weight | -1.378273 .1475561 -9.34 0.000 -1.667478 -1.089069
displacement | -1.108838 .158893 -6.98 0.000 -1.420262 -.7974132
------------------------------------------------------------------------------
Asymmetric 95% CI for untransformed Greiner’s rho
Rho Minimum Maximum
mpg .99654393 .99388566 .99804762
weight -.88056403 -.93121746 -.79653796
displacement -.80365118 -.88965364 -.66258811

The τa of -0.59 between displacement and fuel efficiency (from the unclustered output) is seen to correspond to
a more impressive Pearson correlation of -0.80. The estimated Greiner’s ρ is probably less likely to be oversensitive
to outliers than the usual Pearson coefficient.

3.3 Harrell’s c and Somers’ D in the drugtr data

In this example, we demonstrate the cenind() option with a simple set of survival data distributed by Stata
Press, with 1 observation per subject in a drug trial and data on treatment, age and survival time. We first load the
data, then tabulate the treatment variable drug, then define the new variables youth (representing number of years
to the subject’s 100th birthday) and censind (a censorship indicator equal to 0 for subjects who died and to 1 for
subjects whose survival time is right–censored):
. use http://www.stata-press.com/data/r9/drugtr, clear
(Patient Survival in Drug Trial)
. tab drug, m
Drug type |
(0=placebo) | Freq. Percent Cum.
------------+-----------------------------------
0 | 20 41.67 41.67
1 | 28 58.33 100.00
------------+-----------------------------------
Total | 48 100.00
. gene youth=100-age
. gene byte censind=1-died
. tab died censind, m
1 if |
patient | censind
died | 0 1 | Total
-----------+----------------------+----------
0 | 0 17 | 17
1 | 31 0 | 31
-----------+----------------------+----------
Total | 31 17 | 48

We then use somersd to estimate the Harrell’s c parameters of active treatment and youth with respect to
survival time:
. somersd studytime drug youth, tr(c) cenind(censind)
Somers’ D with variable: studytime
Transformation: Harrell’s c
Valid observations: 48
Symmetric 95% CI for Harrell’s c
------------------------------------------------------------------------------
| Jackknife
studytime | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
drug | .7275986 .0367931 19.78 0.000 .6554855 .7997117
youth | .6415771 .0528314 12.14 0.000 .5380295 .7451246
------------------------------------------------------------------------------
. lincom drug-youth
( 1) drug - youth = 0
------------------------------------------------------------------------------
studytime | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
(1) | .0860215 .0618354 1.39 0.164 -.0351736 .2072166
 Post-publication update 15

------------------------------------------------------------------------------

We see that active drug treatment and youth are both positive survival indicators, as they both have values of
Harrell’s c greater than 0.5. (See Harrell et al., 1982 and Harrell et al., 1996 for more information on Harrell’s c.)
However, when we use lincom to estimate the difference between the two Harrell’s c parameters (equal to half the
difference between the corresponding Somers’ D parameters), we find that the confidence interval for the difference
includes zero. Based on this difference alone, we cannot state that the active treatment is a more or less positive
predictor than being young. However, we can split the sample into 3 age tertiles, and estimate pooled, stratified
Harrell’s c values for youth and treatment (based only on comparisons within age tertiles), and their difference:

. xtile agegp=age, n(3)

. tab agegp, m
3 quantiles |
of age | Freq. Percent Cum.
------------+-----------------------------------
1 | 18 37.50 37.50
2 | 16 33.33 70.83
3 | 14 29.17 100.00
------------+-----------------------------------
Total | 48 100.00
. somersd studytime drug youth, tr(c) cenind(censind) wstrata(agegp)
Somers’ D with variable: studytime
Transformation: Harrell’s c
Within strata defined by: agegp
Valid observations: 48
Symmetric 95% CI for Harrell’s c
------------------------------------------------------------------------------
| Jackknife
studytime | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
drug | .7630597 .0398266 19.16 0.000 .685001 .8411184
youth | .5559701 .0607348 9.15 0.000 .4369321 .6750082
------------------------------------------------------------------------------
. lincom drug-youth
( 1) drug - youth = 0
------------------------------------------------------------------------------
studytime | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
(1) | .2070896 .0660029 3.14 0.002 .0777262 .3364529
------------------------------------------------------------------------------

This time, we see that youth is a less impressive predictor of survival within age tertiles (as the confidence
interval for Harrell’s c contains 0.5), and is a less good predictor than treatment when predicting survival between
subjects in the same age tertile. We can therefore conclude (in very strong terms) that treatment has an effect that
is not entirely caused by confounding by age.
In this analysis, there is only one confounder. There are often many confounders in observational studies in real
life, and this makes stratified analyses more difficult. However, a possible solution might be to define a propensity
score, measuring proneness to allocation to a treatment and dependent on all the confounders, and to use xtile on
the propensity score to define a propensity group variable, which can be used as the wstrata() option by somersd.
The seminal paper on propensity scores is Rosenbaum and Rubin (1983), but a good place to start a literature search
now might be Imai and van Dyk (2004).
Harrell’s c is usually used to compare the predictive power of multivariate predictive scores such as log hazard
ratios from Cox regression models, rather than the univariate predictors assessed here. The estat utility, used after
Cox regression, can calculate Harrell’s c and the corresponding Somers’ D for the log hazard ratio with respect to
survival. It is demonstrated, using the dataset used here, in [ST] stcox postestimation. estat sometimes calculates
values for Harrell’s c and Somers’ D different from those calculated by somersd. This is because estat assumes that,
if two subjects have equal lifetimes and one ends in death and the other ends in censorship, then the second has
survived the first, whereas somersd assumes that neither has survived the other, based on the csign(., ., ., .) function
of (2). If the survival times are expressed in integral numbers of units (such as days), then such ties can be broken
by adding a fraction of a unit to the censored lifetimes and not to the uncensored lifetimes. This will cause somersd
to give the same values as estat.
It must be stressed that confidence limits for the Harrell’s c of a Cox regression score with respect to survival
should be treated with great caution if they are calculated from the same dataset in which the Cox model parameters
16 Post-publication update RBN-1

were fitted. Harrell et al. (1982) and Harrell et al. (1996) stress this, and propose some solutions, using a sub–sampling
method similar to the bootstrap. When these methods are used on large datasets, the search tree algorithm of Newson
(2006), used in default by somersd, will require appreciably less time than the trivial quadratic–time algorithms
used by many other programs (and by somersd if the user specifies notree).
The Wilcoxon–Breslow–Gehan test was introduced by Breslow (1970) and Gehan (1965), and is demonstrated
using Stata in [ST] sts test. It tests the hypothesis of a zero value of the Somers’ D of survival (as the Y –variable)
with respect to membership of a particular group (as the X–variable). Using somersd, we can improve on this by
defining a confidence interval for this Somers’ D parameter:
. sts test drug, wilcoxon
failure _d: died
analysis time _t: studytime
Wilcoxon (Breslow) test for equality of survivor functions
| Events Events Sum of
drug | observed expected ranks
------+--------------------------------------
0 | 19 7.25 385
1 | 12 23.75 -385
------+--------------------------------------
Total | 31 31.00 0
chi2(1) = 22.61
Pr>chi2 = 0.0000
. somersd drug studytime, tr(z) cenind(0 censind)
Somers’ D with variable: drug
Transformation: Fisher’s z
Valid observations: 48
Symmetric 95% CI for transformed Somers’ D
------------------------------------------------------------------------------
| Jackknife
drug | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
studytime | .8297787 .1935732 4.29 0.000 .4503821 1.209175
------------------------------------------------------------------------------
Asymmetric 95% CI for untransformed Somers’ D
Somers_D Minimum Maximum
studytime .68035714 .42221306 .83643191

We see (from the Wilcoxon test) that the treated group has fewer deaths, and the placebo group has more deaths,
than we would expect by chance, assuming population survival distributions to be the same in the two groups. We
also see (from the confidence interval for the untransformed Somers’ D) that, if we sample a subject at random from
each of the two subpopulations (treated and placebo), then the event that the treated subject survives the placebo
subject is 42% to 84% more probable than the event that the placebo subject survives the treated subject. Again,
we can stratify the analysis by age tertile:
. somersd drug studytime, tr(z) cenind(0 censind) wstrata(agegp)
Somers’ D with variable: drug
Transformation: Fisher’s z
Within strata defined by: agegp
Valid observations: 48
Symmetric 95% CI for transformed Somers’ D
------------------------------------------------------------------------------
| Jackknife
drug | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
studytime | .9729551 .2404965 4.05 0.000 .5015905 1.44432
------------------------------------------------------------------------------
Asymmetric 95% CI for untransformed Somers’ D
Somers_D Minimum Maximum
studytime .75 .46336709 .89456394

We see that, if we sample a subject at random from the same age tertile in each of the two treatment groups
(treated and placebo), then it is 46% to 89% more likely that the treated subject survives the untreated subject
than vice versa.
Note that Harrell’s c and the Wilcoxon–Breslow–Gehan test are based on Somers D parameters “in different
directions”, in that the survival variable is the first variable in the list when calculating Harrell’s c and the second
variable in the list when calculating the Wilcoxon–Breslow–Gehan Somers’ D. The two methods are complementary.
The Harrell method is more useful for the “purely scientific” inference of establishing that a predictor has some
 Post-publication update 17

predictive power that is not entirely caused by a confounder. The Wilcoxon–Breslow–Gehan method is more useful
for estimating the size of the “causal effect” on survivorship of choosing one treatment.instead of another.

3.4 Gini coefficients in the womenwage data

The Gini coefficient of inequality is used by econometricians as a measure of the inequality of a distribution
of incomes (or wealth) in a population, on a scale from zero (when everybody has an equal share) to one (when
one person has everything). It is traditionally understood by reference to the Lorenz curve, which is the set of
(X, Y )–points on the unit square such that the richest 100Y percent of the population have 100X percent of the
income (or wealth). The Gini coefficient is equal to the difference between the area above the Lorenz curve and the
area below the Lorenz curve. Note that Gini also invented several other coefficients, which are also referred to in
various contexts as “the Gini coefficient”, including Gini’s mean difference, defined in Chapter 5 of Serfling (1980).
These are summarized in Goodman and Kruskal (1959).
The Gini coefficient of inequality is a special case of Somers’ D. To see this, imagine that two lotteries are
organized in a population, and that, in the first lottery, each member of the population has one ticket, whereas, in the
second lottery, each individual buys a number of tickets proportional to that individual’s income. The first lottery
is equivalent to sampling uniformly from the Y –axis of the Lorenz plot, whereas the second lottery is equivalent
to sampling uniformly from the X–axis of the Lorenz plot. The region above the Lorenz curve corresponds to the
event that the second lottery winner is a higher earner than the first, whereas the region below the Lorenz curve
corresponds to the event that the first lottery winner is a higher earner than the second. Therefore, the Gini coefficient
is a clustered Somers’ D, where the clusters are individuals in the population, the observations are combinations of
individual and lottery (first or second), the Y –variate is income, the X–variate is lottery sequence (1 or 2), and the
importance weights are equal for all individuals in the first lottery and proportional to income for all individuals in
the second lottery. The Lorenz curve is similar to the ROC curve (see [R] roc, Hanley and McNeil (1982), or Newson
(2002)), except that the ROC curve demonstrates the outcome of sampling equiprobably from two sub–populations,
whereas the Lorenz curve demonstrates the outcome of sampling twice from the same population, with different
probability weights.
We can illustrate this principle using the womenwage dataset, distributed by Stata Press and used in [R] intreg.
We first calculate the Gini index (and a few other inequality indices), using the program ineqdeco (Jenkins, 1999).
We then preserve the data, and use the expgen package (an extended version of expand downloadable from SSC) to
replace each observation in the original dataset (containing one observation per woman) with 2 observations (one
per woman per lottery). The new dataset is indexed by the variables womanid (denoting sequence number of the
woman) and lotseq (denoting sequence number of the lottery). We create an importance variable pwt, containing
probability weights equal for all women in the first lottery and equal to a woman’s wage (to the nearest kilodollar)
in the second lottery. We then use somersd twice, first without any tranformation, and second with the normalizing
and/or variance–stabilizing z–transformation, before restoring the old dataset:
. use http://www.stata-press.com/data/r9/womenwage, clear
(Wages of women)
. ineqdeco wage
Percentile ratios for distribution of wage: all valid obs.
------------------------------------------------------------
p90/p10 p90/p50 p10/p50 p75/p25 p75/p50 p25/p50
------------------------------------------------------------
3.222 1.933 0.600 1.909 1.400 0.733
Generalized Entropy indices GE(a), where a = income difference
sensitivity parameter, and Gini coefficient
----------------------------------------------------------------------
All obs | GE(-1) GE(0) GE(1) GE(2) Gini
----------+-----------------------------------------------------------
| 0.14595 0.12947 0.13383 0.16022 0.27984
----------------------------------------------------------------------
Atkinson indices, A(e), where e > 0 is the inequality aversion parameter
----------------------------------------------
All obs | A(0.5) A(1) A(2)
----------+-----------------------------------
| 0.06358 0.12144 0.22594
----------------------------------------------
. preserve
. expgen =2, oldseq(womanid) copyseq(lotseq)
. lab var lotseq "Lottery sequence number"
. gene pwt = (lotseq==1) + wage*(lotseq==2)
. lab var pwt "Probability weight"
. somersd lotseq wage [pwei=pwt], cluster(womanid) funtype(vonmises)
18 Post-publication update RBN-1

Von Mises Somers’ D with variable: lotseq

Transformation: Untransformed
Valid observations: 976
Number of clusters: 488
Symmetric 95% CI
(Std. Err. adjusted for 488 clusters in womanid)
------------------------------------------------------------------------------
| Jackknife
lotseq | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
wage | .2798363 .0105714 26.47 0.000 .2591168 .3005558
------------------------------------------------------------------------------
. somersd lotseq wage [pwei=pwt], cluster(womanid) funtype(vonmises) tr(z)
Von Mises Somers’ D with variable: lotseq
Transformation: Fisher’s z
Valid observations: 976
Number of clusters: 488
Symmetric 95% CI for transformed Somers’ D
(Std. Err. adjusted for 488 clusters in womanid)
------------------------------------------------------------------------------
| Jackknife
lotseq | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
wage | .2875044 .0114695 25.07 0.000 .2650246 .3099843
------------------------------------------------------------------------------
Asymmetric 95% CI for untransformed Somers’ D
Somers_D Minimum Maximum
wage .27983629 .25898919 .30042278
. restore

We see that, if the women in this dataset organized two lotteries amongst themselves, and each woman bought
one ticket in the first lottery and a number of tickets worth a constant fraction of her wages in the second lottery,
then the second lottery winner would be 27.98% more likely than the first lottery winner to be the higher earner of
the two. (The option funtype(vonmises) is necessary because of the small but nonzero probability that the same
woman wins both lotteries.) And, if the same lotteries were organized in the population from which these women were
sampled, then the difference would be from 25.91% to 30.06% (according to the untransformed confidence interval),
or 25.90% to 30.04% (according to the z–transformed confidence interval). In general, we expect normalizing and/or
variance–stabilizing transformations to be more important in populations where the population Gini coefficient is
higher, because then the sampling distribution of the sample Gini coefficient will be more skewed. This principle is
discussed in Daniels and Kendall (1947) and Edwardes (1995).
Income distribution indices are often calculated on datasets in which each observation represents an income
group, rather than an individual. We can create such a dataset using contract on the original womanwage data
to produce a new dataset, with 1 observation per wage group. If we do this, then we can repeat our analysis to
demonstrate the cfweight() option of somersd. We first use expgen to expand the dataset with 1 observation per
wage group to a dataset with 1 observation per wage group per lottery, and create the weight variable as before.
This new dataset contains a cluster of 2 variables per wage group, and each of these clusters represents a number
of duplicate clusters equal to the value of the variable freq created by contract. When calling somersd, we use
freq as a cfweight() variable, and produce the same results as before:
. preserve
. contract wage
. expgen =2, oldseq(wagegp) copyseq(lotseq)
. lab var wagegp "Wage group"
. lab var lotseq "Lottery sequence number"
. gene pwt = (lotseq==1) + wage*(lotseq==2)
. lab var pwt "Probability weight"
. describe
Contains data from http://www.stata-press.com/data/r9/womenwage.dta
obs: 102 Wages of women
vars: 5 3 Mar 2005 18:14
size: 1,530 (99.9% of memory free)
-------------------------------------------------------------------------------
storage display value
variable name type format label variable label
-------------------------------------------------------------------------------
wage float %9.0g wages in 1000s of dollars
_freq byte %12.0g Frequency
wagegp byte %8.0g Wage group
 Post-publication update 19

lotseq byte %8.0g Lottery sequence number

pwt float %9.0g Probability weight
-------------------------------------------------------------------------------
Sorted by: wagegp lotseq
Note: dataset has changed since last saved
. somersd lotseq wage [pwei=pwt], cluster(wagegp) cfweight(_freq) funtype(vonmises)
Von Mises Somers’ D with variable: lotseq
Transformation: Untransformed
Valid observations: 976
Number of clusters: 488
Symmetric 95% CI
(Std. Err. adjusted for 488 clusters in wagegp)
------------------------------------------------------------------------------
| Jackknife
lotseq | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
wage | .2798363 .0105714 26.47 0.000 .2591168 .3005558
------------------------------------------------------------------------------
. somersd lotseq wage [pwei=pwt], cluster(wagegp) cfweight(_freq) funtype(vonmises) tr(z)
Von Mises Somers’ D with variable: lotseq
Transformation: Fisher’s z
Valid observations: 976
Number of clusters: 488
Symmetric 95% CI for transformed Somers’ D
(Std. Err. adjusted for 488 clusters in wagegp)
------------------------------------------------------------------------------
| Jackknife
lotseq | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
wage | .2875044 .0114695 25.07 0.000 .2650246 .3099843
------------------------------------------------------------------------------
Asymmetric 95% CI for untransformed Somers’ D
Somers_D Minimum Maximum
wage .27983629 .25898919 .30042278
. restore

4 Historical note
This document is a post–publication update of an article which appeared in the Stata Technical Bulletin (STB)
as Newson (2000a). The somersd package was later revised in Newson (2000b), Newson (2000c), Newson (2000d),
Newson (2001a) and Newson (2001b). An important upgrade (Newson, 2000d) was the addition to the somersd
package of the program cendif, which calculates robust confidence intervals for Hodges–Lehmann median differences,
other percentile differences, and percentile ratios. A post–publication update of that STB article is distributed with
this document as part of the documentation to the somersd package. After 2001, STB was replaced by The Stata
Journal (SJ), and all subsequent updates only appeared on SSC and on Roger Newson’s homepage at http://www.kcl-
phs.org.uk/rogernewson, which was accessible from within net–aware Stata as of 01 February 2006. However, Newson
(2002) gives a comprehensive review of Somers’ D, Kendall’s τa , median differences, and their estimation in Stata
using the somersd package.

5 Acknowledgements
I would like to thank William Gould of Stata Corporation for suggesting the predict program somers p, and
Nicholas J. Cox of Durham University, UK, for some very helpful discussions on Somers’ D and Kendall’s τa .

6 References
Arvesen, J. N. 1969. Jackknifing U–statistics. Annals of Mathematical Statistics 40: 2076–2100.
Breslow, N. E. 1970. A generalized Kruskal–Wallis test for comparing k samples subject to unequal patterns of censorship. Biometrika 57:
579–594.
Daniels, H. E. and Kendall, M. G. 1947. The Significance of Rank Correlation Where Parental Correlation Exists. Biometrika 34: 197–208.
Edwardes, M. D. deB. 1995. A Confidence Interval for Pr(X <Y) - Pr(X >Y) Estimated From Simple Cluster Samples. Biometrics 51:
571–578.
Fisher, R. A. 1921. On the “Probable Error” of a Coefficient of Correlation deduced from a Small Sample. Metron 1(4): 3–32.
Gayen, A. K. 1951. The Frequency Distribution of the Product–Moment Correlation Coefficient in Random Samples of Any Size Drawn from
Non–Normal Universes. Biometrika 38: 219–247.
Gehan, E. A. 1965. A generalized Wilcoxon test for comparing arbitrarily single–censored samples. Biometrika 52; 203–223.
20 Post-publication update RBN-1
View publication stats

Goodman, L. A. and W. H. Kruskal. 1959. Measures of association for cross classifications. II: Further discussion and references. Journal of
the American Statistical Association 54(285): 123–163.
Hanley, J. A. and B. J. McNeil. 1982. The meaning and use of the area under a receiver operating characteristic (ROC) curve. Radiology
143: 29–36.
Harrell, F. E., R. M. Califf, D. B. Pryor, K. L. Lee and R. A. Rosati. 1982. Evaluating the yield of medical tests. Journal of the American
Medical Association 247(18): 2543–2546.
Harrell, F. E., K. L. Lee and D. B. Mark. 1996. Multivariate prognostic models: issues in developing models, evaluating assumptions and
adequacy, and measuring and reducing errors. Statistics in Medicine 15: 361–387.
Hoeffding, W. 1948. A class of statistics with asymptotically normal distribution. Annals of Mathematical Statistics 19: 293–325.
Imai K. and D. A. van Dyk. Causal inference with general treatment regimes: generalizing the propensity score. 2004. Journal of the
American Statistical Association 99(467): 854–866.
Jenkins, S. P. 1999. Analysis of income distributions. Stata Technical Bulletin 48: 4–18. Reprinted in Stata Technical Bulletin Reprints,
vol. 8, pp. 243–260.
Kendall, M. G. 1949. Rank and Product–Moment Correlation. Biometrika 36: 177–193.
Kendall, M. G. and J. D. Gibbons. 1990. Rank Correlation Methods. th edition. New York: Oxford University Press.
Newson, R. 2000a. snp15: somersd – Confidence intervals for nonparametric statistics and their differences. Stata Technical Bulletin 55:
47–55. Reprinted in Stata Technical Bulletin Reprints, vol. 10, pp. 312–322.
Newson, R. 2000b. snp15.1: Update to somersd. Stata Technical Bulletin 57: 35. Reprinted in Stata Technical Bulletin Reprints, vol.
10, pp. 322–323.
Newson, R. 2000c. snp15.2: Update to somersd. Stata Technical Bulletin 58: 30. Reprinted in Stata Technical Bulletin Reprints, vol.
10, p. 323.
Newson, R. 2000d. snp16: Robust confidence intervals for median and other percentile differences between groups. Stata Technical Bulletin
58: 30–35. Reprinted in Stata Technical Bulletin Reprints, vol. 10, pp. 324–331.
Newson, R. 2001a. snp15.3: Update to somersd. Stata Technical Bulletin 61: 22. Reprinted in Stata Technical Bulletin Reprints, vol.
10, p. 33x.
Newson, R. 2001b. snp16.1: Update to cendif. Stata Technical Bulletin 61: 22. Reprinted in Stata Technical Bulletin Reprints, vol. 10,
p. 33x.
Newson, R. 2002. Parameters behind “nonparametric” statistics: Kendall’s tau, Somers’ D and median differences. The Stata Journal 2: 45–64.
A pre–publication draft can be downloaded from Roger Newson’s website at http://www.kcl-phs.org.uk/rogernewson/ as
of 01 February 2006, using the net command in Stata.
Newson, R. 2006. Efficient calculation of jackknife confidence intervals for rank statistics. Journal of Statistical Software 15(1): 1–10.
Downloadable from http://www.jstatsoft.org/ as of 01 February 2006.
Rosenbaum P. R. and D. B. Rubin. 1983. The central role of the propensity score in observational studies for causal effects. Biometrika
70(1): 41–55.
Serfling, R. J. 1980. Approximation theorems of mathematical statistics. New York: John Wiley & Sons.
Somers, R. H. 1962. A New Asymmetric Measure of Association for Ordinal Variables. American Sociological Review 27: 799–811.