Recap of Causal Inference Bootcamp: Causal Inference Bootcamp


[MUSIC] It’s been a long journey through all
of the different methods that you can use to learn about causality, and here I’m just
going to recap the main ideas that you should walk away with after having watched all of
these modules. First is the definition of a causal effect. A causal effect is the effect
of change in a treatment on an outcome variable, holding all other variables constant. That’s what we’ve been trying to get out all
along, and the way we do this is by using variation in the treatment that we see in
our data to get at these causal effects. And it’s very important that this variation is
exogenous, meaning it is independent of any other variables that may be affecting things,
so we don’t have any confounders going on. That’s what I mean by exogenous variation. Now, randomized experiments delivered us this
exogenous variation for free because the treatments are randomly assigned. They are not related
to anything else that’s going on, so if we just compare mean outcomes in the treatment
group with mean outcomes in the control group, we get our causal effects immediately. Instrumental variables also provide this kind
of variation, although we had to use a much more indirect route to get there. And we saw
that when this variation doesn’t exist in our data like with climate change – remember,
we never observe the planet Earth without humans having existed – then we have to use
other knowledge to try to construct the counterfactual prediction of what would have happened if
humans hadn’t existed on the planet in order to learn about causal effects. And we also
saw that regressions are going to show us correlations in general and not causal effects,
unless we assume that the variation in treatments in the data is exogenous – by assumption.
If we do that, then a regression actually does show a causal effect. When we have data over time – panel data – we
saw that if we do a before-and-after comparison, that’s only going to give us exogenous variation
if nothing except the treatment changed over time. But if we thought that there are things
changing over time that are affecting everybody, then maybe before-and-after comparison isn’t
such a good idea, and in that case talked about differences-in-differences, which removes
all time trends that are affecting all units equally, and therefore giving us some exogenous
variation to learn about treatment effects. And finally we talked about regression discontinuity,
which gives us exogenous variation around a certain cutoff, and so lets us learn about
causal effects for people close to that cutoff. So these are the main ideas behind all of
the methods we’ve talked about. And I hope you’ve enjoyed this sequence of modules and
are excited to learn more about causality. [MUSIC].

Leave a Reply

Your email address will not be published. Required fields are marked *