Should we model $Y$ or change from baseline? $Y$ is measured at baseline and then after 1 and 2 hours ($Y_0$, $Y_1$, $Y_2$).
I want to fit a model against baseline and treatment arm in order to compare the two treatment arms in terms of change from baseline.
Does it make a difference to use $Y_j$ or $Y_j - Y_0$ as response ($j = 1, 2$)? Or both models are equivalent?
 A: The most commonly used approach in randomized controlled trials is to use a model (such as e.g. analysis of covariance) with $y_0$ as a covariate in the model. This is generally regarded as a good practice. In that case, it does not matter, at all, as far as the estimated treatment difference is concerned, whether you look at $y_j-y_0$ or $y_j$. If one does not use $y_0$ as a covariate, then it does make a difference.
A number of authors have discussed this, e.g. Senn in his book "Statistical Issues in Drug Development", in this article

Senn, S. (2006). Change from baseline and analysis of covariance
  revisited. Statistics in medicine, 25(24), 4334-4344.

and probably a lot of other authors, too.
A: I imagine you do an experiment where you prepare some object with a substance, measure the amount of the substance $Y_{0}$ and then apply the treatment (or do nothing) and then measure the amount of the substance $Y_{1}$ and later $Y_{2}$. So you are investigating growth or something comparable.
Here is an example where using $Y_1$ or $Y_{2}$ instead of $Y_{1} - Y_{0}$ or $Y_{2} - Y_{0}$ could go wrong:
Imagine you accidentally applied too much of the substance in the treatment arm samples. If you then only fit $Y_{1}$ or $Y_{2}$, then you might be fitting just that difference in the amount of applied substance instead of the actual growth. You avoid sources of error by fitting $Y_{1} - Y_{0}$ or $Y_{2} - Y_{0}$. (However, this example is a bit flawed: your growth could also depend on the amount of the substance present at $Y_{0}$, then your results are still affected by the different amounts of substance at $Y_{0}$. In that case you would need to control for $Y_{0}$ in the model.)
Also, you state your research question as 

... to compare the two treatment arms in terms of change from baseline.

So, you are actually only interested in the change from baseline after applying a treatment or doing nothing.
I would say it is always better to measure your target directly instead of measuring a proxy.
