Causality: Controlling

The sixth video in a series on causality. This video discusses what it means to control/adjust for a variable in order to close a back door, and walks through the mechanics of how it can be done. See more information about the class at nickchk.com/econ305.html.

[@armansultanov6879]

Thank you for making things more explicit!

[@metfan46]

Thanks for these videos, they're great!

[@bakther]

Excellent as usual.

[@donottraveltoiran2461]

Thank you very much for your explanation, I have got a question for you, if I want to find out the job satisfaction (dependence )how can be affected by other variables (400 independent ) how can I find that ?

[@NickHuntingtonKlein]

With that many predictors I might recommend regularized regression (LASSO)

[@toni_canada]

Thank you for your video! I have a question, ¿how you control a variable when is a continuous one? I mean, in the video W was 0 or 1, but how can we control a variable when it's continuous? Thanks!

[@NickHuntingtonKlein]

There are several ways. You can use the continuous variable to predict both treatment and outcome using any method that accepts a continuous predictor (like regression or a random forest) and subtract out the prediction. Or you can use a matching-based procedure.

[@arehmankhn]

Thank you for explaining the concept so clearly! What is the formula one can use in R when doing linear regression and controlling for several variables? I've already done the simple regression without controlling for the variables, and would now like to run it controlling for several variables.

[@NickHuntingtonKlein]

If your regression formula without controls is, for example, y ~ x, then you can add controls w and z with y ~ x + w + z

[@arehmankhn]

Hi! Where would felm and dummy variables enter the picture? I'm new to R and each web search seems to produce 100 different answers. Until I found this video, I wasn't even clear on the concept of "controlling." I'm looking for the simplest / most intuitive way to control for fixed effects in R.

[@NickHuntingtonKlein]

@@arehmankhn Check out my video on the estimatr package (although felm is good too)

[@larawhite5890]

@@NickHuntingtonKlein I think that if both y and x depend on w and z then interaction terms xw and xz should be added.

[@NickHuntingtonKlein]

@@larawhite5890 You generally want to add interaction terms if the effect of x is modified by w and z. If you just think x depends on w and z, then adding them as controls is sufficient.

[@user-cl2pd2po1t]

Who are you? can you provide some information into the video description?

[@NickHuntingtonKlein]

I'm an economics professor at CSU Fullerton

[@donoiskandar6820]

60% go when sick, 10% go when not sick. thus 60 - 10 = 50% of going to the doctor is explained by being sick. are you assuming that the total sample of all people who are not being sick and those who are already being sick is identical?

[@NickHuntingtonKlein]

Nope! It still works with uneven sample sizes.