Skip to content
Tutorials Center
My learning

Topics

Lesson 22 of 49 ยท Inference and Specification

Chapter 4 review: inference and specification

Review chapter 4 of econometrics: classical assumptions, standard errors, t and F tests, confidence intervals, omitted variable bias and multicollinearity, with a recap table and a ten-question quiz.

You can now do more than fit a line: you can say how much to trust it. This chapter gave you the assumptions behind OLS, the tools for testing hypotheses about one or several coefficients, and a checklist of specification problems that threaten an estimate. Chapter 5 will relax two assumptions (homoskedasticity and independence); Chapter 6 will tackle the hardest one (exogeneity).

Six cards: assumptions, standard error, t test, interval, F test and specification.
Chapter 4 in one picture: six ideas for judging and defending a regression.

The chapter in one table

Topic Key facts
Assumptions MLR.1 linear, MLR.2 random sample, MLR.3 no perfect collinearity, MLR.4 E[u | x] = 0, MLR.5 homoskedasticity, MLR.6 normal errors
Properties MLR.1 to 4: unbiased. MLR.1 to 5: BLUE (Gauss-Markov). MLR.1 to 6: exact t and F
Variance Var(b1) = ฯƒยฒ/ฮฃ(x โˆ’ xฬ„)ยฒ; ฯƒฬ‚ยฒ = RSS/(n โˆ’ k โˆ’ 1); se(b1) = ฯƒฬ‚/โˆšฮฃ(x โˆ’ xฬ„)ยฒ
t test t = (bj โˆ’ c)/se(bj); reject if `
p-value Probability, under H0, of a statistic at least as extreme; not the probability that H0 is true
Interval bj ยฑ t*ยทse(bj); contains exactly the nulls a two-sided test fails to reject
F test F = [(SSRr โˆ’ SSRur)/q]/[SSRur/(n โˆ’ k โˆ’ 1)]; for q = 1, F = tยฒ
Omitted variable bias E[bฬƒ1] = ฮฒ1 + ฮฒ2ยทฮด1; sign it with the sign table
Multicollinearity VIF = 1/(1 โˆ’ Rjยฒ); inflates variance, not bias
Measurement error Error in a regressor attenuates the slope by the reliability ratio

How to think about a regression result

When you read (or write) a regression table, run through a checklist.

  1. What is the estimate? State it with units and in the comparison it describes.
  2. How precise is it? Look at the standard error and the confidence interval, not only the stars.
  3. Is it statistically distinguishable from the relevant benchmark? The t statistic or p-value answers this for a single coefficient; the F test for several.
  4. Is it big enough to matter? Economic significance is a judgment about size.
  5. Could it be biased? Think about what is in the error term and whether it moves with the regressor. Sign the likely bias.
  6. Would it survive a different specification? Check robustness to adding sensible controls, changing the functional form and dropping outliers.

Notice that steps 1 to 4 are about sampling noise and step 5 is about the design. A beautiful standard error cannot rescue a biased estimate.

Common mistakes to avoid

Treating "not significant" as "no effect". Check the interval and the sample size.

Reading p-values as the probability that the null is true.

Using individual t tests when the question is joint. Use an F test.

Fixing collinearity by dropping a relevant variable. You swap imprecision for bias.

Thinking a large sample cures bias. It does not.

Try it

First, a compact function that takes any fitted regression's pieces and produces a full results table with estimates, standard errors, t statistics, p-values and intervals, the numbers you see in every published regression table. It is a good way to check you understand each column.

Python Building a regression table from scratch

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

Second, a visual summary of how the pieces trade off: the width of a 95% interval for a slope as sample size grows, for two levels of noise.

Python Interval width versus sample size

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

Check your understanding

Quick check ยท 10 questions

of 10 answered You scored / 10 ยท perfect!

  1. 1. OLS is unbiased under which set of assumptions?

    Linearity, random sampling, no perfect collinearity and zero conditional mean deliver unbiasedness.
  2. 2. Which assumption does the Gauss-Markov theorem add to MLR.1 to MLR.4 to conclude OLS is BLUE?

    BLUE requires MLR.5, constant error variance.
  3. 3. If b = 2.4, se = 0.8 and df is large, the t statistic for H0: ฮฒ = 0 is

    2.4/0.8 = 3.0, which exceeds about 1.96.
  4. 4. A 95% confidence interval for a coefficient is (0.10, 0.90). A two-sided 5% test of H0: ฮฒ = 0.5 would

    0.5 lies inside the interval, so it is not rejected.
  5. 5. What does a p-value of 0.03 mean?

    A p-value is a probability about the data given the null.
  6. 6. A joint test of two restrictions in a model with 60 observations and 5 regressors uses degrees of freedom

    q = 2 and n โˆ’ k โˆ’ 1 = 60 โˆ’ 5 โˆ’ 1 = 54.
  7. 7. In a regression of wages on schooling, omitted ability that is positively related to both schooling and wages causes

    ฮฒ2 > 0 and ฮด1 > 0 give a positive bias.
  8. 8. Multicollinearity between two controls and the regressor of interest being uncorrelated with them means

    Collinearity among controls inflates only their own variances; the regressor of interest keeps its precision.
  9. 9. If the reliability ratio of a noisily measured regressor is 0.6 and the true slope is 3, the OLS slope will be about

    Attenuation multiplies the slope by the reliability ratio: 0.6 ร— 3 = 1.8.
  10. 10. Which is a good reason NOT to control for a variable?

    Controlling for a mediator removes part of the effect you want to measure.

Summary

  • MLR.1 to MLR.4 give unbiasedness; MLR.5 gives Gauss-Markov efficiency and the usual standard errors; MLR.6 gives exact t and F tests.
  • Use t tests and confidence intervals for one coefficient, F tests for joint hypotheses, and always report the size of the estimate.
  • Omitted variable bias is ฮฒ2ยทฮด1; multicollinearity inflates variance, not bias; measurement error attenuates slopes.
  • Standard errors describe sampling noise and cannot cure a biased specification.
  • Chapter 5 asks what happens when errors have unequal variance or are correlated with each other.

Next: Chapter 5 overview: when the error assumptions break