Week 4 - Econometrics

0.0(0)
Studied by 0 people
call kaiCall Kai
learnLearn
examPractice Test
spaced repetitionSpaced Repetition
heart puzzleMatch
flashcardsFlashcards
GameKnowt Play
Card Sorting

1/23

encourage image

There's no tags or description

Looks like no tags are added yet.

24 Terms

1
New cards

Which of the three questions (what OLS does, fit, inference) need assumptions?

Only inference. What OLS does (FOCs, algebraic properties) and fit (R², SER) need no assumptions; inference is the only one about the population and needs SLR/MLR assumptions

wage1 SLR recap. (1) Residuals sum to zero and are uncorrelated with educ; the line passes through (12.56, 5.90). (2) R² = 0.165, SER = $3.38. (3) se = 0.053, t = 10.2, CI [0.437, 0.646]. Only (3) needs judgement: is SLR.4 believable?

2
New cards

State MLR.1 and MLR.4.

MLR.1: y = β₀ + β₁x₁ + … + βₖxₖ + u (linear in parameters). MLR.4: E(u | x₁,…,xₖ) = 0

Taking expectations given the xs under MLR.1: E(y|x₁,…,xₖ) = β₀ + β₁x₁ + … + βₖxₖ + E(u|x₁,…,xₖ). MLR.4 makes the last term zero. MLR.4 is the key assumption, and it fails if anything in u is correlated with ANY regressor.

3
New cards

Under MLR.4, what is βⱼ in the population?

The difference in E(y | x₁,…,xₖ) for a one-unit difference in xⱼ with the other regressors held fixed: a partial association

This reading rests on MLR.4. If MLR.4 fails, βⱼ is not the slope of the conditional mean in xⱼ.

4
New cards

What does OLS compute with k regressors, and what algebraic properties follow?

It chooses b₀,…,bₖ to minimise Σ(yᵢ − b₀ − b₁xᵢ₁ − … − bₖxᵢₖ)², giving k + 1 FOCs. Residuals sum to 0, have zero sample covariance with every xⱼ, and the fitted equation passes through (x̄₁,…,x̄ₖ, ȳ). SST = SSE + SSR still holds

FOCs: Σûᵢ = 0 and Σxᵢⱼûᵢ = 0 for j = 1,…,k. That's k + 1 equations in k + 1 unknowns, so there's one solution provided no regressor is an exact linear combination of the others (MLR.3).

5
New cards

wage-hat = −2.873 + 0.599·educ + 0.022·exper + 0.169·tenure (n = 526, R² = 0.306). How do you interpret the educ coefficient correctly?

"Comparing two workers with the same experience and tenure whose education differs by one year, the predicted hourly wage is $0.60 higher for the more educated one."

It's a ceteris paribus statement where only the OTHER REGRESSORS (exper, tenure) are held equal. Tenure, by the same template: 'Comparing two workers with the same education and experience whose tenure differs by one year, the predicted hourly wage is $0.17 higher for the one with more tenure.'

6
New cards

What is wrong with each of these? (a) 'raises wages by $0.60, holding everything else constant' (b) 'OLS compares workers with identical exper and tenure' (c) 'one more year is associated with a $0.60 higher predicted wage'

(a) Causal verb, and only exper and tenure are held fixed, not everything. (b) It doesn't compare identical workers: it uses the variation in educ not linearly explained by exper and tenure. (c) The comparison is missing (same exper and tenure); without it, that's the simple regression's question, answered by 0.54

On (b): 198 of the 526 workers share their exper and tenure values with nobody else, and OLS still uses every one of them.

7
New cards

What is the Frisch–Waugh–Lovell (FWL) theorem?

Step 1: regress x₁ on x₂,…,xₖ (with an intercept) and keep the residuals r̂ᵢ₁. Step 2: regress y on r̂ᵢ₁. The slope equals β̂₁ from the full regression: β̂₁ = Σr̂ᵢ₁yᵢ / Σr̂ᵢ₁²

r̂ᵢ₁ = the part of x₁ not linearly explained by the other regressors. The formula has the same form as the simple slope, with r̂ᵢ₁ in place of (xᵢ − x̄). wage1: regressing wage on the residual from educ ~ exper + tenure gives 0.599, the same as the full regression. Wooldridge 3.2 calls this the 'partialling out' interpretation.

8
New cards

Why does FWL hold? (proof sketch)

From the auxiliary regression's FOCs, Σr̂ᵢ₁ = 0, Σr̂ᵢ₁xᵢⱼ = 0 for j ≥ 2, and Σr̂ᵢ₁xᵢ₁ = Σr̂ᵢ₁². Also Σr̂ᵢ₁ûᵢ = 0. Multiply the full fitted equation yᵢ = β̂₀ + β̂₁xᵢ₁ + … + ûᵢ by r̂ᵢ₁ and sum: everything drops except Σr̂ᵢ₁yᵢ = β̂₁Σr̂ᵢ₁²

Σr̂ᵢ₁xᵢ₁ = Σr̂ᵢ₁² because xᵢ₁ = x̂ᵢ₁ + r̂ᵢ₁ and r̂ is uncorrelated with the fitted values x̂. Σr̂ᵢ₁ûᵢ = 0 because r̂ᵢ₁ is a linear combination of 1, x₁,…,xₖ, and the full residuals are uncorrelated with all of them.

9
New cards

Ceteris paribus in multiple regression: what is held fixed, and what isn't?

Held fixed: the other regressors in the equation. In the population (under MLR.4), β₁ is the slope of E(y|x) in x₁ with x₂,…,xₖ fixed. In the sample (FWL), β̂₁ uses the variation in x₁ left after removing its linear association with the others. NOT held fixed: anything in u

'Holding all other factors fixed' is the causal reading, which is Part II. If x₁ is almost a linear function of the other regressors, little variation is left and β̂₁ is imprecise.

10
New cards

wage1: the educ coefficient is 0.5414 alone and 0.5990 with exper and tenure added. Educ on exper gives −1.468; educ on tenure gives −0.147. Why did the coefficient rise?

Identity: β̃₁ = β̂₁ + β̂₂δ̃₂ + β̂₃δ̃₃: 0.5414 = 0.5990 + (0.0223)(−1.468) + (0.1693)(−0.147) = 0.5990 − 0.0328 − 0.0248. More educated workers have less exper and tenure, which both have positive coefficients, so the simple slope = the multiple slope + two negative terms

β̃₁ = simple slope; β̂ⱼ = multiple coefficients; δ̃ⱼ = slope from regressing xⱼ on x₁. This holds exactly in every sample; it's pure algebra. Neither number is 'the right one': (1) answers how predicted wages differ with education, (2) the same among workers with equal exper and tenure.

11
New cards

Why does the identity β̃₁ = β̂₁ + β̂₂δ̃₂ hold? (two regressors)

Substitute yᵢ = β̂₀ + β̂₁xᵢ₁ + β̂₂xᵢ₂ + ûᵢ into β̃₁ = Σ(xᵢ₁ − x̄₁)yᵢ/SST₁. The constant drops (Σ(xᵢ₁ − x̄₁) = 0); the x₁ term gives β̂₁; the x₂ term gives β̂₂·Σ(xᵢ₁ − x̄₁)xᵢ₂/SST₁ = β̂₂δ̃₂; and the û term = 0 by the FOCs

With more regressors there's one β̂ⱼδ̃ⱼ term for each. The two coefficients are equal only if β̂₂ = 0 or δ̃₂ = 0 (x₁ and x₂ uncorrelated in the sample). This is the basis of omitted variable bias (Part II, Week 8).

12
New cards

What happens to R² when you add a regressor, and what is adjusted R²?

R² never falls: adding a regressor can't raise SSR (OLS could set its coefficient to 0). Adjusted R² = 1 − [SSR/(n − k − 1)]/[SST/(n − 1)] charges for each coefficient through the degrees of freedom, so it can fall

wage1: R² 0.165 → 0.306; adjusted R² 0.302; SER falls from $3.38 to $3.08. A higher R² describes a closer fit in this sample. It doesn't make any coefficient more 'correct' (or unbiased), and it isn't a test.

13
New cards

State MLR.2 and MLR.3.

MLR.2: {(xᵢ₁,…,xᵢₖ, yᵢ): i = 1,…,n} is a random sample. MLR.3: no perfect collinearity; in the sample no regressor is constant and none is an exact linear combination of the others

MLR.3 is the new assumption compared with SLR (it replaces SLR.3, sample variation in x). It guarantees Σr̂ᵢⱼ² > 0, so each β̂ⱼ exists.

14
New cards

What does MLR.3 rule out, and what does it allow?

Rules out: the same variable twice in different units (exper in years and months); a dummy for every group plus an intercept (female and male = 1 − female, the dummy variable trap); n < k + 1. Allows: correlation between regressors, however high, as long as it's not exact

wage1: educ and exper have correlation −0.30, which is fine. With female and male in the same regression, R drops one and reports NA for it.

15
New cards

How do you prove β̂ⱼ is unbiased in MLR?

By FWL, β̂ⱼ = Σr̂ᵢⱼyᵢ/Σr̂ᵢⱼ². Substitute MLR.1 to get β̂ⱼ = βⱼ + Σr̂ᵢⱼuᵢ/Σr̂ᵢⱼ². Given X, each r̂ᵢⱼ is a constant, and E(uᵢ|X) = 0 by MLR.2 + MLR.4, so E(β̂ⱼ|X) = βⱼ. MLR.3 ensures Σr̂ᵢⱼ² > 0

The other terms drop because Σr̂ᵢⱼ = 0, Σr̂ᵢⱼxᵢₗ = 0 for l ≠ j, and Σr̂ᵢⱼxᵢⱼ = Σr̂ᵢⱼ². It has the same form as Week 3's Step 1. It uses MLR.1 to MLR.4 only, not MLR.5.

16
New cards

What happens if you (a) include an irrelevant regressor (β = 0), or (b) leave out a relevant one (β ≠ 0)?

(a) MLR.1–4 still hold, so OLS stays unbiased; the cost is precision (it can raise the variance of the other coefficients). (b) Its term moves into u; if it's correlated with an included regressor, MLR.4 fails and OLS is biased (omitted variable bias, Week 8)

Asymmetry: including too much costs precision; leaving out too much can cost unbiasedness.

17
New cards

What is Var(β̂ⱼ | X) under MLR.1 to MLR.5?

σ² / [SSTⱼ(1 − Rⱼ²)], where SSTⱼ = Σ(xᵢⱼ − x̄ⱼ)² and Rⱼ² is the R² from regressing xⱼ on all the other regressors

MLR.5 (homoscedasticity): Var(u | x₁,…,xₖ) = σ². Derivation: β̂ⱼ − βⱼ = Σwᵢuᵢ with wᵢ = r̂ᵢⱼ/Σr̂ᵢⱼ². Independence (MLR.2) and constant variance (MLR.5) give Var = σ²/Σr̂ᵢⱼ². Σr̂ᵢⱼ² is the SSR of the auxiliary regression, which equals SSTⱼ(1 − Rⱼ²). With one regressor Rⱼ² = 0, giving Week 3's σ²/SSTₓ.

18
New cards

What four things set the precision of β̂ⱼ?

Smaller σ²; larger n; more variation in xⱼ (larger sⱼ²); less collinearity (smaller Rⱼ²). All four give a smaller variance

The first three are from Week 3. Collinearity is new: Rⱼ² measures how much of xⱼ the other regressors already explain linearly. SSTⱼ = (n − 1)sⱼ².

19
New cards

How much does collinearity cost? (variance multiplier 1/(1 − Rⱼ²))

Rⱼ² = 0, 0.5, 0.9, 0.99 gives variance × 1, 2, 10, 100 and standard error × 1, 1.41, 3.16, 10. wage1: educ on exper and tenure has R² = 0.101, so variance × 1.11, a small cost

The standard error multiplier is the square root of the variance multiplier: √(1/(1 − Rⱼ²)). This is compared with Rⱼ² = 0, holding σ² and SSTⱼ fixed. (1/(1 − Rⱼ²) is often called the variance inflation factor, VIF.)

20
New cards

What are the three precise statements about (imperfect) collinearity?

(1) It violates no assumption unless it's perfect, so OLS stays unbiased. (2) It raises the variance of the coefficients involved, only through Rⱼ², giving wide CIs. (3) Dropping a correlated regressor is not a fix: the coefficient then answers a different question, and MLR.4 may fail

Classic wrong answer: 'The regressors are correlated, so the estimates are biased.' Exit-ticket answer: educ–exper correlation of −0.30 does NOT bias the educ coefficient; it slightly raises its variance.

21
New cards

How do you estimate σ² in MLR, and what is se(β̂ⱼ)?

σ̂² = SSR/(n − k − 1) (unbiased under MLR.1–5). se(β̂ⱼ) = σ̂/√[SSTⱼ(1 − Rⱼ²)]

df = n − k − 1: n observations minus k + 1 FOCs (k slopes + the intercept). wage1 with 3 regressors: df = 526 − 4 = 522. This uses MLR.5; under heteroscedasticity the formula is wrong (Week 7).

22
New cards

Consistency and Gauss–Markov in MLR: what do they need?

Consistency: MLR.1–3 + E(u) = 0 + Cov(xⱼ, u) = 0 for every j (not MLR.5). More data can't fix Cov(xⱼ, u) ≠ 0. Gauss–Markov: under MLR.1–5, each β̂ⱼ has the smallest variance (given X) among linear unbiased estimators, so OLS is BLUE

It's the same structure as Week 3, with 'Cov(x, u) = 0' now required for every regressor.

23
New cards

What does each result need in MLR?

Partial-association reading: MLR.1 + MLR.4. Unbiased: MLR.1–4. Consistent: MLR.1–3 + E(u) = 0 + Cov(xⱼ,u) = 0. Variance formula and unbiased σ̂²: MLR.1–5. BLUE: MLR.1–5. Exact t and F: MLR.1–6 (normal errors)

The k + 1 FOCs are not an assumption in any row; they only fix the df at n − k − 1.

24
New cards

True or false: "R² rose to 0.31, so the education coefficient is now unbiased."

False. R² describes in-sample fit only. Unbiasedness depends on MLR.1–4, especially MLR.4 (E(u|x) = 0), which R² tells you nothing about; e.g. ability could still be in u and correlated with educ

Fit ≠ unbiasedness ≠ causality. A high R² does not mean MLR.4 holds.