Rather than take a z-score, which is symmetrical, one idea for a small improvement might be to use a critical value from a chi-squared distribution. Since it's

Rather than take a z-score, which is symmetrical, one idea for a small improvement might be to use a critical value from a chi-squared distribution. Since it's asymmetric, it'll probably fit the distribution better and offer a higher signal-to-noise ratio. EDIT: or actually, since this is a count, might as well go straight for the poisson distribution or negative binomial -- same idea.