Model specification
A well-specified model is grounded in economic reasoning, uses the right functional form, is parsimonious, and satisfies the regression assumptions. Common errors:
- Omitted variables: biases the other coefficients if the omitted variable is correlated with them.
- Wrong functional form: for example, a linear model when the relationship is logarithmic.
- Inappropriate scaling: mixing levels and percentages carelessly.
- Pooling data from different regimes or populations.
Heteroskedasticity
The variance of errors changes across observations, often growing with an X.
- Unconditional: not related to the Xs. Not a problem for inference.
- Conditional: related to the Xs. Standard errors are usually too small, so t-statistics are too big (Type I errors).
Breusch-Pagan test: regress squared residuals on the Xs. BP = n × R², chi-square with k degrees of freedom. Example: 60 × 0.12 = 7.2 < 7.81, so no significant problem.
Fix: robust (White-corrected) standard errors, or generalized least squares.
Serial correlation
Errors are correlated across time, common in time series.
- Positive serial correlation makes standard errors too small (false significance).
- If the model includes a lagged dependent variable, serial correlation also makes coefficients biased and inconsistent.
Tests: Durbin-Watson (first-order) and Breusch-Godfrey (several lags). Fix: Newey-West (serial-correlation- and heteroskedasticity-consistent) standard errors, or change the model.
Multicollinearity
Two or more Xs are highly correlated. Coefficients are unbiased but imprecise; standard errors are inflated, so t-tests lack power.
Signs: significant F with insignificant t-statistics; coefficients that swing when variables are added.
Variance inflation factor: VIFj = 1 ÷ (1 − Rj²), where Rj² comes from regressing Xj on the other Xs. Above 5 is a concern, above 10 is serious.
Summary
| Problem | Effect on coefficients | Effect on standard errors | Detection | Fix |
|---|---|---|---|---|
| Conditional heteroskedasticity | Unbiased | Usually too small | Breusch-Pagan | Robust SEs |
| Serial correlation | Unbiased (biased with lagged Y) | Usually too small | Durbin-Watson, Breusch-Godfrey | Newey-West SEs |
| Multicollinearity | Unbiased but unstable | Too large | VIF, significant F with insignificant ts | Drop or combine variables |
Exam focus
Identify each problem from symptoms, compute a Breusch-Pagan statistic and a VIF, and match each problem to its consequences and fix.