Exam 2
Term: 2026F (Fall 2026)
Instructor: Chi-Kuang Yeh
Date: November 11, 2026
Status: Completed
Time: 75 minutes
Total: [100] points
Coverage: Everything through Section 6.3: Outliers, emphasizing material covered after Exam 1.
Emphasis: Conceptual understanding, model interpretation, statistical reasoning, and brief calculations.
Notation: \(p\) counts independent design columns, including any intercept. Vectors and matrices are boldfaced. \(\mathrm{SSE}=\sum_i e_i^2\) denotes the error (residual) sum of squares.
| Question | Format | Points |
|---|---|---|
| 1 | Multiple Choice (10 parts, [5] points each) | [50] |
| 2 | True or False (5 parts, [4] points each) | [20] |
| 3 | Short Answer (3 parts, [10] points each) | [30] |
| Total | [100] |
- Work independently. Only a non-programmable calculator is permitted as an electronic device. No code execution or statistical-table lookup is required.
- Record Questions 1-2 on the separate answer sheet: one letter A-D or T/F per part. Only the grid is graded; no explanations are required.
- For Question 3, show brief calculations and give concise explanations; equations or 1-2 clear sentences are sufficient. Formulas and any needed critical value are supplied.
- Submit only the answer sheet. You may keep the question paper.
Answer sheet: Write all final answers on the separate answer sheet.
Expand the Answer below each question to view the solution and explanation. A compact answer key and grading rubric appear at the end.
Question 1: Multiple Choice [50]
Each part is worth [5] points. Select the single best answer under the stated conditions. Record it on the separate answer sheet. No explanation is required.
1a)
Two OLS mean models are \(M_A:\mu_A(x)=\alpha_0+\alpha_1x\) and \(M_B:\mu_B(z)=\gamma_0+\gamma_1z\). On the same observations, \(\mathbf1\), \(\mathbf x\), and \(\mathbf z\) are linearly independent, and \(\mathrm{SSE}_A=84\), \(\mathrm{SSE}_B=72\). Which statement is correct?
A. The usual partial \(F\) test comparing A directly with B has one added degree of freedom.
B. The lower \(\mathrm{SSE}_B\) establishes that \(z\) contributes after accounting for \(x\).
C. Neither mean space contains the other, so their SSE difference does not give a valid nested-model \(F\) test.
D. The SSE values must be equal because both models have two coefficients.
Answer: C. The independence of \(\mathbf1\), \(\mathbf x\), and \(\mathbf z\) means neither two-dimensional mean space contains the other. The partial \(F\) formula requires nested models, not merely two SSE values. To test \(z\) after \(x\), compare \(\alpha_0+\alpha_1x\) with \(\beta_0+\beta_1x+\beta_2z\).
1b)
Consider \(\mu(x,z)=\beta_0+\beta_xx+\beta_gz+\beta_{xg}xz\), with \(z=0\) or \(1\). Rewrite the same model using \(x_c=x-5\), retaining all terms. Which statement is correct?
A. The coefficient of \(z\) becomes \(\beta_g+5\beta_{xg}\); fitted values and SSE are unchanged.
B. The coefficient of \(z\) must remain \(\beta_g\), even when \(\beta_{xg}\ne0\).
C. Centering changes the fitted mean for each observation.
D. The coefficient of \(x_cz\) becomes \(5\beta_{xg}\).
Answer: A. Substitute \(x=x_c+5\). The coefficient of \(z\) is \(\beta_g+5\beta_{xg}\), the group difference \(\Delta(5)\). The intercept becomes \(\beta_0+5\beta_x\), while the slope difference remains \(\beta_{xg}\). Centering moves the reference point; fitted values, residuals, and SSE stay the same.
1c)
For the same observations, consider \(M_0:\mu=\alpha_0\), \(M_1:\mu=\alpha_0+\alpha_1x_1\), and \(M_2:\mu=\beta_0+\beta_1x_1+\beta_2x_2\). Which comparison tests whether \(x_2\) contributes to the mean after accounting for \(x_1\)?
A. \(M_0\) versus \(M_2\).
B. \(M_1\) versus \(M_2\).
C. \(M_0\) versus \(M_1\).
D. Compare only the two fitted coefficients of \(x_1\).
Answer: B. The reduced model retains \(x_1\) and removes \(x_2\). The partial \(F\) test addresses \(H_0:\beta_2=0\) in the full model: does \(x_2\) contribute after accounting for \(x_1\)? It does not test equality of the two fitted \(x_1\) coefficients.
1d)
Two full-rank nested OLS models use the same observations and include an intercept, with \(\mathrm{SST}>0\) and \(\mathrm{SSE}_F>0\). Adding one coefficient gives partial \(F=0.8\). Which change from the reduced to the full model follows? (The formulas are given in 3a.)
A. Both \(R^2\) and \(R^2_{\mathrm{adj}}\) increase.
B. \(R^2\) decreases, while \(R^2_{\mathrm{adj}}\) increases.
C. Both measures remain unchanged because \(F<1\).
D. \(R^2\) increases, while \(R^2_{\mathrm{adj}}\) decreases.
Answer: D. Here \(F>0\), so \(\mathrm{SSE}_R>\mathrm{SSE}_F\) and \(R^2\) increases. For one added coefficient, \(F<1\) means the SSE reduction is less than \(\mathrm{MSE}_F\), so \(\mathrm{MSE}_R<\mathrm{MSE}_F\). With the same SST, adjusted \(R^2\) therefore decreases.
1e)
Two predictors \(x\) and \(z\) are correlated. In sequential ANOVA on the same observations, the sum of squares for \(z\) is 45 when \(z\) enters first, but 12 when it enters after \(x\). Why can both values be correct?
A. They use the same reduced model but different estimates of the error variance.
B. They use different reduced models: intercept only versus intercept plus \(x\).
C. They estimate the noise variance before and after adjustment for \(x\).
D. They use different final fits because reversing the term order changes fitted values.
Answer: B. A sequential sum of squares is the reduction in SSE when a term is added after the preceding terms. Adding \(z\) first compares intercept only with intercept plus \(z\); adding it after \(x\) compares intercept plus \(x\) with intercept plus \(x\) and \(z\). Correlated predictors can give different reductions, while the final fitted values are unchanged.
1f)
A group-only model uses A as reference: \(\mathbb E(Y_i\mid G_i)=\beta_0+\beta_BD_{Bi}+\beta_CD_{Ci}\), where \(D_{Bi}\) and \(D_{Ci}\) indicate membership in B and C. All three groups are observed. With B as the new reference, which intercept and C coefficient preserve the same group means?
A. Intercept \(\beta_0\); C coefficient \(\beta_C\).
B. Intercept \(\beta_B\); C coefficient \(\beta_C\).
C. Intercept \(\beta_0+\beta_B\); C coefficient \(\beta_C\).
D. Intercept \(\beta_0+\beta_B\); C coefficient \(\beta_C-\beta_B\).
Answer: D. With A as reference, \(\mu_A=\beta_0\), \(\mu_B=\beta_0+\beta_B\), and \(\mu_C=\beta_0+\beta_C\). With B as reference, the intercept is \(\mu_B\) and the C coefficient is \(\mu_C-\mu_B=\beta_C-\beta_B\). The coefficient meanings change, while fitted group means, residuals, and SSE stay the same.
1g)
Let \(z=0\) for group A, \(z=1\) for group B, and \(x_c=x-2\). A fitted interaction model is \(\widehat\mu(x,z)=12+3x_c+5z-x_cz\). Which pair gives the fitted slope for B and the B-minus-A fitted mean difference at \(x=4\), respectively?
A. Slope \(2\); mean difference \(3\).
B. Slope \(-1\); mean difference \(3\).
C. Slope \(2\); mean difference \(5\).
D. Slope \(8\); mean difference \(3\).
Answer: A. Setting \(z=0\) gives slope 3; setting \(z=1\) gives slope \(3-1=2\). The B-minus-A fitted mean difference is \(\widehat\Delta(x)=5-(x-2)\), giving 3 at \(x=4\). The negative interaction means this gap decreases with \(x\), while both fitted slopes remain positive.
1h)
For full-rank OLS, \(\hat{\mathbf Y}=\mathbf H\mathbf Y\) and \(\mathbf e=(\mathbf I-\mathbf H)\mathbf Y\). Keep \(\mathbf X\) fixed and increase only \(Y_i\) by 10. If \(h_{ii}=0.8\), which statement must hold after refitting?
A. The raw residual \(e_i\) increases by 10, and \(h_{ii}\) is unchanged.
B. The raw residual \(e_i\) increases by 2, and \(h_{ii}\) increases.
C. The raw residual \(e_i\) increases by 2, and \(h_{ii}\) is unchanged.
D. The fitted value \(\hat Y_i\) is unchanged, and \(e_i\) increases by 8.
Answer: C. The hat matrix depends only on \(\mathbf X\), so \(h_{ii}\) is unchanged. Row \(i\) of \(\hat{\mathbf Y}=\mathbf H\mathbf Y\) gives a fitted-value increase of \(10h_{ii}=8\). Hence \(e_i=Y_i-\hat Y_i\) increases by \(10-8=2\), rather than by the full response change.
1i)
With an intercept and one nonconstant predictor, \(h_{ii}=1/n+(x_i-\bar x)^2/\sum_j(x_j-\bar x)^2\). Replace every \(x_i\) by \(u_i=10+2x_i\) and refit with an intercept to the same observations. What happens to leverage?
A. Every leverage doubles because the predictor is multiplied by 2.
B. Every leverage is divided by 4.
C. Adding 10 makes every observation have higher leverage.
D. Every leverage is unchanged.
Answer: D. The new mean is \(\bar u=10+2\bar x\), so \(u_i-\bar u=2(x_i-\bar x)\). Both the squared numerator and the sum in the denominator multiply by 4; the ratio and \(1/n\) are unchanged. A change of predictor origin or units does not change leverage.
1j)
An observation has the largest response \(Y_i\) in the sample, but \(Y_i=\hat Y_i\). Assume \(h_{ii}<1\), \(s>0\), and \(s_{(-i)}>0\). Which conclusion follows?
A. It must be a response outlier because its response is largest.
B. \(e_i=r_i=t_i=0\); its response size alone does not make it a response outlier.
C. Its leverage must be zero because its residual is zero.
D. Its standardized residual must be large because \(Y_i\) is large.
Answer: B. A response outlier is unusual given its predictor values and the specified model. Here the raw residual \(e_i=Y_i-\hat Y_i=0\), so the standardized residual \(r_i\) and externally studentized residual \(t_i\) are also zero. An extreme response alone does not make an observation an outlier; zero residual does not imply zero leverage.
Question 2: True or False [20]
Each part is worth [4] points. Record T or F on the separate answer sheet. No explanation is required.
2a)
For a nested normal-model comparison on the same observations that adds one coefficient, the partial \(F\) statistic equals \(T^2\), where \(T\) is the corresponding two-sided \(t\)-test statistic from the full model, regardless of its sign.
Answer: True. For one coefficient restriction, \(F=T^2\), as in Chapter 3. Squaring removes the sign. Both tests address the same null hypothesis; the \(F\) reference degrees of freedom are \((1,n-p_F)\).
2b)
With all four groups observed, an intercept plus all four group indicators gives five identifiable coefficients in a group-only regression model.
Answer: False. The four indicators sum to the intercept column, so the design has rank 4 rather than 5. Use an intercept plus three indicators or all four indicators without a separate intercept.
2c)
For the additive linear mean model \(\mu(x_1,x_2)=\beta_0+\beta_1x_1+\beta_2x_2\), the difference \(\mu(x_1+1,x_2)-\mu(x_1,x_2)\) equals \(\beta_1+\beta_2x_2\).
Answer: False. Holding \(x_2\) fixed, the intercept and unchanged contribution \(\beta_2x_2\) cancel. The mean-response difference is \(\beta_1\). Additivity means this difference does not depend on \(x_2\); it does not require independent predictors.
2d)
For a full-rank design with \(p\) columns, \(\sum_i h_{ii}=p\) implies that every observation has leverage \(p/n\).
Answer: False. Because \(\sum_i h_{ii}=p\), average leverage is \(p/n\). Individual leverages can differ because they describe the positions of the predictor rows. The identity does not make each leverage equal to the average.
2e)
Convert every response from centimeters to meters and refit the same OLS model, keeping \(\mathbf X\) fixed. If \(s>0\) and all \(h_{ii}<1\), every standardized residual \(r_i\) is unchanged.
Answer: True. OLS fitted values, raw residuals, and \(s\) all divide by 100, while \(\mathbf H\) and \(h_{ii}\) stay the same. Thus \(r_i=(e_i/100)/\{(s/100)\sqrt{1-h_{ii}}\}\) is unchanged. The numerical response units cancel.
Question 3: Short Answer [30]
Each part is worth [10] points. Show brief calculations and concise reasoning on the separate answer sheet.
3a) Better fit versus evidence [10]
Two full-rank nested OLS models with intercepts use the same \(n=28\) observations, with \(\mathrm{SST}>0\). The reduced model has \(p_R=2\), \(\mathrm{SSE}_R=84\); the full model has \(p_F=4\), \(\mathrm{SSE}_F=72\). Assume independent normal errors with common variance under the full model. Use
\[ F=\frac{(\mathrm{SSE}_R-\mathrm{SSE}_F)/(p_F-p_R)}{\mathrm{SSE}_F/(n-p_F)},\qquad R^2_{\mathrm{adj}}=1-\frac{\mathrm{SSE}/(n-p)}{\mathrm{SST}/(n-1)}. \]
3a(i) [4]
Compute \(F\) and give its null reference degrees of freedom.
Answer: \(F=\{(84-72)/(4-2)\}/\{72/(28-4)\}=6/3=2\) [2]. Under the reduced-model null, \(F\sim F_{2,24}\) [2].
3a(ii) [6]
A classmate claims, “Adjusted \(R^2\) increases, so we must reject the reduced model at 5%.” Assess both parts using the formulas; only the direction of change in adjusted \(R^2\) is needed. The 5% upper-tail critical value is 3.40. Explain why the two criteria need not agree.
Answer: The same positive \(\mathrm{SST}/(n-1)\) is used for both fits. Since \(\mathrm{MSE}_R=84/26\approx3.231>3=\mathrm{MSE}_F\), adjusted \(R^2\) increases [2]. However, \(2<3.40\), so fail to reject the reduced model at 5% [2]. Adjusted \(R^2\) compares residual mean squares; the test asks whether improvement is sufficiently large relative to sampling variation under the null [2]. In fact, adjusted \(R^2\) increases when \(F>1\), a weaker condition than \(F>3.40\). The classmate’s implication is false.
3b) Agreement at one point or everywhere [10]
Let \(z=0\) for group A, \(z=1\) for group B, and \(x_c=x-8\). Consider the full-rank interaction model
\[ \mu(x,z)=\beta_0+\beta_xx_c+\beta_gz+\beta_{xg}x_cz. \]
3b(i) [4]
Write \(\Delta(x)=\mu_B(x)-\mu_A(x)\) and the null restriction for equal group means at \(x=10\).
Answer: Subtracting the two group mean functions gives \(\Delta(x)=\mu_B(x)-\mu_A(x)=\beta_g+\beta_{xg}(x-8)\) [2]. Equality at \(x=10\) is the single restriction \(\beta_g+2\beta_{xg}=0\) [2].
3b(ii) [6]
Give and verify one numerical pair \((\beta_g,\beta_{xg})\) for which the group means agree at \(x=10\) but the two lines are not identical. Then state and justify the coefficient restrictions implied by agreement at both \(x=6\) and \(x=10\).
Answer: For example, \(\beta_g=-2\) and \(\beta_{xg}=1\) give \(\Delta(10)=0\) but \(\Delta(8)=-2\), so the lines intersect without being identical [2]. Equality at 6 and 10 gives \(\beta_g-2\beta_{xg}=0\) and \(\beta_g+2\beta_{xg}=0\). Subtraction yields \(4\beta_{xg}=0\); then \(\beta_g=0\) [2 for the restrictions, 2 for the reasoning]. Thus two distinct points of agreement force these two straight mean lines to be identical.
3c) Which residual answers the question? [10]
Use the Section 6.3 standardized and externally studentized residuals:
\[ r_i=\frac{e_i}{s\sqrt{1-h_{ii}}},\qquad t_i=\frac{e_i}{s_{(-i)}\sqrt{1-h_{ii}}}. \]
One full fit has \(n=9\), \(p=4\), \(s=2\), and the following diagnostics:
| Observation | Raw residual \(e_i\) | Leverage \(h_{ii}\) |
|---|---|---|
| A | 2 | 0.75 |
| B | 3 | 0.36 |
3c(i) [6]
Compute \(r_A\) and \(r_B\). Which is more unusual by absolute standardized residual? Explain why the larger raw residual gives the opposite ranking.
Answer: \(r_A=2/(2\sqrt{0.25})=2\) [2]; \(r_B=3/(2\sqrt{0.64})=1.875\) [2]. A has the larger absolute standardized residual [1], despite its smaller raw residual: A’s larger leverage makes its denominator smaller [1].
3c(ii) [4]
Refitting without A gives \(s_{(-A)}=1\) and retains full column rank. Compute \(t_A\). A student attributes \(|t_A|>|r_A|\) to a deleted-fit residual in the numerator. Correct the explanation: identify what changes and what stays the same.
Answer: \(t_A=2/(1\sqrt{0.25})=4\) [2]. The claim is incorrect: the numerator remains the full-fit raw residual \(e_A\), and leverage also comes from the full fit [1]. Only the residual SD is replaced, from \(s=2\) to \(s_{(-A)}=1\), doubling the scaled residual [1].
Answer Key and Grading Rubric
Question 1: Multiple Choice [50]
| Part | 1a | 1b | 1c | 1d | 1e | 1f | 1g | 1h | 1i | 1j |
|---|---|---|---|---|---|---|---|---|---|---|
| Answer | C | A | B | D | B | D | A | C | D | B |
Award [5] points for each correct answer and [0] points otherwise. No explanations are required.
Question 2: True or False [20]
| Part | 2a | 2b | 2c | 2d | 2e |
|---|---|---|---|---|---|
| Answer | True | False | False | False | True |
Award [4] points for each correct answer and [0] points otherwise. No explanations are required.
Question 3: Short Answer [30]
Accept equivalent reasoning, concise equations, and any verified counterexample in 3b. Grade statistical content, not English style. A short correct explanation earns full reasoning credit. Deduct an arithmetic error once and give follow-through credit. No proof of the \(F>1\) equivalence, residual-deletion identity, or formal outlier test is required.
| Part | Grading criteria | Points |
|---|---|---|
| 3a(i) | \(F=2\) [2]; null reference \(F_{2,24}\) [2]. | [4] |
| 3a(ii) | \(\mathrm{MSE}_R=84/26>\mathrm{MSE}_F=72/24=3\), so adjusted \(R^2\) increases [2]; fail to reject since \(2<3.40\) [2]; improving adjusted \(R^2\) is a fit criterion, not sufficient evidence at the 5% level [2]. | [6] |
| 3b(i) | \(\Delta(x)=\beta_g+\beta_{xg}(x-8)\) [2]; equality at \(x=10\): \(\beta_g+2\beta_{xg}=0\) [2]. | [4] |
| 3b(ii) | Verified counterexample, e.g., \(\beta_g=-2\), \(\beta_{xg}=1\) [2]; equality at both 6 and 10 implies \(\beta_g=\beta_{xg}=0\) [2]; subtracting the restrictions gives \(4\beta_{xg}=0\), hence identical lines [2]. | [6] |
| 3c(i) | \(r_A=2\) [2]; \(r_B=1.875\) [2]; A has larger \(\lvert r\rvert\) [1]; its smaller leverage factor outweighs its smaller raw residual [1]. | [6] |
| 3c(ii) | \(t_A=4\) [2]; \(e_A\) and \(h_{AA}\) still come from the full fit [1]; only the residual SD changes from \(s\) to \(s_{(-A)}\) [1]. | [4] |
| Total | [30] |