STATISTICS IN MEDICINE

Statist. Med. 2009; 28:541–557

Published online 8 December 2008 in Wiley InterScience
(www.interscience.wiley.com) DOI: 10.1002/sim.3491

Sample size calculations for evaluating mediation

E. Vittinghoff∗, † , Ś. Sen and C. E. McCulloch

Department of Epidemiology and Biostatistics, University of California San Francisco,
San Francisco, CA, U.S.A.

SUMMARY
This paper presents sample size calculations for testing the mediation of the effect of a primary predictor
by an intermediate variable. This problem is related to validating surrogate markers and to testing the
effect of a primary predictor in the presence of confounders. For those problems, proposals for sample
size calculation exist in the literature and can be adapted to the problem of mediation. Methods based
on the variance inflation factor in linear regression provide exact sample size calculations for the linear
model and approximations for the logistic, Poisson, and Cox models. We propose another procedure based
on simulation of the underlying data structure, with applications to the logistic and Cox models. For
the Poisson model, a new analytic method is also proposed. The behavior of the different proposals is
investigated by means of simulation studies. Copyright q 2008 John Wiley & Sons, Ltd.

KEY WORDS: sample size; mediation; Poisson model; logistic model; Cox model

1. INTRODUCTION

This paper presents sample size calculations for testing the mediation of the effect of a primary
predictor by an intermediate variable, a common issue in epidemiologic and clinical research.
For example, adkipocytokine levels may mediate the relationship between visceral adiposity and
diabetes in older adults [1]; similarly, physician perceptions of patient behaviors and characteristics
may mediate the association between patient race/ethnicity and treatment recommendations for
coronary artery disease [2]. This problem is related to validating surrogate markers and to testing
the effect of a primary predictor in the presence of confounders. For those problems, proposals
for sample size calculation exist in the literature and can be adapted to the problem of mediation.
Methods based on the variance inflation factor in linear regression provide exact sample size
calculations for the linear model and approximations for the logistic, Poisson, and Cox models. We
propose another procedure based on simulation of the underlying data structure, with applications

∗ Correspondence to: Eric Vittinghoff, 185 Berry Street, Suite 5700, San Francisco, CA 94107, U.S.A.
†
E-mail: [email protected]

Contract/grant sponsor: San Francisco VA Medical Center

Received 24 February 2008

Copyright q 2008 John Wiley & Sons, Ltd. Accepted 2 October 2008
542 E. VITTINGHOFF, Ś. SEN AND C. E. MCCULLOCH

to the logistic and Cox models. For the Poisson model, a new analytic method is also proposed.
The behavior of the different proposals is investigated by means of simulation studies.
In Section 2, we review the background of our proposed methods. In Sections 3–6 we develop
sample size methods for the linear, logistic, Poisson, and Cox models. Section 7 describes a
simulation study validating our methods, and Section 8 briefly discusses sample size calculations
for testing special null hypotheses. We conclude with a discussion in Section 9; the analytic method
for the Poisson model is derived in the Appendix.

2. BACKGROUND

Mediation is commonly assessed by the relative change in the strength of association of the primary
predictor with the outcome after adjustment for the mediator. In terms of regression coefficients,
∗ ∗ ∗
this is (ˆ 1 − ˆ 1 )/ˆ 1 , where ˆ 1 and ˆ 1 are the estimates for the primary predictor X 1 before and
after adjustment for the mediator X 2 . To assess mediation, we would estimate 1 by fitting a full
model including both predictors, and ∗1 by fitting a reduced model omitting X 2 ; both the full and
reduced models might also include a common set of confounding variables.
∗ ∗
In the context of clinical trials, (ˆ 1 − ˆ 1 )/ˆ 1 is called the proportion of treatment effect explained,
or PTE. In developing methods for validating surrogate endpoints, Freedman et al. [3] showed how
to assess PTE using the logistic model, while Lin et al. [4] developed methods for the Cox model.
Prentice set out criteria for evaluating surrogate endpoints in clinical trials, specifying that for a
true surrogate, PTE = 1 [5]. In this context, inference has focused on testing the null hypothesis
that PTE is less than some fraction h = 0.5 or 0.75. This can be done by constructing a confidence
∗
interval for PTE using the -method or Fieller’s theorem and estimates of Var(ˆ 1 ), Var(ˆ 1 ), and
∗
Cov(ˆ 1 , ˆ 1 ) [3, 4].
∗ ∗
Two related methods have been proposed for estimating Var(ˆ 1 ), Var(ˆ 1 ), and Cov(ˆ 1 , ˆ 1 ). In
GLMs with the canonical link function, including the logistic model, the asymptotic covariance
matrix of the parameter vector estimate b̂ is (X VX)−1 , where X is the predictor matrix and V
is a diagonal matrix with Vii = Var(Yi ) [6]. In constructing a confidence interval for PTE in the
logistic model, Freedman et al. [3] proposed using model-based estimates of the variance of ˆ 1
∗ ∗
and ˆ 1 from the full and reduced models, respectively, and obtaining Cov(ˆ 1 , ) ˆ as the appropriate
 −1   −1
element of (X1 V̂1 X1 ) (X1 V̂X)(X V̂X) , where X and X1 are the predictor matrices from the
full and reduced models, V̂ii = p̂i (1− p̂i ) is based on the fitted probabilities from the full model,
and V̂1 is defined analogously using the fitted probabilities from the reduced model.
Lin et al. [4] extended Freedman’s approach to the Cox model, using robust estimators for
∗ ∗
Var(ˆ 1 ) and Var(ˆ 1 ) as well as for Cov(ˆ 1 , ). ˆ They also showed how software capable of fitting
∗
the Cox model with robust standard errors [7] could be used to estimate Cov(ˆ 1 , ˆ 1 ). Specifically,
two modified copies of the original data set are concatenated, with resulting predictor matrix
 
X1 0 0
(1)
0 X1 X 2
with the predictor matrix X1 from the reduced model potentially including the common set of
confounding variables as well as the primary predictor X 1 . The Cox model is run with stratification

Copyright q 2008 John Wiley & Sons, Ltd. Statist. Med. 2009; 28:541–557
DOI: 10.1002/sim
 SAMPLE SIZE CALCULATIONS FOR EVALUATING MEDIATION 543

of the baseline hazard on the first and second records for each observation, and robust standard
errors accounting for the link between the two. The coefficient estimate corresponding to the first
∗
copy of X 1 is ˆ 1 , while the estimate corresponding to the second copy is ˆ 1 . The robust estimate of
∗
Cov(ˆ 1 , ˆ 1 ) correctly reflects the fact that these parameters are estimated using the same data. This
method is easily extended to the logistic and other GLMs, as well as to the linear model. For these
models, independence working correlation must be specified and a column of ones corresponding
to the intercept included in both copies of X1 .
In contrast to surrogate markers for clinical trials, with an ideal PTE of 1, in the epidemiologic
context the question is more often whether the mediator explains any of the effect of the primary
predictor. Thus, the focus can shift to testing PTE = 0, or equivalently ∗1 = 1 . In the next section
we show that if the correlation, , between the primary predictor and mediator is non-zero, then
testing ∗1 = 1 is equivalent to testing for an independent effect of the mediator X 2 in the full model
including both predictors. As a result, we can apply existing approximate methods for the logistic
[8] and Cox [9–11] models. Based on a Wald test of 2 = 0, these methods use 1/(1−2 ), the
variance inflation factor from linear regression, to inflate sample size estimates to account for
the loss of precision due to adjustment for X 1 . We show in Section 3 that this adjustment is exact for
the linear model, but simulation studies reported in Section 7 show that it performs less accurately
for the other models. For the logistic and Cox models, these problems can be conveniently overcome
using Monte Carlo integration to approximate the expected information matrix. This provides a
better estimate of Var(ˆ 2 ), leading to a more accurate sample size calculation for the Wald test
of 2 = 0. For the Poisson model we derive analytic expressions for the information matrix under
four common data configurations.

3. SAMPLE SIZE FOR THE LINEAR MODEL

Consider the simplest case of the linear model with a single primary predictor X 1 and a single
mediator X 2 , so that E(Y ) = 0 +1 X 1 +2 X 2 . Then provided the full model is correctly specified,
∗
E(ˆ 1 ) = 1 +2 2 /1 , where  is the correlation of X 1 and X 2 and 21 and 22 are their variances.
If   = 0, testing the null hypothesis PTE = 0 is equivalent to testing ∗1 −1 = 2 2 /1 = 0, or
equivalently 2 = 0, a completely standard problem. As a result, we can use

(z  + z  )2 2e
n= (2)
(2 2 )2 (1−2 )
to estimate the required sample size, accounting for the correlation between X 1 and X 2 ; in (2), z 
and z  are the quantiles of the standard normal distribution corresponding to the specified one-sided
type-I and type-II error rates, and 2e is the residual variance of the outcome.
Formula (2) can be derived using the fact that under standard assumptions, Cov(b̂) = (X X)−1 2e .
Assuming that X 1 and X 2 are centered to have mean 0,
⎛ ⎞
1 0 0
⎜ ⎟
E(X X) = n ⎜
⎝ 0  2
1  1  2
⎟
⎠ (3)
0 1 2 22

Copyright q 2008 John Wiley & Sons, Ltd. Statist. Med. 2009; 28:541–557
DOI: 10.1002/sim
544 E. VITTINGHOFF, Ś. SEN AND C. E. MCCULLOCH

This also holds exactly when X 1 , X 2 , or both are centered indicators rather than continuous.
Taking the inverse of (3), we obtain Var(ˆ 2 ) = 2e [n22 (1−2 )]−1 , and thus (2) to calculate sample
size for the Wald test of 2 = 0. For binary X 2 , 22 = f 2 (1− f 2 ), with f 2 = Pr(X 2 = 1) before
centering. If confounders need to be included in both the full and reduced models, this could be
accommodated by redefining  in (2) as the multiple correlation of X 2 with the confounders as well
as X 1 .
Formula (2) can also be specified in terms of PTE, ∗1 , and 1 . Using PTE = 2 2 /∗1 1 , we
obtain:

(z  + z  )2 2 2e
n= (4)
(∗1 1 PTE)2 (1−2 )

(z  + z  )2 2e
= (5)
(2 2 )2 −(∗1 1 PTE)2

(z  + z  )2 2e
= (6)
(2 2 )2 −[(∗1 −1 )1 ]2

If either of the predictors is binary, its variance is 2j = f j (1− f j ), j = 1, 2, where f j is its marginal
prevalence. With continuous predictors, considerable simplification results if we assume e = 1 =
2 = 1, posing the problem in terms of standardized effects: for example, 2 would in that case
denote the increase, in standard deviations, in E(Y ) for each standard deviation increase in X 2 .
Formulas (2) and (4)–(6) give users a choice depending on which two or three parameters can be
most easily or reliably specified. For example, (4) would be useful when estimates of the overall
independent effect of X 1 and its correlation with X 2 are available, and a minimum substantively
meaningful value for PTE can be specified.
As an example, suppose that both X 1 and X 2 are continuous with 21 = 22 = 2e = 1.0, 1 = 0.5,
2 = 0.1, and  = 0.3. These parameters are consistent with ∗1 = 0.53 and PTE ≈ 6 per cent. Then to
provide 80 per cent power ( = 0.2) in a one-sided test with type-I error of 2.5 per cent ( = 0.025),
we obtain n = 863 using (2) or (4)–(6). In simulation studies of attained power reported in Section 7,
this sample size provided an estimated 82 per cent power to reject 2 = 0 and 81 per cent power
to reject ∗1 = 1 .

4. SAMPLE SIZE FOR THE LOGISTIC MODEL

For the logistic model, the expression analogous to (2) is

(z  + z  )2
n= (7)
(2 2 )2 (1−2 ) p(1− p)

where p is the marginal prevalence of the outcome, and 2 , 2 , and  are defined as before [8].
This can be derived using the approximation V ≈ p(1− p)I, so that (X VX) ≈ p(1− p)(X X). Since
E(X X) has the same form as (3), Var(ˆ 2 ) ≈ [n22 (1−2 ) p(1− p)]−1 .

Copyright q 2008 John Wiley & Sons, Ltd. Statist. Med. 2009; 28:541–557
DOI: 10.1002/sim
 SAMPLE SIZE CALCULATIONS FOR EVALUATING MEDIATION 545

This approximation might suffice in cases where covariate effects are relatively weak, so that
p does not vary much across observations. Then, under the further assumption that the fitted
values, and thus the linear predictor, are approximately the same in the full and reduced models,
∗1 ≈ 1 +2 2 /1 , so that formulas analogous to (4)–(6), substituting 1/[ p(1− p)] for 2e , would
also hold approximately.
However, simulation studies of attained power reported in Section 7 show that this approach
is not always accurate. For the logistic model, E(X VX)/n, the expected contribution of a single
observation to X VX is analytically intractable except in the case where both the primary predictor
and mediator are discrete. Numeric integration is possible for the other cases, but a more easily
implemented alternative is to use Monte Carlo integration to estimate E(X VX)/n. Specifically,
we simulate n s = 10 000 observations under the assumed joint distribution of X 1 and X 2 , calculate
pi (1− pi ), i = 1, . . . , n s , under the assumed logistic model, and finally compute n s (X VX)−1 . This
procedure does require a simple sub-routine to compute the value of the intercept parameter 0
consistent with 1 , 2 , and the marginal outcome prevalence, p. Then

(z  + z  )2 ˜ 2
n= 2
(8)
22

where ˜ 2 is the diagonal element of n s (X VX)−1 corresponding to 2 . This approach has the
2
additional attraction that if pilot data including the primary predictor and mediator are available,
then it would be straightforward to use the existing data rather than simulation to approximate X;
relevant confounders could also be included in this procedure.
As an example, suppose that p = 0.5, X 1 is binary with f 1 = 0.5, 1 = log 2, X 2 is continuous
with 22 = 1, 2 = log 1.5, and  = 0.5, values that are consistent with ∗1 = log 3 and PTE ≈ 37
per cent on the linear predictor scale. Using (7) we obtain n = 255, which in our simulations
provided an estimated 74 per cent power to reject 2 = 0 and 76 per cent power to reject ∗1 = 1 .
Using Monte Carlo integration and (8), we obtain n = 294, which provided 81 and 84 per cent
power, respectively.

5. SAMPLE SIZE FOR THE POISSON MODEL

For the Poisson model, E(X VX) can be derived analytically. We show in the Appendix that if X 1
and X 2 are bivariate Normal, or if X 2 is conditionally Normal within levels of binary X 1 , then the
analytic sample size estimator is

(z  + z  )2
n= (9)
(2 2 )2 (1−2 )E(Y )
where E(Y ) is the marginal mean of the outcome, 2 is the log rate ratio per unit increase in
X 2 , and 2 and  are defined as before. The sample size resulting from (9) should be inflated in
proportion to any expected over-dispersion, or Var(Y )/E(Y ).
Note that (9), derived analytically for the two configurations with a continuous mediator, has the
same form as (7), with p(1− p) replaced by E(Y ) = Var(Y ). It is also the approximate estimator
for configurations with a binary mediator, derived like (7) using the approximation V ≈ E(Y )I. In
contrast, the analytic solutions are more complicated for those configurations.

Copyright q 2008 John Wiley & Sons, Ltd. Statist. Med. 2009; 28:541–557
DOI: 10.1002/sim
546 E. VITTINGHOFF, Ś. SEN AND C. E. MCCULLOCH

Consider first the configuration with binary mediator X 2 with prevalence f 2 = Pr(X 2 = 1), contin-
uous primary predictor X 1 with mean zero and variance 21 , and correlation  between X 1 and X 2 .
We again assume that within levels of X 2 , X 1 is Normal with conditional variance 21.2 = 21 (1−2 ),
√ √
and means 0 = −1 f 2 /(1− f 2 ) when X 2 = 0 and 1 = 1 (1− f 2 )/ f 2 when X 2 = 1. In this
case, we obtain
(z  + z  )2 [(B + D)2 21.2 +BD(1 −0 )2 ]
n= (10)
22 B D21.2 E(Y )
where B = f 2 exp(1 1 +2 ), D = (1− f 2 ) exp(1 0 ), and 1 and 2 are the log rate ratios for X 1
and X 2 ; (10) is also derived in the Appendix from E(X VX)/n for this model.
In the configuration where both the primary predictor X 1 and the mediator X 2 are binary,
with prevalences √ f 1 and f 2 and correlation , the expected proportion with both exposures is
f 11 = f 1 f 2 + f 1 (1− f 1 ) f 2 (1− f 2 ), with values of  subject to the constraint that 0< f 11 < f 1 ∧ f 2 .
Then f 10 , the proportion with X 1 = 1 and X 2 = 0 is f 1 − f 11 ; similarly, f 01 = f 2 − f 11 , and f 00 =
1− f 01 − f 10 − f 11 . Finally, we obtain
(z  + z  )2 (B +C + D + E)(B + D)(C + E)
n= (11)
22 (BC D + BC E + B D E +C D E)E(Y )
where B = f 00 , C = f 10 exp(1 ), D = f 01 exp(2 ), and E = f 11 exp(1 +2 ); (11) is derived in the
Appendix from E(X VX)/n for this model. Again, the sample size should be inflated by any
expected over-dispersion.
Monte Carlo integration can also be used for the Poisson model. In brief, we would simulate
n s = 10 000 observations under the assumed joint distribution of X 1 and X 2 , calculate Var(Y ) =
E(Y ) for each observation under the assumed Poisson model, and finally compute n s (X VX)−1 .
Expressions for the intercept parameter 0 are given in the Appendix. Then

(z  + z  )2 ˜ 2
n= 2
(12)
22

where ˜ 2 is the diagonal element of n s (X VX)−1 corresponding to 2 . This method gives sample
2
size estimates that vary from those given by (9)–(11) only by a small margin of simulation error.
As an example, suppose that E(Y ) = 0.5, X 1 is continuous with 21 = 1, 1 = log 1.4, X 2 is binary
with f 2 = 0.25, 2 = log 1.35, and  = 0.5, values that are consistent with ∗1 ≈ log 1.5 and PTE
≈ 16 per cent on the linear predictor scale. Using (9) we obtain n = 1239, which in our simulations
provided an estimated 87 per cent power to reject 2 = 0 and 86 per cent power to reject ∗1 = 1 .
Using Monte Carlo integration and (10), we obtain n = 1037, which provided 79 and 78 per cent
power, respectively.

6. SAMPLE SIZE FOR THE COX MODEL

For the Cox model, the approximation analogous to (7) is

(z  + z  )2
n= (13)
(2 2 )2 (1−2 )

Copyright q 2008 John Wiley & Sons, Ltd. Statist. Med. 2009; 28:541–557
DOI: 10.1002/sim
 SAMPLE SIZE CALCULATIONS FOR EVALUATING MEDIATION 547

where is the probability that an observation is uncensored, so that the number of events d = n
[9–11]. When X 2 is binary with prevalence f 2 , 22 = f 2 (1− f 2 ).
As with the logistic model, simulation studies reported in Section 7 suggest some room for
improvement, in this case for the configurations with binary mediators. This problem may result
from our focus on mediators with relatively powerful effects; in contrast, (13) is generally motivated
under the approximation of no effect [10, 12]. Schmoor et al. [10] provide expressions for the
information matrix for the configuration with binary primary predictor and mediator, which could
be extended to the other configurations and then solved using numeric integration.
As an alternative, we again propose Monte Carlo integration. Our implementation simulates
n s = 10 000 exponential failure times under the assumed joint distribution of the X 1 and X 2 and the
specified proportional hazards model, censors all but the shortest fraction of the failure times,
a form of type-II independent censoring [13], and fits a Cox model to the simulated data. This
gives

(z  + z  )2 ˜ 2
n= 2
(14)
22

ˆ ˆ 2 )—that is, n s times the estimated variance of ˆ 2 from the Cox model fit to
where ˜ 2 = n s Var(
2
the simulated data.
As an example, suppose that = 0.2, X 1 is binary with f 1 = 0.5, 1 = log 2, X 2 is binary with
f 2 = 0.25, 2 = log 1.5, and  = 0.3, values that are consistent with ∗1 ≈ log 2.26 and PTE ≈ 15
per cent on the linear predictor scale. Using (13) we obtain n = 1399 and d = 280, which in our
simulations provided an estimated 88 per cent power to reject both 2 = 0 and ∗1 = 1 . Using
Monte Carlo integration and (14), we obtain n = 1123 and d = 225, which provided 81 per cent
power for both tests.
Our implementation assumes what is sometimes called administrative censoring, but does not
incorporate earlier censoring, for example, from loss to follow-up or extended accrual. This simpli-
fication is reasonable because the number of events d = n is critical for power in the Cox model,
with little sensitivity to n and when d is held fixed [12]. In Section 7 we comment briefly on the
results of two additional simulation studies, the first allowing for early drop-out and the second
holding n rather than d fixed.

7. SIMULATION STUDIES OF SAMPLE SIZE ESTIMATES

We used simulations to assess the performance of our sample size estimators and to compare them
with the simpler approximate methods. In general, we computed sample sizes supposed to provide
80 per cent power in tests with a one-sided type-I error rate of 2.5 per cent, using (2) and (7)–(14),
for a range of the relevant parameters. For binary predictors, we considered prevalence of 0.25
and 0.5, while for continuous X 1 and X 2 , we set 1 = 2 = 1;  was variously set at 0.3, 0.5, and
for the linear model 0.7. For the linear model 2e = 1, for the logistic model p = 0.2 or 0.5, for the
Poisson model, E(Y ) = 0.5 or 1.0, and for the Cox model, = 0.2 or 0.5, allowing us to examine
the influence of n and d. We used values of 1 large enough so that there would at least 80 per
cent power to detect an effect of X 1 in the reduced model with the sample size selected to detect
the independent effect of X 2 , since PTE can only be well-estimated in this case [3, 4]. Finally,

Copyright q 2008 John Wiley & Sons, Ltd. Statist. Med. 2009; 28:541–557
DOI: 10.1002/sim
548 E. VITTINGHOFF, Ś. SEN AND C. E. MCCULLOCH

values of 2 were chosen so that the resulting sample sizes ranged from less than 100 to more
than 1000.
For each model and set of parameters, we simulated 1000 data sets of the specified size, with
both predictors as well as the outcome randomly generated under the assumed model in each data
set. We first ran the Wald test of 2 = 0 in the full model. Then, to validate our claim that a test of
2 = 0 is equivalent to a test of ∗1 = 1 , we used the approach of Lin et al. [4], constructing the
concatenated data set specified by (1), fitting the model with robust standard errors, then running
the Wald test of ∗1 = 1 . For both tests, power was estimated by the proportion of data sets in
which the null hypothesis was rejected.
In Figures 1–4, we summarize the results by plotting the estimated power of both tests against
sample size or the number of events. For the logistic, Poisson, and Cox models, the four panels in
the left column of the figure use the sample size approximations based on the variance inflation
factor 1/(1−2 ), while those on the right are based on simulation of the information matrix for
the logistic and Cox models or on our analytic results for the Poisson model. In each panel, the
symbol ‘1’ marks the estimated power of the test of 2 = 0, while ‘2’ marks the power of the test
of ∗1 = 1 . The two dashed horizontal lines give approximate 95 per cent point-wise error margins
for the simulation results with 1,000 repetitions if the true value were 80 per cent. Tabular results
corresponding to the figures are available at http://www.biostat.ucsf.edu/mediation.html.
Figure 1 shows that (2) performs very well for the test of 2 = 0 in the linear model with all
four configurations of primary predictor and mediator. However, power to reject ∗1 = 1 is lower
than specified when the sample size estimate is small.
Figure 2 shows that the approximation (7) performs less satisfactorily for the logistic model
with continuous mediator and small sample sizes (two upper left panels). However, use of (8),

100 100
Simulated power (%)

Simulated power (%)

90 90

80 1
2 2
1 80 1
212
21 1
111 1 1
2 11121121 1
2 2 2
1
2 12
12
70 2 2 70 2 2

60 60
continuous primary predictor, continuous mediator binary primary predictor, continuous mediator
50 2 50

0 500 1000 1500 0 500 1000 1500

Estimated sample size Estimated sample size

100 100
Simulated power (%)

Simulated power (%)

90 90
1 1 12 1 2 1
80 11 2 1 2
1 2 1 2 1 80 121 1 11 1
2 2 1 1
2
2 1 1 2
1 2 1
2 2 2 2
1 2
2 2
70 2 70
2
2
60 60
2
continuous primary predictor, binary mediator binary primary predictor, binary mediator
50 50

0 500 1000 1500 0 500 1000 1500

Estimated sample size Estimated sample size

Figure 1. Estimated power for linear model, using (2): (1) test of 2 = 0 and (2) test of ∗1 = 1 .

Copyright q 2008 John Wiley & Sons, Ltd. Statist. Med. 2009; 28:541–557
DOI: 10.1002/sim
 SAMPLE SIZE CALCULATIONS FOR EVALUATING MEDIATION 549

Approximation Simulation
100 100
Simulated power (%)

Simulated power (%)

90 90
2 21 2
11 2
1 12 1 2 2 1
2
80 2 2 1 12 2 2 80 1 1
2 1 1 2
1
2 2
1 1 2
1 1
1
21 1 1
11
70 70
2 2
60 60
continuous primary predictor, continuous mediator 2
continuous primary predictor, continuous mediator
50 2 50

0 500 1000 1500 0 500 1000 1500

Estimated sample size Estimated sample size

100 100
Simulated power (%)

Simulated power (%)

90 90
2
11 211 2 1 2
1 1
2 2
1
80 1 2 2 80 1 21 2 2 2
1
2 2 2 1
1 2
1 2
1 1 2
1
1
21 1 1
70 11 70 2
2
60 2 60 2
binary primary predictor, continuous mediator binary primary predictor, continuous mediator
50 50

0 500 1000 1500 0 500 1000 1500

Estimated sample size Estimated sample size

100 100
Simulated power (%)

Simulated power (%)

90 90
2
11
2 1 1 21
80 2 1
1 2
2 2
80 1 111 2
1 2
1
2 1
11 2
2 2
1 1
2
11
2 22 1
2 22
1 2
1 1 1
2 2
1
11121 1 22
70 70 2
2 2
60 60
2continuous primary predictor, binary mediator
2 continuous primary predictor, binary mediator
50 50

0 500 1000 1500 0 500 1000 1500

Estimated sample size Estimated sample size

100 100
Simulated power (%)

Simulated power (%)

90 90
11 1
2
2 1 12
2 1 112
2 11 2
1 12
211 2
1
80
11
2122 2
2
2 1
1
21
12 2
1
80 1 12 2
1 2
1 2
1
2
1
11 12 22
70 70
2
60 2 60
2
binary primary predictor, binary mediator binary primary predictor, binary mediator
50 50

0 500 1000 1500 0 500 1000 1500

Estimated sample size Estimated sample size

Figure 2. Estimated power for logistic model, using (7)–(8): (1) test of 2 = 0 and (2) test of ∗1 = 1 .

Copyright q 2008 John Wiley & Sons, Ltd. Statist. Med. 2009; 28:541–557
DOI: 10.1002/sim
550 E. VITTINGHOFF, Ś. SEN AND C. E. MCCULLOCH

Approximation Analytic
100 100
Simulated power (%)

Simulated power (%)

90 90

80 1 1
2 80 1 1 1
11 111
21 11
2 2 1
2 1
2 1 111
21 21
2
1
2 1
2 1
2
2 2 2 2 2 2
70 70

60 22 60 22
continuous primary predictor, continuous mediator continuous primary predictor, continuous mediator
50 50

0 500 1000 1500 0 500 1000 1500

Estimated sample size Estimated sample size

100 100
Simulated power (%)

Simulated power (%)

90 90

80 1 1 11 1
2 1
2 1
2 80 1 2
1 1
2
1 1112
12
2 2 22 2
11 1111
2 2 22
1 1 1
2
2
70 2 70 2
2 2
60 60

50 2 binary primary predictor, continuous mediator 50 2

binary primary predictor, continuous mediator

0 500 1000 1500 0 500 1000 1500

Estimated sample size Estimated sample size

100 100
Simulated power (%)

Simulated power (%)

90 90
1 1 1 1
1 1 1
2 1 2
2 2 2 11 112 2
80 1 2 2
1 2 2 80 12 1 1 12 1
2 1 1
22 2 1
1 12 1 2 21 12 1 2
1 11
12 2 2 2
70 70
2
60 60
2 2 2
continuous primary predictor, binary mediator continuous primary predictor, binary mediator
50 50

0 500 1000 1500 0 500 1000 1500

Estimated sample size Estimated sample size

100 100
Simulated power (%)

Simulated power (%)

90 90
11 1
12 1 1 1
2 1 2 2
2 2 2 1 2 2 2
80
2 1 80 111 121 1 21 1 1
1 21
1 1
2
1
2 2
1 2 1 2 2
12 2 2
1 12 1 1 2 2
70 1 2 70
2
60 60 2
2
binary primary predictor, binary mediator binary primary predictor, binary mediator
50 2 50

0 500 1000 1500 0 500 1000 1500

Estimated sample size Estimated sample size

Figure 3. Estimated power for Poisson model, using (9)–(12): (1) test of 2 = 0 and (2) test of ∗1 = 1 .

Copyright q 2008 John Wiley & Sons, Ltd. Statist. Med. 2009; 28:541–557
DOI: 10.1002/sim
 SAMPLE SIZE CALCULATIONS FOR EVALUATING MEDIATION 551

Approximation Simulation

100 100
Simulated power (%)

Simulated power (%)

90 90
1
2 1 2 1
80 22
2
1 1
2
2
1
1 1
2
1 2
1 2
2
1 80 2 21
1
1
1
21
2 1
2
1 22
1 21
12
1 1 2
70 70

60 60
continuous primary predictor, continuous mediator continuous primary predictor, continuous mediator
50 50

0 50 100 150 200 250 300 0 50 100 150 200 250 300

Estimated number of events Estimated number of events

100 100
Simulated power (%)

Simulated power (%)

90 90

12
2 2
1 2 1
1 2 12
1 1
80 12
2 1 2
1
2
1
2
1
2 2
1 1
2 80 2
2 1
1 21
1 2 21
2 21
12
1
70 70

60 60
binary primary predictor, continuous mediator binary primary predictor, continuous mediator
50 50

0 50 100 150 200 250 300 0 50 100 150 200 250 300

Estimated number of events Estimated number of events

100 100
Simulated power (%)

Simulated power (%)

90 2
1 90
1
2 2
1 2
1
2 1
21 22 22 1
1
80 2
1 1
2 1
2
2
80 21
12 1
1 21 1 1
2 2 21
1 21
2 1
2
1
1
2
70 70

60 60
continuous primary predictor, binary mediator continuous primary predictor, binary mediator
50 50

0 50 100 150 200 250 300 0 50 100 150 200 250 300

Estimated number of events Estimated number of events

100 100
Simulated power (%)

Simulated power (%)

90 1
2
2
1 2
1 90
1
2
1 1
2 1
1
221 1
2 2
80 2
1 1
2 1
2 80 21
21 21 21
1 212 21
1 21
212
1 1
2 2
1 21
1
1
2 2
1
70 70

60 60
binary primary predictor, binary mediator binary primary predictor, binary mediator
50 50

0 50 100 150 200 250 300 0 50 100 150 200 250 300

Estimated number of events Estimated number of events

Figure 4. Estimated power for Cox model, using (13)–(14): (1) test of 2 = 0 and (2) test of ∗1 = 1 .

Copyright q 2008 John Wiley & Sons, Ltd. Statist. Med. 2009; 28:541–557
DOI: 10.1002/sim
552 E. VITTINGHOFF, Ś. SEN AND C. E. MCCULLOCH

which depends on Monte Carlo integration to approximate E(X VX)/n, corrects these deficiencies
(two upper right panels), although slight conservatism is in evidence. Both methods perform well
for binary mediators (four lower panels). As with the linear model, smaller estimated sample sizes
provide less than the specified power to reject ∗1 = 1 .
Figure 3 shows results for the Poisson model. In general, there is some under-estimation of the
needed sample size for tests of ∗1 = 1 when the estimate is small, in particular when n E(Y ), the
total expected outcome count was less than 100. Equation (9) performs well for the configurations
with a continuous mediator (four upper panels), for which it is both the approximate and analytic
sample size estimator; results on the left and right represent two runs of the same program, as
a gauge of sampling variability. However (9) does less well for binary mediators (two lower left
panels), for which it is only approximate. The prevalence of the mediator was most influential,
with conservatism when f 2 = 0.25, and underestimation of sample size when f 2 = 0.5. The two
lower right panels show that (10) and (11) achieve modest improvement for the configurations
with binary mediators.
Figure 4 shows results for the Cox model, plotted against the number of events d = n . Paired
results for = 0.2 and 0.5 are almost indistinguishable, confirming that the influence of sample size
is negligible after accounting for the numbers of events. For both configurations with a continuous
mediator, the approximation (13) and the simulation method using (14) are equally good (four
upper panels). However, results using (13) are noisier and in several cases slightly conservative
in the two configurations with a binary mediator (two lower left panels); this bias reflected the
prevalence of the mediator as with the Poisson model. Using (14), based on simulation of the
information matrix, resolves these difficulties (two lower right panels). Furthermore, in contrast
to the linear, logistic, and Poisson models, tests of ∗1 = 1 appear to have roughly the specified
power even with relatively small numbers of events.
In a sensitivity analysis allowing for censoring of 20 per cent of observations by ‘drop-out’
in generation of the test data as well as in the sample size calculation using an extension of the
simulation approach, essentially no changes in performance were noted for either the approximation
(13) or the simulation approach. Performance was also unaffected in a second sensitivity analysis
where we fixed the sample size and the expected number of events at the target levels specified by
(13) or (14), but let the actual number of events vary across the 1000 simulated data sets. Results of
both sensitivity analyses are plotted and tabulated at http://www.biostat.ucsf.edu/mediation.html.

8. TESTING NON-ZERO NULL HYPOTHESES

Freedman and Schatzkin [14] provide an approximate method for sample size calculation for
testing the null hypothesis that PTE is less than some fraction h = 0.5 or 0.75 under the alternative
hypothesis that PTE = 1, as relevant for assessing the adequacy of a surrogate endpoint for clinical
trials. In contrast, our methods are primarily applicable for h = 0, the more relevant value for
studying mediation in epidemiologic studies. This holds because 2 = 0 implies PTE = 0 in this
case; otherwise, PTE depends on 1 or ∗1 as well as 2 and . Our methods could be used to
determine the sample size required to test the null hypothesis that the absolute (rather than the
relative) difference between ∗1 and 1 is less than some non-zero amount k. The null value of 2
corresponding to k = ∗1 −1 is b = sign(2 )|k1 /2 |. Then we would substitute (2 −b) for 2
in (2), (7), and (8)–(14).

Copyright q 2008 John Wiley & Sons, Ltd. Statist. Med. 2009; 28:541–557
DOI: 10.1002/sim
 SAMPLE SIZE CALCULATIONS FOR EVALUATING MEDIATION 553

Furthermore, our simulation approach could be extended to testing PTE<h for values of h>0.
Specifically, the simulation algorithm would include the following steps:

1. Simulate n s observations from the assumed distribution of X 1 , X 2 , and the outcome Y .

2. Concatenate two copies of the resulting data set with predictor matrix in the form of (1),
then fit the assumed linear, logistic, Poisson, or Cox model with robust standard errors, and
independence working correlation for the first three models.
∗ ∗
3. Obtain estimates of Var(ˆ 1 ), Var(ˆ 1 ), and Cov(ˆ 1 , ˆ 1 ) from the model fit to the simulated
∗ ∗
data, and compute the -method variance of PTE = (ˆ 1 − ˆ 1 )/ˆ 1 .
4. Set

(z  + z  )2 ˜ 2pte
n= (15)
(PTE−h)2

where ˜ 2pte is n s times the -method variance for PTE.

9. DISCUSSION

In this paper we propose methods for estimating sample sizes for testing the mediation of a primary
predictor, based on the insight that when the primary predictor and mediator are correlated, showing
the inequality of the effects of the primary predictor before and after adjustment for the mediator
is equivalent to showing an independent effect of the mediator. Moreover, this equivalence makes
our methods applicable to sample size problems focusing on the independent effect of a primary
predictor in the presence of confounding. We provide new analytic results for Poisson models,
and our simulation-based approach improves on the performance of current approximations for the
logistic and Cox models. Our methods can be inverted to compute power or minimum detectable
effects for a given sample size.
Many approaches have been proposed for computing power and sample size for generalized linear
models with multiple predictors. In contrast to methods based on score [15, 16] and likelihood-ratio
tests [16–18], our proposed methods, like those of several other authors [11, 19–24], are based on
Wald tests, in part because these tests are more commonly used in practice, although they may be
less reliable in small samples and at alternatives far from the null [25].
Our proposed methods are built on approximation of the expected information matrix from the
full model. Analytic results for the information matrix have previously been obtained for a limited
number of cases, including Demidenko’s result for the logistic model when both predictors are
binary [24]. Our analytic results extend Signorini’s solution for the Poisson model with multivariate
Normal predictors [21], which in turn takes advantage of the fact that moment-generating func-
tions, used earlier by Whittemore to approximate the expected information matrix for the logistic
model with small response probability [19], are exact for the Poisson model. Similarly, Bernardo
et al., derive the information matrix for the exponential survival model with a binary predictor and
arbitrary covariates, resulting in an expression equivalent to (13), and show using simulations that
it works well for the Cox model in a limited number of cases [11].
We also follow others [23, 24] in estimating the information matrix only under the alternative,
in contrast to some other work in which it is also evaluated under the null hypothesis [20, 22].

Copyright q 2008 John Wiley & Sons, Ltd. Statist. Med. 2009; 28:541–557
DOI: 10.1002/sim
554 E. VITTINGHOFF, Ś. SEN AND C. E. MCCULLOCH

The rationale is that the information matrix evaluated at the maximum likelihood estimates of the
parameters is used in Wald tests as implemented in most statistical packages.
Computation of the information matrix using numeric integration was recently advanced for
the logistic and Cox models by Schoenfeld and Borenstein [23], who have implemented their
complex results in commercial software. The simulation approach we propose is simpler, but
not without limitations [16, 26]. Among them is the potential difficulty of generating data from
multivariate distributions other than the relatively simple cases we consider, although re-sampling
from suitable data might work. However, analogous difficulties arise with analytic approaches or
numeric integration. Furthermore, the simulation approach is flexible, and could be extended, for
example, to the more difficult problem of sample size calculations for the Cox model involving
mediation by a surrogate event, modeled as a time-dependent binary covariate.
The simulation estimates of attained power reported in Section 7 generally support the validity
of our proposed methods, but do show that relatively small estimated sample sizes may provide
slightly less than the specified power for the linear, logistic, and Poisson models. Arising mainly
with tests of ∗1 = 1 , this problem is also seen in a milder form at larger sample sizes estimated
using the analytic formulas (10) and (11) for the Poisson model with a binary mediator. In both
cases, simulations of power could be used to guide inflation of the initial sample size estimate.
For the logistic, Poisson, and Cox models, the standard errors are asymptotic, and their plug-in
estimators are likely noisy in small samples. Similar considerations may explain the shortfalls in
tests of ∗1 = 1 in the linear, logistic, and Poisson models. Error may also arise from the fact
that the full and reduced models cannot both hold exactly. Although the reduced model may be a
good approximation if the effect of the primary predictor is small [4], we focused on cases where
the effect of the primary predictor is large enough to be easily detectable, which we see as more
appropriate for studies of mediation.
An additional limitation is that our methods are not generally useful for sample size estimation
for testing non-zero null hypotheses about PTE, as are relevant in assessing the adequacy of
surrogate endpoints in clinical trials. Although our methods can be modified for testing non-zero
null hypotheses about the absolute rather than relative difference between ∗1 and 1 , this quantity
is more difficult to interpret than PTE unless the absolute value of ∗1 or 1 is well established.
However, we do show how simulation could be used to address the more difficult case.
Finally, we have made the assumption throughout that   = 0. Certainly, the evidence for mediation
is stronger if we can reject  = 0 as well as ∗1 = 1 or 2 = 0. It would generally be easy using
standard methods to determine whether the sample size computed using the methods proposed in
this paper would also provide adequate power to reject  = 0 under the assumptions made about
the joint distribution of the primary predictor and mediator, and to increase the sample size as
necessary if this does not hold.
In this paper, we have proposed easy-to-use methods for calculating sample sizes for detecting
mediation in linear, logistic, Poisson, and Cox models. Our methods are also applicable to
detecting the independent effect of a primary predictor in the presence of substantial confounding.
They have been validated using simulation and are implemented in R [27] code available at
http://www.biostat.ucsf.edu/mediation.html. In conclusion, analysts needing to estimate sample
size for evaluating mediation or to account for loss of power due to covariate adjustment should
use formulas (2) or (4)–(6) for the linear model, and (9)–(11) for the Poisson model. For the
logistic and Cox models, we recommend use of our easily implemented simulation methods, which
afford worthwhile improvements over existing approximations based on the variance inflation
factor.

Copyright q 2008 John Wiley & Sons, Ltd. Statist. Med. 2009; 28:541–557
DOI: 10.1002/sim
 SAMPLE SIZE CALCULATIONS FOR EVALUATING MEDIATION 555

APPENDIX A

For the Poisson model with bivariate Normal predictors X 1 and X 2 , both assumed to have mean
zero, variances 21 and 22 , and correlation ,
⎡ ⎤
1 B C
⎢ ⎥
E(X VX)/n = exp(A) ⎣ B B 2 +21 BC +1 2 ⎦ (A1)
C BC +1 2  C 2 +22

where A = 0 +(21 21 +22 22 )/2+1 2 1 2 , B = 1 21 +2 1 2 , and finally C = 1 1 2 +
2 22 . The intercept parameter is given by

0 = log E(Y )−(21 21 +22 22 )/2−1 2 1 2 

To derive (9) for this model, we obtained the lower right element of the inverse of (A1) as
−1
(S22 − S21 S11 S12 )−1 , where the {Si j } are the corresponding partitioned sub-matrices of (A1).
Equation (9) follows directly.
For the Poisson model with binary primary predictor X 1 with prevalence f 1 = Pr(X 1 = 1), and
mediator X 2 conditionally Normal within levels of X 1 ,
⎡ ⎤
B+D B BC + D E
⎢ ⎥
E(X VX)/n = exp(A) ⎣ B B BC ⎦ (A2)
BC + D E BC B F + DG

where A = 0 +22 22.1 /2, B = f 1 exp(1 +2 1 ), C = 1 +2 22.1 , D = (1− f 1 ) exp(2 0 ), E = 0 +
2 22.1 , F = 22.1 +C 2 , and G = 22.1 + E 2 . The intercept parameter is given by

0 = log E(Y )−22 22.1 /2−log( f 1 exp(1 +2 1 )+(1− f 1 ) exp(2 0 ))

Obtaining the lower right element of the inverse of (A2) as before using the formula for the inverse
of a partitioned matrix, (9) follows after simplification for this case also.
For the corresponding Poisson model with binary mediator X 2 with prevalence f 2 = Pr(X 2 = 1),
and primary predictor X 1 conditionally Normal within levels of X 2 ,
⎡ ⎤
B+D BC + D E B
⎢ ⎥
E(X VX)/n = exp(A) ⎣ BC + D E B F + DG BC ⎦ (A3)
B BC B

In this case A = 0 +21 21.2 /2, B = f 2 exp(1 1 +2 ), C = 1 +1 21.2 , D = (1− f 2 ) exp(1 0 ), E =
0 +1 21.2 , F = 21.2 +C 2 , and G = 21.2 + E 2 . The intercept parameter is given by

0 = log E(Y )−21 21.2 /2−log( f 2 exp(1 1 +2 )+(1− f 2 ) exp(1 0 ))

Obtaining the lower right element of the inverse of (A3) as before, (10) follows after simplification.

Copyright q 2008 John Wiley & Sons, Ltd. Statist. Med. 2009; 28:541–557
DOI: 10.1002/sim
556 E. VITTINGHOFF, Ś. SEN AND C. E. MCCULLOCH

For the Poisson model with binary predictors X 1 and X 2 ,

⎡ ⎤
B +C + D + E C + E D+E
⎢ ⎥
E(X VX)/n = exp(A) ⎣ C+E C+E E ⎦ (A4)
D+E E D+E
where A = 0 , B = f 00 , C = f 10 exp(1 ), D = f 01 exp(2 ), and E = f 11 exp(1 +2 ). The intercept
parameter, in this case the log rate in the baseline group with X 1 = X 2 = 0, is given by
0 = log E(Y )−log(B +C + D + E)
After obtaining the lower right element of the inverse of (A4) as before, (11) follows after some
simplification.

ACKNOWLEDGEMENTS
We gratefully acknowledge the substantial contributions of both referees. Ś. Sen was partially supported
by the Health Services Research Enhancement Award Program (REAP) of the San Francisco VA Medical
Center.

REFERENCES
1. Kanaya AM, Harris T, Goodpaster BH, Tylavsky F, Cummings SR. Adipocytokines attenuate the association
between visceral adiposity and diabetes in older adults. Diabetes Care 2004; 27:1375–1380.
2. van Ryn M, Burgess D, Malat J, Griffin J. Physicians’ perceptions of patients’ social and behavioral characteristics
and race disparities in treatment recommendations for men with coronary artery disease. American Journal of
Public Health 2006; 96:351–357. DOI: 10.2105/AJPH.2004.041806.
3. Freedman LS, Graubard BI, Schatzkin A. Statistical validation of intermediate endpoints for chronic diseases.
Statistics in Medicine 1992; 11:167–178. DOI: 10.1002/sim.478.
4. Lin DY, Fleming TR, De Gruttola V. Estimating the proportion of treatment effect explained by a surrogate
marker. Statistics in Medicine 1997; 16:1515–1527. DOI: 10.1002/(SICI)1097-0258(19970715)16:13<1515::AID-
SIM572>3.0.CO;2-1.
5. Prentice RL. Surrogate endpoints in clinical trials: definition and operational criteria. Statistics in Medicine 1989;
8:431–440. DOI: 10.1002/sim.4780080407.
6. McCulloch CE, Searle SR. Generalized, Linear, and Mixed Models. Wiley: New York, 2001.
7. Lin DY, Wei LJ. The robust inference for the Cox proportional hazards model. Journal of the American Statistical
Association 1989; 84:1074–1078.
8. Hsieh FY, Bloch DA, Larsen MD. A simple method of sample size calculation for linear and logistic regression.
Statistics in Medicine 1998; 17:1623–1634. DOI: 10.1002/(SICI)1097-0258(19980730)17:14<1623::AID-
SIM871>3.0.CO;2-S.
9. Hsieh FY, Lavori PW. Sample-size calculations for the Cox proportional hazards regression model with nonbinary
covariates. Controlled Clinical Trials 2000; 21:552–560. DOI: 10.1016/S0197-2456(00)00104-5.
10. Schmoor C, Sauerbrei W, Schumacher M. Sample size considerations for the evaluation of prognostic
factors in survival analysis. Statistics in Medicine 2000; 19:441–452. DOI: 10.1002/(SICI)1097-
0258(20000229)19:4<441::AID-SIM349>3.0.CO;2-N.
11. Bernardo MVP, Lipsitz SR, Harrington DP, Catalano PJ. Sample size calculations for failure time random
variables in non-randomized studies. Journal of the Royal Statistical Society, Series D, The Statistician 2000;
49:31–40. DOI: 10.1111/1467-9884.00217.
12. Schoenfeld DA. Sample-size formula for the proportional-hazards regression model. Biometrics 1983; 39:499–503.
13. Kalbfleisch JD, Prentice RL. The Statistical Analysis of Failure Time Data. Wiley: New York, 1980; 41.
14. Freedman LS, Schatzkin A. Sample size for studying intermediate endpoints within intervention trials or
observational studies. American Journal of Epidemiology 1992; 136:1148–1159.

Copyright q 2008 John Wiley & Sons, Ltd. Statist. Med. 2009; 28:541–557
DOI: 10.1002/sim
 SAMPLE SIZE CALCULATIONS FOR EVALUATING MEDIATION 557

15. Self SG, Mauritsen RH. Power/sample size calculations for generalized linear models. Biometrics 1988; 44:79–86.
16. Lyles RH, Lin H-M, Williamson JM. A practical approach to computing power for generalized linear models
with nominal, count, or ordinal responses. Statistics in Medicine 2007; 26:1632–1648. DOI: 10.1002/sim.2617.
17. Self SG, Mauritsen RH, Ohara J. Power calculations for likelihood ratio tests in generalized linear models.
Biometrics 1992; 48:31–39.
18. Shieh G. On power and sample size calculations for likelihood ratio tests in generalized linear models. Biometrics
2000; 56:1192–1196. DOI: 10.1111/j.0006-341X.2000.01192.x.
19. Whittemore AM. Sample size for logistic regression with small response probability. Journal of the American
Statistical Association 1981; 76:27–32.
20. Wilson SR, Gordon I. Calculating sample sizes in the presence of confounding variables. Applied Statistics 1986;
35:207–213.
21. Signorini DF. Sample size for Poisson regression. Biometrika 1991; 78:446–450. DOI: 10.1093/biomet/78.2.446.
22. Shieh G. On power and sample size calculations for Wald tests in generalized linear models. Journal of Statistical
Planning and Inference 2005; 128:43–59. DOI: 10.1016/j.jspi.2003.09.017.
23. Schoenfeld DA, Borenstein M. Calculating the power or sample size for the logistic and proportional
hazards models. Journal of Statistical Computation and Simulation 2005; 75:771–785. DOI: 10.1080/
00949650410001729445.
24. Demidenko E. Sample size determination for logistic regression revisited. Statistics in Medicine 2007; 26:
3385–3397. DOI: 10.1002/sim.2771.
25. Hauck WW, Donner A. Wald’s test as applied to hypotheses in logit analyses. Journal of the American Statistical
Association 1977; 72:851–853.
26. Glueck DH, Muller KE. Adjusting power for a baseline covariate in linear models. Statistics in Medicine 2003;
22:2535–2551. DOI: 10.1002/sim.1341.
27. R Development Core Team. R: a language and environment for Statistical Computing. R Foundation for Statistical
Computing, Vienna, Austria, 2008. ISBN 3-900051-07-0. http://www.R-project.org.

Copyright q 2008 John Wiley & Sons, Ltd. Statist. Med. 2009; 28:541–557
DOI: 10.1002/sim