Essentially, I have some covariate data X, and a dependent variable Y consisting of proportions of a sample that shown a certain response (i.e. between 0 and 1). I suspect I want to proceed via a GLM approach, but the thing is, I don't know the sizes of each of those samples!

My thoughts are to proceed by a quasibinomial methodology, estimating the dispersion parameter. Assuming the sizes of each sample are not too different, I can thus keep the logit link between the linear predictor and the proportion, but disregard the contribution of n in the usual binomial variance of np(1-p). Then I can do hypothesis testing the usual way?

Does this make any sense?

Some R code:

#simulate some data

X = rnorm(500)
Z = rnorm(500)
p = exp(X*0.1 + 2)/(1+exp(X*0.1 + 2))
n = 50
for (i in 1:length(X)){
    Y = c(Y,sum(runif(n) < p[i])/n)
Y2 = cbind(Y*n, n-Y*n)

#glm, binomial model is 'true' ?
summary(glm(Y~X+Z, family = "quasibinomial"))
summary(glm(Y2~X+Z, family = "binomial"))
anova(glm(Y~X+Z, family = "quasibinomial"), test = "Chisq")
anova(glm(Y2~X+Z, family = "binomial"), test= "Chisq")

Seems to work, but am I missing something? Surely someone's done something like this before?


Your Answer

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

Browse other questions tagged or ask your own question.