advanced 13 min read
Finance & economics · Topic
Panel Data and Fixed Effects
linear algebra · hypothesis testing · probability theory
Panel data observe the same $N$ units (individuals, firms, countries) across $T$ time periods, yielding $NT$ observations with a rich structure that cross-sectional data cannot provide. The central challenge is unobserved heterogeneity — unit-specific characteristics correlated with regressors that bias OLS. Fixed effects estimation eliminates this heterogeneity by demeaning within each unit, enabling consistent estimation under weaker assumptions than random effects.

Panel Data Structure and the Unobserved Heterogeneity Problem

A balanced panel dataset consists of observations $(y_{it}, x_{it})$ for units $i = 1, \ldots, N$ and periods $t = 1, \ldots, T$. The general linear panel model is:

\[y_{it} = x_{it}^\top \beta + \alpha_i + u_{it}\]

where $\alpha_i$ is an unobserved, unit-specific effect (fixed or random) and $u_{it}$ is an idiosyncratic error with $\mathbb{E}[u_{it} \mid x_{it}, \alpha_i] = 0$.

The problem with pooled OLS: If $\alpha_i$ is correlated with $x_{it}$ (i.e., $\text{Cov}(\alpha_i, x_{it}) \neq 0$), pooled OLS regressing $y_{it}$ on $x_{it}$ and a constant is inconsistent. The OLS estimator confounds the true effect $\beta$ with the spurious relationship driven by $\alpha_i$.

Example: Estimating the wage return to education. Firms with more productive workers (high $\alpha_i$) may also pay higher wages and hire more educated workers. OLS overstates the return to education.

The panel regression can be written in matrix form. Let $\mathbf{y}i = (y{i1}, \ldots, y_{iT})^\top$, $\mathbf{X}_i$ be the $T \times k$ matrix of regressors, and $\mathbf{u}_i$ the error vector. Stacking across units:

\[\mathbf{y} = \mathbf{X}\beta + \mathbf{D}\alpha + \mathbf{u}\]

where $\mathbf{D}$ is the $NT \times N$ matrix of unit dummies ($D_{it,j} = \mathbf{1}[i=j]$) and $\alpha = (\alpha_1, \ldots, \alpha_N)^\top$.

The Within (Fixed Effects) Estimator

The within estimator eliminates $\alpha_i$ by demeaning each unit’s observations. Define unit-mean deviations:

\[\tilde{y}_{it} = y_{it} - \bar{y}_i, \qquad \tilde{x}_{it} = x_{it} - \bar{x}_i\]

where $\bar{y}i = T^{-1}\sum_t y{it}$. Subtracting the unit mean from both sides of the model:

\[\tilde{y}_{it} = \tilde{x}_{it}^\top \beta + (u_{it} - \bar{u}_i)\]

The individual effect $\alpha_i$ drops out. The FE (within) estimator is:

\[\hat{\beta}_{\text{FE}} = \left(\sum_i \sum_t \tilde{x}_{it}\tilde{x}_{it}^\top\right)^{-1}\!\sum_i \sum_t \tilde{x}_{it}\tilde{y}_{it}\]

This is algebraically equivalent to OLS on the demeaned data, or equivalently OLS on the original data including $N-1$ unit dummy variables (the “least squares dummy variable” estimator). Both give the same slope estimates but the dummy approach is computationally expensive when $N$ is large.

Identification: $\hat{\beta}_{\text{FE}}$ is identified only from within-unit variation over time. Time-invariant regressors (gender, race, country) are perfectly collinear with the unit dummies and are absorbed — their effects cannot be estimated.

Consistency: Under strict exogeneity $\mathbb{E}[u_{it} \mid x_{i1}, \ldots, x_{iT}, \alpha_i] = 0$, the FE estimator is consistent as $N \to \infty$ with $T$ fixed.

Random Effects and the Hausman Test

Random effects (RE) assumes $\alpha_i \perp x_{it}$, treating $\alpha_i$ as part of the composite error $v_{it} = \alpha_i + u_{it}$. Because $v_{it}$ and $v_{is}$ share $\alpha_i$, the error is equicorrelated:

\[\text{Cov}(v_{it}, v_{is}) = \sigma^2_\alpha \quad (t \neq s), \qquad \text{Var}(v_{it}) = \sigma^2_\alpha + \sigma^2_u\]

GLS exploits this structure. The RE estimator is a matrix-weighted combination of the within (FE) estimator and the between estimator (OLS on unit means):

\[\hat{\beta}_{\text{RE}} = \lambda\,\hat{\beta}_{\text{FE}} + (1-\lambda)\,\hat{\beta}_{\text{between}}, \qquad \lambda = 1 - \frac{\sigma_u}{\sqrt{T\sigma^2_\alpha + \sigma^2_u}}\]

RE is more efficient than FE when $\alpha_i \perp x_{it}$ because it uses both within and between variation. But if $\text{Cov}(\alpha_i, x_{it}) \neq 0$, RE is inconsistent.

The Hausman test formally tests $H_0: \text{Cov}(\alpha_i, x_{it}) = 0$ by comparing $\hat{\beta}{\text{FE}}$ and $\hat{\beta}{\text{RE}}$. Under $H_0$, both are consistent but RE is efficient; under $H_1$, only FE is consistent. The test statistic is:

\[H = (\hat{\beta}_{\text{FE}} - \hat{\beta}_{\text{RE}})^\top \left[\widehat{\text{Var}}(\hat{\beta}_{\text{FE}}) - \widehat{\text{Var}}(\hat{\beta}_{\text{RE}})\right]^{-1} (\hat{\beta}_{\text{FE}} - \hat{\beta}_{\text{RE}}) \xrightarrow{d} \chi^2_k\]

under $H_0$, where $k$ is the number of time-varying regressors. Rejection suggests FE is preferred; failure to reject supports RE (with the caveat that the test has low power in small samples).

Two-Way Fixed Effects

Two-way FE adds time fixed effects $\lambda_t$ to control for aggregate shocks common to all units:

\[y_{it} = x_{it}^\top \beta + \alpha_i + \lambda_t + u_{it}\]

The within transformation now demeans by unit means, time means, and the grand mean:

\[\tilde{y}_{it} = y_{it} - \bar{y}_i - \bar{y}_t + \bar{y}\]

This uses only variation that is neither purely cross-sectional ($\alpha_i$ absorbed) nor purely time-series ($\lambda_t$ absorbed). Two-way FE is the standard specification in difference-in-differences designs and controls for any time-invariant unit heterogeneity and any unit-invariant time shocks simultaneously.

Identification under two-way FE requires variation in $x_{it}$ that is not fully explained by unit means or time means. A regressor that is purely a unit-specific trend, for instance, is not identified.

Cluster-Robust Standard Errors

Within a unit, errors $u_{it}$ across $t$ are likely correlated (e.g., a firm that had a good year tends to have correlated residuals). OLS standard errors that assume independence across all $it$ pairs are invalid. Cluster-robust standard errors allow arbitrary correlation within unit $i$ across time:

\[\widehat{\text{Var}}_{\text{CR}}(\hat{\beta}) = (X^\top X)^{-1} \left(\sum_{i=1}^N \tilde{X}_i^\top \hat{u}_i \hat{u}_i^\top \tilde{X}_i\right) (X^\top X)^{-1}\]

where $\hat{u}i = (\hat{u}{i1}, \ldots, \hat{u}{iT})^\top$ is the vector of within-unit residuals and $\tilde{X}_i$ is the demeaned regressor matrix for unit $i$. This is the panel analog of heteroskedasticity-robust (White) standard errors. Cluster-robust SEs are consistent as $N \to \infty$ and are nearly always preferred in panel settings. With few clusters ($N < 50$), a small-sample correction using $t{N-1}$ or wild cluster bootstrap is recommended.

Nickell Bias in Dynamic Panels

When lagged outcomes enter as regressors, strict exogeneity fails and the FE estimator is biased. The dynamic panel model is:

\[y_{it} = \rho\,y_{i,t-1} + x_{it}^\top \beta + \alpha_i + u_{it}\]

After within-demeaning, $\tilde{y}{i,t-1}$ is correlated with the demeaned error $\tilde{u}{it}$ because $\bar{y}i$ contains $y{i,t-1}$ and $\bar{u}i$ contains $u{i,t-1}$. The resulting Nickell bias is of order $O(1/T)$:

\[\text{plim}_{N\to\infty}(\hat{\rho}_{\text{FE}} - \rho) \approx -\frac{1+\rho}{T-1}\]

For small $T$ (e.g., $T = 5$), this bias can be substantial. Solutions include:

Method Approach
Anderson-Hsiao Instrument $\Delta y_{i,t-1}$ with $y_{i,t-2}$
Arellano-Bond GMM using all available lags as instruments
Blundell-Bond System GMM adding level equations
Bias correction Analytical correction of order $O(1/T)$ (Kiviet)

The Arellano-Bond estimator uses first differences to eliminate $\alpha_i$ and then instruments $\Delta y_{i,t-1}$ with levels $y_{i,t-2}, y_{i,t-3}, \ldots$, which are valid instruments under the assumption that $u_{it}$ is serially uncorrelated.