How to deal with very low within-cluster variance in a two-level model? option "residuals"?

[Gianfranco Di Gennaro]

Dear all,
I have a dataset in which I have a binary variable as an outcome (adverse events, var "event") detected at different timepoints (var "followup") and the patients ("id") are clustered in different hospitals (var "center").

The objective of my analysis is to evaluate how some covariates affect the possibility of an adverse event.

Model
melogit event i.cov1 i.cov2 ....... || center: || id:
does not converge.

I believe that this is due to a very low within-patient variability (almost always a patient always has adverse events, or never).

I would try to assume an independence matrix of the within-patient residuals (to reduce the model to a logit model with the within-hospital clustering level only) but I can't figure out how to do it. I thought you could use a "residuals" option but that's not the case.

Anyone have an idea?

Thank you

Gianfranco

[Carlo Lazzaro]

Gianfranco:
I'd consider -xtlogit,re- with standard errors clustered at patient level (shared frailty for repeated adverse events) and -i.hospital- as a predictor.
If allowed by the number of hospitals, it may be interesting to investigate whether (or not) the type of hospital (I assume your dataset relates to the Italian setting) such as teaching hospitals, self-governing hospitals, Local Health Authorities-managed hospitals and their property (public or private) have any bearing, coeteris paribus, on the regressand.

[Gianfranco Di Gennaro]

Dear Carlo Lazzaro , thank you for your suggestion.
However I have 23 hospitals and estimating all these parameters would be difficult. I tried but the model has a lot of trouble converging.

xi: xtlogit Event i.cov1 i.cov2..... i.hospital , re vce(cluster id)

I'm trying to categorize into fewer categories, like you suggested (University Hospital, private Hospitals and so on).

My question: I've tried using a marginal estimation (GEE) via xtgee and corr(ind). I used the 23 hospitals as fixed factor and it easily works.

xtset id time
xtgee Event i.cov1 i.cov2..... i.hospital, corr(ind)


However, Do you think it is possible to decompose the correlation between two levels, hospital and id in the xtgee model?
So to avoid using Hospital as fixed factor.

Thanks again Carlo Lazzaro , have a nice weekend.

Gianfranco

[Carlo Lazzaro]

Gianfranco:
admittedly, I've never challenged myself with this topic.
As far as I know -corr()- option available form -xtgee- does not allow the decomposition you're interested in (but I could be wrong).
That said, I'd take a look at https://journals.sagepub.com/doi/10....36867X19854021.
Have a great W_E you too.