Ega Kurnia Yazid

Ega Kurnia Yazid

Subsidy Policy Specialist. MA at Columbia University. BSc in Econ. Former CSIS. Interested in navigating data science to address energy, environmental, and health problems

Are fixed-effects models sufficient for causal inference?

4 minute read

Photo by Robert Bye on Unsplash

Doing an experiment is by far the gold standard for conducting causal inference research. One common practice, particularly in quantitative social science, is doing randomized control trials (RCTs).

But here’s the scoop: if you’re a young researcher and just starting out, doing RCTs can be super expensive. In fact, for young researchers looking to build their portfolios early in their careers and pave the way for future success, the high expenses associated with RCTs can pose a significant challenge.

Pivoting the Gold Standard

One plausible strategy is they ought to join a legit RA-ship program to get the opportunity of doing RCTs. Otherwise, those who are less lucky should think of an alternative such as conducting quasi-experiment research.

Several methods are available to execute such studies. Popular instances are Difference-in-Difference (Diff-in-Diff), Propensity Score Matching (PSM), Regression Discontinuity Design (RDD), and finally Fixed-Effects (FE).

While the first three methods required specific conditions such as pre-post and treatment-control groups, I personally think that FE is one of the most versatile when it comes to natural experiment design - this opinion may not be correct universally but I think it’s true for most of the cases.

But one major question arises when we prefer this pivot. Is adopting these models sufficient for claiming causal inference?

This article, thus attempts to answer such question by exploring the possible reasonings and cautions in implementing the FE for natural experiment design with panel datasets.

Revisiting FE models

Photo by Bekky Bekks on Unsplash

A. General Setup

To begin with, let’s check the following general setup for FE model using typical panel datasets:

where y is the expected outcome, predicted by vector covariates x with coefficient β, fixed-effects term α, and error-term ϵ indexed at unit i time t.

B. How FE Models Help Identify Causal Effects?

You’ve probably heard: “Correlation is not causation.” That’s true. But fixed-effects models offer a partial fix - they difference out all unchanging confounders over time.

If you believe that the omitted variables are constant over time (e.g., innate ability, geographic features), the FE transformation eliminates them. This reduces omitted variable bias and brings us closer to estimating the causal effect of x​ on y​.

For example, suppose you’re evaluating the impact of internet rollout on school test scores across districts over a 10-year period. District-level differences in wealth, governance, or culture may bias your results - but if those differences are constant over time, a fixed-effects model accounts for them.

So, FE is valuable - if your theoretical framework justifies it.

C. Beware the pitfalls

FE models sound great. But they’re not a silver bullet. They require strong assumptions, and failing to meet them can lead to misleading conclusions. Below are key issues to watch out for:

Time-Varying Confounders: FE only controls for time-invariant factors. If there are unobserved variables that change over time and also affect both treatment and outcome (e.g., economic policy shifts, shocks, trends), your estimates will still be biased.

Reverse Causality: What if y also affects x? This two-way relationship (endogeneity) is common in social science and violates the assumption of exogeneity. FE models don’t solve this.

Low Within-Unit Variation: FE estimates are identified only from changes within units over time. If your treatment variable doesn’t change much (e.g., a policy that’s adopted once and never changes), you’ll have little identifying variation, and your standard errors may explode.

Measurement Error: Differencing can amplify measurement error, making estimates noisy and biased toward zero.

Bottomline

Here’s a quick decision guide:

✅ Use FE when:

  • You have panel data with enough time periods and variation within units.
  • You believe time-invariant unobserved heterogeneity is a major concern.
  • You suspect confounding, but not by time-varying factors.

❌ Avoid relying solely on FE when:

  • Key confounders vary over time.
  • Your treatment assignment is endogenous or persistent.
  • There’s little variation in treatment within units.

In such cases, you should consider combining FE with:

  • Instrumental Variables (IV) for endogeneity
  • Event study or dynamic DiD to model pre-trends
  • Matching or control function approaches

So, are fixed-effects models sufficient for causal inference? Sometimes - but not always.

FE models are powerful tools, especially when used carefully with rich panel data. But their success in identifying causal effects depends on the assumptions you make and your ability to test them.

A good rule of thumb? Think of FE as a useful method - not a magical one. It’s strongest when supported by a solid theoretical framework, robustness checks, and complementary methods.

This post was written by Ega Kurnia Yazid.

A more detailed technical explanation please check: Fixed effects and Causal Inference

I welcome feedback and constructive criticism. To discuss further you can comment on each post, or you can reach me through any social media platforms you can find on my webpage’s profile.

You May Also Enjoy

How should we respond to electability polling?

Photo by Elliott Stallion on Unsplash

Approaching Indonesia’s presidential election on this year’s Valentine’s Day, pollsters are racing to publish the presidential candidates’ electability statistics. Some argued that the polling results did not represent the actual data, some said that the pollsters got paid by one of the candidates, and some just agreed with the polling results. Thus, the difference between public opinions and vague perception of presidential electability polling results raises a question of how the public should respond to the polling results.

And this article will discuss the issue further.

TL;DR, it’s a very simple question, but to answer it requires a complex thought process. For me, we first need to answer the questions by asking further questions such as Why we need the electability polling in the very first place? Why we cannot entirely trust polling results? When can we trust it? And what to do about it?

Setting up a virtual environment for a Python project

Photo by Kelvin Han on Unsplash

I need to write this on a post because I keep forgetting this by the time I need to create a new python project. Plus, I always get confused when figuring out to set up a Python virtual environment, so I did my research and created new blog posts to help other students in understanding the virtual environment.

What I wish I knew when I learned Data Science

I started to learn Data Science in college but began to actively use it when working as Research Assistant. During my learning journey, I met lots of unnecessary mistakes that ended up in the learning block. In this article, I will share with you several mistakes that you can avoid so that you can experience a fruitful and smooth learning process.