Question about dropout and fixed effect in panel data

[Jianhong Xu]

Dear Stata community

I have an imbalanced panel data about the survey of the health related quality of life of patients with COVID. There are several waves for them to complete the survey. However, some participants may dropout as they are recovered from the COVID and do want to state their HRQoL. For this kind of imbalanced panel data, how can I calculate the weight to deal with dropout to get accurate estimation?

Another question is that I decided to use fixed effect to find the predictors for their HRQoL like income, age, sex, etc. However, after I doing the fixed effect, the time-invariant variables like sex are droped due to collinearity. Although random effect can show the coefficients, hausman test suggests fixed effect. What should I do?

Thanks!

[Clyde Schechter]

What are your research goals. If your goal is to estimate the effects of time-invariant variables, then you cannot use a fixed effects model, regardless of what Hausman or any other test has to say about it. It is mathematically impossible to estimate effects of time-invariant variables in a fixed effects model. However, if your research goals do not require that you estimate the effects of these variables, and you are just doing it out of curiosity, or for fun, or for non particular reason, then just don't do it.

On the assumption that you really do need to estimate these effects, you need to use a random-effects model to do that. You might also consider using a hybrid model: see -xthybrid-, available from SSC.

[Jianhong Xu]

What are your research goals. If your goal is to estimate the effects of time-invariant variables, then you cannot use a fixed effects model, regardless of what Hausman or any other test has to say about it. It is mathematically impossible to estimate effects of time-invariant variables in a fixed effects model. However, if your research goals do not require that you estimate the effects of these variables, and you are just doing it out of curiosity, or for fun, or for non particular reason, then just don't do it.

On the assumption that you really do need to estimate these effects, you need to use a random-effects model to do that. You might also consider using a hybrid model: see -xthybrid-, available from SSC.

Hello Clyde

Thanks for your reply. It is really helpful. As for my first question about weight, do you have some suggestions about it?

[Clyde Schechter]

The general approach that I am familiar with is to develop a propensity model for attrition, typically a logistic or probit model. Then you calculate the predicted probability of being in-ample for each person at each time. Each observation is then weighted by the inverse of the probability of still being in the study at the time of that observation. It's explained a bit better than this at https://www.laterite.com/blog/managi...-with-weights/ (though in the context of having only two time periods--but the same principles apply).

But I'm not sure this approach can be applied to your data. If you are planning to use -xt- commands for your analysis, those require that weights be constant within panels, whereas this approach assigns to each panel a possibly different weight at each timepoint.

Sorry, but this is not an area with which I have in-depth knowledge. Perhaps others are following a long who can give you better advice.

[Jianhong Xu]

The general approach that I am familiar with is to develop a propensity model for attrition, typically a logistic or probit model. Then you calculate the predicted probability of being in-ample for each person at each time. Each observation is then weighted by the inverse of the probability of still being in the study at the time of that observation. It's explained a bit better than this at https://www.laterite.com/blog/managi...-with-weights/ (though in the context of having only two time periods--but the same principles apply).

But I'm not sure this approach can be applied to your data. If you are planning to use -xt- commands for your analysis, those require that weights be constant within panels, whereas this approach assigns to each panel a possibly different weight at each timepoint.

Sorry, but this is not an area with which I have in-depth knowledge. Perhaps others are following a long who can give you better advice.

Thank you so much!