Your Panel Data Estimates Are Probably Wrong
How Cross-Sectional Dependence Breaks Standard Inference
Introduction
Panel data methods such as pooled OLS, fixed effects, and random effects are commonly used in applied work. As data accessibility has grown, it’s become easier to build panels for analysis, and these estimators have become widely used starting point for empirical analysis.
However, they rely on an important assumption that is easy to overlook. After controlling for observed variables and fixed effects, the remaining error terms are assumed to be independent across units. In many practical applications, this assumption does not hold.
When this assumption is violated, standard inference can become unreliable. Estimated standard errors may be too small, test statistics may be overstated, and statistical significance can be misleading.
This issue is known as cross-sectional dependence, and it is especially common in macroeconomic and financial panel data.
In this post, we’ll dig deeper into the problem of cross-sectional dependence and explain why it matters for applied panel data analysis. In particular, we will:
- Define cross-sectional dependence and describe how it arises in practice.
- Explain why standard estimators such as fixed effects do not fully address the issue.
- Discuss how ignoring cross-sectional dependence affects inference.
- Introduce a practical solution based on the Common Correlated Effects (CCE) estimator.
In a follow-up post, we will implement these ideas in GAUSS using the dccelib library and walk through a full example.
What Is Cross-Sectional Dependence?
Cross-sectional dependence occurs when the error term for one unit is correlated with the error term for another unit at the same point in time. Intuitively, this occurs when there are common factors that impact multiple groups in the panel during the same time period.
This can arise from several sources:
- Global shocks such as recessions, policy changes, or commodity price movements.
- Unobserved common factors that affect all units but with different magnitudes.
- Spillovers due to geographic proximity or economic linkages.
These features are common in many datasets, particularly those involving countries, regions, or firms operating in related environments.
What does cross-correlation look like?
It can be helpful to think about cross-sectional dependence visually.
Imagine plotting the residuals from a standard fixed effects model across countries at a given point in time.
- If there is no cross-sectional dependence, the residuals should appear randomly scattered.
- If common factors are present, the residuals will move together across units.
For example, during a global recession, many countries will simultaneously experience negative shocks. This creates a pattern where residuals are correlated across units rather than independent.
After applying the CCE approach, this shared movement is largely removed. The residuals become more dispersed and less synchronized across units, indicating that the common factor has been accounted for.
Why It Matter For Inference
When cross-sectional dependence is present and not accounted for, standard inference procedures break down.
In particular:
- Standard errors are typically underestimated.
- Test statistics are inflated.
- Hypothesis tests reject too often.
As a result, estimated relationships may appear more precise and more statistically significant than they actually are.
Get Eric Clower’s stories in your inbox
Join Medium for free to get updates from this writer.
Evidence from simulation studies, such as Pesaran (2004), shows that these effects can be substantial even in panels with a moderate number of time periods.
Why Fixed Effects Are Not Sufficient
A common approach is to include time fixed effects to capture common shocks. While this helps remove the average impact of those shocks, it does not account for differences in how individual units respond.
After removing the average effect, the remaining variation can still be correlated across units. This means that cross-sectional dependence can persist even after applying standard fixed effects transformations.
A Diagnostic First Step
Before choosing an estimation strategy, it is useful to test for cross-sectional dependence directly.
Pesaran’s CD test provides a simple and widely used diagnostic. The test statistic follows an asymptotic standard normal distribution under the null hypothesis of cross-sectional independence.
In many applied datasets, the null is strongly rejected, indicating that alternative estimation methods should be considered.
A Practical Solution: Common Correlated Effects
Pesaran (2006) proposed the Common Correlated Effects (CCE) estimator as a way to address cross-sectional dependence.
The key idea is to include cross-sectional averages of the dependent variable and regressors as additional controls. These averages serve as proxies for unobserved common factors.
This approach:
- Does not require specifying the number of factors.
- Allows for heterogeneous responses across units.
- Is straightforward to implement.
Mathematically, if the unobserved common factors are driving correlation across units, we can think of the error term as having a structure like:
While this approach is theoretically intuitive, it is challenging in practice because these common factors are not observed.
Using Cross-Sectional Averages as Proxies
CCE provides a practical way to deal with this.
Instead of trying to estimate the unobserved factors directly, we use cross-sectional averages of the data:
- The average of the dependent variable across units.
- The average of the regressors across units.
These averages contain information about the common factors because they reflect the shared movements across units.
By including these cross-sectional averages as additional regressors, we can control for the unobserved common shocks.
What This Does in Practice
Including cross-sectional averages has two important effects:
- Removes bias from omitted common factors : If the common factors are correlated with the regressors, omitting them leads to biased coefficient estimates. Including cross-sectional averages helps account for this.
- Improves inference: Once the common factors are accounted for, the remaining error term is closer to being cross-sectionally independent. This leads to more reliable standard errors and test statistics.
An important advantage of this approach is that it does not require specifying the number of common factors or estimating them explicitly.
Two Common Implementations
In practice, CCE is often combined with a mean group framework.
- Mean Group (MG) estimates a separate regression for each unit and averages the coefficients.
- CCE Mean Group (CCE-MG) augments each unit’s regression with cross-sectional averages.
CCE-MG allows for heterogeneous slopes while addressing cross-sectional dependence.
References
- Pesaran, M.H. (2006). Estimation and inference in large heterogeneous panels with a multifactor error structure. *Econometrica*, 74(4), 967–1012.
- Pesaran, M.H. (2004). General diagnostic tests for cross-section dependence in panels. *CESifo Working Paper* №1229.
