In his engrossing book "Naked Statistics" Charles Wheelan begins to explain how controlling for variables works by stratifying the sample. However, he stops short of explaining the reaggregation, leaving one intrigued by the stratification analogy but confused as to whether- and if so, how- it can be applied.

To be more specific, he is looking to isolate the effect of education on weight, by controlling for gender, height, and income. He imagines an experiment in which a number of people (constituting the sample) all gather in one place. Then the stratification begins:

  1. Men and women are separated
  2. The men and women are separately subdivided by height. At this point, as he says it "There will be a room of 6-foot tall men. Next door, there will be a room of 6-foot 1-inch men"
  3. Finally, each of these rooms (identified by combinations of gender and height) can be further subdivided by income. As he puts it: "Eventually we will have lots of rooms, each of which contains individuals who are identical in all respects except for education and weight, which are the two variables we care about. There would be a room of forty-five-year old 5-foot 5-inch men who earn 30,000 to 40,000 a year".

While we can regress weight on education in each of these rooms, how useful is that given that we are looking to end up with a single coefficient that encapsulates the relationship between weight and education observed in each of these rooms?

While I understand that data insufficiency for "rooms" obtained by stratifying by a large set of variables, and continuous variables, is a challenge, I am hoping that someone can come up with a simple example with made up numbers in which we stratify by 2 binary variables (edit: or use my example below), to keep things simple.

What magic analysis would need to be performed using data of the people in each of these rooms, which then could be aggregated to obtain the regression coefficient which can be interpreted as the effect of the independent variable of interest on the dependent variable, after we control for the two binary variables? How exactly would the aggregation work?

While I found this question/discussion How exactly does one “control for other variables”? interesting, controlling for a variable by its inclusion in a regression is insightful but not a very intuitive answer. Stratification and Reaggregation, if it is possible, shows quite a bit of promise of intuitive understanding.

Edit: To put some numbers behind it, with the completely made up data below, I obtain the following regression results: On the left are the results of the simple regression of weight on education (after stratifying by income and gender) in each of the 4 rooms (each room with 15 people/observations), whereas on the right are the results of the multiple regression of weight on education, income, and gender (60 people/observations combined).

Is it possible to aggregate the results of the regressions obtained in each of the 4 rooms (stratified by income and gender) to obtain the coefficient on education obtained in the multiple regression (-3.28), in which we control for income and gender?

enter image description here enter image description here

  • $\begingroup$ Question 1: You already has an example. Question 2: I think it is impossible, at least I do not find a way to combine them. $\endgroup$ – user158565 Dec 7 '18 at 4:42
  • $\begingroup$ Question 1: The example was a later addition. Feel free to use your example or use mine. Question 2: @Fomite in his answer to this related question stats.stackexchange.com/questions/17336/… says "The easiest, and one you came up with, is to stratify your data so you have sub-groups with similar characteristics - there are then methods to pool those results together to get a single "answer". This works if you have a very small number of variables you want to control for.." His answer was liked 110 times. Sounds like he says it is possible. $\endgroup$ – ColorStatistics Dec 7 '18 at 5:09
  • 1
    $\begingroup$ When you say "he is looking to isolate the effect of education on weight", what does he mean by effect? Does he mean causal effect? $\endgroup$ – Carlos Cinelli Dec 11 '18 at 2:46
  • $\begingroup$ Thank you for the question, Carlos. No, the context here is not causal, but merely one of association. $\endgroup$ – ColorStatistics Dec 11 '18 at 2:50
  • 1
    $\begingroup$ @ColorStatistics so what do you mean by "isolating the effect of education on weight"? $\endgroup$ – Carlos Cinelli Dec 11 '18 at 2:54

You can use a technique known as Mantel-Haenszel pooling to get a single summary estimate out of a stratified estimates.

There's a good treatment here.

  • $\begingroup$ Thank you very much for the reference. I will study it and follow up. $\endgroup$ – ColorStatistics Dec 11 '18 at 15:10

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.