AI summary by TwinMind
Lecture 19: Practical Issues in Running Regressions
MIT OpenCourseWare
1h 20m 2s
T-Tests in Linear Regression
- T-test used when substituting estimated error variance (σ̂²) for true variance (σ²)
- Distribution becomes t-distribution rather than normal
- T-distribution has fatter tails, more conservative critical values
- As sample size increases, t-distribution approaches normal distribution
- T-test formula: (β̂ - C) / SE(β̂)
- For complex hypotheses: (Rβ̂ - C) / SE(Rβ̂)
- For simple linear hypotheses, F-test and T-test are equivalent
- T-statistic = √F-statistic
- One-sided tests require T-test (cannot use F-test framework)
- Default regression output includes:
- F-test for all coefficients equal zero
- T-values for individual coefficients testing each equals zero
- Confidence intervals constructed as: coefficient ± critical value × standard error
Dummy Variables on Right-Hand Side
- Dummy variable = 1 if observation belongs to group A, 0 if group B
- Examples: treatment/control, male/female, before/after policy change
- Without controls: β̂ = mean(Group A) - mean(Group B)
- Standard error differences between Neyman vs OLS:
- Neyman: accounts for different group sizes in variance calculation
- OLS: uses pooled variance with total N in denominator
- Differences matter in small samples or unequal group sizes
Categorical Variables with Multiple Categories
- Cannot directly include categorical variable (1,2,3,4) as regressor
- Must create separate dummy variables for each category
- Drop one category to avoid multicollinearity with constant
- Choose reference category deliberately (typically control group)
- Coefficients represent differences from omitted category
- Can test equality of coefficients across categories
Dummy Variables with Other Regressors
- Model: Yi = α + βDi + γXi + εi
- β represents intercept shift between groups
- Assumes treatment shifts intercept but not slope
- Common in RCTs with baseline controls
- Example: wage differences controlling for education, experience
Interaction Terms and Difference-in-Difference
- Model: Yi = α + βDi + γMi + δ(Di × Mi) + εi
- α = average for omitted group (female, control)
- β = treatment effect for females
- γ = male-female difference in control group
- δ = additional treatment effect for males
- Total treatment effect for males = β + δ
Difference-in-Difference Applications
Mariel Boatlift Study (Card)
- Cuban refugees suddenly arrived in Miami
- Compared Miami wages to similar control cities
- Before/after refugee arrival
- Found no negative wage effects on native workers
New Jersey-Pennsylvania Minimum Wage Study
- Compared fast food employment across state border
- Before/after minimum wage increase in New Jersey
- Difference-in-difference design controls for unobserved trends
Indonesian School Construction Study
- Oil boom funded 62,000 school construction (1973-1978)
- Schools allocated to regions with low initial education
- Compared education outcomes by:
- Region intensity (high vs low program regions)
- Age cohort (young enough vs too old to benefit)
- Results: 0.12 additional years education per school per 1000 children
- Placebo test: no effect for older cohorts who couldn't benefit
Advanced Difference-in-Difference Techniques
- Synthetic control method (Abadie): weighted average of control units
- Event study design: separate effects by years since treatment
- Fixed effects regression: region and time fixed effects plus interaction
- Robustness checks: control for pre-existing trends, other concurrent programs
Functional Form Issues
Log Transformations
- Log-log: log(Y) = α + β log(X) → β = elasticity
- Log-linear: log(Y) = α + βX → β = percentage change per unit change in X
- Common in wage equations: each year education increases wages by β%
Other Transformations
- Box-Cox transformations
- Discrete choice models (logit): log(p/(1-p)) = linear function
- Used when theory suggests non-linear relationships
