Final Exam 2023 — Moed A (Feb 15, 2023) · Dr. Aluma Dembo · worked-solution

2023 Moed A — Flight Prices & Omitted Variables

Dataset: Part A: 101,336 TLV departures, June 2021 (price, duration, stops, weekday). Part B: N=78 anonymised sample, Y on X1 and X2

2023 Moed A — Flight Prices & Omitted Variables

Part of: Intro Econometrics Final Exam 2023 — Moed A — Introduction to Econometrics, Dr. Aluma Dembo (15 Feb 2023, 3 hours, open notes, 100 points) Key concepts: Log-Log Model, Elasticity, t-test, p-value, F-test, Dummy Variable, Base Category, Interaction Term, Chow Test, Change of Variables, R-squared, Adjusted R-squared, Multicollinearity, Omitted Variable Bias, Classical Assumptions, Quadratic Model, Marginal Effect, Heteroskedasticity, Residuals Builds on: Lec_08-Goodness of Fit & Multicollinearity, Lec_10-Dummy Variables, Lec_11-Interaction Terms, Lec_12-Heteroskedasticity, Lec_13-Exam Review (covers the whole course, including Lectures 1 to 7 and 9, which have no notes), Assign_02-Functional Form & the Money-Supply Log-Log Model, Assign_03-Change of Variables, Regression Through the Origin & the Housing Model, Assign_05-Quadratics, Dummy Variables, Interactions & the Chow Test

Fully official

The source PDF contains the paper and the official answer key for all 18 items. Every answer below follows the key. Where the key looks inconsistent with the question wording, the answer is kept and a > [!warning] shows the check (see Q4, Q5, Q6, Q16, Q17). No question is AI-solved. The one place where I finished a conclusion the key left open is Q16 (flagged there).

Exam rules that shape every answer
  • Multiple-choice answers need a 1 to 2 sentence justification or they score 0, so each MC item below gives the verdict, the reason, and why each wrong option is wrong.
  • Significance level is 5% unless stated. Numbers may be rounded to 3 significant figures. Use the printed tables, not R's qt() or qf() (see Lec_13-Exam Review).
  • Course assumption labels used by the key: A1 mean-zero error, A2 exogeneity E[u∣x]=0E[u \mid x]=0, A3 no autocorrelation, A4 homoskedasticity.

What this paper tests

Two datasets, one toolkit

Part A (Q1 to Q8), flight prices in logs:

  1. Q1 to Q3: reading a log-log model (constant elasticity) and the t-test, including testing H0:β1=1H_0:\beta_1=1 by rebuilding the t-statistic from the reported SE.
  2. Q4 to Q6: dummy variables and a base category, a joint F-test written with R2R^2, and what a change of units does to the coefficients.
  3. Q7 to Q8: interactions as a test for "different slopes", and a Chow test from separate regressions.

Part B (Q9 to Q18), a small anonymised sample (N=78N=78): 4. Q9 to Q12: an insignificant variable, multicollinearity, omitted variable bias and the bias formula. 5. Q13 to Q16: adding a quadratic, nested F-test with anova(), marginal effects, one-sided t-test. 6. Q17 to Q18: residual plots, heteroskedasticity/misspecification, and what OLS guarantees about residuals.

The unlocking idea for Part B

Compare the three fits. ols1 (Y on X1, X2) has a curved residual plot, ols3 (adds X12X1^2) has a shapeless one and adjusted R2R^2 jumps from 0.215 to 0.810. The true relationship is quadratic in X1, so everything in ols1 and ols2 suffers from an omitted variable.


Part A — Flights from TLV (101,336 observations)

Data: Price (NIS) and Duration (minutes) for every departing flight in June 2021.

Q1 — Which model has constant elasticity? [5 pts]

Question

The researcher believes the price elasticity with respect to duration is constant. Which model does he use? [2 pts] select one, [3 pts] explain.

Option Model
A Price=β0+β1 Duration+uPrice = \beta_0 + \beta_1\,Duration + u
B log⁡(Price)=β0+β1 Duration+u\log(Price) = \beta_0 + \beta_1\,Duration + u
C log⁡(Price)=β0+β1log⁡(Duration)+u\log(Price) = \beta_0 + \beta_1\log(Duration) + u
D Duration=β0+β1 Price+uDuration = \beta_0 + \beta_1\,Price + u
E log⁡(Duration)=β0+β1 Price+u\log(Duration) = \beta_0 + \beta_1\,Price + u
F log⁡(Duration)=β0+β1log⁡(Price)+u\log(Duration) = \beta_0 + \beta_1\log(Price) + u
G None of the above / more than one correct
Answer: C

It is a log-log model with price as the outcome, so β1\beta_1 is the constant elasticity of price with respect to duration.

Why C. In a log-log model dlog⁡Pricedlog⁡Duration=β1\frac{d\log Price}{d\log Duration}=\beta_1, which is the elasticity (%ΔPrice\%\Delta Price per 1%ΔDuration1\%\Delta Duration) and does not depend on the level of either variable (see Lec_13-Exam Review, section 3.3, and Assign_02-Functional Form & the Money-Supply Log-Log Model).

Why the others are wrong.

  • A (level-level): β1\beta_1 is NIS per minute, so the elasticity β1⋅Duration/Price\beta_1\cdot Duration/Price changes along the line.
  • B (log-level): β1\beta_1 is a percent change in price per extra minute, not per percent of duration; the elasticity is β1⋅Duration\beta_1\cdot Duration, which grows with duration.
  • D, E, F: price is on the right-hand side (duration is the outcome). F is log-log but it measures the elasticity of duration with respect to price, the opposite direction.
  • G: only one option fits, so G is wrong.

Q2 — p-values from t-statistics [6 pts]

Question

The estimated slope is 0.958 (t = 249.17) and the intercept is 0.6 (t = 22.97). What can you say about the p-values of the significance tests for β1\beta_1 and β0\beta_0? [2 pts] select one, [4 pts] explain.

Option Statement
A Both p-values are less than 0.01
B β1\beta_1 p-value below 0.01, β0\beta_0 p-value above 0.01
C β1\beta_1 p-value above 0.01, β0\beta_0 p-value below 0.01
D Both above 0.01
E Cannot say with the information provided
Answer: A

If ∣t∣|t| exceeds the critical value at level α\alpha, then the p-value is below α\alpha. With N=101,336N=101{,}336 the 1% two-sided critical value is 2.576. Both 249.17249.17 and 22.9722.97 are far above it, so both p-values are below 0.01.

Why A. The p-value is the probability of a ∣t∣|t| at least this large under H0H_0. With about 101,000 observations the t distribution is the standard normal for practical purposes (critical value 2.576 at 1%). A t-statistic of 22.97 is already roughly 9 times the cutoff.

Why the others are wrong. B, C and D each claim that at least one p-value exceeds 0.01, which would need ∣t∣<2.576|t|<2.576; neither statistic is close. E is wrong because the table value is all that is needed: a t-statistic plus the degrees of freedom pins down which side of any significance level the p-value falls.


Q3 — Testing H0:β1=1H_0:\beta_1 = 1 [6 pts]

Question

From earlier studies the researcher believes the slope is 1. At which level(s) of significance can H0:β1=1H_0:\beta_1=1 be rejected? Use the estimates from Q2. [2 pts] select all correct: α=1%\alpha=1\%, α=5%\alpha=5\%, α=10%\alpha=10\%, not enough information. [4 pts] show your work.

Answer: rejected at 1%, 5% and 10%

The reported t=249.17t=249.17 tests H0:β1=0H_0:\beta_1=0, so first recover the standard error, then build the new t-statistic against 1.

Working

tβ1=0=β^1SE(β^1) ⇒ SE(β^1)=0.958249.17=0.00385t_{\beta_1=0}=\frac{\hat\beta_1}{SE(\hat\beta_1)}\ \Rightarrow\ SE(\hat\beta_1)=\frac{0.958}{249.17}=0.00385

tβ1=1=β^1−1SE(β^1)=0.958−10.00385=−10.9t_{\beta_1=1}=\frac{\hat\beta_1-1}{SE(\hat\beta_1)}=\frac{0.958-1}{0.00385}=-10.9

∣−10.9∣|{-10.9}| exceeds 2.576 (1%), 1.96 (5%) and 1.645 (10%), so reject at all three levels.

Why "not enough information" is wrong. The standard error is not printed, but it is recoverable from β^1\hat\beta_1 and its t-statistic. Common trap: reusing 249.17 as the statistic for H0:β1=1H_0:\beta_1=1; that number is only for H0:β1=0H_0:\beta_1=0.

Intuition

Statistical significance is not economic significance: 0.958 and 1 are close, but with 100,000 flights the SE is tiny (0.004), so even a 0.04 gap is "ten standard errors" away.


Model (1) used in Q4 to Q6

Variables: Day (date in June), Stop1, Stop2, Stop3 (dummies, base category = non-stop), day_of_week2 to day_of_week7 (dummies, base category = Sunday).

log⁡(Price)=β0+β1log⁡(Duration)+β2 Day+β3 Stop1+β4 Stop2+β5 Stop3+∑k=27γk dowk+u\log(Price)=\beta_0+\beta_1\log(Duration)+\beta_2\,Day+\beta_3\,Stop1+\beta_4\,Stop2+\beta_5\,Stop3+\sum_{k=2}^{7}\gamma_k\,dow_k+u

R output (transcribed):

Variable Estimate Std. Error t value Pr(>abs t)
(Intercept) 0.698029 0.029452 23.70 below 2e-16
lg.duration 1.025149 0.004984 205.67 below 2e-16
day -0.000160 0.000247 -0.65 0.517
stop1 -0.601653 0.012850 -46.82 below 2e-16
stop2 -0.488720 0.014081 -34.71 below 2e-16
stop3 -0.544587 0.017891 -30.44 below 2e-16
day_of_week2 -0.020264 0.008166 -2.48 0.013
day_of_week3 -0.031064 0.007748 -4.01 6.1e-05
day_of_week4 -0.005644 0.007678 -0.74 0.462
day_of_week5 0.002648 0.008208 0.32 0.747
day_of_week6 0.003238 0.008087 0.40 0.689
day_of_week7 -0.068507 0.008163 -8.39 below 2e-16
R output
Residual standard error: 0.674 on 101324 degrees of freedom
Multiple R-squared: 0.398,   Adjusted R-squared: 0.398
F-statistic: 6.09e+03 on 11 and 101324 DF, p-value: < 2e-16

Q4 — Cheapest day of the week [6 pts]

Question

According to model (1), on which day of the week is it cheapest to fly? Consider equal-duration non-stop flights in the same calendar week (e.g. Sunday 6 June to Saturday 12 June 2021). Explain.

Answer: Monday

Compare the γk\gamma_k coefficients: each is the log-price difference relative to the Sunday base. The most negative one, γ^7=−0.0685\hat\gamma_7=-0.0685, belongs to Monday in the key's labelling.

Working (as in the key)

For a non-stop Sunday flight: log⁡(Price)Sun=0.698+1.03log⁡(Duration)−0.0001 Day\log(Price)_{Sun}=0.698+1.03\log(Duration)-0.0001\,Day.

The same flight on the next day (Monday):

log⁡(Price)Mon=log⁡(Price)Sun−0.0001−0.068507\log(Price)_{Mon}=\log(Price)_{Sun}-0.0001-0.068507

and the day after (Tuesday): log⁡(Price)Tue=log⁡(Price)Sun−0.0001⋅2−0.020264\log(Price)_{Tue}=\log(Price)_{Sun}-0.0001\cdot 2-0.020264, and so on. Since DayDay rises by one each calendar day, the DayDay term only shifts every weekday by a further −0.00016-0.00016 per day, which is tiny and insignificant (p = 0.517).

Largest weekday discount: −0.0685-0.0685 (Monday), then −0.0311-0.0311 (Wednesday), −0.0203-0.0203 (Tuesday). Monday is cheapest by a wide margin (roughly 7% below Sunday, from 100×γ^7100\times\hat\gamma_7).

Weekday labelling in the key

The question says day_of_week2 is Tuesday, day_of_week3 is Wednesday and Sunday is the base, which would make day_of_week7 a Sunday again if the numbering simply continued. The key resolves this by reading day_of_week7 as Monday. I kept the key's answer (Monday); the dummy numbering in the stem is the sloppy part. The method, picking the most negative dummy coefficient, is what the examiner is testing.


Q5 — Does the number of stops matter? [8 pts]

Question

The researcher checks the claim that the number of stops does not affect the price (assuming at least one layover). (A) [2] Null hypothesis? (B) [2] Restricted model? (C) [4] Given R2R^2 of the restricted model =0.3947=0.3947, is the claim correct? Show the test statistic and critical value.

a. Null hypothesis

Answer

H0: β3=β4=β5=0H_0:\ \beta_3=\beta_4=\beta_5=0 (the coefficients on Stop1, Stop2, Stop3).

b. Restricted model

Answer

Impose H0H_0 by dropping the three stop dummies:

log⁡(Price)=β0+β1log⁡(Duration)+β2 Day+γ2 dow2+γ3 dow3+γ4 dow4+γ5 dow5+γ6 dow6+γ7 dow7+u\log(Price)=\beta_0+\beta_1\log(Duration)+\beta_2\,Day+\gamma_2\,dow_2+\gamma_3\,dow_3+\gamma_4\,dow_4+\gamma_5\,dow_5+\gamma_6\,dow_6+\gamma_7\,dow_7+u

c. The F-test, written with R2R^2

Answer: F=185.14F=185.14 versus critical value 2.62.6, so reject H0H_0; the claim is not correct

F=(RUR2−RR2)/q(1−RUR2)/(N−M)=(0.398−0.3947)/3(1−0.398)/101324=185.14F=\frac{(R^2_{UR}-R^2_{R})/q}{(1-R^2_{UR})/(N-M)}=\frac{(0.398-0.3947)/3}{(1-0.398)/101324}=185.14

With q=3q=3 restrictions and N−M=101,324N-M=101{,}324 degrees of freedom the 5% critical value is about 2.6. Since 185.14≫2.6185.14\gg 2.6, the stop dummies are jointly significant: the number of stops affects price.

Why the R2R^2 form works. RSS=(1−R2)⋅TSSRSS=(1-R^2)\cdot TSS and both models share the same TSSTSS (same dependent variable), so the TSSTSS cancels in the ratio (see Lec_13-Exam Review, section 1.7 and 1.8, and Lec_08-Goodness of Fit & Multicollinearity).

Wording versus the key's hypothesis

The question says "assuming there is at least one layover stop". Read literally, that is a claim that 1, 2 and 3 stops cost the same (H0:β3=β4=β5H_0:\beta_3=\beta_4=\beta_5, q=2q=2, restricted model = one "any stop" dummy), which would need a different restricted R2R^2. The key instead tests that all three stop effects are zero relative to non-stop (q=3q=3) and supplies RR2=0.3947R^2_R=0.3947 for that restricted model. Following the key (the given RR2R^2_R only fits the "drop all three dummies" model). The coefficients (-0.60, -0.49, -0.54) are all large and negative, so the conclusion is not in doubt either way.


Q6 — Change of units [6 pts]

Question

How do the estimates in model (1) change if duration is measured in hours instead of minutes and price in USD instead of NIS? Use an exchange rate of 3.5 USD for every 1 NIS.

Answer (official)

The slope on log-Duration, the Day, Stop and weekday dummy coefficients do not change; only the intercept changes. The key's new intercept is 3.064 (or 1.976 if the exchange rate is applied the other way round; both accepted).

Working

Duration in hours = Duration in minutes divided by 60. Price in USD is the NIS price scaled by the exchange rate. Because both variables are in logs, a rescaling just adds a constant:

log⁡(Pricenew)=log⁡(Price)+log⁡(3.5)log⁡(Durationnew)=log⁡(Duration)−log⁡(60)\log(Price_{new})=\log(Price)+\log(3.5)\qquad \log(Duration_{new})=\log(Duration)-\log(60)

Substituting into model (1), only the constant picks up the extra terms, and the new intercept β0new\beta_0^{new} satisfies

β0new=β0+β1log⁡(60)+log⁡(3.5)\beta_0^{new}=\beta_0+\beta_1\log(60)+\log(3.5)

Using the key's numbers (log⁡60=1.778\log 60 = 1.778, log⁡3.5=0.544\log 3.5 = 0.544):

β0new=0.698+1.025×1.778+0.544=3.064\beta_0^{new}=0.698+1.025\times 1.778+0.544=\mathbf{3.064}

With the opposite convention for the exchange rate: 0.698+1.025×1.778−0.544=1.9760.698+1.025\times1.778-0.544=\mathbf{1.976}.

Slopes are unchanged because a log turns a multiplicative rescaling into an additive shift: log⁡(cX)=log⁡c+log⁡X\log(cX)=\log c+\log X (see Assign_03-Change of Variables, Regression Through the Origin & the Housing Model and Lec_13-Exam Review, section 3.1).

Which logarithm?

The key's constants 1.7781.778 and 0.5440.544 are log⁡10(60)\log_{10}(60) and log⁡10(3.5)\log_{10}(3.5). If the course's lg. variables are natural logs, the constants are ln⁡60=4.094\ln 60 = 4.094 and ln⁡3.5=1.253\ln 3.5 = 1.253 and the intercepts would be 0.698+1.025(4.094)±1.2530.698+1.025(4.094)\pm1.253, that is about 6.15 or 3.64. My check: 0.698+1.025×log⁡1060+log⁡103.5=3.0650.698 + 1.025\times\log_{10}60 + \log_{10}3.5 = 3.065 (key rounds to 3.064). The structure (only the intercept moves) is the same either way. I kept the official numbers.


Q7 — Is the effect of duration different for Europe? [7 pts]

Question

The researcher considers flights to Europe and wants to test whether prices there are less affected by duration than prices to other continents. He has dest_Europe (=1 if the destination is in Europe). (A) [3] Which variable(s) should be added? (B) [4] He adds them and the p-value is below 0.05: what does this say about the coefficient on duration in model (1)? State the hypothesis being tested.

a. Variables to add

Answer

Add two variables: the dummy dest_Europe and the interaction duration_Europe=Duration×dest_Europeduration\_Europe = Duration\times dest\_Europe.

The dummy shifts the intercept for Europe; the interaction lets the slope on duration differ (see Lec_11-Interaction Terms). A different slope is the hypothesis, but keeping the dummy alongside the interaction preserves the hierarchy of the model and avoids forcing both groups through the same intercept.

b. Reading the result

Answer

The hypothesis tested is H0: βduration_Europe=0H_0:\ \beta_{duration\_Europe}=0 (the duration slope is the same for Europe and elsewhere). A p-value below 0.05 rejects it, so model (1) omits an important variable. That means model (1) has omitted variable bias: the estimate of the duration coefficient (and possibly others) is biased, and it is not efficient.

Intuition

Model (1) forces a single duration slope on every flight. If Europe flights genuinely have a different slope, the single estimate is a blend, correct for neither group.


Q8 — Chow test: Europe versus North America [6 pts]

Question

Model (1) is estimated separately for the 6,000 flights to Europe and the 6,000 to North America. Pooled RSS =23,599=23{,}599; anova(fit_EU) residual SS =16,835=16{,}835; anova(fit_N_A) residual SS =5,638=5{,}638. Test whether the two destinations share the same pricing model. (A) [2] Degrees of freedom? (B) [2] F-statistic? (C) [2] Conclusion at α=1%\alpha=1\%?

a. Degrees of freedom

Answer: q=12q=12 and N−M=11,976N-M=11{,}976

The restricted (pooled) model has 12 parameters; the unrestricted model fits all 12 separately for each group. Hence q=12q=12 restrictions (all 12 coefficients equal across groups), and N−M=(6000+6000)−(12+12)=11,976N-M=(6000+6000)-(12+12)=11{,}976.

b. F-statistic

Answer: F=50.00F=50.00

F=(RSSpooled−(RSSEU+RSSNA))/q(RSSEU+RSSNA)/(N−M)=(23599−(16835+5638))/12(16835+5638)/11976=50.00F=\frac{\big(RSS_{pooled}-(RSS_{EU}+RSS_{NA})\big)/q}{(RSS_{EU}+RSS_{NA})/(N-M)}=\frac{\big(23599-(16835+5638)\big)/12}{(16835+5638)/11976}=50.00

Check: numerator =(23599−22473)/12=93.83=(23599-22473)/12=93.83, denominator =22473/11976=1.8765=22473/11976=1.8765, ratio =50.00=50.00.

c. Conclusion

Answer: reject H0H_0

50>2.3250>2.32, where 2.32 is the closest critical value in the printed table at 1%. Europe and North America do not share the same pricing model.

Version 2 of the Chow test

This uses the separate-regressions form of the Chow test, which the lecture notes prefer when groups may have different error variances: the unrestricted RSS is simply the sum of the two group RSSs (Lec_13-Exam Review, section 3.6; Assign_05-Quadratics, Dummy Variables, Interactions & the Chow Test).


Part B — A colleague's project (N=78N=78, variables Y, X1, X2)

Summary statistics (from summary(sample.data[, c('X1','X2','Y')])):

X1 X2 Y
Min 0.8835 -28.14 0
1st Qu. 22.4332 22.43 962
Median 55.6736 52.39 1389
Mean 51.5469 49.20 1279
3rd Qu. 81.0683 72.86 1645
Max 99.8656 126.95 2075

Model (1): Y=β0+β1X1+β2X2+uY=\beta_0+\beta_1X1+\beta_2X2+u (ols1)

Estimate Std. Error t value Pr(>abs t)
(Intercept) 914.954 94.271 9.706 6.85e-15
X1 10.944 2.806 3.900 0.000208
X2 -4.074 2.456 -1.659 0.101303

Residual SE 425.3 on 75 df; R2=0.2353R^2=0.2353, adjusted R2=0.2149R^2=0.2149; F = 11.54 on 2 and 75 DF, p = 4.275e-05.

Model (2): Y=β0+β1X1+uY=\beta_0+\beta_1X1+u (ols2)

Estimate Std. Error t value Pr(>abs t)
(Intercept) 913.309 95.347 9.579 1.05e-14
X1 7.088 1.590 4.457 2.82e-05

Residual SE 430.2 on 76 df; R2=0.2072R^2=0.2072, adjusted R2=0.1968R^2=0.1968; F = 19.87 on 1 and 76 DF. cor(X1, X2) =0.8283643=0.8283643.


Q9 — Is X2 significant? [4 pts]

Question

The colleague thinks X2 is not significant and should be omitted. [1] Which number(s) did she use? [3] Explain.

Answer

The p-value for X2: 0.101303 (the t-statistic −1.659-1.659 is also accepted).

It is greater than 0.05, so we cannot reject H0:β2=0H_0:\beta_2=0: the estimate is not statistically significant at the 5% level. If she cites the t-statistic she must compare it with the critical value (∣−1.659∣<|{-1.659}|< about 1.99 for 75 df, so same conclusion).


Q10 — Why does the fit get worse when X2 is dropped? [6 pts]

Question

After removing X2, goodness-of-fit for model (2) is worse than for model (1). [2] What numbers show this? [4] Normally dropping an insignificant variable improves fit. What might explain the opposite?

Answer

Numbers: adjusted R2R^2 is 0.1968 in ols2 versus 0.2149 in ols1.

Explanation: X1 and X2 are correlated at 0.828 (above the 0.7 rule of thumb). Including both created partial multicollinearity, which inflates the standard error of X2 and makes it look insignificant. But if X2 truly affects Y, then omitting it from model (2) worsens the fit.

Why the adjusted R2R^2. Plain R2R^2 can never rise when a variable is dropped (0.2353 to 0.2072). Adjusted R2R^2 penalises the extra regressor, so it is the fair comparison. Here even the penalised measure falls, so X2 is carrying real explanatory power (Lec_08-Goodness of Fit & Multicollinearity).

Intuition

X1 and X2 move together, so the data cannot say how much of the effect is X1 and how much is X2. The pair is jointly useful (overall fit) but individually imprecise (large SE on X2). Insignificance from collinearity is not evidence that the variable is irrelevant.


Q11 — What happens to β^1\hat\beta_1 when X2 is omitted? [5 pts]

Question

Based on the outputs, what happens to the estimate of β1\beta_1 when X2 is omitted, and why? Select one.

Option Statement
A biased downwards and A1 is violated
B biased upwards and A2 is violated
C biased downwards and A3 is violated
D biased upwards and A4 is violated
E None of the above
Answer: E

Omitting X2 violates A2 (exogeneity), but the bias is downward: the estimate falls from 10.944 to 7.088. No option pairs "downward" with A2, so none of A to D is right.

Direction of the bias. The bias of the short-regression slope is β2⋅S12S11\beta_2\cdot\frac{S_{12}}{S_{11}} with β2=−4.074<0\beta_2=-4.074<0 (effect of the omitted variable) and S12>0S_{12}>0 (X1 and X2 are positively correlated, 0.828). Negative times positive is negative, so the estimate is biased down, which matches the drop in the estimates (Lec_13-Exam Review, section 2.3).

Why each option fails.

  • A: right direction, wrong assumption. A1 (mean-zero errors) is not the one broken; the error now contains β2X2\beta_2 X2 and is correlated with X1, which is a failure of A2.
  • B: right assumption, wrong direction (upward).
  • C: right direction, wrong assumption (A3 is no autocorrelation; the data are an independent random sample).
  • D: wrong direction and wrong assumption (A4 is homoskedasticity, not the issue here).

Q12 — Recovering the variance of X1 [4 pts]

Question

If Cov(X1,X2)=899.626Cov(X1,X2)=899.626, what is the variance of X1?

Answer: S11≈950S_{11}\approx 950 (key: 950.5, "around 950 with rounding errors")

Use the omitted variable bias formula: short slope = long slope + β2 S12/S11\beta_2\,S_{12}/S_{11}.

7.088=10.944−4.074⋅899.626S117.088=10.944-4.074\cdot\frac{899.626}{S_{11}}

Solve for S11S_{11}:

S11=4.074×899.62610.944−7.088=3665.13.856=950.5S_{11}=\frac{4.074\times 899.626}{10.944-7.088}=\frac{3665.1}{3.856}=950.5

The key writes the coefficients rounded (10.94 and 4.07) and gets the same 950.5; rounding differences move the third digit only.


Q13 — Adding a quadratic term and the nested F-test [6 pts]

Question

The colleague suspects a quadratic relationship and uses model (3): Y=β0+β1X1+β2X2+β3(X1)2+uY=\beta_0+\beta_1X1+\beta_2X2+\beta_3(X1)^2+u. She emails the output below. (A) [2] What null hypothesis does the R code test? (B) [2] Degrees of freedom? (B') [1] Critical value at 5%? (C) [1] Conclusion.

R output
> ols3 = lm(Y ~ X1 + X2 + X1.sq, sample.data)
> anova(ols2, ols3)
Model 1: Y ~ X1
Model 2: Y ~ X1 + X2 + X1.sq
  Res.Df      RSS Df Sum of Sq      F    Pr(>F)
1     76 14063548
2     74  3233078  2  10830470 123.95 < 2.2e-16 ***

summary(ols3):

Estimate Std. Error t value Pr(>abs t)
(Intercept) 124.53213 69.19819 1.800 0.076
X1 59.72394 3.45884 17.267 below 2e-16
X2 -2.55818 1.21108 -2.112 0.038
X1.sq -0.50026 0.03253 -15.379 below 2e-16

Residual SE 209 on 74 df; R2=0.8178R^2=0.8178, adjusted R2=0.8104R^2=0.8104; F = 110.7 on 3 and 74 DF.

a. Null hypothesis

Answer

H0: β2=0, β3=0H_0:\ \beta_2=0,\ \beta_3=0: the two variables that ols3 adds to ols2 (X2 and X12X1^2) have no effect.

b. Degrees of freedom and critical value

Answer

Numerator q=2q=2; denominator N−M=78−4=74N-M=78-4=74. Critical value at 5%: 3.15 (3.10 also accepted).

c. Conclusion

Answer: reject H0H_0

The test statistic 123.95 is far above 3.15 (equivalently, the p-value 2.2×10−162.2\times10^{-16} is far below 0.05).

Check of the F-statistic from the RSS values

F=(14063548−3233078)/23233078/74=541523543690=123.95F=\frac{(14063548-3233078)/2}{3233078/74}=\frac{5415235}{43690}=123.95

anova(ols2, ols3) is the R version of the nested F-test: the restricted model is the smaller one (Lec_13-Exam Review, section 1.7).


Q14 — Is the marginal effect of X1 always zero? [5 pts]

Question

The colleague claims that, based on ols3, we can reject the null that the marginal effect of X1 on Y is always 0. True, false or cannot say? [1] choose, [4] explain.

Answer: TRUE or CANNOT SAY (both accepted)

The marginal effect of X1 is 59.73−1.00⋅X159.73-1.00\cdot X1. "Always zero" means β1=0\beta_1=0 and β3=0\beta_3=0 jointly, which needs an F-test that ols3's output does not report directly. However, both coefficients are individually highly significant (t = 17.267 and -15.379), so the joint test would certainly reject.

Why not a plain t-test. A t-test handles one coefficient. The effect is zero at every X1 only if both pieces vanish, so the hypothesis is joint. When each piece alone is already significant, the joint F-test is also significant (Lec_13-Exam Review, section 1.7).

PP01_q14_marginal_effect_quadratic

The estimated marginal effect of X1 in ols3 is a straight line that crosses zero at X1≈59.7X1\approx59.7 (=59.72394/(2×0.50026)=59.72394/(2\times0.50026)). It is zero at one point only, but nonzero everywhere else, so "always zero" is rejected.

Intuition

Do not confuse "the effect is zero somewhere" (at X1≈59.7X1\approx 59.7 the effect changes sign: this is the turning point of the parabola) with "the effect is zero everywhere".


Q15 — Is the rate of return of X2 always negative? [5 pts]

Question

The colleague claims that, based on ols3, the rate of return of X2 on Y is always negative. True, false or cannot say?

Answer: TRUE

In a level-level model the rate of return is the marginal effect divided by the outcome: β^2/Y\hat\beta_2/Y. The coefficient is negative (−2.558-2.558) and the outcome is positive, so the rate of return is always negative.

Why. Rate of change =ΔY/ΔXY=\frac{\Delta Y/\Delta X}{Y}, which in level-level form is β^2⋅1Y\hat\beta_2\cdot\frac{1}{Y} (Lec_13-Exam Review, section 3.3). X2 enters only linearly in ols3, so its marginal effect is the constant −2.558-2.558 and the sign cannot change.

Edge case

The summary table shows min⁡Y=0\min Y=0, where the ratio is undefined. The key ignores this and treats Y as strictly positive. I kept the official TRUE.


Q16 — One-sided test on X2 [4 pts]

Question

Test the null hypothesis that the effect of X2 in model (3) is negative at the α=5%\alpha=5\% level.

Answer (official method)

The key's guidance: compare half the p-value to 0.05, or run the full test with the one-tailed 5% critical value. From ols3: t=−2.112t=-2.112, two-sided p=0.038p=0.038, so the one-sided p-value is 0.038/2=0.019<0.050.038/2=0.019<0.05.

Conclusion not stated in the key (my completion)

The key only states the method, not a verdict. With the one-sided method it describes (half the p-value, 0.019<0.050.019<0.05), the evidence supports the effect of X2 being negative: we reject H0:β2≥0H_0:\beta_2\ge 0 in favour of H1:β2<0H_1:\beta_2<0. The critical value for a one-tailed 5% test with 74 df is about −1.67-1.67, and −2.112<−1.67-2.112<-1.67 agrees. Wording caveat: if the null is taken literally as "β2<0\beta_2<0", the point estimate is already negative and the test would fail to reject; the key's half-p-value instruction only makes sense when the claim to be established is placed in the alternative. I followed the key's method.

PP01_q16_one_sided_t_test

The shaded lower tail beyond t=−2.112t=-2.112 holds half of the two-sided p-value (0.019).


Q17 — What can reasonably be said about ols1? [6 pts]

Question

Based on the residual plots and the outputs of ols1 and ols3, which statements are reasonable? [2] select one, [4] explain.

Option Statement
A The estimate of β1\beta_1 in ols1 is unbiased
B The standard error of β1\beta_1 in ols1 is biased
C Both A and B are reasonable
D Neither A nor B is reasonable

PP01_q17_residual_plots_ols1_ols3

Real data, taken from the exam PDF. Left: ols1 residuals against X1 show a clear arch (curvature) and a wider spread at the ends. Right: ols3 residuals are a shapeless cloud.

Answer: B

ols3 fits far better, so it is likely closest to the correct model. Then in ols1 the estimate of β1\beta_1 is biased because X12X1^2 is omitted, which breaks A2. The plot also suggests heteroskedasticity, which means the standard error of β1\beta_1 is biased too. Only (B) is reasonable.

Why each option.

  • A is wrong: ols1 omits the relevant X12X1^2 (the squared term is highly significant in ols3), and X1 is correlated with X12X1^2, so the exogeneity assumption fails and the estimate is biased.
  • B is right: under heteroskedasticity (A4 fails) the usual SE formula is wrong; the key reads the changing spread of residuals in the ols1 plot that way (Lec_12-Heteroskedasticity).
  • C is wrong because A is wrong. D is wrong because B is reasonable.
Small discrepancy in the key

The key quotes the adjusted R2R^2 of ols3 as "0.83". The output shows adjusted R2=0.8104R^2=0.8104 (R2=0.8178R^2=0.8178), so the figure is slightly off. The argument and the answer (B) do not change.

Intuition

The arch in the left plot is as much a sign of a missing curvature term (misspecification) as of heteroskedasticity. Either way, the standard errors from ols1 should not be trusted.


Q18 — Residuals in ols1 versus ols3 [5 pts]

Question

CLAIM 1: the sum of residuals is higher in ols3 than in ols1. CLAIM 2: the sum of the squared residuals is lower in ols3 than in ols1. [2] select one, [3] explain.

Option Statement
A Only Claim 1 is true
B Only Claim 2 is true
C Both are true
D Neither is true
Answer: B

Claim 1 is false: with an intercept, OLS residuals always sum to exactly 0, so the sums are equal in ols1 and ols3. Claim 2 is true: ols3 has the higher R2R^2 (0.8178 versus 0.2353), and since RSS=(1−R2) TSSRSS=(1-R^2)\,TSS with the same TSSTSS, its RSS is lower.

Numbers behind Claim 2. From the outputs, RSSols1=425.32×75≈13.57×106RSS_{ols1}=425.3^2\times75\approx13.57\times10^6 and RSSols3=3,233,078RSS_{ols3}=3{,}233{,}078 (from anova), which is about 4 times smaller. (Also RSSols2=14,063,548RSS_{ols2}=14{,}063{,}548, consistent with R2R^2 falling when X2 is dropped.)

Why the others are wrong. A and C rely on Claim 1, which is false because the first-order condition for the intercept forces ∑u^i=0\sum \hat u_i=0 in every OLS model that has a constant. D denies Claim 2, which the R2R^2 comparison confirms.


Recap table

Q Topic Tool Answer
1 Constant elasticity Log-log C
2 p-values from t t vs 2.576 A (both below 0.01)
3 H0:β1=1H_0:\beta_1=1 Rebuild t: SE=0.00385 t=−10.9t=-10.9, reject at 1%, 5%, 10%
4 Cheapest weekday Compare dummy coefs Monday (−0.0685-0.0685)
5 Stops claim F via R2R^2, q=3q=3 F=185.14>2.6F=185.14>2.6, reject
6 Change of units Logs shift the intercept only 3.064 (or 1.976)
7 Europe slope Dummy plus interaction Reject, OVB in (1)
8 Chow EU vs NA q=12q=12, df=11976df=11976 F=50.00>2.32F=50.00>2.32, reject
9 X2 significance p = 0.101 Not significant
10 Fit worsens Adj R2R^2, corr 0.828 Multicollinearity
11 OVB direction β2S12/S11\beta_2 S_{12}/S_{11} E (down, A2)
12 Variance of X1 Bias formula ≈950.5\approx 950.5
13 anova(ols2, ols3) Nested F, q=2q=2, df=74df=74 F=123.95>3.15F=123.95>3.15, reject
14 Marginal effect zero Joint F TRUE or CANNOT SAY
15 Rate of return of X2 β^2/Y\hat\beta_2/Y TRUE (negative)
16 One-sided X2 Half p-value 0.019<0.050.019<0.05
17 ols1 reliability Residual plot B
18 Residuals Sum is 0; RSS lower B