I hope this is the the right video for calculating minimum detectable effect size if I see an observational study published in a paper and I am reviewing it for discussing in journal club? My main concern is not to jump to an erroneous conclusion of equivalence based on an underpowered observational study which did not even mention any power analysis. This misunderstanding has a potential for negatively impacting patient care. Is there an article and is there a calculator?
What if the reason that the treatment group was given the treatment (i.e, not random assignment) is correlated with the treatment group's trend? In other words, what if assignment of treatment was done non-randomly precisely because a certain group in the population was identified to have a different trend than other groups? Is there any economic/statistical check for that?
Thank you so much, very helpful! Regarding the subgroup analysis: if the interacton coefficient is not significant, would that mean that the subgroup are different in the sample at hand, but that there is no statistical significance for it? Thanks in advance for the clarification :)
Best explanation series in the RU-vid I have watched so far. Thanks Professor for each video on this serie. Worth to note, the videos are really underrated.
Thank you professor for your nice and detailed presentations! If we are using Probit model and there will be a hetroskedasticity , can we report the marginal effect coefficients or totally leave the model use only the results of LPM? Thank you!
Professor, thank you so much for a very clear explanation of DID. I was having a hard time to understand this method but through your video, it helps me to understand it clearly.
Wow! What an elaborate way of explaining the DiD concept. This is the best lecture so far. Thanks so much sir, i have learnt alot. kindly help me understand, incase there are three groups (treated, control and pure control) in an RCT experiment, how do you estimate the DiD?
Thank you!!! Can we have two stratification variables? In my case, I need to stratify my control and treatment groups based on gender and governorate to have equal representation of both. Is it doable?
Great video! Question though: with training2 we saw that there was good evidence for common support. However, we also saw that ~50% of treatment population had a pscore2 of <0.4. This means that the Probit model is not doing a great job of fitting the data. So while we have strong evidence for common support, does it come at the cost of putting our selection on observables assumption under question?
Great video! Thanks! Two quick questions: first, isn't there redundancy in saying Average Treatment Effect "on the treated"? Treatment effect will _always_ be on the treated by definition, isn't it? It makes me think if there can be something like Average Treatment Effect on the untreated which is absurd. Second, if we have a large enough sample, and the observable covariates balance, does it also guarantee that the unobservables would also balance?