Week 3 - Econometrics

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

1/26

encourage image

There's no tags or description

Looks like no tags are added yet.

27 Terms

1
New cards

What is the difference between the OLS estimator and an OLS estimate?

Estimator = the OLS formula applied to a random sample, which is a random variable before the sample is drawn. Estimate = the number one particular sample gives (e.g. 0.54 in wage1)

Because β̂₁ is a random variable, it has a sampling distribution. A histogram of slopes from many samples approximates it. The three questions this week: where is it centred (unbiasedness), how spread out is it (variance), and what happens as n grows (consistency)?

2
New cards

SLR.1 and SLR.2: what do they say?

SLR.1: linear in parameters, y = β₀ + β₁x + u. SLR.2: {(xᵢ, yᵢ): i = 1,…,n} is a random sample from the population

SLR.1: the model must be linear in the βs, though x and y can be transformed (logs etc.). SLR.2: each observation is an independent draw from the same population. It fails if, for example, you only sample people with y below 5.

3
New cards

SLR.3 and SLR.4: what do they say, and which is the key assumption?

SLR.3: sample variation in x (not all xᵢ are equal, so SSTₓ > 0). SLR.4: zero conditional mean, E(u | x) = 0. SLR.4 is the key one, and the one most likely to fail

SLR.4 means that knowing x tells you nothing about the average of the unobservables. It fails with observational data when, for example, unobserved ability is in u and is correlated with education. Part II of the course is about exactly this.

4
New cards

Why can't the FOCs (Σûᵢ = 0, Σxᵢûᵢ = 0) give unbiasedness?

They hold for the residuals by construction in every sample, even when OLS is badly biased. Unbiasedness is about the errors u, which are unobserved, so it needs an extra assumption on u (SLR.4)

Residuals ûᵢ are computed and always satisfy the FOCs. Errors uᵢ are never observed, and in any one sample Σ(xᵢ − x̄)uᵢ ≠ 0. That sum is exactly the sampling error of the slope. Residuals are not errors.

5
New cards

β̂₁ = β₁ + Σ(xᵢ − x̄)uᵢ / SSTₓ, where SSTₓ = Σ(xᵢ − x̄)²

Start from β̂₁ = Σ(xᵢ − x̄)yᵢ/SSTₓ and substitute yᵢ = β₀ + β₁xᵢ + uᵢ. The three sums: β₀Σ(xᵢ − x̄) = 0; β₁Σ(xᵢ − x̄)xᵢ = β₁SSTₓ; and Σ(xᵢ − x̄)uᵢ is what's left. Dividing by SSTₓ gives the result. This is pure algebra, true in every sample whatever we assume about u. The second term is the sampling error.

6
New cards

Steps 2 to 4: how do you show E(β̂₁ | x) = β₁?

Condition on the x's, so SSTₓ and each (xᵢ − x̄) are constants. Then E(β̂₁|x) = β₁ + Σ(xᵢ − x̄)E(uᵢ|x)/SSTₓ, and E(uᵢ|x) = 0 by SLR.2 + SLR.4, so E(β̂₁|x) = β₁

Linearity of E pulls out the constants. The key fact is E(uᵢ | x₁,…,xₙ) = 0, using random sampling (SLR.2) and zero conditional mean (SLR.4). Unbiasedness uses SLR.1 to SLR.4 only, not SLR.5. The same steps give E(β̂₀) = β₀.

7
New cards

Define bias, and what does unbiasedness mean in practice?

Bias = E(β̂₁) − β₁. Unbiased means the sampling distribution is centred on the truth: a property of the estimator across samples, not of any one estimate

Your one sample's slope can be far from the truth even when OLS is unbiased (e.g. 0.48 vs a true 0.68). One sample cannot reveal its own bias.

8
New cards

Lab 2 designs (true β₁ = 0.68): which assumption fails, and what happens, for (i) ability in u correlated with x and (ii) sampling only people with y below 5?

(i) SLR.4 fails (E(u|x) ≠ 0), mean slope 0.78, upward bias. (ii) SLR.2 fails (not a random sample), mean slope about 0.59, downward bias

By contrast, the low-x random sample (x only from 0.5 to 3) is still unbiased (mean 0.68) but has a larger variance. Surprising result: in about half of draws, the biased ability design gives a slope closer to 0.68 than the unbiased low-x design. One draw tells you nothing about bias.

9
New cards

What is SLR.5, and what does it affect?

Homoscedasticity: Var(u | x) = σ², the error variance doesn't depend on x. It affects the variance of OLS, not its expected value

Heteroscedastic means the error variance changes with x (e.g. it fans out as x grows). wage1 example: the residuals are more dispersed at higher education, so SLR.5 likely fails. Unbiasedness survives, but the usual variance formula σ²/SSTₓ is wrong. Robust standard errors come in Week 7.

10
New cards

What is Var(β̂₁ | x) under SLR.1 to SLR.5?

σ² / SSTₓ = σ² / Σ(xᵢ − x̄)²

Derivation: β̂₁ − β₁ = Σwᵢuᵢ with weights wᵢ = (xᵢ − x̄)/SSTₓ. Given x, the uᵢ are independent (SLR.2), so all the covariance terms vanish: Var = Σwᵢ²Var(uᵢ|x). Each Var(uᵢ|x) = σ² (SLR.5), so Var = σ²Σwᵢ². And Σwᵢ² = Σ(xᵢ − x̄)²/SSTₓ² = 1/SSTₓ.

11
New cards

Var(β̂₀ | x) under SLR.1 to SLR.5 (stated, not derived)?

σ² · (n⁻¹Σxᵢ²) / SSTₓ

You just need to recognise it. Like the slope, it gets smaller with smaller σ² and more variation in x.

12
New cards

What three things make the slope estimate more precise (smaller Var(β̂₁))?

Smaller error variance σ²; more observations n; more variation in x (larger sₓ²)

SSTₓ = (n − 1)sₓ², so Var(β̂₁) = σ²/[(n − 1)sₓ²]. Adding an observation never lowers SSTₓ. In the low-x lab, small sₓ² gave slopes with a larger variance but the same mean.

13
New cards

What does it mean for an estimator to be consistent?

As n grows, the probability that it lands more than ε from β₁ goes to zero, for every ε > 0 (it converges in probability to β₁: plim β̂₁ = β₁)

Sketch of why OLS is consistent: it's unbiased and Var(β̂₁) = σ²/SSTₓ = (σ²/n)/(SSTₓ/n). SSTₓ/n converges to Var(x) > 0, so the variance goes to 0 and the sampling distribution collapses onto β₁.

14
New cards

Give an estimator that is unbiased but not consistent, and one that is consistent but biased.

Unbiased, not consistent: the OLS slope using only the first 2 (or 5) observations, whose variance doesn't shrink as n grows. Consistent, biased: the OLS slope + 1/n, biased for every n but the bias goes to 0

Unbiasedness is about the centre at a fixed n. Consistency is about what happens as n → ∞. They're different properties. OLS with ability in u is biased AND inconsistent.

15
New cards

What does consistency of the OLS slope require? Can more data fix Cov(x, u) ≠ 0?

It needs E(u) = 0 and Cov(x, u) = 0 (weaker than SLR.4), plus SLR.1 to SLR.3, but not SLR.5. No: if Cov(x, u) ≠ 0, then plim β̂₁ = β₁ + Cov(x, u)/Var(x) ≠ β₁

Why Cov(x, u) = 0 is weaker than SLR.4: E(u|x) = 0 implies zero covariance, but not the other way round. The inconsistency term Cov(x, u)/Var(x) doesn't shrink with n. A larger sample just converges more tightly to the wrong value (e.g. ability bias).

16
New cards

What does the Gauss–Markov theorem say?

Under SLR.1 to SLR.5, OLS is BLUE: the Best (smallest variance, given the x's) Linear (a weighted sum of the yᵢ) Unbiased Estimator

Linear: β̃₁ = Σaᵢyᵢ, where the weights may depend on the x's. Unbiased: E = β₁ whatever β₀ and β₁ are. Best: smallest variance among linear unbiased estimators only. If you drop SLR.5, OLS is still unbiased but not necessarily best, and the usual standard errors are wrong.

17
New cards

Sketch of the proof that OLS is BLUE in SLR.

A linear β̃₁ = Σaᵢyᵢ is unbiased iff Σaᵢ = 0 and Σaᵢxᵢ = 1. Its variance is σ²Σaᵢ². Write aᵢ = wᵢ + dᵢ with OLS weights wᵢ = (xᵢ − x̄)/SSTₓ. Then Σaᵢ² = Σwᵢ² + Σdᵢ² ≥ Σwᵢ², with equality only if every dᵢ = 0 (OLS itself)

Substituting yᵢ = β₀ + β₁xᵢ + uᵢ gives E(β̃₁|x) = β₀Σaᵢ + β₁Σaᵢxᵢ, hence the two conditions. Since both aᵢ and wᵢ satisfy them, Σdᵢ = 0 and Σdᵢxᵢ = 0, so the cross term Σwᵢdᵢ = 0. Var(β̃₁) = Var(β̂₁) + σ²Σdᵢ² ≥ Var(β̂₁). Notation: a tilde marks any other estimator; the hat is reserved for OLS.

18
New cards

How do you estimate σ², and why divide by n − 2?

σ̂² = SSR/(n − 2) = Σûᵢ²/(n − 2), which is unbiased under SLR.1 to SLR.5. The residuals obey 2 restrictions (the two FOCs), so only n − 2 are free (the degrees of freedom). Dividing by n would understate σ²

σ² is the variance of the unseen errors, so we use the residuals instead. We can compute SSTₓ (4025 in wage1) but not σ², which is why it has to be estimated. The FOCs aren't assumptions; they only fix the degrees of freedom at n − 2.

19
New cards

What is the standard error of the regression (SER), and what is se(β̂₁)?

SER = σ̂ = √[SSR/(n − 2)] (R calls it the 'residual standard error'). se(β̂₁) = σ̂/√SSTₓ, which estimates the SD of β̂₁ across samples

Chain: square and sum the residuals (SSR) → divide by n − 2 (σ̂²) → take the square root (SER, σ̂) → divide by √SSTₓ to get se(β̂₁). In wage1: se = 0.0532 with SSTₓ = 4025, so σ̂ ≈ 0.0532 × √4025 ≈ 3.38. The standard error is what R prints next to each coefficient. It uses SLR.5, so it's wrong under heteroscedasticity (Week 7).

20
New cards

What is SLR.6, and what distribution does the t-statistic follow under it?

SLR.6: u is independent of x and u ~ N(0, σ²). Then (β̂₁ − β₁)/se(β̂₁) ~ t with n − 2 df exactly

It's t rather than normal because we divide by an estimated se (σ̂ replaces σ). Without SLR.6, the same statistic is approximately normal in large samples (CLT), under SLR.1 to SLR.5. Small samples give the t heavier tails and so larger critical values. With n − 2 = 524 the t and normal are almost identical: 5% two-sided critical values of 1.965 vs 1.960.

21
New cards

wage1: β̂₁ = 0.5414, se = 0.0532, n = 526. Test H₀: β₁ = 0 vs H₁: β₁ ≠ 0 at 5%.

t = 0.5414/0.0532 ≈ 10.2. Critical value (t₅₂₄, 2.5% each tail) ≈ 1.965. |t| > 1.965, so reject H₀; p ≈ 3 × 10⁻²². Statistically significant at any conventional level

Four steps: (1) hypotheses; (2) statistic t = (β̂₁ − β₁,₀)/se(β̂₁); (3) critical value from t with n − 2 df; (4) decide (reject if |t| > c, or p < α) and conclude in words. Testing β₁ = 0 asks whether x has any linear association with y.

22
New cards

Test H₀: β₁ = 0.5 (two-sided, 5%) in wage1 (β̂₁ = 0.5414, se = 0.0532).

t = (0.5414 − 0.5)/0.0532 ≈ 0.78 < 1.965, so do not reject H₀

Always subtract the hypothesised value: t = (β̂₁ − β₁,₀)/se. Consistent with the CI: 0.5 lies inside [0.437, 0.646].

23
New cards

Statistically significant vs economically important: what's the difference, using wage1?

Statistical: t = 10.2, very unlikely if β₁ = 0. Economic: $0.54/hour per year ≈ 9% of the mean wage ($5.90), which is the size. Always report the size: with huge samples, tiny slopes become significant. And still not causal: a t-test can't fix a failure of SLR.4

Significance is about precision relative to the null. Importance is about magnitude in context. Neither implies causality (ability bias would make the estimate biased however large the t is).

24
New cards

Construct the 95% CI for β₁ in wage1 (β̂₁ = 0.5414, se = 0.0532, c = 1.965).

β̂₁ ± c·se = 0.5414 ± 1.965 × 0.0532 = 0.5414 ± 0.1045 = [0.437, 0.646]

General form: [β̂₁ − c·se(β̂₁), β̂₁ + c·se(β̂₁)], where c = the 97.5th percentile of t with n − 2 df. Where it comes from: P(−c ≤ (β̂₁ − β₁)/se ≤ c) = 0.95; rearrange to isolate β₁.

25
New cards

Correct interpretation of a 95% confidence interval?

Intervals built this way contain the true β₁ in 95% of samples. β₁ is fixed; the interval endpoints are random

The probability is over repeated samples. Once you've computed a specific interval, it either contains β₁ or it doesn't. Saying '95% probability β₁ lies in [0.437, 0.646]' is the classic wrong answer.

26
New cards

How do confidence intervals and hypothesis tests relate?

A 95% CI is the set of values a two-sided 5% test would not reject. A hypothesised value outside the CI is rejected; one inside is not

wage1: 0 is outside [0.437, 0.646], so reject H₀: β₁ = 0 (t = 10.2). 0.5 is inside, so don't reject H₀: β₁ = 0.5 (t = 0.78).

27
New cards

What does each result need? (unbiased, consistent, variance formula, unbiased σ̂², BLUE, exact t, approximate t)

Unbiased: SLR.1–4. Consistent: SLR.1–3 + E(u) = 0 + Cov(x,u) = 0. Var = σ²/SSTₓ: SLR.1–5. Unbiased σ̂²: SLR.1–5. BLUE: SLR.1–5. Exact t: SLR.1–6. Approximate t/normal: SLR.1–5 + large n

Pattern: SLR.4 is about the centre (bias). SLR.5 is about the spread (variance, standard errors, BLUE). SLR.6 is about the exact shape (the t distribution). The FOCs are not an assumption for any of these; they only fix the df at n − 2.