[sesqu]

Noprocrast caught out my attempt to edit. The normal variance is too large.

Here's the plot. Black for discretized(n=1000) binomial likelihood, red for normal approximation. The effect is clear, but a t-test won't show it. I'm not familiar with the theory behind the g-test, but there's clearly a lot of room for improvement at these sample sizes.

http://img693.imageshack.us/img693/4880/bindiff.png

[btilly]

It looks like you forgot to rescale the binomial distribution.

If X_i is a series of independent, identically distributed random variables with mean m and variance v, then X_1 + X_2 + ... + X_n is approximately a normal variable with mean nm and variance vn. Therefore

(X_1 + X_2 + ... + X_n - nm)/sqrt(vn)

is approximately a standard normal.

If you draw that graph, visually the two lines should lie right on top of each other. To see the problem you need to zoom in on the tail and blow it up, and only then will you see the issues with the convergence.

[sesqu]

You're right, I did have the wrong scale, and should have realized. I believe this plot to be correct. Black is binomial, red is normal, both for the difference in conversion rates.

http://img267.imageshack.us/img267/6691/binnormdifflikelihoo...

The normal approximation seems reasonable, and so a t-test shouldn't cause problems. I made a plot of the relative error of the normal approximation; the tails are indeed too fat. In particular, there's a bump to the left of 0, so a t-test would slightly overestimate p.

http://img205.imageshack.us/img205/6691/binnormdifflikelihoo...

[btilly]

The problem is that when you're near confidence you're in the tail, and the student t-test is extremely sensitive to the shape of said tail. If n is large enough this difference will be washed away as everything converges to normals. But with smaller sample sizes the difference can be quite significant.