Causal Effects via Propensity Scores | Introduction & Python Code

This is the 2nd video in a series on causal effects. Here I introduce the Propensity Score and discuss 3 ways we can use it to compute causal effects from observational data. At the end, I share a concrete example with code of what using these methods might look like in practice.

Resources:
- An Introduction to Propensity Score Methods for Reducing the Effects of Confounding in Observational Studies by Peter C. Austin

--

Introduction - 0:00
Observational vs Interventional Studies - 0:32
Propensity Score - 3:25
3 Propensity Score-based Methods - 4:56
1) Matching - 5:18
2) Stratification - 9:07
3) Inverse Probability of Treatment Weighting - 10:37
Example: ATE of Grad on Income - 12:29
Word of Caution - 15:46

[journey-in-pixels]

Thanks for taking time to put this video, appreciate the working example in addition to theory.

[ifycadeau]

Great video as always! Looking forward to more. 😃

[youtubeuser]

Thank you for this. Can you please do a similar video or series on uplift modeling. There are lots of videos and literature explaining the concept but not enough examples of practical application.

[christopherosariemeniyangb]

Thanks for explaining Propensity Score Matching. How do I get the complete python code (Jupyter notebook) to run the nearest neighbourhood matching method. I already have my data

[FallenJakarta]

thank you for the great video! i have question, as you mention the original data i found the original data is different from what you use in this video. why did you make feature enginering to the data? please answer my question, thank you in advance

[programmingwithjackchew]

hi sir what if i totally don't know the variables and want to measure the causal effect sub kpi to main kpi, in this case, I don't know what sub kpi should be classified as confounder . Can you gimme suggestion

[varun]

Hey, is there any dataset that you could recommend, where I can work on this method?

[surajjha]

While working with observational data, how we decide how much sample size for results to be statistically significance?

[emilyqian]

do you need to test if the result is statistically significant?

[shadiaelgazzar]

thanks for your amazing video, but i have a question can i use this line
to load dataset instead of
df=pickle.load(open('df_prospenty_score.p', 'rb'))

[mauriciomandujano]

Thank you for such digestible and concise video!
Do you mind if I shoot you an email with a couple q's?

[pachakhan]

Is there any email that I can contact you? I am working on PSM!

[EVUTube]

This is really wonderful, but I'd like to offer a suggestion. The suggestion has to do with cadence (i.e., the musical quality of speech and where you place emphasis in the sentence).

When you start out the video, you are speaking in a fairly natural way and you use emphasis in an appropriate way. In other words, if you emphasize what needs to be emphasized. However, by the half-way point in the video (especially in the section about programming), you slip into an ossified, mechanical cadence which is repetitive, static, and which is divorced from what needs to be emphasized. The emphasis is no longer falling on the term / idea which needs emphasizing, you're just emphasizing because you happen to be near the end of the sentence. It begins to feel like you're trying to hypnotize us - like a tour guide who has offered the tour too many times to be awed by what is being shown. The programming part is really important - please don't drone thru it. Use cadence and emphasis to signal to us that you're engaged, that this is important, and that we should be paying attention.

To me, cadence and emphasis are important signalers in a presentation - that is why parents (when reading to children) use cadence to amplify what is happening in the story. If you're placing emphasis on words which don't need them, you are effectively confusing or misdirecting the listener. It is like reading a story to a child, but using a spooky-voice when telling a story about a beach trip and a happy voice when talking about monsters under the bed. If the programming is important, don't use a cadence reminiscent of someone snoring.

In your case, your cadence during these hypnotic periods is characterized by low, flat, fairly rapid rolling speech terminating in a loud WORD trailed by a deflation. It is almost like you are bored of this part or want to get thru it quickly and that it is not very important. blah-blah-blah-BLAHhh. blah-blah-blah-BLAHhh.

Please embrace that your cadence is a powerful tool for you to wield wisely. Don't let your cadence confess to us that you're bored. As silly as it sounds, your cadence helps students retain information.

[EVUTube]

I'll offer some constructive criticism regarding the cadence of your delivery. Obviously, you can take it or leave it. When starting to speak a sentence you speak fairly rapidly with a low volume, and low variability, but near the end of each sentence, you select one of the penultimate words upon which you will vastly increase the volume and slow the pace. Then start the next sentence very low and rapid and again end slow and loud. The challenge is that it is not as though you placed an emphasis on a term that is particularly worthy of emphasis - it just happens to be one of the words at the end of a sentence. It is just a rote pattern played over and over and over. It reminds me of a tour-bus operator who has been giving the same tour for the last decade and is bored to tears. For me, personally, I find this pattern very irritating and a little like Chinese water torture and since the emphasis is placed on terms not requiring emphasis, my attention is diverted to low-information terms.
This is probably the natural way you speak and not something you can change. But just compare you delivery to the delivery of news casters, actors, comedians, any other entity that presents information verbally - I would argue that this terminal-spiking patter pattern is not optimal for the listener.