[Md Bayzidur Rahman]

Hi Clyde, Thanks for the classical reference. One disadvantage that I am facing the interpretation of the adjusted rate. It is the average of the predicted rates from the model which is not same as the predicted rate for average covariate distribution (or for an average person).
My context: I am receiving individual level rate data (numerator and denominator separately) on a monthly basis for multiple aged care facilities. I need to calculate adjusted rates (adjusting for age, sex and 6 other comorbidities) for each facilities so that they can be compared with each other after removing the effect of those covariates. For this, the best approach is direct standardization. However, the arithmetic method of direct standardization using the dstdize command would not give reliable estimates because many of the specific rates for the cross-classification of the 8 covariates will be zero due to no observed events. In this case, the regression based direct standardization would be a better alternative as described in this paper (https://bmchealthservres.biomedcentr...472-6963-8-275). I am looking for any Stata code or routine to implement this with Poisson regression.

[Clyde Schechter]

which is not same as the predicted rate for average covariate distribution (or for an average person)

That is true. But if what you want is the predicted average rate for the average covariates, use -margins-' -atmeans- option (or the -at()- option if you only want to standardize at the means of some of the model variables.)

However, the arithmetic method of direct standardization using the dstdize command would not give reliable estimates because many of the specific rates for the cross-classification of the 8 covariates will be zero due to no observed events.

I don't understand this. To the extent I do, it is not correct. If the sample has no observed events in some cells of the 8-way cross classification, then the rates for those cells are zero, and they just get averaged in like any other cells' rates. Nothing wrong with that. What would be a problem is if some of the cells of the 8-way cross classification are empty: there are no observations at all. Then the rates for those cells are undefined due to denominator 0, and that makes the standardized rate not estimable.

I have not read the reference you linked in detail. But it looks off hand like you can get those results using the -at()- option of -margins-.

[Md Bayzidur Rahman]

Thanks for confirming that the arithmetic method is acceptable in absence of the numerator if denominator is present.
For the regression method, the at() option in margins would give the value of the covariate, but we need weight of that category instead.