Fixed Effects vs Random Effects: Which Should You Use?

If you work with panel data—the same units observed repeatedly over time—you will sooner or later face the fixed-effects-versus-random-effects decision. It is a genuine fork in the road: the two approaches can give different estimates, rest on different assumptions, and answer subtly different questions. Yet it is often decided mechanically, by running a single test and following whatever it says, without asking what the test actually means. That habit produces analyses that are hard to defend when a reviewer probes them.

This article explains what fixed and random effects really do, what separates them, and how to make the choice deliberately—grounded in your question and your data rather than a reflex. It sits alongside our broader guide on choosing the right statistical method, and it reflects how we approach panel work within our Longitudinal, Panel & Multilevel Research practice.

What panel data lets you do

The reason this choice exists at all is that panel data has a structure ordinary cross-sectional data does not. When you observe many units—firms, individuals, countries—each over several time periods, you can separate two kinds of variation. There is within-unit variation: how a unit changes over time relative to its own average. And there is between-unit variation: how units differ from one another on average. This separation is the great advantage of panel data, and the fixed-versus-random choice is fundamentally a choice about which of these variations your estimate should use.

The concern that motivates the whole discussion is unobserved heterogeneity: stable characteristics of each unit that you cannot measure but that affect the outcome. A firm's management culture, an individual's innate ability, a country's institutional history—these are hard to quantify, yet they shape outcomes and are often correlated with the variables you can measure. If they are ignored, they contaminate your estimates. Fixed and random effects are two different strategies for dealing with exactly this problem.

What fixed effects do

A fixed-effects model gives each unit its own intercept, effectively asking: as a unit changes over time, how does its outcome change with the explanatory variables? Because the comparison is always within the same unit, everything about that unit that does not change over time—observed or unobserved—is automatically controlled for. The management culture, the innate ability, the institutional history: all of it drops out, because it is constant within the unit and the model only uses within-unit variation.

This is the great strength of fixed effects. It controls for all time-invariant characteristics of each unit, even ones you never measured and could not name. For that reason, fixed effects are often the more defensible choice when you are making a causal claim and worry that unobserved, stable traits of your units might be biasing the relationship—a concern at the heart of credible causal inference. The cost is twofold: fixed effects cannot estimate the effect of anything that does not vary over time (a variable that is constant within a unit is absorbed along with the intercept), and by using only within-unit variation they are less statistically efficient—they discard the between-unit information entirely.

What random effects do

A random-effects model treats each unit's individual effect not as a separate parameter to be estimated, but as a random draw from a distribution. In doing so it uses both within-unit and between-unit variation, which makes it more efficient—it extracts more information from the same data and produces smaller standard errors. It can also estimate the effects of time-invariant variables, which fixed effects cannot.

But this efficiency comes with a significant condition. Random effects rest on a crucial assumption: that the unit-specific effects are uncorrelated with the explanatory variables in the model. In plain terms, it assumes the unobserved characteristics of each unit are not systematically related to the things you are measuring. When that assumption holds, random effects are the better choice—more efficient, and able to answer questions fixed effects cannot. When it fails—when the unobserved traits are correlated with your predictors, which is common in observational social-science and management data—random-effects estimates are biased, and fixed effects are the safer choice.

The Hausman test—and its limits

The standard tool for this decision is the Hausman test, which compares the two sets of estimates. The logic is straightforward: if the random-effects assumption holds, the two models should give similar coefficients; if the assumption is violated, they will diverge. A significant Hausman result is usually read as evidence that random effects are inconsistent and fixed effects should be preferred.

The test is useful, but treating it as the sole arbiter is a mistake, for several reasons. It is an aid to judgement, not a replacement for it. Like any test, its result depends on sample size and power, and it answers a narrow statistical question—whether the two estimators differ—rather than the substantive question of which model your research actually calls for. A reviewer who sees a decision justified purely by “the Hausman test said so,” with no reasoning about the data or the question, will rightly be unconvinced. The test should inform the decision; it should not make it alone.

How to actually decide

A sound decision weighs three things together, of which the test is only one.

The first is your research question. Are you interested in how units change over time (a within-unit question), or in explaining differences between units, including the effects of characteristics that do not change (a between-unit question)? If a time-invariant variable is central to your hypothesis—a person's sex, a country's legal origin, a firm's founding conditions—fixed effects cannot estimate it, and that alone may decide the matter.

The second is the plausibility of the random-effects assumption in your specific context. Ask honestly whether the unobserved, stable traits of your units are likely to be correlated with your predictors. In much observational data—especially in management, economics, and the social sciences—they plausibly are, which tilts the balance toward fixed effects. This is a substantive judgement about your data, informed by the Hausman test but not dictated by it.

The third is what your field expects. Conventions differ across disciplines and even across sub-fields, and reviewers have expectations about how the choice should be justified and reported. In many applied microeconomic and management contexts, fixed effects are the default expectation for observational panel data precisely because they are more robust to the endogeneity concerns those fields care about; elsewhere, random or mixed approaches are standard. Knowing the norm in your target journal is part of making—and defending—the choice.

Beyond the binary

It is worth remembering that fixed versus random is not the whole universe of options. When your data has more than two levels—students within classes within schools, employees within teams within firms—multilevel (hierarchical) models generalise the random-effects idea to nested structures and are often the appropriate tool. Correlated random-effects and hybrid specifications can combine the robustness of fixed effects with the ability to include between-unit information. And when the outcome depends on its own past values, dynamic panel methods become relevant. The binary choice is the right starting point, but it is a starting point; the best specification depends, as always, on the question and the data.

The bottom line

Fixed versus random effects is not a decision to delegate to a single test. Fixed effects protect you from unobserved, time-invariant confounders at the cost of efficiency and the ability to study time-invariant variables. Random effects are more efficient and more flexible, but only when their assumption about unobserved effects holds. The Hausman test informs the choice; your research question, an honest assessment of your data, and your field's conventions should decide it. Made that way, the choice is defensible—and defensibility, under review and after, is the whole point.