This question is in response to an answer given by @Greg Snow in regards to a question I asked concerning power analysis with logistic regression and SAS Proc GLMPOWER.

If I am designing an experiment and will analze the results in a factorial logistic regression, how can I use simulation ( and here ) to conduct a power analysis?

Here is a simple example where there are two variables, the first takes on three possible values {0.03, 0.06, 0.09} and the second is a dummy indicator {0,1}. For each we estimate the response rate for each combination (# of responders / number of people marketed to). Further, we wish to have 3 times as many of the first combination of factors as the others (which can be considered equal) because this first combination is our tried and true version. This is a setup like given in the SAS course mentioned in the linked question.

enter image description here

The model that will be used to analyze the results will be a logistic regression, with main effects and interaction (response is 0 or 1).

mod <- glm(response ~ Var1 + Var2 + I(Var1*Var2))

How can I simulate a data set to use with this model to conduct a power analysis?

When I run this through SAS Proc GLMPOWER (using STDDEV =0.05486016 which corresponds to sqrt(p(1-p)) where p is the weighted average of the shown response rates):

data exemplar;
  input Var1 $ Var2 $ response weight;
    3 0 0.0025  3
    3 1 0.00395 1
    6 0 0.003   1
    6 1 0.0042  1
    9 0 0.0035  1
    9 1 0.002   1;

proc glmpower data=exemplar;
  weight weight;
  class Var1 Var2;
  model response = Var1 | Var2;

Note: GLMPOWER only will use class (nominal) variables so 3, 6, 9 above are treated as characters and could have been low, mid and high or any other three strings. When the real analysis is conducted, Var1 will be used a numeric (and we will include a polynomial term Var1*Var1) to account for any curvature.

The output from SAS is

enter image description here

So we see that we need 762,112 as our sample size (Var2 main effect is the hardest to estimate) with power equal to 0.80 and alpha equal to 0.05. We would allocate these so that 3 times as many were the baseline combination (i.e. 0.375 * 762112) and the remainder just fall equally into the other 5 combinations.

  • $\begingroup$ This is easy to do in R. 1st question: am I correct that you want 75% of all cases to be {var1=.03, var2=0} & 25% for all other combos, & not 3 units there for every 1 unit in each of the other combos (ie, 37.5%)? 2nd question, can you specify the effects you are interested in detecting? Ie, what would be the log odds of 1 vs 0? How should the log odds of success change if var1 goes up by .01? Do you think there might be an interaction (if so, how big is it)? (NB, these Q's can be hard to answer, 1 strategy is to specify the proportion of 1's you think might be in each combo.) $\endgroup$ – gung Sep 9 '12 at 0:38
  • $\begingroup$ 1st: The weight of 3 for the baseline case is that there is 3 times as many cases where {var1=0.03, var2=0}. So the results from SAS (which says that we need 762,112 total sample size to have 80% power of rejecting main effect var2=0, so that is the total sample size we need) would be allocated 37.5% to this baseline case. $\endgroup$ – B_Miner Sep 9 '12 at 0:48
  • $\begingroup$ 2nd: Well all we have is the response rates (which is the expected ratio of the number of success over number of trials). So, if we send 1,000 letters with Var1=0.03 and Var2=0 which could correspond to an interest rate offer on a credit card direct mail offer of 0.03 (3%) and no sticker on the envelope (where Var2=1 means there is a sticker), we expect 1000*0.0025 responses. $\endgroup$ – B_Miner Sep 9 '12 at 0:52
  • $\begingroup$ 2nd cont: We do expect an interaction - hence the response rates. Note there is a different response rate for Var2=0 depending on the value of Var1. I am not sure how to translate these to log odds and then the a simulated data set. $\endgroup$ – B_Miner Sep 9 '12 at 0:54
  • $\begingroup$ One last thing, though. I notice that the response rates are linear for var1 when var2=0 (ie, .25%, .30%, .35%). Did you intend for this to be a linear effect or curvilinear? You should know that probabilities can look fairly linear for small subsets of their range, but cannot actually be linear. Logistic regression is linear in log odds, not probability (I discuss stuff like that in my answer here). $\endgroup$ – gung Sep 9 '12 at 1:47


  • As discussed in the G*Power manual, there are several different types of power analyses, depending on what you want to solve for. (That is, $N$, the effect size $ES$, $\alpha$, and power exist in relation to each other; specifying any three of them will let you solve for the fourth.)

    • in your description, you want to know the appropriate $N$ to capture the response rates you specified with $\alpha=.05$, and power = 80%. This is a-priori power.
    • we can start with post-hoc power (determine power given $N$, response rates, & alpha) as this is conceptually simpler, and then move up
  • In addition to @GregSnow's excellent post, another really great guide to simulation-based power analyses on CV can be found here: Calculating statistical power. To summarize the basic ideas:

    1. figure out the effect you want to be able to detect
    2. generate N data from that possible world
    3. run the analysis you intend to conduct over those faux data
    4. store whether the results are 'significant' according to your chosen alpha
    5. repeat many ($B$) times & use the % 'significant' as an estimate of (post-hoc) power at that $N$
    6. to determine a-priori power, search over possible $N$'s to find the value that yields your desired power
  • Whether you will find significance on a particular iteration can be understood as the outcome of a Bernoulli trial with probability $p$ (where $p$ is the power). The proportion found over $B$ iterations allows us to approximate the true $p$. To get a better approximation, we can increase $B$, although this will also make the simulation take longer.

  • In R, the primary way to generate binary data with a given probability of 'success' is ?rbinom

    • E.g. to get the number of successes out of 10 Bernoulli trials with probability p, the code would be rbinom(n=10, size=1, prob=p), (you will probably want to assign the result to a variable for storage)
    • you can also generate such data less elegantly by using ?runif, e.g., ifelse(runif(1)<=p, 1, 0)
    • if you believe the results are mediated by a latent Gaussian variable, you could generate the latent variable as a function of your covariates with ?rnorm, and then convert them into probabilities with pnorm() and use those in your rbinom() code.
  • You state that you will "include a polynomial term Var1*Var1) to account for any curvature". There is a confusion here; polynomial terms can help us account for curvature, but this is an interaction term--it will not help us in this way. Nonetheless, your response rates require us to include both squared terms and interaction terms in our model. Specifically, your model will need to include: $var1^2$, $var1*var2$, and $var1^2*var2$, beyond the basic terms.

  • Although written in the context of a different question, my answer here: Difference between logit and probit models has a lot of basic information about these types of models.
  • Just as there are different kinds of Type I error rates when there are multiple hypotheses (e.g., per-contrast error rate, familywise error rate, & per-family error rate), so are there different kinds of power* (e.g., for a single pre-specified effect, for any effect, & for all effects). You could also seek for the power to detect a specific combination of effects, or for the power of a simultaneous test of the model as a whole. My guess from your description of your SAS code is that it is looking for the latter. However, from your description of your situation, I am assuming you want to detect the interaction effects at a minimum.

  • For a different way to think about issues related to power, see my answer here: How to report general precision in estimating correlations within a context of justifying sample size.

Simple post-hoc power for logistic regression in R:

Let's say your posited response rates represent the true situation in the world, and that you had sent out 10,000 letters. What is the power to detect those effects? (Note that I am famous for writing "comically inefficient" code, the following is intended to be easy to follow rather than optimized for efficiency; in fact, it's quite slow.)


repetitions = 1000
N = 10000
n = N/8
var1  = c(   .03,    .03,    .03,    .03,    .06,    .06,    .09,   .09)
var2  = c(     0,      0,      0,      1,      0,      1,      0,     1)
rates = c(0.0025, 0.0025, 0.0025, 0.00395, 0.003, 0.0042, 0.0035, 0.002)

var1    = rep(var1, times=n)
var2    = rep(var2, times=n)
var12   = var1**2
var1x2  = var1 *var2
var12x2 = var12*var2

significant = matrix(nrow=repetitions, ncol=7)

startT = proc.time()[3]
for(i in 1:repetitions){
  responses          = rbinom(n=N, size=1, prob=rates)
  model              = glm(responses~var1+var2+var12+var1x2+var12x2, 
  significant[i,1:5] = (summary(model)$coefficients[2:6,4]<.05)
  significant[i,6]   = sum(significant[i,1:5])
  modelDev           = model$null.deviance-model$deviance
  significant[i,7]   = (1-pchisq(modelDev, 5))<.05
endT = proc.time()[3]

sum(significant[,1])/repetitions      # pre-specified effect power for var1
[1] 0.042
sum(significant[,2])/repetitions      # pre-specified effect power for var2
[1] 0.017
sum(significant[,3])/repetitions      # pre-specified effect power for var12
[1] 0.035
sum(significant[,4])/repetitions      # pre-specified effect power for var1X2
[1] 0.019
sum(significant[,5])/repetitions      # pre-specified effect power for var12X2
[1] 0.022
sum(significant[,7])/repetitions      # power for likelihood ratio test of model
[1] 0.168
sum(significant[,6]==5)/repetitions   # all effects power
[1] 0.001
sum(significant[,6]>0)/repetitions    # any effect power
[1] 0.065
sum(significant[,4]&significant[,5])/repetitions   # power for interaction terms
[1] 0.017

So we see that 10,000 letters doesn't really achieve 80% power (of any sort) to detect these response rates. (I am not sufficiently sure about what the SAS code is doing to be able to explain the stark discrepancy between these approaches, but this code is conceptually straightforward--if slow--and I have spent some time checking it, and I think these results are reasonable.)

Simulation-based a-priori power for logistic regression:

From here the idea is simply to search over possible $N$'s until we find a value that yields the desired level of the type of power you are interested in. Any search strategy that you can code up to work with this would be fine (in theory). Given the $N$'s that are going to be required to capture such small effects, it is worth thinking about how to do this more efficiently. My typical approach is simply brute force, i.e. to assess each $N$ that I might reasonably consider. (Note however, that I would typically only consider a small range, and I'm typically working with very small $N$'s--at least compared to this.)

Instead, my strategy here was to bracket possible $N$'s to get a sense of what the range of powers would be. Thus, I picked an $N$ of 500,000 and re-ran the code (initiating the same seed, n.b. this took an hour and a half to run). Here are the results:

sum(significant[,1])/repetitions      # pre-specified effect power for var1
[1] 0.115
sum(significant[,2])/repetitions      # pre-specified effect power for var2
[1] 0.091
sum(significant[,3])/repetitions      # pre-specified effect power for var12
[1] 0.059
sum(significant[,4])/repetitions      # pre-specified effect power for var1X2
[1] 0.606
sum(significant[,5])/repetitions      # pre-specified effect power for var12X2
[1] 0.913
sum(significant[,7])/repetitions      # power for likelihood ratio test of model
[1] 1
sum(significant[,6]==5)/repetitions   # all effects power
[1] 0.005
sum(significant[,6]>0)/repetitions    # any effect power
[1] 0.96
sum(significant[,4]&significant[,5])/repetitions   # power for interaction terms
[1] 0.606

We can see from this that the magnitude of your effects varies considerably, and thus your ability to detect them varies. For example, the effect of $var1^2$ is particularly difficult to detect, only being significant 6% of the time even with half a million letters. On the other hand, the model as a whole was always significantly better than the null model. The other possibilities are arrayed in between. Although most of the 'data' are thrown away on each iteration, a good bit of exploration is still possible. For example, we could use the significant matrix to assess the correlations between the probabilities of different variables being significant.

I should note in conclusion, that due to the complexity and large $N$ entailed in your situation, this was not as simple as I had suspected / claimed in my initial comment. However, you can certainly get the idea for how this can be done in general, and the issues involved in power analysis, from what I've put here. HTH.

  • $\begingroup$ Gung - WOW thank you very much for such a detailed and thoughtful answer! In writing my own and playing with your code, the quadratic terms appear to be the issue - as at least 80% power is achieved with a much smaller sample size without considering it in the model. $\endgroup$ – B_Miner Sep 10 '12 at 13:23
  • 1
    $\begingroup$ That's great, @B_Miner, that's the kind of thing you want to do. Moreover, it's the reason I think the simulation-based approach is superior to analytical software that just spits out a number (R has this also, the pwr package). This approach gives you the opportunity to think much more clearly (&/or refine your thinking) about what you expect to happen, how you would deal w/ that, etc. NB, however, that you do need the quadratic terms, or something analogous, if your posited rates are right, b/c they are not linear, & the interaction alone doesn't let you capture curvilinear relationships. $\endgroup$ – gung Sep 10 '12 at 13:43
  • $\begingroup$ I think you should be demonstrating the use of poly rather than showing new users of R the more error-prone strategy of squaring raw values. I think the full model should have been posed as glm(responses~ poly(var1, 2) * var2, family=binomial(link="logit"). It would be both less prone to statistical error in interpretation and much more compact. Might not be important in this exact instance when you are only looking at an overall fit, but could easily mislead less sophisticated users who might be tempted to look at individual terms. $\endgroup$ – DWin May 3 '16 at 19:22
  • $\begingroup$ @DWin, when I use R to illustrate things here on CV, I do it in a very non-R manner. The idea is to be as transparent as possible for those who aren't familiar w/ R. Eg, I'm not using the vectorized possibilities, am using loops, =, etc. People will be more familiar with squaring variables from a basic regression class, & less aware of what poly() is, if they aren't R users. $\endgroup$ – gung May 3 '16 at 20:15

@Gung's answer is great for understanding. Here is the approach that I would use:

mydat <- data.frame( v1 = rep( c(3,6,9), each=2 ),
    v2 = rep( 0:1, 3 ), 
    resp=c(0.0025, 0.00395, 0.003, 0.0042, 0.0035, 0.002) )

fit0 <- glm( resp ~ poly(v1, 2, raw=TRUE)*v2, data=mydat,
    weight=rep(100000,6), family=binomial)
b0 <- coef(fit0)

simfunc <- function( beta=b0, n=10000 ) {
    w <- sample(1:6, n, replace=TRUE, prob=c(3, rep(1,5)))
    mydat2 <- mydat[w, 1:2]
    eta <- with(mydat2,  cbind( 1, v1, 
                v1^2, v2,
                v1^2*v2 ) %*% beta )
    p <- exp(eta)/(1+exp(eta))
    mydat2$resp <- rbinom(n, 1, p)

    fit1 <- glm( resp ~ poly(v1, 2)*v2, data=mydat2,
    fit2 <- update(fit1, .~ poly(v1,2) )
    anova(fit1,fit2, test='Chisq')[2,5]

out <- replicate(100, simfunc(b0, 10000))
mean( out <= 0.05 )
abline(v=0.05, col='lightgrey')

This function tests the overall effect of v2, the models can be changed to look at other types of tests. I like writing it as a function so that when I want to test something different I can just change the function arguments. You could also add a progress bar, or use the parallel package to speed things up.

Here I just did 100 replications, I usually start around that level to find the approximate sample size, then up the itterations when I am in the right ball park (no need to waste the time on 10,000 iterations when you have 20% power).

  • $\begingroup$ Thanks Greg! I was wondering about this same approach (if I am understanding correctly what you did). To confirm: Are you creating a data set (in effect, but doing it with weights instead of brute force creating individual records of the values of Var1 and Var2 and then 1's and 0's for the response) that is very large based on "mydat", fitting a logistic regression and then using those coefficients to sample from the proposed model in the simulation? It seems this is a general way to come up with the coefficients - then its just like your response about ordinal regression power I linked to. $\endgroup$ – B_Miner Sep 10 '12 at 17:44
  • $\begingroup$ The initial model uses weights to get the coefficients to use, but in the simulation it is creating a data frame with n rows. It may be more efficient to do weights in the function as well. $\endgroup$ – Greg Snow Sep 10 '12 at 18:00
  • $\begingroup$ I am correct that you are using the data initially (making it big to get very good estimates) for the purpose of getting the coefficients that are used? $\endgroup$ – B_Miner Sep 10 '12 at 18:20
  • $\begingroup$ Yes, well the big is so that the proportion times the weight gives an integer. $\endgroup$ – Greg Snow Sep 10 '12 at 21:00
  • 2
    $\begingroup$ @B_Miner, I am planning on an article, I don't know that there is enough for a full book or not. But thanks for the encouragement. $\endgroup$ – Greg Snow Jan 11 '17 at 16:45

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.