@JorgeTeixeira see inline below:
I am not sure that there is complete consensus on anything in statistics.
That being said, the use of the baseline measurement as a covariate, that is, as an “adjustment” for any baseline imbalance between the groups, even in a randomized study, is, in my mind, the preferred approach and there have been numerous publications on this topic. In essence, in a repeated measures setting, it is an extension of the traditional ANCOVA approach, where the usual OLS model formulation would be:
FU ~ Group * Baseline
Where FU is the single follow up measurement (one record per patient), Group is the treatment group factor and Baseline is the baseline measurement.
There is a recent thread regarding the use of change from baseline in this forum that raises related issues and has numerous references to the use of ANCOVA:
Changing the “Status Quo” analyses for Regulatory Submission - Change from Baseline
There is one reference listed therein:
https://www.tandfonline.com/doi/abs/10.1080/10543406.2011.550113
that you may find of particular interest in the context of a repeated measures setting, such as you have.
As per the ANCOVA formulation above, it is the explicitly stated interaction term between the treatment group and the baseline measurement. Rather than using, in R syntax, the more concise:
Group * Baseline
which embodies both the main effects and the interaction terms, it is using the longer syntax of:
Group + Basline + Group:Baseline
to more explicitly express the model terms.
My own preference would be M4, for the reasons we have been discussing, where, as I noted in my reply, you have two, not three, records per patient, given the timing of your observations.