Final Exam 2025 — Moed A (25 Feb 2025) · Dr. Aluma Dembo · worked-solution

2025 Moed A — Wage Growth, Dummies, Chow Test & Log Models

2025 Moed A — Wage Growth, Dummies, Chow Test & Log Models

Part of: Intro Econometrics Final Exam 2025, Moed A (25 Feb 2025, 3 hours, open notes, 100 points) — Introduction to Econometrics, Dr. Aluma Dembo Key concepts: t-test, F-test, p-value, Confidence Interval, Dummy Variable, Base Category, Interaction Term, Quadratic Model, Marginal Effect, Chow Test, Change of Variables, Log-Level Model, Elasticity, R-squared, Adjusted R-squared, Multicollinearity, Heteroskedasticity, White Correction, Regression Through the Origin Builds on: Lec_08-Goodness of Fit & Multicollinearity, Lec_10-Dummy Variables, Lec_11-Interaction Terms, Lec_12-Heteroskedasticity, Lec_13-Exam Review, Assign_03-Change of Variables, Regression Through the Origin & the Housing Model, Assign_05-Quadratics, Dummy Variables, Interactions & the Chow Test

No official key: every answer in this note is AI-written

This paper came without a solutions file. All 18 questions (Q1 to Q18) are solved here from the course notes, the formula sheet logic and the R output printed in the paper. Each answer carries its own warning. Check before relying on it. Arithmetic was re-derived in Python. Critical values come from exact scipy quantiles; the exam expects the printed tables, whose nearest row can differ in the second decimal but never changes a conclusion below.

Format reminder

Open notes. Every multiple-choice answer needs a 1 to 2 sentence justification, so each MC question below gives the verdict, the justification, and why every other option is wrong. Rounding to 3 significant figures is allowed.


What this paper tests

The one idea behind each half
  • Part I (52 pts), Thai engineers' wages (N = 403). Everything is about categorical and non-linear regressors: an interaction (male.educ) gives men and women different education slopes (Q2, Q3); a squared term makes marginal effects depend on the sample point (Q4); the joint F-test is the right test for a whole dummy block (Q7) and for "no growth" (Q5); the Chow test compares separate regressions (Q8); and rescaling units moves coefficients predictably (Q9).
  • Part II (48 pts), sample.data (N = 125). The outcome is log(Y), so every coefficient is a log-level one. Themes: a one-sided t-test (Q10), why R2R^2 cannot compare a no-intercept model (Q11, Q12), detecting partial multicollinearity (Q13), confidence intervals (Q14), log-level and elasticity readings (Q15 to Q17), and what heteroskedasticity / White standard errors do to OLS (Q18).
  • Two traps worth remembering: the base category is the dummy that is missing (Q6), and the R2R^2 of a model with no intercept is not comparable (Q11).

Part I: Wage growth of engineers (52 pts)

Data (N = 403, Thailand, July 1998)

wage (monthly THB), yrs.exp (years at current job), yrs.exp.sq, start.wage, prev.yrs.exp (+ square), educ, is.male, male.educ (= educ × is.male), and degree dummies hs.dropout, hs.grad, polytech, college, post.college.

Q1 — Is prev.yrs.exp.sq irrelevant in Model [1]? (5 pts)

Question. Model [1]:

start.wage=β0+β1 is.male+β2 educ+β3 male.educ+β4 prev.yrs.exp+β5 prev.yrs.exp.sq+ustart.wage = \beta_0 + \beta_1\, is.male + \beta_2\, educ + \beta_3\, male.educ + \beta_4\, prev.yrs.exp + \beta_5\, prev.yrs.exp.sq + u

The R output wage.1 estimates Model [1]. What number lets you conclude prev.yrs.exp.sq does not affect starting wage at α=1%\alpha = 1\%? (1 pt for the number, 4 pts for the explanation.)

R output wage.1 (N = 403, 397 df):

Term Estimate Std. Error t Pr(>abs t)
(Intercept) 3393.344 2136.125 1.589 0.1130
is.male -17752.432 2827.402 -6.279 8.96e-10
educ 669.883 168.133 3.984 8.05e-05
male.educ 1556.837 208.436 7.469 5.18e-13
prev.yrs.exp 197.521 78.891 2.504 0.0127
prev.yrs.exp.sq -2.759 2.254 -1.224 0.2217

Residual SE 4812 on 397 df; R2=0.607R^2 = 0.607, adj. R2=0.6021R^2 = 0.6021; F = 122.6 on 5 and 397 df.

AI-written answer, not official. Check before relying on it.
Answer

The p-value 0.2217 (equivalently the t-statistic −1.224-1.224).

Explanation. Test H0:β5=0H_0:\beta_5 = 0 against H1:β5≠0H_1:\beta_5 \neq 0. The p-value 0.2217>0.010.2217 > 0.01, so we fail to reject H0H_0 at 1%. Equivalent route: ∣t∣=1.224|t| = 1.224 is below the two-sided 1% critical value t0.005(397)≈2.59t_{0.005}(397) \approx 2.59 (about 2.576 from the table's large-df row). So the squared term is statistically indistinguishable from zero.

Intuition

"Fail to reject" is not "proved irrelevant". With β^5=−2.759\hat\beta_5 = -2.759 and SE 2.2542.254 the data simply cannot tell the coefficient from zero. The question phrases it as "does not affect", which is the conventional reading of failing to reject.


Q2 — Gender-specific models from Model [2] (6 pts)

Question. Model [2]: start.wage=β0+β1 is.male+β2 educ+β3 male.educ+β4 prev.yrs.exp+ustart.wage = \beta_0 + \beta_1\, is.male + \beta_2\, educ + \beta_3\, male.educ + \beta_4\, prev.yrs.exp + u. Use the output wage.2.

R output wage.2 (398 df):

Term Estimate Std. Error t Pr(>abs t)
(Intercept) 3652.6 2126.9 1.717 0.0867
is.male -17489.3 2821.0 -6.200 1.42e-09
educ 671.2 168.2 3.989 7.88e-05
male.educ 1553.3 208.6 7.448 5.93e-13
prev.yrs.exp 105.3 23.4 4.500 8.92e-06

Residual SE 4815 on 398 df; R2=0.6055R^2 = 0.6055, adj. R2=0.6016R^2 = 0.6016.

AI-written answer, not official. Check before relying on it.

a. Estimated model for men and for women (2 pts)

Answer
  • Women (is.male=0is.male = 0, male.educ=0male.educ = 0): start.wage^=3652.6+671.2 educ+105.3 prev.yrs.exp\widehat{start.wage} = 3652.6 + 671.2\, educ + 105.3\, prev.yrs.exp
  • Men (is.male=1is.male = 1, male.educ=educmale.educ = educ): start.wage^=(3652.6−17489.3)+(671.2+1553.3) educ+105.3 prev.yrs.exp=−13836.7+2224.5 educ+105.3 prev.yrs.exp\widehat{start.wage} = (3652.6 - 17489.3) + (671.2 + 1553.3)\, educ + 105.3\, prev.yrs.exp = -13836.7 + 2224.5\, educ + 105.3\, prev.yrs.exp

Why. The dummy is.maleis.male shifts the intercept by −17489.3-17489.3; the interaction male.educmale.educ shifts the education slope by +1553.3+1553.3. The experience slope is common to both groups.

b. Man, 0 years previous experience, 12 years of education (1 pt)

Answer

−13836.7+2224.5(12)=−13836.7+26694=12,857.3-13836.7 + 2224.5(12) = -13836.7 + 26694 = \mathbf{12{,}857.3} THB

c. Woman, same characteristics (1 pt)

Answer

3652.6+671.2(12)=3652.6+8054.4=11,707.03652.6 + 671.2(12) = 3652.6 + 8054.4 = \mathbf{11{,}707.0} THB

d. Man minus woman at 16 years of education, same previous experience (2 pts)

Answer

β^1+β^3(16)=−17489.3+1553.3(16)=−17489.3+24852.8=7,363.5\hat\beta_1 + \hat\beta_3(16) = -17489.3 + 1553.3(16) = -17489.3 + 24852.8 = \mathbf{7{,}363.5} THB higher for the man.

Check

Man at 16 years: −13836.7+2224.5(16)=21755.3-13836.7 + 2224.5(16) = 21755.3. Woman at 16 years: 3652.6+671.2(16)=14391.83652.6 + 671.2(16) = 14391.8. Difference =7363.5= 7363.5. The previous-experience term cancels, which is why "same previous years of experience" lets you ignore it.

PP05_q2_men_women_lines

Figure (theory diagram, numbers from the solution): the two fitted lines at prev.yrs.exp = 0. The male line starts far below the female line at low education (negative intercept shift) but is steeper, so the gap opens up from about 12 years of education onward: 1150.3 at 12 years (12857.3 vs 11707.0) and 7363.5 at 16 years.

See Lec_11-Interaction Terms and Assign_05-Quadratics, Dummy Variables, Interactions & the Chow Test.


Q3 — Same marginal effect of education for everyone? (5 pts)

Question. True or false: "In Model [2], the marginal effect of education on starting wage is the same for all observations." (1 pt verdict, 4 pts explanation)

AI-written answer, not official. Check before relying on it.
Answer

FALSE.

Explanation. Because of the interaction male.educ=educ×is.malemale.educ = educ \times is.male:

∂ start.wage∂ educ=β2+β3 is.male\frac{\partial\, start.wage}{\partial\, educ} = \beta_2 + \beta_3\, is.male

For women this is β^2=671.2\hat\beta_2 = 671.2 THB per extra year; for men it is β^2+β^3=671.2+1553.3=2224.5\hat\beta_2 + \hat\beta_3 = 671.2 + 1553.3 = 2224.5 THB per extra year. It depends on gender (and the interaction is highly significant, p=5.93×10−13p = 5.93\times10^{-13}), so it differs across observations. It would be constant only if β3=0\beta_3 = 0.


Q4 — Marginal effect of experience in Model [3] (5 pts)

Question. Model [3]: wage=β0+β1 start.wage+β2 yrs.exp+β3 yrs.exp.sq+uwage = \beta_0 + \beta_1\, start.wage + \beta_2\, yrs.exp + \beta_3\, yrs.exp.sq + u. What is the estimated marginal effect of years of experience on wage? If not constant, evaluate it at the relevant sample average. (1 pt number, 4 pts working)

R output growth.3 (399 df) and means:

Term Estimate Std. Error t Pr(>abs t)
(Intercept) 3.974e+04 1.791e+04 2.219 0.02707
start.wage 1.985e+00 5.178e-02 38.336 < 2e-16
yrs.exp -6.466e+03 2.654e+03 -2.436 0.01529
yrs.exp.sq 2.599e+02 9.774e+01 2.659 0.00814

Residual SE 7877 on 399 df; R2=0.788R^2 = 0.788, adj. R2=0.7864R^2 = 0.7864; F = 494.3 on 3 and 399 df.

Sample means: wage 34683, yrs.exp 13.51, yrs.exp.sq 185.7, start.wage 17158.

AI-written answer, not official. Check before relying on it.
Answer

ME^=−6466+2(259.9)(13.51)≈556\widehat{ME} = -6466 + 2(259.9)(13.51) \approx \mathbf{556} THB per extra year of experience (at the mean yrs.exp=13.51yrs.exp = 13.51).

Working. yrs.exp appears in two terms, so

∂ wage∂ yrs.exp=β2+2β3 yrs.exp\frac{\partial\, wage}{\partial\, yrs.exp} = \beta_2 + 2\beta_3\, yrs.exp

depends on the level of experience. Evaluate at the sample mean of yrs.exp (13.51):

−6466+2(259.9)(13.51)=−6466+7022.5≈556.5-6466 + 2(259.9)(13.51) = -6466 + 7022.5 \approx 556.5
Exam trap

Use the mean of yrs.exp (13.51) inside 2β3 yrs.exp2\beta_3\, yrs.exp. Do not plug in the mean of yrs.exp.sq (185.7): the derivative of yrs.exp2yrs.exp^2 is 2 yrs.exp2\,yrs.exp, and the mean of the square is not the square of the mean (13.512=182.5≠185.713.51^2 = 182.5 \neq 185.7). Do not use the mean of wage either; it is not part of the derivative.

PP05_q4_marginal_effect_yrs_exp

Figure (theory diagram, numbers from the solution): the marginal effect is a straight line in yrs.exp. It is negative below the turning point 6466/(2×259.9)≈12.446466/(2 \times 259.9) \approx 12.44 years and positive above it; at the sample mean of 13.51 years it is about 556 THB.

See Lec_13-Exam Review (quadratic terms) and Assign_05-Quadratics, Dummy Variables, Interactions & the Chow Test.


Q5 — Testing "no wage growth" (5 pts)

Question. Define growth=wage−start.wagegrowth = wage - start.wage. How do you test that there is no wage growth in Model [3]? (A) the null in terms of Model [3]'s parameters (3 pts). (B) the restricted model under the null (2 pts).

AI-written answer, not official. Check before relying on it.

a. Null hypothesis (3 pts)

Answer

H0: β0=0, β1=1, β2=0, β3=0H_0:\ \beta_0 = 0,\ \beta_1 = 1,\ \beta_2 = 0,\ \beta_3 = 0 (four restrictions, q=4q = 4). H1H_1: at least one of these fails.

Why. No growth means wage=start.wagewage = start.wage exactly (plus noise): growth=0growth = 0 on average whatever the experience. Rewrite Model [3] in growth form:

growth=β0+(β1−1) start.wage+β2 yrs.exp+β3 yrs.exp.sq+ugrowth = \beta_0 + (\beta_1 - 1)\, start.wage + \beta_2\, yrs.exp + \beta_3\, yrs.exp.sq + u

Growth is zero for every value of the regressors only if the intercept, the coefficient on start.wagestart.wage minus one, and both experience coefficients are all zero. Note the restriction on β1\beta_1 is =1=1, not =0=0.

b. Restricted model (2 pts)

Answer

wage=start.wage+uwage = start.wage + u, equivalently growth=ugrowth = u (no parameters at all).

Imposing H0H_0 removes every coefficient (the start.wagestart.wage coefficient is fixed at 1, so it moves to the left side). Because β1=1≠0\beta_1 = 1 \neq 0 the dependent variable changes from wagewage to growthgrowth.

Intuition

Estimate the unrestricted model with growthgrowth as the outcome and test that all coefficients are zero: F=(RSSR−RSSUR)/4RSSUR/(N−4)F = \dfrac{(RSS_R - RSS_{UR})/4}{RSS_{UR}/(N-4)} with RSSR=∑growthi2RSS_R = \sum growth_i^2 and q=4q = 4.

See Lec_13-Exam Review (F-test steps: new dependent variable when the null fixes a coefficient at a non-zero value).


Q6 — Model [4]: degree dummies (5 pts)

Question. Model [4]:

wage=β0+β1 start.wage+β2 yrs.exp+β3 yrs.exp.sq+β4 hs.grad+β5 polytech+β6 college+β7 post.college+uwage = \beta_0 + \beta_1\, start.wage + \beta_2\, yrs.exp + \beta_3\, yrs.exp.sq + \beta_4\, hs.grad + \beta_5\, polytech + \beta_6\, college + \beta_7\, post.college + u

R output growth.4 (395 df):

Term Estimate Std. Error t Pr(>abs t)
(Intercept) 4.370e+04 1.718e+04 2.544 0.0113
start.wage 1.581e+00 8.777e-02 18.014 < 2e-16
yrs.exp -6.554e+03 2.533e+03 -2.588 0.0100
yrs.exp.sq 2.604e+02 9.327e+01 2.792 0.0055
hs.grad 1.130e+03 1.450e+03 0.779 0.4365
polytech 3.875e+03 1.535e+03 2.524 0.0120
college 1.025e+04 1.915e+03 5.355 1.46e-07
post.college 1.122e+04 2.422e+03 4.634 4.88e-06

Residual SE 7510 on 395 df; R2=0.8092R^2 = 0.8092, adj. R2=0.8058R^2 = 0.8058; F = 239.2 on 7 and 395 df.

AI-written answer, not official. Check before relying on it.

a. Base category (1 pt)

Answer

High-school dropout (hs.dropout, yrs.educ<12yrs.educ < 12).

Why. Five mutually exclusive degree categories, so only four dummies enter. The omitted one is the Base Category, where all four dummies equal 0. Including all five with an intercept would be the dummy-variable trap (perfect multicollinearity).

b. Interpret β^4\hat\beta_4 and β^6\hat\beta_6 (4 pts)

Answer
  • β^4=1130\hat\beta_4 = 1130 (hs.grad): holding start.wage and experience fixed, a high-school graduate earns on average 1,130 THB per month more than a high-school dropout. Not statistically significant (p=0.4365p = 0.4365, t=0.779t = 0.779): no evidence of a real wage difference between graduates and dropouts at 5% (or 10%).
  • β^6=10,250\hat\beta_6 = 10{,}250 (college): holding start.wage and experience fixed, a 4-year college graduate earns on average 10,250 THB per month more than a high-school dropout. Highly significant (p=1.46×10−7p = 1.46\times10^{-7}, below even 0.1%).
Exam trap

Each coefficient is a difference relative to the base category (dropouts), not relative to the previous level and not relative to each other. And you must interpret each on its own t-test here; the question explicitly says "not joint".

See Lec_10-Dummy Variables.


Q7 — Joint test that highest degree matters in Model [4] (9 pts)

Question. Test that highest degree earned affects wage in Model [4] at α=1%\alpha = 1\%. (A) null (1 pt); (B) unrestricted and restricted models (2); (C) degrees of freedom (2); (D) test statistic (2); (E) conclusion with the critical value (2).

AI-written answer, not official. Check before relying on it.

a. Null hypothesis (1 pt)

Answer

H0: β4=β5=β6=β7=0H_0:\ \beta_4 = \beta_5 = \beta_6 = \beta_7 = 0 (highest degree has no effect). H1H_1: at least one of them is non-zero.

b. Unrestricted and restricted models (2 pts)

Answer
  • Unrestricted: Model [4] itself.
  • Restricted (this is Model [3]): wage=β0+β1 start.wage+β2 yrs.exp+β3 yrs.exp.sq+uwage = \beta_0 + \beta_1\, start.wage + \beta_2\, yrs.exp + \beta_3\, yrs.exp.sq + u.

c. Degrees of freedom (2 pts)

Answer

Numerator q=4q = 4 (four restrictions). Denominator N−M=403−8=395N - M = 403 - 8 = 395 (Model [4] has 8 parameters). So F∼F4, 395F \sim F_{4,\,395}.

d. Test statistic (2 pts)

Both models use the same outcome wage, so the R2R^2 form of the F-statistic works. Restricted R2R^2 is Model [3]'s 0.7880.788, unrestricted R2R^2 is Model [4]'s 0.80920.8092:

F=(RUR2−RR2)/q(1−RUR2)/(N−M)=(0.8092−0.788)/4(1−0.8092)/395=0.00530.000483≈10.97F = \frac{(R^2_{UR} - R^2_R)/q}{(1 - R^2_{UR})/(N-M)} = \frac{(0.8092 - 0.788)/4}{(1 - 0.8092)/395} = \frac{0.0053}{0.000483} \approx 10.97
Answer

F≈10.97F \approx \mathbf{10.97} (about 11.0).

Cross-check with the residual sums of squares

RSSURRSS_{UR} for Model [4] =75102×395=2.228×1010= 7510^2 \times 395 = 2.228\times10^{10} (this is the "22280" in Q8, printed in millions). RSSRRSS_R for Model [3] =78772×399=2.476×1010= 7877^2 \times 399 = 2.476\times10^{10}. Then F=(24757−22278)/422278/395≈10.98F = \dfrac{(24757 - 22278)/4}{22278/395} \approx 10.98. Same answer; the R2R^2 route is exposed to rounding of the printed R2R^2 (0.788 is only 3 significant figures), so quote about 11.

e. Conclusion (2 pts)

Answer

Reject H0H_0 at 1%. Critical value F0.01(4,395)≈3.37F_{0.01}(4, 395) \approx 3.37 (table: about 3.3 to 3.5 depending on the nearest row). Since 10.97>3.3710.97 > 3.37, highest degree earned jointly affects wage.

Intuition

Joint significance here is driven mainly by college and post.college (both highly significant individually), even though hs.grad alone is not (Q6b). That is why a dummy block is tested with an F-test, not several t-tests.

See Lec_13-Exam Review (testing a categorical variable jointly) and Lec_08-Goodness of Fit & Multicollinearity (F in R2R^2 form).


Q8 — Chow test: separate models for men and women? (7 pts)

Question. Using the R output below, test whether Model [4] should be estimated separately for men and women. (A) F-statistic (5 pts); (B) conclusion with critical value at 5% (2 pts).

R output (sums of squared residuals, as printed):

Regression Sum of squared residuals
Pooled Model [4] (growth.4) 22280
Men only (is.male == 1) 17917
Women only (is.male < 1) 3126

(The printed values are scaled; the pooled value matches 75102×395≈22280×1067510^2 \times 395 \approx 22280\times 10^6. The units cancel in FF.)

AI-written answer, not official. Check before relying on it.

a. F-statistic (5 pts)

Chow test, version 2 (separate regressions). Under H0H_0 all coefficients are the same for men and women, so the pooled model is the restricted one:

  • RSSR=22280RSS_R = 22280
  • RSSUR=RSSm+RSSf=17917+3126=21043RSS_{UR} = RSS_m + RSS_f = 17917 + 3126 = 21043
  • q=8q = 8: the number of parameters in Model [4] (intercept plus 7 slopes), each allowed to differ by gender
  • N−M=403−2(8)=387N - M = 403 - 2(8) = 387 (two sets of 8 parameters)
F=(RSSR−RSSUR)/qRSSUR/(N−M)=(22280−21043)/821043/387=154.654.4≈2.84F = \frac{(RSS_R - RSS_{UR})/q}{RSS_{UR}/(N-M)} = \frac{(22280 - 21043)/8}{21043/387} = \frac{154.6}{54.4} \approx 2.84
Answer

F≈2.84F \approx \mathbf{2.84} with (8, 387)(8,\ 387) degrees of freedom.

b. Conclusion (2 pts)

Answer

Reject H0H_0 at 5%. F0.05(8,387)≈1.96F_{0.05}(8, 387) \approx 1.96 (table rows around 120 to infinity: about 1.94 to 2.02). Since 2.84>1.962.84 > 1.96, the model differs by gender, so estimate it separately for men and women.

Exam trap

The parameter count MM is the full parameter count of Model [4] (8, including the intercept), so q=8q = 8 and the denominator df is N−2M=387N - 2M = 387. Using 7 (slopes only) or forgetting the doubling are the usual slips.

See Assign_05-Quadratics, Dummy Variables, Interactions & the Chow Test and Lec_11-Interaction Terms.


Q9 — Converting THB to USD (5 pts)

Question. Model [5]: wage=β0+β1 start.wage+β2 yrs.exp+β3 yrs.exp.sq+β4 educ+uwage = \beta_0 + \beta_1\, start.wage + \beta_2\, yrs.exp + \beta_3\, yrs.exp.sq + \beta_4\, educ + u. If both wage and start.wage are converted from THB to USD (1 THB = 0.03 USD), what happens to the estimated marginal effect of education on wage? (1 pt answer, 4 pts working)

R output growth.5 (398 df):

Term Estimate Std. Error t Pr(>abs t)
(Intercept) 2.978e+04 1.753e+04 1.699 0.09002
start.wage 1.742e+00 7.035e-02 24.763 < 2e-16
yrs.exp -6.277e+03 2.580e+03 -2.433 0.01541
yrs.exp.sq 2.519e+02 9.501e+01 2.651 0.00834
educ 9.518e+02 1.927e+02 4.939 1.16e-06

Residual SE 7655 on 398 df; R2=0.8002R^2 = 0.8002, adj. R2=0.7982R^2 = 0.7982; F = 398.5 on 4 and 398 df.

AI-written answer, not official. Check before relying on it.
Answer

It is multiplied by 0.03: β^4=951.8\hat\beta_4 = 951.8 THB per year of education becomes 951.8×0.03≈28.55951.8 \times 0.03 \approx \mathbf{28.55} USD per year of education.

Working. Let c=0.03c = 0.03 and the new variables be wage∗=c⋅wagewage^* = c \cdot wage, start.wage∗=c⋅start.wagestart.wage^* = c \cdot start.wage. Multiply the whole model by cc:

c⋅wage=cβ0+β1(c⋅start.wage)+cβ2 yrs.exp+cβ3 yrs.exp.sq+cβ4 educ+c uc \cdot wage = c\beta_0 + \beta_1 (c \cdot start.wage) + c\beta_2\, yrs.exp + c\beta_3\, yrs.exp.sq + c\beta_4\, educ + c\,u

So wage∗=cβ0+β1 start.wage∗+cβ2 yrs.exp+cβ3 yrs.exp.sq+cβ4 educ+u∗wage^* = c\beta_0 + \beta_1\, start.wage^* + c\beta_2\, yrs.exp + c\beta_3\, yrs.exp.sq + c\beta_4\, educ + u^*. The coefficients on regressors that are not rescaled (yrs.exp, yrs.exp.sq, educ) and the intercept are multiplied by cc; the coefficient on start.wage stays at 1.742 because both yy and that xx are scaled by the same cc (the cc in the numerator and the 1/c1/c from the rescaled xx cancel).

Standard errors scale with their coefficients, so t-statistics and p-values (and R2R^2) do not change.

See Assign_03-Change of Variables, Regression Through the Origin & the Housing Model and Lec_13-Exam Review (unit-change table).


Part II: sample.data (48 pts)

Data (N = 125)

YY and log⁡(Y)\log(Y) (the R variable is ln.Y), X∈[2.5,5]X \in [2.5, 5], W∈[1.5,6]W \in [1.5, 6], ZZ a 0/1 dummy. Note: the printed ln.Y values run from about −80.6-80.6 to 28.628.6, so the "log" is large in magnitude; the log-level readings in Q15 to Q17 are therefore only approximate (see Q15).

Q10 — Test β2\beta_2 against 2 in Model [A] (5 pts)

Question. Model [A]: log⁡(Y)=β0+β1X+β2W+β3Z+u\log(Y) = \beta_0 + \beta_1 X + \beta_2 W + \beta_3 Z + u. Test the null that β2\beta_2 is less than 2 at 10%. (A) test statistic (2 pts); (B) df and critical value (2); (C) conclusion (1).

R output model.A (121 df; the df cell is redacted in the paper but the F line says 3 and 121):

Term Estimate Std. Error t Pr(>abs t)
(Intercept) -6.098 10.247 -0.595 0.55288
X -6.733 5.610 -1.200 0.23240
W 4.470 4.819 0.927 0.35552
Z -12.285 4.140 -2.968 0.00362

Residual SE 20.74; R2=0.07855R^2 = 0.07855, adj. R2=0.05571R^2 = 0.05571; F = 3.438 on 3 and 121 df, p = 0.01909.

AI-written answer, not official. Check before relying on it.
Which hypothesis?

The wording "the null that β2\beta_2 is less than 2" is read as a one-sided test with the equality in the null: H0:β2≤2H_0:\beta_2 \le 2 versus H1:β2>2H_1:\beta_2 > 2. (If you instead set it up as H0:β2≥2H_0:\beta_2 \ge 2 vs H1:β2<2H_1:\beta_2 < 2, the statistic is the same and the conclusion is also "fail to reject", see below.)

a. Test statistic (2 pts)

t=β^2−2s.e.⁡(β^2)=4.470−24.819=2.4704.819≈0.513t = \frac{\hat\beta_2 - 2}{\operatorname{s.e.}(\hat\beta_2)} = \frac{4.470 - 2}{4.819} = \frac{2.470}{4.819} \approx 0.513
Answer

t≈0.513t \approx \mathbf{0.513}.

b. Degrees of freedom and critical value (2 pts)

Answer

df=N−M=125−4=121df = N - M = 125 - 4 = \mathbf{121}. One-sided 10% critical value t0.10(121)≈1.289t_{0.10}(121) \approx \mathbf{1.289} (table row 120: 1.289).

c. Conclusion (1 pt)

Answer

Fail to reject H0H_0 at 10%: 0.513<1.2890.513 < 1.289. No evidence that β2\beta_2 exceeds 2. (With the opposite orientation, H1:β2<2H_1:\beta_2 < 2, the rejection region is t<−1.289t < -1.289; 0.5130.513 is not in it either: also fail to reject.)

Exam trap

The null is not "β2=0\beta_2 = 0", which is the t-value R prints (0.927). You must subtract the hypothesised value 2 first. Use a one-tailed critical value because the hypothesis is one-sided.

See Lec_13-Exam Review (t-test steps, one- vs two-sided).


Q11 — Which comparison shows Model [B] fits better? (5 pts)

Question. Model [B]: log⁡(Y)=β1X+β2W+β3Z+u\log(Y) = \beta_1 X + \beta_2 W + \beta_3 Z + u (no intercept). Comparing the output for [B] with [A], you conclude [B] fits better. Which comparison supports that?

R output model.B (122 df):

Term Estimate Std. Error t Pr(>abs t)
X -8.417 4.832 -1.742 0.084048
W 4.749 4.783 0.993 0.322720
Z -13.117 3.886 -3.376 0.000988

Residual SE 20.68 on 122 df; R2=0.5896R^2 = 0.5896, adj. R2=0.5795R^2 = 0.5795; F = 58.42 on 3 and 122 df.

AI-written answer, not official. Check before relying on it.
Answer

(v) cannot say / none of the above.

Explanation. Model [B] has no intercept. R computes R2R^2 for a no-intercept model against the uncentred total sum of squares (∑yi2\sum y_i^2) instead of ∑(yi−yˉ)2\sum (y_i - \bar y)^2. Since log⁡(Y)\log(Y) here is far from zero (mean about −23.8-23.8), the uncentred benchmark is huge, which mechanically inflates R2R^2 from 0.079 to 0.590. The two numbers are on different scales, so neither R2R^2 nor adjusted R2R^2 can compare a model with an intercept to one without (Lec_08-Goodness of Fit & Multicollinearity: "models with no intercept").

The only comparable evidence points the other way

Compare residual sums of squares, which are on the same scale: Model [A] 20.742×121≈52,04820.74^2 \times 121 \approx 52{,}048 versus Model [B] 20.682×122≈52,17520.68^2 \times 122 \approx 52{,}175 (the anova in Q13 prints 52181). Model [A] (more flexible, nested) fits at least as well in-sample, and [B]'s tiny residual-SE advantage comes only from using one more degree of freedom.

Each option.

Option Verdict Why
(i) 0.5896 > 0.07855 Wrong Compares an uncentred R2R^2 (no intercept) with a centred R2R^2. Not comparable.
(ii) 0.5795 > 0.05571 Wrong Same problem for adjusted R2R^2: it inherits the uncentred R2R^2, so the penalty for fewer parameters cannot repair it.
(iii) 0.23240 > 0.084048 Wrong These are the p-values of XX in [A] and [B]. A smaller p-value for one coefficient says nothing about overall fit. (And "0.2324 > 0.084" would mean [A] is less significant, not that [B] fits better.)
(iv) 4.749 > 4.470 Wrong Compares estimates of β2\beta_2. Coefficient size is not a goodness-of-fit measure.
(v) cannot say / none Correct No listed comparison is valid.
Intuition

Dropping the intercept forces the line through the origin. That changes what R2R^2 measures, so its jump is an artefact, the same trap as in Assign_03-Change of Variables, Regression Through the Origin & the Housing Model (Q5c).


Q12 — R code for Model [B] (3 pts)

AI-written answer, not official. Check before relying on it.
Answer
r
model.B = lm(ln.Y ~ 0 + X + W + Z, data = sample.data)

(Equivalent: lm(ln.Y ~ X + W + Z - 1, data = sample.data).)

Why. Model [B] drops the intercept, and in R's formula syntax 0 + (or - 1) removes it. The outcome variable is called ln.Y in sample.data (as in the model.A call). The output has no (Intercept) row, which confirms it. See Regression Through the Origin.


Q13 — Evidence of partial multicollinearity between XX and WW (8 pts)

Question. After seeing Model [B] you suspect XX and WW are irrelevant and estimate Model [C]: log⁡(Y)=β1Z+u\log(Y) = \beta_1 Z + u. Three pieces of evidence together indicate partial multicollinearity between XX and WW. (3 pts list; 5 pts how they combine.)

R output (Model [C], F-test and correlations):

Item Value
Model [C] Z coefficient -27.202 (SE 2.329, t = -11.68)
Model [C] residual SE 22.1 on 124 df; R2=0.5238R^2 = 0.5238, adj. R2=0.52R^2 = 0.52
anova(model.C, model.B) Res.Df 124 vs 122; RSS 60540 vs 52181; Df 2; Sum of Sq 8359.1; F = 9.7719; p = 0.0001158
cor⁡(X,W)\operatorname{cor}(X, W) 0.89534
cor⁡(X,Z)\operatorname{cor}(X, Z) / cor⁡(W,Z)\operatorname{cor}(W, Z) -0.0555 / -0.0654
AI-written answer, not official. Check before relying on it.
Answer: the three pieces
  1. Individually insignificant: in Model [B], neither XX (p=0.084p = 0.084) nor WW (p=0.323p = 0.323) is significant at 5%.
  2. Jointly significant: the F-test of Model [C] against Model [B] (H0:βX=βW=0H_0:\beta_X = \beta_W = 0) gives F=9.77F = 9.77, p=0.0001p = 0.0001. Reject: XX and WW together do matter.
  3. High correlation: cor⁡(X,W)=0.895\operatorname{cor}(X, W) = 0.895, far above the rule-of-thumb 0.7. (XX and WW are barely correlated with ZZ, so the problem is only between those two.)

How they combine (5 pts). The recipe in Lec_13-Exam Review is: neither variable significant on its own, yet jointly significant, and the two are strongly correlated.

  • The F-test (evidence 2) says the pair carries real explanatory power, so "X and W are irrelevant" is false. Irrelevance would have shown up as both t-tests and the joint F-test being insignificant.
  • Yet the t-tests (evidence 1) say neither one matters when the other is in the model. These two statements are only compatible if the variables share the same information: each is a near-substitute for the other, so the model cannot attribute the effect to either separately.
  • Evidence 3 supplies the mechanism: with RX∣W2≈0.8952≈0.80R^2_{X\mid W} \approx 0.895^2 \approx 0.80, the variance inflation factor is 1/(1−0.80)≈5.01/(1 - 0.80) \approx 5.0. From Var⁡(β^j)=σu2/[(1−Rj2)∑(xji−xˉj)2]\operatorname{Var}(\hat\beta_j) = \sigma_u^2 / [(1 - R_j^2)\sum(x_{ji} - \bar x_j)^2] the standard errors of β^X\hat\beta_X and β^W\hat\beta_W are inflated (by about 5.0≈2.2\sqrt{5.0} \approx 2.2), pushing their t-statistics toward zero. Estimates stay unbiased; only precision is lost.
  • Hence do not drop both (that would throw away real information, as the F-test shows); keep at least one of them.
Intuition

Model [D] (used in Q14 to Q18) keeps only XX and ZZ: XX becomes highly significant (t=−4.31t = -4.31) and adj. R2R^2 stays at 0.5795, exactly as in [B]. That is a fourth, supporting symptom of the same collinearity: once WW stops competing with it, XX's precision returns.

See Lec_08-Goodness of Fit & Multicollinearity and Multicollinearity.


Model [D] for Q14 to Q17

Model [D]: log⁡(Y)=β1X+β2Z+u\log(Y) = \beta_1 X + \beta_2 Z + u (no intercept again).

R output model.D (123 df):

Term Estimate Std. Error t Pr(>abs t)
X -3.6952 0.8577 -4.308 3.33e-05
Z -13.3992 3.8752 -3.458 0.000749

Residual SE 20.68 on 123 df; R2=0.5863R^2 = 0.5863, adj. R2=0.5795R^2 = 0.5795; F = 87.14 on 2 and 123 df.

Sample means: Xˉ=3.761\bar X = 3.761, Wˉ=3.689\bar W = 3.689, Zˉ=0.72\bar Z = 0.72, ln⁡Y‾=−23.78\overline{\ln Y} = -23.78.

Q14 — 95% confidence interval for β1\beta_1 (5 pts)

AI-written answer, not official. Check before relying on it.
Answer

[-5.39, -2.00] (lower ≈−5.39\approx -5.39, upper ≈−2.00\approx -2.00).

Working. β^1±t0.025(123)×s.e.⁡\hat\beta_1 \pm t_{0.025}(123) \times \operatorname{s.e.} with df=125−2=123df = 125 - 2 = 123 and t0.025(123)≈1.98t_{0.025}(123) \approx 1.98 (table row 120: 1.980):

−3.6952±1.98×0.8577=−3.6952±1.698-3.6952 \pm 1.98 \times 0.8577 = -3.6952 \pm 1.698

Lower =−5.393= -5.393, upper =−1.997= -1.997. Zero lies outside the interval, consistent with XX being significant at 5% (Lec_13-Exam Review: CI and two-sided test duality).


Q15 — Interpreting β^1=−3.7\hat\beta_1 = -3.7 (5 pts)

AI-written answer, not official. Check before relying on it.
Answer

(ii) A 1 unit increase in XX decreases YY by 370%.

Explanation. This is a log-level model: the outcome is log⁡(Y)\log(Y), the regressor XX is in levels. In a log-level model a one-unit increase in XX changes YY by approximately 100×β^1100 \times \hat\beta_1 percent (Lec_13-Exam Review, log-level row): 100×(−3.6952)≈−370%100 \times (-3.6952) \approx -370\%.

Option Verdict Why
(i) decreases YY by 3.70 units Wrong That is the level-level reading. β^1\hat\beta_1 is the change in log⁡(Y)\log(Y), not in YY.
(ii) decreases YY by 370% Correct Log-level reading: ΔY/Y≈100β^1 %\Delta Y/Y \approx 100\hat\beta_1\,\%.
(iii) a 1% increase in XX decreases YY by 370% Wrong That is a log-log reading and would need log⁡(X)\log(X) as the regressor. Here XX is in levels.
(iv) decreases log⁡(Y)\log(Y) by 370% Wrong log⁡(Y)\log(Y) falls by 3.70 log points, not by 370%; the percentage belongs to YY.
(v) cannot say Wrong A standard reading exists (ii).
Caveat on the approximation

100β^100\hat\beta is a good approximation only for small coefficients. For β^=−3.70\hat\beta = -3.70 it breaks down: the exact change is 100(e−3.6952−1)≈−97.5%100(e^{-3.6952} - 1) \approx -97.5\%, and a "370% decrease" is impossible for a positive YY. The alternative (v) is defensible only if the marker insisted on exact percentages. Treat (ii) as the intended answer under the course's approximate rule, and mention the exact figure in your justification if space allows.


Q16 — Interpreting β^2=−13.4\hat\beta_2 = -13.4 (5 pts)

AI-written answer, not official. Check before relying on it.
Answer

(ii) YY is 1,340% lower for observations in category Z.

Explanation. ZZ is a dummy and the outcome is log⁡(Y)\log(Y), so 100β^2=100×(−13.3992)≈−1340%100\hat\beta_2 = 100 \times (-13.3992) \approx -1340\% is the approximate percentage difference in YY between category Z and the rest, holding XX fixed (log-level with a dummy).

Option Verdict Why
(i) YY is 13.4 units lower Wrong Level-level reading; the effect is on log⁡(Y)\log(Y).
(ii) YY is 1,340% lower Correct 100β^2100\hat\beta_2 percent, log-level with a dummy.
(iii) a 1% increase in ZZ decreases YY by 13.4% Wrong ZZ is a 0/1 indicator: it cannot rise by 1%. Its coefficient compares two groups.
(iv) log⁡(Y)\log(Y) is 13.4% lower Wrong log⁡(Y)\log(Y) is 13.4 log points lower; the percentage applies to YY and is 100×13.4=1340%100 \times 13.4 = 1340\%, not 13.4%.
(v) cannot say Wrong A standard reading exists.
Same caveat as Q15

The exact difference is 100(e−13.4−1)≈−100%100(e^{-13.4} - 1) \approx -100\%. The approximation is only valid for small β^\hat\beta, so (ii) is the answer under the course's rule, not a literal 1,340% fall.


Q17 — Elasticity of YY with respect to XX (5 pts)

AI-written answer, not official. Check before relying on it.
Answer

ε=β^1Xˉ=−3.6952×3.761≈−13.9\varepsilon = \hat\beta_1 \bar X = -3.6952 \times 3.761 \approx \mathbf{-13.9} (a 1% increase in XX is associated with about a 13.9% fall in YY).

Working. In a log-level model log⁡Y=β1X+…\log Y = \beta_1 X + \dots the marginal effect is β1Y\beta_1 Y, the rate of change is β1\beta_1, and the elasticity is β1X\beta_1 X (the log-level row of the moment table in Lec_13-Exam Review). It depends on XX, so evaluate at the sample mean Xˉ=3.761\bar X = 3.761:

ε=−3.6952×3.761=−13.90\varepsilon = -3.6952 \times 3.761 = -13.90
Exam trap

The elasticity is not β^1=−3.7\hat\beta_1 = -3.7 (that is the log-log answer, wrong here because XX is not logged). Use the mean of XX (3.761), not of WW or ZZ.


Q18 — Heteroskedasticity in Model [D] (8 pts)

Question. You suspect heteroskedasticity in Model [D]. All other classical assumptions hold. (A) What does this mean for the OLS estimates? (1 pt MC + 4 pts explanation). (B) You run coeftest(model.D, vcov = hccm, type = "hc0"). (3 pts MC)

AI-written answer, not official. Check before relying on it.

a. Effect on OLS (5 pts)

Answer

(iii) The OLS estimator of β2\beta_2 is unbiased and not efficient.

Explanation. Heteroskedasticity violates A4 (constant error variance) but not A1 or A2 (mean-zero, exogenous errors). Unbiasedness needs only A1 and A2, so β^\hat\beta stays unbiased. Gauss–Markov efficiency needs A1 to A4, so OLS is no longer BLUE (not efficient): a weighted estimator has smaller variance. In addition the usual standard errors are wrong, so t- and F-tests are invalid (Lec_12-Heteroskedasticity).

Option Verdict Why
(i) β2\beta_2 biased and not efficient Wrong Bias needs A2 to fail (for example an omitted variable), not A4.
(ii) β1\beta_1 biased and not efficient Wrong Same: not biased.
(iii) β2\beta_2 unbiased and not efficient Correct A4 fails, so efficiency is lost, but unbiasedness survives.
(iv) β1\beta_1 unbiased and efficient Wrong Unbiased, but efficiency is lost (the efficiency claim fails).
(v) cannot say Wrong The consequence is well defined.

b. After White standard errors (3 pts)

Answer

(iii) The resulting OLS estimator of β2\beta_2 is unbiased but not efficient.

Explanation. coeftest(..., vcov = hccm, type = "hc0") only recomputes the standard errors (White / heteroskedasticity-consistent). The coefficient estimates are the identical OLS ones, so they remain unbiased and still not efficient; White fixes inference (the SEs, t and p-values), not efficiency. Only weighted least squares, which needs the form of the variance, can restore efficiency (Lec_12-Heteroskedasticity).

Option Verdict Why
(i) β2\beta_2 biased and not efficient Wrong The point estimates are unchanged and still unbiased.
(ii) β1\beta_1 biased and not efficient Wrong Same.
(iii) β2\beta_2 unbiased but not efficient Correct OLS point estimates are untouched; only the SEs change.
(iv) β1\beta_1 unbiased and efficient Wrong White does not make OLS efficient.
(v) cannot say Wrong The answer is determined.
Option (iv) is a trap in both parts

Option (iv) in (A) and (B) says β1\beta_1 is "unbiased and efficient". Half of it is true (unbiased); the exam tests whether you notice that efficiency is the casualty in both cases.


Recap table

Q Pts Tool Answer (AI-solved)
1 5 t-test on β5\beta_5 p = 0.2217 > 0.01, fail to reject
2 6 Interaction model Women: 3652.6 + 671.2 educ + 105.3 prev; men: -13836.7 + 2224.5 educ + 105.3 prev; 12857.3; 11707.0; 7363.5
3 5 Marginal effect with interaction False: β2+β3is.male\beta_2 + \beta_3 is.male (671.2 women, 2224.5 men)
4 5 Quadratic marginal effect −6466+2(259.9)(13.51)≈556-6466 + 2(259.9)(13.51) \approx 556 THB per year
5 5 F-test restrictions H0H_0: β0=0\beta_0 = 0, β1=1\beta_1 = 1, β2=β3=0\beta_2 = \beta_3 = 0; restricted: wage=start.wage+uwage = start.wage + u
6 5 Dummies Base = hs.dropout; β^4=1130\hat\beta_4 = 1130 (not sig.), β^6=10250\hat\beta_6 = 10250 (sig.)
7 9 Joint F-test F=10.97F = 10.97 on (4, 395) > 3.37: reject
8 7 Chow test F=2.84F = 2.84 on (8, 387) > 1.96: reject
9 5 Change of units β^4\hat\beta_4 becomes ×0.03=28.55\times 0.03 = 28.55
10 5 One-sided t-test t=0.513t = 0.513, df 121, crit 1.289: fail to reject
11 5 R2R^2 comparability (v): no-intercept R2R^2 not comparable
12 3 R code lm(ln.Y ~ 0 + X + W + Z, data = sample.data)
13 8 Multicollinearity Insignificant t's + joint F = 9.77 + cor = 0.895
14 5 Confidence interval [-5.39, -2.00]
15 5 Log-level interpretation (ii) about -370% (approximation)
16 5 Log-level dummy (ii) about -1340% (approximation)
17 5 Elasticity β^1Xˉ≈−13.9\hat\beta_1 \bar X \approx -13.9
18 8 Heteroskedasticity (iii) in A and B: unbiased, not efficient