Page 1
STATISTICS IN MEDICINE
Statist. Med. 17, 1623—1634 (1998)
A SIMPLE METHOD OF SAMPLE SIZE CALCULATIONFOR
LINEAR AND LOGISTIC REGRESSION
F. Y. HSIEH?*, DANIEL A. BLOCH? AND MICHAEL D. LARSEN?
?CSPCC, Department of Veterans Affairs, Palo Alto Health Care System (151-K), Palo Alto, California 94304, U.S.A.
?Division of Biostatistics, Department of Health Research and Policy, Stanford University, Stanford, California 94305, U.S.A.
?Department of Statistics, Stanford University, Stanford, California 94305, U.S.A.
SUMMARY
A sample size calculation for logistic regression involves complicated formulae. This paper suggests use of
sample size formulae for comparing means or for comparing proportions in order to calculate the required
samplesize for a simplelogisticregression model.One can thenadjust the requiredsample sizefor a multiple
logisticregression model by a variance inflation factor. This method requires no assumption of low response
probability in the logistic model as in a previous publication. One can similarly calculate the sample size for
linear regression models. This paper also compares the accuracy of some existing sample-size software for
logistic regression with computer power simulations. An example illustrates the methods. ? 1998 John
Wiley & Sons, Ltd.
INTRODUCTION
In a multiple logistic regression analysis, one frequently wishes to test the effect of a specific
covariate, possibly in the presence of other covariates, on the binary response variable. Owing to
the nature of non-linearity, the sample size calculation for logistic regression is complicated.
Whittemore? proposed a formula, derived from the information matrix, for small response
probabilities. Hsieh? simplified and extended the formula for general situations by using the
upper bound of the formula. Appendix I presents a simple closed form, based on an information
matrix, to approximate the sample size for both continuous and binary covariates in a simple
logistic regression. In a different approach, Self and Mauritsen? used generalized linear models
and the score tests to estimate the sample size through an iterative procedure. These published
methods are complicated and may not be more accurate than the conventional sample size
formulae for comparing two means or a test of equality of proportions. In the next section, we
present a simple formula for the approximate sizes of the sample required for simple logistic
regression by using formulae for calculating sample size for comparing two means or for
* Correspondence to: F. Y. Hsieh, CSPCC, Department of Veterans Affairs, Palo Alto Health Care System (151-K),
Palo Alto, California 94304, U.S.A.
Contract/grant sponsor: Department of Veterans Affairs Cooperative Studies Program
Contract/grant sponsor: NIH
Contract/grant number: AR20610
Contract/grant sponsor: National Institute on Drug Abuse
Contract/grant number: Y01-DA-40032-0
CCC 0277—6715/98/141623—12$17.50
? 1998 John Wiley & Sons, Ltd.
Received February 1997
Revised October 1997
Page 2
comparing two proportions. We can then adjust the sample size requirement for a multiple
logistic regression by a variance inflation factor. This approach applies to multiple linear
regression as well.
SIMPLE LOGISTIC REGRESSION
Ina simple logisticregressionmodel,we relateacovariate X?to the binaryresponse variable½ in
a model log(P/(1!P))"??#??X?where P"prob(½"1). We are interested in testing the
null hypothesis H?:??"0 against the alternativeH?:??"?*, where ?*O0, that the covariate is
related to the binary response variable. The slope coefficient ??is the change in log odds for an
increase of one unit? in X?. When the covariate is a continuous variable with a normal
distribution, the log odds value ??is zero if and only if the group means, assuming equal
variances, between the two response categories are the same. Therefore we may use a sample size
formula for a two-sample t-test to calculate the required sample size. For simplicity, we use
a normal approximation instead, as the sample size formula (see formula (7) in Appendix I) may
be easily changed to include t-tests if required:
n"(Z?????#Z???)?/[P1(1!P1)?*?]
where n is the required total sample size, ?* is the effect size to be tested, P1 is the event rate at the
mean of X, and Z?is the upper uth percentile of the standard normal distribution.
When the covariate is a binary variable, say X"0 or 1, the log odds value ??"0 if and only if
the two event rates are equal. The sample size formula for the total sample size required for
comparing two independent event rates has the following form (see formula (10)):
n"?Z?????[P(1!P)/B]???#Z???[P1(1!P1)#P2(1!P2)(1!B)/B]?????
/[(P1!P2)?(1!B)]
where: P("(1!B)P1#BP2) is the overall event rate; B is the proportion of the sample with
X"1; P1 and P2 are the event rates at X"0 and X"1, respectively. For B"0)5, the required
sample size is bounded by the following simple form (see formula (11)):
n(4P(1!P)(Z?????#Z???)?/(P1!P2)?.
Appendix I presents two simpler forms, formulae (12) and (13), than formula (2). A later section
presents the comparisons of these formulae with computer power simulations.
(1)
(2)
(3)
MULTIPLE LOGISTIC REGRESSION
When there is more than one covariate in the model, a hypothesis of interest is the effect of
a specific covariate in the presence of other covariates. In terms of log odds parameters, the null
hypothesis for multiple logistic regression is H?: [??,??,2,??]"[0,??,2,??] against the
alternative [?*,??,2,??]. Let b?be the maximum likelihood estimate of ??. Whittemore? has
shown that, for continuous, normal covariates X, the variance of b?in the multivariate setting
with p covariates, var?(b?), can be approximated by inflating the variance of b?obtained from the
one parameter model, var?(b?), by multiplying by 1/(1!??????2p) where ?1.232pis the multiple
correlation coefficient relating X?with X?,2,X?. That is, approximately
var?(b?)"var?(b?)/(1!??1.232p)
1624
F. HSIEH, D. BLOCH AND M. LARSEN
Statist. Med. 17, 1623—1634 (1998)
? 1998 John Wiley & Sons, Ltd.
Page 3
The squared multiple correlation coefficient ??1.232p, also known as R?, is equal to the proportion
of the variance of X?explained by the regression relationship with X?,2,X?. The term
1/(1!??1.232p) will be referred to as a variance inflation factor (VIF). The required sample size for
the multivariate case can also be approximated from the univariate case by inflating it with the
samefactor 1/(1!??1.232p).Followingthe
n?"n?/(1!??1.232p) where n?and n?are the sample sizes required for a logistic regression model
with p and 1 covariates, respectively. The same VIF seems to work well for binary covariates (see
Appendix III).
relationshipof thevariances, wehave
MULTIPLE LINEAR REGRESSION
For multiple linear regression models, we can easily derive the same VIF for p covariates (see
Appendix II). Therefore, we can adjust similarly the sample size for a regression model with
p covariates. It is known that in a simple linear regression model, the correlation coefficient ? and
the regression parameter ??have the relationship ?"????/??. Hence ?"0 if and only if ??"0.
When both X and ½ are standardized, testing the hypotheses that ?"0 and that ??"0 are
equivalent and the required sample sizes are the same.
Let r be the estimate of the correlation coefficient between X and ½. The sample size formula
(see Sokal and Rohlf?) for testing H?: ?"0 against the alternative H?: ?"r is
n?"(Z?????#Z???)?/C(r)?#3
where the Fisher’s transformation C(r)"??log((1#r)/(1!r)). If we add p!1 covariates to the
regression model, the required sample size for testing H?: [??,??,2,??]"[0,??,2,??]
againstthe alternative[?*,??,2,??] is n?"n?/(1!??1.232p), approximately.If we already have
q covariates in the model and would like to expand the model to p('q) covariates, then, from
Appendix II, n?"n?((var?(b?)/var?(b?))"n?/(1!??(1q#12p))(232q)) where the partial correlation
coefficient ?(1q#12p))(232q) measures the linear association between covariates X?
X???,2,X?when the values of covariates X?,2,X?are held fixed.
and
COMPARISON OF SAMPLE-SIZE SOFTWARE
There are at least two computer programs available that use formula (4) (see AppendixI): nQuery
from Dr. Janet Elashoff,? and SSIZE? from the first author. One program, EGRET SIZ from
SERC,?uses the approachofSelf and Mauritsen.?For logisticregression,thecomputerprograms
nQuery and SSIZE provide sample sizes only for continuous covariates while EGRET SIZ only
provides estimates for discrete covariates. Both nQuery and EGRET SIZ are commercial
software. Note that the sample size calculation for logistic regression is only one of the many
features provided by the above three computer programs.
Table I presents sample size examples for a binary covariate using formula (4) and software
EGRET SIZ as well as the corresponding sample size for comparing two proportions (without
continuitycorrection from formulae (2), (3), (12) and (13)), and the results of power simulations. In
the table, P1 and P2 are event rates at X"0 and X"1, respectively; B is the proportion
of the sample with X"1; OR is the odds ratio of X"1 versus X"0 such that
OR"P2(1!P1)/(P1(1!P2)); P"(1!B)P1#BP2 is the overall event rate or case fraction.
Table I is designed to show the relationshipof sample sizes for different study designs. It is known
that a balanced design (B"0)5) requires less sample size than an unbalanced design
SAMPLE SIZE CALCULATION
1625
Statist. Med. 17, 1623—1634 (1998)
? 1998 John Wiley & Sons, Ltd.
Page 4
Table I. Results of sample size calculations for a binary covariate from six different methods,
power"95 per cent, two-sided significance level 5 per cent
DesignSample size Power simulation
Balanced design with high event rates
(4):
P1"0)4, P2"0)5, B"0)5
(2):
P"0)45, P1"0)4, P2"0)5, B"0)5
(3):
P"0)45, P1"0)4, P2"0)5, B"0)5
(12)
P"0)45, P1"0)4, P2"0)5, B"0)5
(13)
P1"0)4, P2"0)5, B"0)5
SIZ:OR"1)5, case fraction P"0)45,
sampling fraction 50/50
1367
1282
1287
1287
1274
1285
96)0$0)63%
95)4$0)66%
94)7$0)71%
94)7$0)71%
94)6$0)71%
95)9$0)63%
Balanced design with low odds ratio
(4):
P1"0)5, P2"0)2, B"0)5
(2):
P"0)35, P1"0)5, P2"0)2, B"0)5
(3):
P"0)35, P1"0)5, P2"0)2, B"0)5
(12):
P"0)35, P1"0)5, P2"0)2, B"0)5
(13):
P1"0)5, P2"0)2, B"0)5
SIZ: OR"0)25, case fraction P"0)35,
sampling fraction 50/50
141
126
131
131
119
129
96)3$0)60%
95)0$0)69%
96)6$0)57%
96)6$0)57%
94)9$0)70%
96)1$0)61%
Balanced design with high odds ratio
(4):
P1"0)2, P2"0)5, B"0)5
(2):
P"0)35, P1"0)2, P2"0)5, B"0)5
(3):
P"0)35, P1"0)2, P2"0)5, B"0)5
(12):
P"0)35, P1"0)2, P2"0)5, B"0)5
(13):
P1"0)2, P2"0)5, B"0)5
SIZ:OR"4)0, case fraction P"0)35,
sampling fraction 50/50
166
126
131
131
119
129
99)0$0)31%
95)0$0)69%
96)6$0)57%
96)6$0)57%
92)9$0)81%
95)4$0)66%
Balanced design with high odds ratio
(4):
P1"0)05, P2"0)1, B"0)5
(2):
P"0)075, P1"0)05, P2"0)1, B"0)5
(3):
P"0)075, P1"0)05, P2"0)1, B"0)5
(12):
P"0)075, P1"0)05, P2"0)1, B"0)5
(13):
P1"0)05, P2"0)1, B"0)5
SIZ: OR"2)111, case fraction P"0)075,
sampling fraction 50/50
¸ow prevalence rate
(4):
P1"0)05, P2"0)1, B"0)2
(2):
P"0)06, P1"0)05, P2"0)1, B"0)2
(12):
P"0)06, P1"0)05, P2"0)1, B"0)2
(13):
P1"0)05, P2"0)1, B"0)2
SIZ: OR"2)111, case fraction P"0)06,
sampling fraction 80/20
1818
1437
1443
1443
1430
1417
98)2$0)42%
94)4$0)73%
95)8$0)63%
95)8$0)63%
94)4$0)73%
94)5$0)72%
2612
2186
1833
2648
2070
97)4$0)50%
94)9$0)70%
91)2$0)90%
97)4$0)50%
94)6$0)71%
High prevalence rate
(4):
P1"0)05, P2"0)1, B"0)8
(2):
P"0)09, P1"0)05, P2"0)1, B"0)8
(12):
P"0)09, P1"0)05, P2"0)1, B"0)8
(13):
P1"0)05, P2"0)1, B"0)8
SIZ: OR"2)111, case fraction P"0)09,
sampling fraction 20/80
3060
2257
2661
1820
2347
98)3$0)41%
95)0$0)69%
97)8$0)46%
89)5$0)97%
97)2$0)52%
1626
F. HSIEH, D. BLOCH AND M. LARSEN
Statist. Med. 17, 1623—1634 (1998)
? 1998 John Wiley & Sons, Ltd.
Page 5
(B"0)2 or 0)8); a low prevalence rate (B"0)2) requires less sample size than a high prevalence
rate (B"0)8); sample size remains the same if the odds ratio is reversed. In addition to the
significance level and the power of the test, the values of the following parameters, listed after the
sample size methods, are specified in the table:
Formula (4): P1, P2 and B.
Formula (2): tests of proportions: P, P1, P2 and B.
Formula (3): simple form for a balanced design: P, P1 and P2.
Formula (12): simple form for an unbalanced design: P, P1, P2 and B.
Formula (13): simple form for an unbalanced design: P1, P2 and B.
SIZ: OR, sampling fractions 1!B and B, and overall case fraction P.
The power simulations, obtained from SIZ with 1000 replications, use the likelihood ratio test for
the logistic regression model. The simulations show that the sample sizes obtained from testing
two proportions (formulae (2) and (3)) have statistical power within one standard deviation of the
expected power of 95 per cent. Also, formulae (2) and (3) are more stable than the other four
methods. Note that formula (4) calculates the required total number of events based on the event
rate corresponding to X"0, then inflates the number of events to obtain the total sample size.
Therefore, formula (4) produces a larger sample size if the lower event rate is assigned to P1
instead of P2. Formula (4) tends to overestimate the required sample sizes especially when the
event ratesare low (seeTable I). Formula(3) is a specialcase of formula(12) for a balanceddesign.
As shown in Table I, formula (3) gives the same sample sizes as formula (12) when B"0)5, but
slightly larger sample size than formula (2). Since formula (3) is designed for B"0)5, no sample
sizes for formula (3) are given for low or high prevalence rate. Formulae (12) and (13) are simpler
than formula (2), but lack accuracy when the sample size ratio is not close to 1 (say'2 or (0)5),
and should not be used when the accuracy of sample size calculation is important. It is known
thata designwithlow prevalencerate requiresless samplesize than highprevalencerate. InTable
I, formula (13) does not show this relationship which indicates that the formula overestimates the
sample size for low prevalence rate and underestimates high prevalence rate.
Table II presents the results for a continuous covariate from sample size programs nQuery and
SSIZE. The corresponding sample sizes from a two-sample t-test (formula (6) with Z-values
replaced by t-values) and from formula (1) are also listed for comparison. The table specifies the
following parameters indicated after the sample size methods:
Formula (1): P1, effect size"log(OR)"?*.
Two-sample t-test: effect size"log(OR),
sample size ratio"prob(½"1)/prob(½"0)"(1!P1)/P1.
nQuery: P1(event rate at the mean of X),
P2(event rate at one standard deviation above the mean of X).
SSIZE: P1(event rate at the mean of X),
OR (odds ratio at one standard deviation above the mean of X)
"P2(1!P1)/(P1(1!P2)).
Table II also provides power simulations obtained from 1000 replications generated by assuming
a normally distributed variable X. We used the Wald test in the simulation of the logistic
regression model. The results show that the sample sizes estimated by using the two-sample t-test
formula and formula (1) seem to be more conservative, but still large enough to achieve the
SAMPLE SIZE CALCULATION
1627
Statist. Med. 17, 1623—1634 (1998)
? 1998 John Wiley & Sons, Ltd.
Page 6
Table II. Results of sample size calculations for a continuous covariate from four different
methods, power"95 per cent, two-sided significance level 5 per cent
DesignSample size Power simulation
Balanced design
(1):
t-test:
nQuery: P1"0)5, P2"0)6
SSIZE:
P1"0)5, OR"1)5
ºnbalanced design, high event rates
(1):
P1"0)4, effect size ?*"0)405
t-test: effect size"0)405, sample size ratio"1)5
nQuery: P1"0)4, P2"0)5
SSIZE:
P1"0)4, OR"1)5
ºnbalanced design, low event rates
(1):
P1"0)1, effect size ?*"0)405
t-test: effect size"0)405, sample size ratio"9
nQuery: P1"0)1, P2"0)143
SSIZE:
P1"0)1, OR"1)5
P1"0)5, effect size ?*"0)405
effect size"0)405, sample size ratio"1
317
320
342
341
95)0$0)69%
95)5$0)66%
96)1$0)61%
95)3$0)67%
330
333
380
379
94)4$0)73%
94)8$0)70%
96)7$0)56%
96)7$0)56%
880
890
951
950
95)5$0)66%
96)1$0)61%
96)6$0)57%
96)6$0)57%
desired power. In other words, Table II seems to indicate that the t-test is a good estimate of
sample size which preserves power. Since we used to upper bound of the required sample size in
the formulae in both nQuery and SSIZE,both programs provide samplesizes slightly higher than
those required. When the odds ratio is fixed, a balanced design (that is, response rate P1"0)5)
requires less sample size than an unbalanced design (for example, P1"0)4 or 0)1). Note that due
to the exponential nature of the correction term (see Appendix I), we do not recommended use of
either software for logistic regression when the odds ratio is large (say*3).
EXAMPLE
We use a Department of Veterans Affairs Cooperative Study entitled ‘A Psychophysiological
Study of Chronic Post-Traumatic Stress Disorder’? to illustrate the preceding sample size
calculation for logistic regression with continuous covariates. The study developed and validated
alogistic regressionmodelto explorethe use ofcertainpsychophysiologicalmeasurementsfor the
prognosis of combat-related post-traumatic stress disorder (PTSD). In the study, patients’ four
psychophysiological measurements — heart rate, blood pressures, EMG and skin conductance
—were recordedwhile patientswere exposedto video tapescontainingcombat and neutralscenes.
Among the psychophysiologicalvariables, the differenceof the heart rates obtained while viewing
the combat and the neutral tapes (DCNHR) is considered a good predictor of the diagnosis of
PTSD. The prevalence rate of PTSD among the Vietnam veterans was assumed to be 20 per cent.
Therefore, we assumed a four to one sample size ratio for the non-PTSD versus PTSD groups.
The effect size of DCNHR is approximately0)3 which is the difference of the group means divided
by the standard deviation. With a two-sided significance level of 0)05 and a power of 95 per cent,
the required sample size based on a two-sample t-test is 905. The squared multiple correlation
1628
F. HSIEH, D. BLOCH AND M. LARSEN
Statist. Med. 17, 1623—1634 (1998)
? 1998 John Wiley & Sons, Ltd.
Page 7
coefficient of DCNHR versus the other three psychophysiological variables was estimated to be
0)1 and thus the VIF is 1)11. After adjusting for the VIF, a sample size of 1005 was needed for
fitting a multiple logistic regression model.
CONCLUSION
The proposed simple methods to calculate sample size for linear and logistic regression models
have several advantages. The formulae for the simple methods are well known and do not require
specialized software. This paper also provides simple forms of the formulae for easy hand
calculation. Compared to more accurate, but more complicated formulae, formulae (1) and (3)
have high degrees of accuracy. Computer simulations suggest that the proposed sample size
methods for comparing means and for comparing proportions are more accurate than SSIZE,
nQueryand EGRET SIZ. This paper suggests not to use SSIZE or nQuery when the odds ratio is
large (say*3) and Liu and Liang’s formula (13) when the sample size ratio is not close to 1
(say'2 or (0)5). This paper derives the variance inflation factor (VIF) for the linear regression
model and also shows, through computer simulations, that the same VIF applies to the logistic
regression model with binary covariates. The usage of the VIF to expand the sample size
calculation from one covariate to more than one covariate appears very useful and can be
extended to other multivariate models. In conclusion, this paper presents more accurate and
simple formulae for sample size calculation with extensions to multivariate models of various
types.
APPENDIX I
In a simple logistic regression model log(P/(1!P))"??#??X?, where P"prob(½"1), the
hypothesis H?: ??"0 against H?: ??"?* is of interest. A power of 1!? and a two-sided
significancelevel? are usually prespecifiedto calculatethe sample sizefor the hypothesistest. The
followingsample size formula, used in both SSIZE and nQuery, is a combinationof Whittemore?
formulae (6) and (16):
n"(»(0)???Z?????#»(?*)???Z???)?(1#2P1?)/(P1?*?)
where the log odds value ?*"log(P2(1!P1)/(P1(1!P2))), and Z???and Z?????are standard
normal variables with a tail probability of ? and ?/2, respectively.
For a continuous covariate, »(0)"1, »(?*)"exp(!?*?/2), P1 and P2 are the event rates at
the mean of X and one SD above the mean, respectively. The value of ? for continuous covariates
is from Hsieh? formula (3): ?"(1#(1#?*?)exp(5?*?/4))(1#exp(!?*?/4))??.
For a binary covariate, the overall event rate P"(1!B)P1#BP2, where P1 and P2 are the
event rates at X"0 and X"1, respectively; B is the proportion of the sample with X"1,
»(0)"1/(1!B)#1/B, and »(?*)"1/(1!B)#1/(Bexp(?*)). The value of ? for binary covari-
ates is from Whittemore? formula (14): ?"(»(0)???#»(?*)???R)/(»(0)???#»(?*)???) where R is
from Whittemore? formula (15): R"»(?*)B(1!B)exp(2?*)/(Bexp(?*)#(1!B))?. Note that
R"?"1 when ?*"0.
The proposed method is to use a two-sample test instead of a one-sample test for sample
size calculation. The popular sample size formula for testing the equality of two independent
sample means with equal sample sizes from two normally distributed groups has the familiar
(4)
SAMPLE SIZE CALCULATION
1629
Statist. Med. 17, 1623—1634 (1998)
? 1998 John Wiley & Sons, Ltd.
Page 8
form (see Rosner??):
n"2(???#???)(Z?????#Z???)?/??
(5)
where n is the total sample size and ? is the difference of the two group means to be detected;
???and ???are the variances of the two groups. For an unequal-sample-size design with a sample
size ratio of k, the required total sample size should be inflated by a factor of (k#1)?/(4k).
Assuming equal variances, the test statistic employs the common variance of the two groups and
formula (5) reduces to
n"??(Z?????#Z???)?[(k#1)?/k]/??
In a simple logistic regression model with a continuous covariate, the sample size ratio is
k"(1!P1)/P1 where P1 is the event rate of the response at X"0. Therefore, P1 is also the
overall event rate when X is standardized to have mean 0 and variance 1. By replacing the effect
size ?/? by ?*, formula (6) becomes
n"(Z?????#Z???)?/[P1(1!P1)?*?].
As derived by Whittemore,? 1"»(0)*»(?*), and therefore formula (4) can be bounded by
n)(Z?????#Z???)?(1#2P1?)/(P1?*?).
Formula (7) is more general than the formula derived by Whittemore,? who assumed that P1 is
small and therefore 1/(1!P1) is negligible. Note that Hsieh? formula (3) implies that one should
not use formula (4) when the odds ratio is large (say*3).
Whenthe covariateis a binary variable,say X"0 or 1, the log oddsvalues ??"0 if and onlyif
the two event rates are equal. We can calculate the total sample size from the formula for
comparing the two independent event rates (see Rosner??):
n"(1#k)?Z?????[P(1!P)(k#1)/k]???#Z???[P1(1! P1)#P2(1!P2)/k]?????/(P1!P2)?
(6)
(7)
(8)
(9)
where: k"B/(1!B) is the sample size ratio; B is the proportion of the sample with X"1;
P"(1!B)P1#BP2 is the overall event rate; P1 and P2 are the event rates at X"0 and X"1,
under the alternative hypothesis, respectively. By replacing k by B/(1!B), formula (9) becomes
n"?Z?????[P(1! P)/B]???#Z???[P1(1!P1)#P2(1!P2)(1!B)/B]?????/[(P1!P2)?(1!B)].
(10)
For a balanced design, k"1 or B"0)5, formula (10) is bounded by
n(4P(1!P)(Z?????#Z???)?/(P1!P2)?.
For an unbalanced design, similar to (6), we inflate formula (11) by a factor of 1/[4B(1!B)] to
obtain a simple approximation:
n"P(1!P)(Z?????#Z???)?/[B(1!B)(P1!P2)?].
In a recent publication, Liu and Liang?? extended Self and Mauritsen’s method for correlated
observations. As a special case, they provided a closed form for a logistic regression model with
(11)
(12)
1630
F. HSIEH, D. BLOCH AND M. LARSEN
Statist. Med. 17, 1623—1634 (1998)
? 1998 John Wiley & Sons, Ltd.
Page 9
one binary covariate. Their closed form, without the adjustment of the design effect for correlated
observations, is very similar to (12):
n"(Z?????#Z???)?[BP1(1!P1)#(1!B)P2(1!P2)]/[B(1!B)(P1!P2)?].
Examples and comparisons of these formulae are provided in Table I.
(13)
APPENDIX II
Let var?(b?) and var?(b?) equal the variances of the parameter estimate obtained from multiple
linear regression models with p and 1 covariates, respectively. We show that, most often, the ratio
var?(b?)/var?(b?) is bounded by 1/(1!??1.232p). In addition, var?(b?)/var?(b?) is bounded by
1/(1!??(1q#12p))(232q)) where the partial correlation coefficient ?(1q#12p))(232q)measures the
linear association between covariates X?and X???,2,X?when the values of covariates
X?,2,X?are held fixed.
We begin with one covariate in a linear regression model ½"??#??X?#e where the error
term e is distributed as Normal (0,???) and, for simplicity, the sample mean of X?is 0. The
variance of the least squares estimate b?is known to equal
var?(b?)"???/?X??.
Whentherearetwo covariatesX?and X?withsamplemeans0, thevariance-covariancematrixof
the estimates of the parameters is
var?(b?,b?)"???(X?X)??"????
?X??
?X?X?
?X?X?
?X???
??
whereX is thematrix of covariates.Throughthe inverseof the 2?2 X?Xmatrix, we can obtainthe
variance of b?as
var?(b?)"????X??/(?X???X??!(?X?X?)?)
"(???/???)var?(b?)/(1!????).
Thevalue of ???/???, in most cases, is less than 1 and close to 1. Since the additional covariate in the
model also takes away a degree of freedom from the error term, the estimate of the variance ratio
???/???may sometimes slightly exceed 1. The squared multiple correlation coefficient, in this case
the same as the simple correlation coefficient, is ????"(?X?X?)?/(?X???X??).
Whenthere are three covariates,the multiple correlationcoefficient?????can be obtainedfrom
the matrix operation
??????"[?X?X?, ?X?X?]?
"(2?X?X??X?X??X?X?!?X??(?X?X?)?!?X??(?X?X?)?]/??X??[?X???X??
!(?X?X?)?]?.
With three covariates in the regression model, the variance-covariance matrix of the estimates of
the parameters can be obtained from the inverse of the 3?3 X?X matrix through the formula
?X??
?X?X?
?X?X?
?X???
???
?X?X?
?X?X????X??.
SAMPLE SIZE CALCULATION
1631
Statist. Med. 17, 1623—1634 (1998)
? 1998 John Wiley & Sons, Ltd.
Page 10
var?(b?,b?,b?)"???(X?X)??. Therefore
var?(b?)"???[?X???X??!(?X?X?)?]/[?X???X???X??#2?X?X??X?X??X?X?
!?X??(?X?X?)?!?X??(?X?X?)?!?X??(?X?X?)?]
"(???/???)var?(b?)/(1!??????).
Usually, ???/???)1 and var?(b?))var?(b?)/(1!??????). In a linear regression model with p para-
meters,var?(b?,b?,2,b?)"???(X?X)??"????. By applying a result of Anderson??(equation 20)
that ???
var?(b?)"??????"???/?X??(1!??1.232p)
"(???/???)var?(b?)/(1!??????2?).
Again, in most situations, ???/???)1
VIF"1/(1!??1.232p) is the approximate upper bound of the ratio var?(b?)/var?(b?). The upper
bound does not hold in the rare situation when ???/???'1 but the approximation is still good
enough. When p is not too large, the bound is tight; when p is large and ?1.232pis near 1, the
bound is inaccurate. A similar result holds for nested models. We would like to expand the model
from the situation of q covariates to p covariates where p'q. Then, reasoning as above
var?(b?)/var?(b?)"(???/???)(1!??1.232p)/(1!??1.232p))(1!??1.232q)/(1!??1.232p)
"1/(1!??(1q#12p))(232q)).
??"?X??(1!??1.232p), we obtain
andvar?(b?))var?(b?)/(1!??1.232p). Then, the
where the partial corelation coefficient ?(1q#12p))(232q)measures the linear association between
covariates X?and X???,2,X?when the values of covariates X?,2,X?are held fixed. The
value of the ratio ???/???should be closer to 1 than ???/???.
APPENDIX III
We use simulations to investigate, in a multiple logistic regression model with p independent
binary covariates, how well the ratio of the maximum likelihood estimates of the variances
var?(b?)/var?(b?) is approximated by 1/(1!??1.232p), where the multiple correlation coefficient
relating binary covariates X?with X?,2,X?has the same formula as continuous covariates
with a normal distribution:
??1.232p"[?X?X?,?X?X?2,?X?X?]
?X??
?X?X?
2
?X?X?
?X?X?
?X??
2
?X?X?
2
2
2
2
?X?X?
?X?X?
2
?X??
??
?X?X?
?X?X?
2
?X?X???X??
.
The 80 computer simulations each use a sample size of 1000 with eight binary covariates. When
all eight covariates are generated independently, the estimate values of ??1.2327are near zero. In
order that the response variable ½ and the covariates X’s be somewhat correlated, and the
estimates of ??1.2327have a broad range of values, say from 0 to 0)7, the generation of the eight
covariates requires some special care.
1632
F. HSIEH, D. BLOCH AND M. LARSEN
Statist. Med. 17, 1623—1634 (1998)
? 1998 John Wiley & Sons, Ltd.
Page 11
Figure 1. Results of 80 simulations: estimates of var?(b?)/var?(b?) versus 1/(1!??
1.2328)
Let º, »?, »?,2, and »?be uniform random variates obtained from a generator in SAS.??
The response variable ½ is Bernoulli with a parameter value 0)5. The eight covariates X?,X?,2,
andX?arealso Bernoulliwith parametersB?,B?,2, andB?which have values 0)5, 0)6, 0)65, 0)7,
0)75, 0)8, 0)85 and 0)9, respectively. The response variable ½ is generated such that ½"1 when
º'0)5 and ½"0, otherwise. In the first simulation, the covariates are generated such that
X?"1 when 0)1º#0)9»?'B?and X?"0, otherwise, for i"1,2,2,8. The same process was
repeated for the second simulation except for the generation of X?where the same random value
forX?was used: X?"1 when 0)1º#0)9»?'B?and X?"0, otherwise. In the third simulation,
the same random value for X?was used for X?. The similar process continued until the
completion of the eighth simulation. After finishing the first eight simulations, the whole process
was then repeated ten times to obtain a total of 80 simulations.
In practice, the estimated values of var?(b?) and ??1.232p(same as R?) can be obtained from SAS
PROC LOGISTIC and PROC REG,?? respectively. The estimates of var?(b?)/var?(b?) versus
1/(1!??1.232p) from the simulations are plotted in Figure 1. The simulation results show that, for
binary covariates, the estimates of 1/(1!??1.232p) closely approximate the value of the estimates
of the ratio var?(b?)/var?(b?). Figure 1 shows that the estimates of 1/(1!??1.232p) very slightly
underestimate the variance ratio var?(b?)/var?(b?).
ACKNOWLEDGEMENTS
The authors thank Drs. Philip Lavori, Kelvin Lee, a referee and the editor for valuable comments
andeditorial suggestions which strengthenedthe content. This work was supported in part by the
DVA Cooperative Studies Program of the Veteran Health Administration, NIH grant AR20610
(Multipurpose Arthritis Center) and Y01-DA-40032-0 (National Institute on Drug Abuse), the
latter to the VA Cooperative Studies Program.
SAMPLE SIZE CALCULATION
1633
Statist. Med. 17, 1623—1634 (1998)
? 1998 John Wiley & Sons, Ltd.
Page 12
REFERENCES
1. Whittemore, A. ‘Sample size for logistic regression with small response probability’, Journal of the
American Statistical Association, 76, 27—32 (1981).
2. Hsieh, F. Y. ‘Sample size tables for logistic regression’, Statistics in Medicine, 8, 795—802 (1989).
3. Self, S. G. and Mauritsen, R. H. ‘Power/sample size calculations for generalized linear models’,
Biometrics, 44, 1, 79—86 (1988)
4. Hosmer, D. W. and Lemeshow, S. Applied ¸ogistic Regression, Wiley, New York, 1989, p. 56.
5. Sokal, R. R. and Rohlf, F. J. Biometry, W. H. Freeman and Company, New York, 1995, p. 578.
6. Elashoff, J. nQuery Advisor Sample Size and Power Determination, Statistical Solutions Ltd., Boston,
MA, 1996.
7. Hsieh, F. ‘SSIZE: A sample size program for clinical and epidemiologic studies’, American Statistician,
45, 338 (1991).
8. SERC. EGRE¹ SIZ sample size and power for nonlinear regression models, Statistics and Epidemiology
Research Corp. Seattle, WA, 1992.
9. Keane, T. M., Kolb, L. C. and Thomas, R. G. ‘A Psychophysiological Study of Chronic Post-Traumatic
Stress Disorder’, Cooperative Study No. 334, Cooperative Studies Program Coordinating Center, VA
Medical Center, Palo Alto, California, U.S.A., 1988.
10. Rosner, B. Fundamentals of Biostatistics, 4th edn, PWS-KENT Publishing Company, 1995, p. 283 and
384.
11. Liu, G. and Liang, K. Y. ‘Sample size calculation for studies with correlated observations’, Biometrics,
53, 537—547 (1997).
12. Anderson, T. W. An Introduction to Multivariate Statistical Analysis, Wiley, New York, 1958, p. 32.
13. SAS Institute Inc. SAS/S¹A¹ ºser’s Guide, »ersion 6 (»ol. 1 and 2), Cary, NC, 1990.
1634
F. HSIEH, D. BLOCH AND M. LARSEN
Statist. Med. 17, 1623—1634 (1998)
? 1998 John Wiley & Sons, Ltd.
Download full-text