Fixed Effects vs. Random Effects: When to Use What in Panel Econometrics
If you've worked with panel data — repeated observations of the same units over time — you've faced this decision: should I use fixed effects (FE) or random effects (RE)? The answer matters. Choosing the wrong estimator introduces bias and inflates standard errors. Yet the choice hinges on a single untestable assumption: whether unobserved unit heterogeneity is correlated with your covariates.
This guide cuts through the statistical fog with a practical decision framework. The goal isn't to make you a panel econometrician — it's to help you reason through the choice with your domain knowledge.
The Core Problem
The panel data model has an unobserved term α_i that captures all time-invariant characteristics of unit i (ability, location, genetic traits). The question is: is α_i correlated with your observed covariates?
If you're studying wages and education, does unobserved ability correlate with education? Almost certainly yes — smarter people both get more education and earn higher wages. The education coefficient is confounded by ability.
This single correlation drives the entire choice between estimators.
Fixed Effects: Robust but Restrictive
Assumption: α_i may be correlated with covariates. No restrictions.
What it does: Eliminates α_i by demeaning the data — subtracting each unit's average from all its observations. This wipes out all time-invariant confounds, whether observed or unobserved. In the wage example, ability differences across workers disappear; you estimate the wage effect of education from within-worker variation as workers' years of education change over time.
The problem: Any covariate that doesn't vary over time vanishes. You cannot estimate effects of gender, race, or static treatment status.
When to use it:
- You believe the omitted heterogeneity is correlated with your covariates
- You don't care about time-invariant effects (or they're not your focus)
- You have enough time-series data that within-unit variation is informative
FE is the defensible default in observational data. It trades statistical efficiency for robustness to unobserved confounding.
Random Effects: Efficient but Fragile
Assumption: α_i is uncorrelated with all covariates. The heterogeneity is "random noise."
What it does: Uses both between-unit and within-unit variation. RE is more efficient than FE — it has smaller standard errors — but only if its assumption holds.
When it breaks: In most real applications, α_i IS correlated with covariates. Workers with higher ability both earn more and get more education. Firms with higher productivity both invest more and hire more skilled workers. If you assume α_i is uncorrelated but it's actually correlated, RE is biased — and you won't know it from any test result.
When to use it:
- Units are random draws from a population (survey data, randomly sampled schools)
- Economic theory strongly implies exogeneity of α_i
- You're willing to report both FE and RE and justify why you choose RE over FE
The Hausman Test: Why It Often Fails You
The standard practice is to run the Hausman test: compare FE and RE estimates. If they differ significantly, reject RE and use FE. If they don't differ significantly, use RE.
This sounds logical but often backfires. Here's why:
The test has low power: In realistic settings with small time periods (T=3–10 years) and moderate samples, the Hausman test fails to reject the exogeneity assumption even when α_i IS correlated with covariates. The test is too conservative. You might falsely conclude "FE and RE give similar answers, so RE is fine" when in fact both are biased and only one direction cancels out by luck.
The test is fragile: Add or remove a single covariate and the test result flips — even though the underlying endogeneity problem hasn't changed.
Bottom line: Don't let the Hausman test drive your decision. Use it as a diagnostic, not a decision rule.
Correlated Random Effects / Hausman-Taylor: A Middle Path
When FE is undesirable (you need time-invariant effects) but RE seems risky (you worry about endogeneity), the Correlated Random Effects (CRE) / Hausman-Taylor approach offers a compromise.
The idea: Let α_i be correlated with covariates, but assume that correlation is explained by the time-averages of the time-varying covariates. This is less restrictive than RE but more flexible than FE.
Hausman-Taylor specifically: Partition covariates into exogenous and endogenous groups. Use lagged and averaged values of exogenous covariates as instruments to identify all parameters, including time-invariant effects.
Requirements:
- You must correctly partition covariates into endogenous/exogenous (this is hard)
- Time-series length must be at least 3
- You must accept that some unobserved heterogeneity remains (the Hausman-Taylor assumption)
Hausman-Taylor is more work but pays off when you need to estimate time-invariant effects and can't afford the "no endogeneity" assumption.
A Practical Decision Framework
-
Do I need time-invariant covariate effects?
- No → Use FE (unless very strong theory says exogeneity)
- Yes → Move to step 2
-
Do I believe unobserved heterogeneity is exogenous?
- Definitely yes (survey design, experimental assignment) → RE
- Probably no, or uncertain → FE or Hausman-Taylor
- Depends on which covariates → Hausman-Taylor with sensitivity checks
-
Run the Hausman test as a diagnostic, not a decision rule. Large differences between FE and RE → use FE. Small differences + domain reasoning suggests exogeneity → RE.
-
Report multiple specifications. Show FE, RE, and Hausman-Taylor (if applicable). Check that your key coefficient is stable across methods. Large swings = red flag that endogeneity is unresolved.
Key Takeaways
- FE is the endogeneity-robust default in observational panel data. It sacrifices efficiency for bias-robustness.
- RE requires strong, often unrealistic, assumptions. Use it only when exogeneity is theoretically justified (random samples, experimental designs).
- The Hausman test is conservative and fragile. Don't rely on it alone.
- Hausman-Taylor / CRE offers a middle path for cases where FE is too restrictive and RE too risky — but it requires care in specification.
- Always justify your choice with domain knowledge, not just statistical tests. Your audience will trust the reasoning more than the test result.
The bottom line: Panel econometrics forces you to make a substantive assumption about what drives unobserved heterogeneity. No test can verify it. Own that assumption, state it clearly, and show robustness across methods. Further Reading:
- Wooldridge, J. M. (2010). Econometric Analysis of Cross-Section and Panel Data, 2nd ed. Chapters 10–15.
- Hausman, J. A., & Taylor, W. E. (1981). "Panel Data and Unobservable Individual Effects." Econometrica 49(6): 1377–1398.