Grokbase Groups R r-help May 2012
Dear list,

I'm a bit perplexed why the 95% confidence bands for the predicted probabilities for units where x=0 and x=1 overlap in the following instance.

I've simulated binary data to which I've then fitted a simple logistic regression model, with one covariate, and the coefficient on x is statistically significant at the 0.05 level. I've then used two different methods to generate 95% confidence bands for predicted probabilities, for each of two possible values of x. Given the result of the model, I expected the bands not to overlap? but they do.

Can anyone please explain where I've gone wrong with the following code, OR why it makes sense for the bands to overlap, despite the statistically significant beta coefficient? There may be a good statistical reason for this, but I'm not aware of it.

Many thanks,
Malcolm Fairbrother

n <- 120
x <- rbinom(n, 1, 0.5)
dat <- within(data.frame(x), ybe <- rbinom(n, 1, plogis(-0.5 + x)))
mod1 <- glm(ybe ~ x, dat, family=binomial)
summary(mod1) # coefficient on x is statistically significant at the 0.05 level? almost at the 0.01 level

pred <- predict(mod1, newdata=data.frame(x=c(0,1)),
with(pred, cbind(low = plogis(fit - 1.96*, est = plogis(fit), up = plogis(fit + 1.96* # confidence bands based on SEs

# simulation-based confidence bands:
sims <- t(replicate(200, coef(glm(simulate(mod1)$sim_1 ~ x, data=dat, family=binomial))))
pred0 <- plogis(quantile(sims%*%c(1,0), c(0.025, 0.5, 0.975)))
pred1 <- plogis(quantile(sims%*%c(1,1), c(0.025, 0.5, 0.975)))
rbind(pred0, pred1)

# the upper bound of the prediction for x=0 is greater than the lower bound of the prediction for x=1, using both methods

Search Discussions

Discussion Posts

Follow ups

Related Discussions

Discussion Navigation
viewthread | post
posts ‹ prev | 1 of 3 | next ›
Discussion Overview
groupr-help @
postedMay 3, '12 at 4:37p
activeMay 6, '12 at 6:18p



site design / logo © 2022 Grokbase