Should I use a Poisson regression for overdispersed count data?

[Scarlett_C]

I’m trying to decide if I should use a Poisson regression for my count data, but the variance is about 1.8 times the mean. I’ve read that this slight overdispersion might still be okay for the model, but I’m worried my standard errors will be underestimated.

[AmeliaVJ]

I’d start from the fact that a variance 1.8 times the mean means mild overdispersion. If you’re using Poisson regression, the standard errors are likely underestimated. Realistically you want to switch to something that handles extra-Poisson variation, like a negative binomial model or a quasi-Poisson approach—the latter gives standard errors that align better with the data.

[Mason.M]

Last time I ran into that, I fit a quasi-Poisson and looked at the dispersion estimate (Pearson chi-square over df). It was around 1.7–1.9 here, SEs jumped a bit, and the p-values softened. Not dramatic, but enough to rethink the inference.

[ZoeyG]

Could it be that the problem isn’t overdispersion but something else in the data, like zero-inflation or influential points? I’d peek at residuals and maybe a hurdle or zero inflated model if there are lots of zeros.

[Sofia35]

I started a quick NB run once and then got pulled into a messy data tweak, so I never finished the comparison. The dispersion mattered more for some predictors than others, and I left the rest as is.