Handling Classical Assumption Problems in Panel Data Regression
A comprehensive guide to detecting and resolving classical assumption violations.
Classical Assumption Testing is a main pillar in OLS (Ordinary Least Squares) regression. If these assumptions are violated, the estimators (regression coefficients) you produce are not BLUE (Best Linear Unbiased Estimator), meaning your research conclusions cannot be relied upon. In Panel Data specifically, the approach differs slightly from regular time-series data regression.
1. Residual Normality Test
In large-scale panels (N > 30), the Residual Normality Assumption is often ignored, referring to the Central Limit Theorem. However, if your professor insists, use the Jarque-Bera (JB) Test. If the JB probability value is > 0.05, then the residuals are normally distributed.
2. Multicollinearity Test (VIF)
Multicollinearity tests whether there is a very strong correlation among independent variables. The easiest way is to look at the Variance Inflation Factor (VIF) value. In EViews, after performing regression, select View > Coefficient Diagnostics > Variance Inflation Factors.
If the Centered VIF value is > 10, your model is certainly affected by serious multicollinearity. The solution is to remove one of the problematic variables or use dimensionality reduction techniques like PCA (Principal Component Analysis).
3. Heteroscedasticity Test (White / Glejser Test)
This is the most common affliction in panel data (Cross-section heteroscedasticity). The variance of the error term is not constant (differs between large vs. small-scale companies). In EViews, you can run the Glejser Test (regressing the absolute value of the residual against independent variables).
Tactical Solution Without Deleting Data:
If heteroscedasticity is proven, don't panic! You don't need to look for new data. You simply need to change your standard error estimation method to Robust Standard Errors. In the EViews estimation window, open the Options tab, then under Covariance Method, select White cross-section. This automatically neutralizes the bias from heteroscedasticity.
4. Autocorrelation Test (Durbin-Watson / Breusch-Godfrey)
Autocorrelation occurs when the error term in the current year (t) is influenced by the error term in the previous year (t-1). This is very common in panel data because we observe the same companies consecutively. If the Durbin-Watson value (DW Stat) is far from the number 2 (e.g., below 1.5 or above 2.5), it means autocorrelation is present.
How to Overcome:
Similar to the Heteroscedasticity trick, you can directly use covariance matrix adjustments (HAC - Heteroskedasticity and Autocorrelation Consistent). In EViews (Options), select the Panel M.A. / Arellano option or use a more dynamic regression model (like GMM - Generalized Method of Moments).