Module contents

In this module, we cover

The potential outcomes framework
Treatment effects
Randomized experiments

Slides

Potential_outcomes_and_experiments_2022.pdf

Follow-up slides:

Potential_outcomes_and_experiments_Follow-up_2022.pdf

Clarification:

At the end of the lecture, I asked if a certain scenario was an example of Simpson's paradox. I was too vague in my answer. No, it is not an example of Simpson's paradox. Because we are comparing the interventional distributions directly, the effect can't be positive on average at the same time as it is negative in all subgroups. Since $LaTeX: \mathbb{E}[Y(1)-Y(0)] = \sum_x p(x) \mathbb{E}[Y(1)-Y(0) \mid x]$ by the law of total expectation, and since LaTeX: p(x)>0 , it can't be that the average effect is positive, but the conditional effect is negative for all LaTeX: x . I've left this part out of the recording to avoid unnecessary confusion.

Recordings:

Follow-up recording:

Recommended reading / watching

Brady Neal's videos on Potential outcomes and Chapter 2 in ICI
Brady Neal's videos on Experiments and Chapter 5 in ICI
Hernan & Robins, Causal Inference: What if..., Chapters 1–2

Previous year's course [2021]

Make note that the previous year's slides and videos may not completely reflect the content of this year's course.

Slides:

Slides [2021], Follow-up slides [2021]

Corrections:

In the slides above, inconsistencies the notation of \Delta, \tau and Y(0), Y(1) remarked upon in the video have been corrected.

Part 1: A history of clinical trials

Part 2: Potential outcomes

Part 3: Causal effects

Part 4: Randomized experiments

Part 5: Limitations of experiments