What Is a Fixed Effect vs Random Effect?

When analyzing data with repeated measurements or grouped observations—students within schools, patients within hospitals, or countries over time—researchers face a fundamental modeling decision: should group-level differences be treated as fixed or random effects? This choice shapes everything from parameter estimation to the conclusions a study can support.

Fixed effects treat each group’s influence as a distinct, unknown constant to be estimated directly, making them ideal when the groups themselves are the entire population of interest. Random effects, by contrast, treat group-level variation as drawn from a larger probability distribution, allowing researchers to generalize findings beyond the specific groups sampled.

Mixed models, which combine both approaches, have become standard tools in fields ranging from psychology to econometrics. Understanding when to apply each method—and why the wrong choice can produce misleading standard errors or biased estimates—is essential for anyone working with hierarchical or longitudinal data.

What Is a Fixed Effect?

A fixed effect is a variable whose specific levels are of direct interest to the researcher and are treated as constant, unknown parameters to be estimated. Rather than assuming these levels come from a broader population, fixed effects models assume the categories observed in the data are the only categories that matter for the analysis.

The Core Idea

Consider a study examining test scores across four specific teaching methods: lecture-based, flipped classroom, project-based, and hybrid. If a researcher is only interested in comparing these four particular methods—not generalizing to “all possible teaching methods”—each method is modeled as a fixed effect.

The general form of a fixed effects regression model is:

$$Y_{it} = \beta_0 + \beta_1 X_{it} + \alpha_i + \epsilon_{it}$$

Where:

• $Y_{it}$ is the outcome for unit $i$i at time $t$t
• $X_{it}$​ represents the explanatory variables
• $\alpha_i$ is the fixed effect for unit $i$i (a constant unique to that unit)
• $\epsilon_{it}$​ is the error term

Each $\alpha_i$αi​ absorbs all time-invariant characteristics of unit $i$i, whether or not those characteristics are measured.

Worked Example

Suppose a researcher tracks quarterly sales performance for five specific regional offices over two years. Because these are the only five offices the company operates—not a sample representing hundreds of possible offices—the office identifier is treated as a fixed effect. Each office gets its own intercept term ($\alpha_1$ through $\alpha_5$​), capturing unmeasured, office-specific factors like local market conditions or management style that remain stable over time.

When Fixed Effects Make Sense

Fixed effects are appropriate when:

Condition	Explanation
Groups are the entire population	Not a sample from a larger population
Interest lies in specific group differences	You want to know how Office A differs from Office B
Controlling for unobserved heterogeneity	Removes bias from time-invariant confounders
Small number of groups	Estimating many fixed effects with few groups is more stable

Fixed effects models are widely implemented using tools like R’s plm package or Stata’s xtreg, fe command, both of which handle the within-group transformation automatically.

What Is a Random Effect?

A random effect is a variable whose levels are treated as a random sample drawn from a larger population of possible levels, rather than as fixed, unique quantities to be individually estimated. Instead of assigning each group its own separate parameter, random effects models assume group-level differences follow a probability distribution—typically normal—centered around an overall mean.

The Core Idea

Random effects models are built on the assumption that the observed groups are just one possible draw from a much larger set of groups that could have been sampled. This allows conclusions to extend beyond the specific groups in the dataset to the broader population they represent.

The general form of a random effects regression model is:

$$Y_{it} = \beta_0 + \beta_1 X_{it} + u_i + \epsilon_{it}$$

Where:

• $Y_{it}$ is the outcome for unit $i$i at time $t$t
• $X_{it}$​ represents the explanatory variables
• $u_i$​ is the random effect for unit $i$i, assumed to follow $u_i \sim N(0, \sigma_u^2)$
• $\epsilon_{it}$​ is the idiosyncratic error term, assumed to follow $\epsilon_{it} \sim N(0, \sigma_\epsilon^2)$

Unlike $\alpha_i$αi​ in the fixed effects model, $u_i$ui​ is not a fixed constant but a random variable, and the model estimates its variance rather than a separate value for every group.

Worked Example

Suppose a researcher studies quarterly sales performance, but this time across fifty regional offices randomly selected from a national chain of several hundred locations. Since these fifty offices represent a sample rather than the complete population of interest, office identity is treated as a random effect. The model estimates how much variance in sales is attributable to differences between offices ($\sigma_u^2$σu2​), without spending a separate parameter on each one.

When Random Effects Make Sense

Random effects are appropriate when:

Condition	Explanation
Groups are a sample	Represent a larger population of possible groups
Interest lies in overall variance, not specific groups	You want to know how much offices vary, not how Office A differs from Office B
Generalizing beyond the sample	Findings should extend to unobserved groups
Large number of groups	Estimating one variance term is more efficient than many intercepts

Random effects models are commonly fit using R’s lme4 package or Stata’s xtreg, re command, both of which use maximum likelihood or restricted maximum likelihood (REML) to estimate the variance components.

Fixed Effect vs Random Effect: Key Differences

Comparison Table

Aspect	Fixed Effect	Random Effect
Nature of groups	Entire population of interest	Sample from a larger population
Parameter treatment	Each group gets its own estimated intercept	Group differences modeled as a variance component
Statistical goal	Estimate specific group effects	Estimate variance across groups
Generalizability	Limited to groups in the dataset	Extends to the broader population
Degrees of freedom	Uses one per group (can be costly with many groups)	Uses one parameter for variance, regardless of group count
Handles time-invariant confounders	Yes, automatically absorbed	Only if uncorrelated with predictors
Best suited for	Small number of groups, specific comparisons	Large number of groups, hierarchical or nested data
Common software	plm (R), xtreg, fe (Stata)	lme4 (R), xtreg, re (Stata)

The Correlation Assumption

The most consequential technical difference lies in an assumption about correlation. Random effects models assume that the group-level effect ($u_i$) is uncorrelated with the explanatory variables ($X_{it}$). Fixed effects models make no such assumption—they allow $\alpha_i$ to correlate freely with the predictors, which is precisely why they’re effective at controlling for unobserved, group-specific confounders.

This distinction matters in practice. If an unmeasured characteristic of a group (say, management quality at a regional office) influences both the outcome and the predictor variables, a random effects model will produce biased estimates, while a fixed effects model will not.

Testing Which Model Fits: The Hausman Test

Researchers rarely choose between fixed and random effects on theoretical grounds alone. The Hausman test offers a formal statistical comparison:

$$H = (\hat{\beta}_{FE} – \hat{\beta}_{RE})’ [\text{Var}(\hat{\beta}_{FE}) – \text{Var}(\hat{\beta}_{RE})]^{-1} (\hat{\beta}_{FE} – \hat{\beta}_{RE})$$

The null hypothesis states that the random effects estimator is consistent (i.e., no correlation between $u_i$​ and $X_{it}$​). A statistically significant result (typically $p < 0.05$p<0.05) suggests the random effects assumption is violated, favoring the fixed effects model instead. This test is available in both R (via the phtest function in the plm package) and Stata (hausman command).

Practical Trade-offs

Consideration	Favors Fixed Effects	Favors Random Effects
Sample size of groups	Small (fewer than ~20-30)	Large
Interest in group-level predictors	No (these get absorbed)	Yes (can be included)
Concern about omitted variable bias	High	Low
Efficiency of estimates	Lower (more parameters)	Higher (fewer parameters)

Neither approach is universally superior—the right choice depends on the research question, the structure of the data, and the assumptions a researcher is willing to defend.

Fixed Effects Model Explained

Mathematical Foundation

The fixed effects model builds on a simple but powerful idea: subtract each unit’s average from its observations, removing anything that doesn’t vary within that unit. Starting with the basic model:

$$Y_{it} = \beta_0 + \beta_1 X_{it} + \alpha_i + \epsilon_{it}$$

Taking the average over time for each unit $i$:

$$\bar{Y_i} = \beta_0 + \beta_1 \bar{X_i} + \alpha_i + \bar{\epsilon_i}$$

Subtracting the second equation from the first eliminates $\alpha_i$​ entirely, since it doesn’t change over time:

$$Y_{it} – \bar{Y_i} = \beta_1 (X_{it} – \bar{X_i}) + (\epsilon_{it} – \bar{\epsilon_i})$$

This is known as the within transformation, and it’s the mathematical mechanism that allows fixed effects models to control for unobserved heterogeneity without ever directly estimating $\alpha_i$.

Three Ways to Estimate Fixed Effects

Method	Description	Best Used When
Least Squares Dummy Variable (LSDV)	Includes a dummy variable for each unit directly in the regression	Small number of units
Within (Demeaning) Estimator	Subtracts unit-specific means before running the regression	Large number of units
First-Differencing	Uses changes between consecutive time periods instead of demeaning	Two time periods, or when serial correlation is a concern

All three methods produce mathematically identical coefficient estimates for $\beta_1$​ under standard conditions, though standard errors can differ slightly.

Worked Example

Returning to the five regional offices tracked over two years: a fixed effects model would estimate the effect of, say, advertising spend on sales while automatically controlling for anything unique and unchanging about each office—local competition, store layout, or customer demographics. The within transformation removes each office’s average sales and average advertising spend, leaving only the year-to-year fluctuations to estimate the relationship.$$\text{Sales}_{it} – \overline{\text{Sales}_i} = \beta_1(\text{Ad Spend}_{it} – \overline{\text{Ad Spend}_i}) + (\epsilon_{it} – \bar{\epsilon_i})$$

Strengths and Limitations

Strengths	Limitations
Controls for all time-invariant confounders, measured or not	Cannot estimate effects of time-invariant variables (e.g., gender, region)
No distributional assumptions about $\alpha_i$αi​	Reduces degrees of freedom with many units
Consistent even when $\alpha_i$αi​ correlates with predictors	Less efficient than random effects when assumptions hold
Well-suited for panel data with few, specific units	Cannot generalize findings beyond the units studied

Implementation

Fixed effects models are straightforward to implement across common statistical software:

• R: The plm package with model = "within"
• Stata: xtreg depvar indepvars, fe
• Python: The PanelOLS function from the linearmodels package
• SPSS: Achieved through GLM with unit dummy variables, though less streamlined than dedicated panel data software

Diagnostic checks—such as testing for serial correlation or heteroskedasticity in the residuals—are typically recommended after fitting, since panel data often violates the classical regression assumption of independent errors.

Random Effects Model Explained

Mathematical Foundation

The random effects model starts from the same basic structure as the fixed effects model, but treats the group-specific term differently:

$$Y_{it} = \beta_0 + \beta_1 X_{it} + u_i + \epsilon_{it}$$

Here, $u_i$ui​ is not a fixed constant but a random variable drawn from a distribution:

$$u_i \sim N(0, \sigma_u^2), \quad \epsilon_{it} \sim N(0, \sigma_\epsilon^2)$$

Because both $u_i$ui​ and $\epsilon_{it}$ϵit​ contribute to the total error, the composite error term becomes:

$$\nu_{it} = u_i + \epsilon_{it}$$

This composite structure creates correlation between error terms within the same group over time, since $u_i$ is shared across all observations of unit $i$. The model accounts for this using Generalized Least Squares (GLS), which down-weights observations based on the ratio of between-group to within-group variance, captured by the parameter $\theta$:$$\theta = 1 – \sqrt{\frac{\sigma_\epsilon^2}{\sigma_\epsilon^2 + T\sigma_u^2}}$$

Where $T$T is the number of time periods per unit. When $\theta$ approaches 1, the model behaves similarly to the within (fixed effects) estimator; when $\theta$ approaches 0, it behaves like simple pooled OLS.

Estimation Methods

Method	Description	Best Used When
Generalized Least Squares (GLS)	Weights observations using estimated variance components	Variance components are known or estimated in a first stage
Maximum Likelihood (ML)	Estimates variance components and coefficients jointly	Sample sizes are moderate to large
Restricted Maximum Likelihood (REML)	Adjusts ML to reduce bias in variance estimates	Preferred default in most modern software

REML has become the standard approach in packages like lme4 because it produces less biased estimates of variance components, particularly when the number of groups is small.

Worked Example

Consider the fifty regional offices sampled from a national chain of several hundred locations. A random effects model estimates the relationship between advertising spend and sales while also estimating $\sigma_u^2$, the variance in baseline sales attributable to differences between offices. Because these offices are treated as a sample, the model’s conclusions about the advertising-sales relationship are understood to generalize to the full population of offices, not just the fifty observed.

$$\text{Sales}_{it} = \beta_0 + \beta_1 \text{Ad Spend}_{it} + u_i + \epsilon_{it}$$

Unlike the fixed effects version, this model can also include time-invariant, office-level predictors, such as whether an office is located in an urban or rural area, since $u_i$ui​ is not absorbing all office-specific variation into a separate intercept.

Strengths and Limitations

Strengths	Limitations
Can include time-invariant predictors	Requires $u_i$ui​ to be uncorrelated with regressors
More efficient than fixed effects when assumptions hold	Biased if unobserved heterogeneity correlates with predictors
Generalizes findings to the broader population	Relies on distributional assumptions about $u_i$ui​
Fewer parameters estimated, more degrees of freedom	Less appropriate with a small number of groups

Implementation

Random effects models are supported across the same major platforms used for fixed effects:

• R: The plm package with model = "random", or lme4 for more flexible hierarchical structures
• Stata: xtreg depvar indepvars, re
• Python: The RandomEffects function from the linearmodels package, or mixedlm in statsmodels
• SPSS: Available through the Mixed Models procedure (Analyze > Mixed Models > Linear)

As with fixed effects, it’s good practice to check the residual structure after fitting, since violations of the normality or independence assumptions for $u_i$ can undermine the validity of standard errors and confidence intervals.

Fixed Effects vs Random Effects in Panel Data

FAQs