Dear Nevine,

First of all, I think you are on the right track: if you have panel (or longitudinal) data with multiple observations nested within persons and aim to explore the population-averaged effect of a covariate on an outcome, GEE analysis is the way to go. This allows you to estimate the effect in the entire population by ignoring (or not taking into account) individual risk relationships or trajectories. If you, however, are more interested in individual-specific effects of a predictor on an outcome, you should better make use of mixed or random-effect models that allow you to model intra-individuality specifically by including random terms (intercept / coefficients) into your model. For more information on GEE vs. mixed models, please have a look at:

https://stats.stackexchange.com/questions/16390/when-to-use-generalized-estimating-equations-vs-mixed-effects-models

https://www.researchgate.net/post/Where_can_I_find_good_material_on_the_difference_between_mixed_models_and_gee_models

Second, with regard to the time perspective, each of your three models has its advantages. However, it strictly depends on what you want to know. Assuming you perform some sort of randomized clinical trial (RCT), it usually defines a specific endpoint (one outcome at a specific follow-up time). For example, in your case, the endpoint could be the severity of depression at six months postpartum in order to evaluate the effectiveness of some kind of treatment at the end of the observation period. If you, however, are interested in short-term vs. long-term effects, it seems appropriate to separate the analysis by time and present the results separately (0-3 months vs. 0-6 months). Please note, however, that there may well be differences between the effects, depending on the time perspective you are looking at. For example, the short-term effect can be significant in itself if the trial was allowed to end after 3 months (see Fig A). However, if you consider the entire period and assume a linear effect over time, the short-term effect, relative to the overall trend, may no longer be significant. Thus, if looking at the entire period of 6 months, the effect at 3 months is still present, but in comparison to the overall trend it is “not large enough” to be statistically significant and is masked by the overall trend (see Fig. B).

In summary, you should look at your main hypothesis and present the results accordingly. Secondary results can be presented in a subordinate way or in the form of sensitivity analyzes.

Hope, this helps. Good luck for your analysis!

Thank you so much Alexander Pabst for your answer. This is very helpful, I really appreciate it! Indeed, I am interested in short-term vs. long-term effects of predictors on postpartum depression that's why I performed two separate GEE analysis in addition to the full model of the entire 6-month period. Also, as you mentioned certain effects were still present in the separate analyses whereas others were not large enough to be statistically significant and as they were masked by the overall trend.

I found this quite interesting as it showed the predictors with larger effects.