Post estimation in ipdmetan: error: last estimate not found

[Bisal Kadariya]

Hello everyone,

I am trying to compute population attributable fraction and population attributable risk of low birth weight (LBW) due to anemia using punaf and regpar after a logistic regression with ipdmetan

ipdmetan, study(study) or random keepall: logit LBW i.anaemia i.malaria en_age i.low_BMI i.gravidity

punaf, (at anaemia==0), subpop(if anaemia==1)

However, i receive error message r(301) last estimation result not found, nothing to store.

How do i run post estimation command for estimation in ipdmetan?

[David Fisher]

Hi Bisal,

The commands punaf and regpar appear to be user-written commands contributed by Roger Newson. In particular, they are described as "a suite of programs that call margins and nlcom to calculate scenario prevalences and means, their differences, their ratios, and other comparison statistics". I am not familiar with these commands, so please bear that in mind when reading the rest of my message!

The command ipdmetan is also a user-written command (contributed by myself). It is a two-stage meta-analysis command, which means that it estimates a single pooled effect estimate and standard error from the data of multiple separate studies. However, it is not what Stata terms an "e-class" command; that is, it does not estimate a set of model coefficients and a corresponding variance-covariance matrix. As such, the commands margins and nlcom are not applicable. This is, indirectly, what your error message is telling you.

To put this another way: in your message you talk about fitting "a logistic regression with ipdmetan". In fact, that is not what you have done. You have (via ipdmetan) fitted a series of separate logistic regressions to the data from each of your studies, taken the "anaemia" model coefficient from each, and pooled them in a meta-analysis to find their weighted average.

If you truly have IPD meta-analysis data here, then you have two main options as far as I can see: either (a) to estimate your attributable risk fractions (or whatever) within each study separately, and then pool them together (assuming they can be transformed to an appropriate scale for pooling); or (b) to fit a one-stage meta-analysis model -- that is, a panel or mixed logistic model -- and then apply punaf and regpar to this model.

I hope that helps!

Best wishes,

David.