Cluster-Robust Regression in Python

This is a short blog post about handling scenarios where one is investigating a correlation across some domain while controlling for an expected correlation across some other domain. If that sentence made no sense to you don't worry, here is a simple example:

Research Question:

To determine if Instagram users are more likely to also post on Facebook or to Twitter.

Analysis Plan:

Perform a t-Test to look at the difference in means of how often Instagram users are posting on Facebook vs. Instagram users posting on Twitter.

Problem:

Some users have Facebook accounts (Group A), some users have Twitter accounts (Group B) and some users have both (Group A/B). So we can't really use a Related Samples t-Test or an Independent Samples t-Test. Also, some users post many photos on Instagram and others only make the occasional post. We could therefore expect that the difference between users may be larger than the difference between what we actually want to measure which is Facebook vs. Twitter posting.

Solution:

Cluster-Robust Regression! We will perform a simple regression where standard error is calculated on a "Per-User" basis.

You can read much more in-depth about this technique here.

The Simple Method

Clustered Regression is supported and very well documented in Stata using the vce(cluster) argument.


The (Free) Python Method

After scouring the web and reading countless confusing and poorly documented pages I discovered that the Python StatsModels Library actually does support clustering in its OLS Regression Package, even though it can be extremely difficult to find.

Sample Data

We will use the UCI School Performance Dataset where each line is has information about a specific student and their grades in a course. This dataset is relevant because the students came from 2 different schools, so it is reasonable to suspect that performance may be similar within the 2 schools. We will see if the number of absences is correlated with final-grade mark and cluster based on "school attended"

Writing Code

Lets begin by loading in the dataset and printing the first few lines:

Next lets isolate our dependent and independent variables and run a "normal" un-clustered regression:
Output:
A coefficient of 0.0196 and a p-value of 0.497 indicates no significant correlation

Lets now perform the robust regression with clustering on "school" to control for the similarities within schools.

Output:
A coefficient of 0.0196 and a p-value of 0.143 indicates no significant correlation

Still, there is no significant correlation, however; there was a notable change in p-value.

Further Reading

I would highly recommend anyone who is interested in this type of modelling to read A Practitioner’s Guide to Cluster-Robust Inference as it goes into much more detail on the statistics behind the code and gives a thorough explanation of performing clustered regression in Stata.

Source Code

Full project repository available here

Comments

Post a Comment