Proc GLM vs Proc Genmod

[Kyra]

Hi,

I am wondering if i should do Proc GLM (linear regression) vs Proc GENMOD (poisson regression) for my outcome below.

Days to goal feed is my outcome which is positive integers (no decimals).

My exposure is type of feed which is bolus vs continuous.

Thanks,

[Reeza]

---

@Kyra wrote:

Hi,

I am wondering if i should do Proc GLM (linear regression) vs Proc GENMOD (poisson regression) for my outcome below.

Days to goal feed is my outcome which is positive integers (no decimals).

My exposure is type of feed which is bolus vs continuous.

Thanks,

---

Graph your data. If you lambda is higher it starts to approximate a normal distribution anyways and doesn't matter too much. If it's definitely not normal, or the lambda isn't high enough to approximate a continuous distribution, then I would recommend Poisson. Also, Poisson measures the number of events within a specific time period, so I think a GLM would work better here as well. But I'll move this to the Stat procedures forum and someone can provide a more robust answer. 

[PaigeMiller]

The response variable does not have to be (approximately) normal for GLM to work. 

The residuals from the fitted GLM model have to be (approximately) iid normal for the GLM hypothesis tests to be valid. Plot the residuals from the GLM model. 

[Reeza]

I was thinking of normal in terms of allowing it to go below zero, not the distribution requirement for regression. That part I remember 🙂

Is your data time to event analysis or number of events analysis?

[StatDave]

While relying on the normal approximation and using PROC GLM is certainly reasonable if conditions allow (more likely if the mean is not small), the positive integer nature of the response makes it clearly nonnormal making it generally safer to use PROC GENMOD. Also, if the data are counts, the assumption of equal variances needed by PROC GLM is also probably violated. If you fit a Poisson model, you should check for evidence of overdispersion which is common and consider instead using the negative binomial distribution or other alternative if overdispersion exists. See this note. Also, when you say the response is positive, if you mean it cannot be zero, then you might want to fit a truncated Poisson or negative binomial model as discussed in this note.