GUJARATI & PORTER · BASIC ECONOMETRICS · 5TH ED.

Multiple Regression Analysis:

The Problem of Inference
Hypothesis testing, F tests, structural breaks, and functional form

Chapter 8 Pages 233–274 High Priority Part I — Single-Equation Models Detailed + Examples

CONTENTS

1. Normality & t-distributions in MLR

2. Types of Hypothesis Tests
3. Testing Individual Coefficients (t-test)
4. Overall Significance: The F-test & ANOVA
5. R² and F: The Link
6. Incremental Contribution of a Variable
7. Testing Equality of Coefficients (§8.5)
8. Restricted Least Squares (§8.6)
9. Chow Test for Structural Break (§8.7)
10. MWD Test: Linear vs Log-Linear (§8.10)
11. LR, Wald, LM Tests (§8.9)
12. Revision Summary

Normality Assumption & t-Distributions in MLR

8.1

normality t-distribution df = n−k Jarque-Bera

KEY CONCEPT

Under the classical assumption that uᵢ ~ N(0, σ²), the standardized OLS estimators follow t-
distributions. In a k-variable model, the degrees of freedom are n − k (one df lost per
estimated parameter).

For the three-variable model (Y, X₂, X₃), each of the following is t-distributed with n − 3 df:

t = (β̂₁ − β₁) / se(β̂₁) ~ t(n−3)

t = (β̂₂ − β₂) / se(β̂₂) ~ t(n−3)
t = (β̂₃ − β₃) / se(β̂₃) ~ t(n−3)

Checking the Normality Assumption

We cannot observe uᵢ directly, but we can examine the residuals ûᵢ as a proxy. The Jarque-Bera
(JB) test formally tests normality using skewness and kurtosis of the residuals.

JB = n × [S²/6 + (K−3)²/24] ~ χ²(2)

S = skewness (0 for normal)

K = kurtosis (3 for normal)
H₀: residuals are normally distributed → do not reject if p-value is large

EXAMPLE 8.1 — Child Mortality Regression

RE G R E S S I O N O U T P U T

DEPENDENT VARIABLE: CHILD MORTALITY (CM), N = 64 COUNTRIES

CMᵢ = 263.64 − 0.0056 PGNPᵢ − 2.2316 FLRᵢ

se = (11.59) (0.0019) (0.2099)
t = (22.74) (−2.82) (−10.63)
p = (0.0000) (0.0065) (0.0000)
R² = 0.7077 R̄² = 0.6981

JB test on residuals: JB = 0.5594, p-value = 0.76 → cannot reject normality. The normality
assumption appears to hold.

▸ PGNP = per capita GNP (income proxy)

▸ FLR = female literacy rate

▸ Both coefficients are statistically significant at the 1% level

▸ Together they explain ~71% of variation in child mortality

8.2 Six Types of Hypothesis Tests in MLR

taxonomy six tests

# HYPOTHESIS TYPE TOOL SECTION

1 Individual coefficient significance (βⱼ = 0?) t-test §8.3

2 All slope coefficients jointly = 0? F-test / ANOVA §8.4

3 Two coefficients equal each other (β₃ = β₄)? t-test on difference §8.5

4 Linear restrictions on coefficients (e.g., β₂ + β₃ = 1)? Restricted LS + F-test §8.6

 4 Linear restrictions on coefficients (e.g., β₂ + β₃ = 1)? Restricted LS + F-test §8.6

5 Structural stability over time or groups? Chow test (F-test) §8.7

6 Linear vs. log-linear functional form? MWD test (t-test) §8.10

⚠ Critical: You cannot test a joint hypothesis (e.g., β₂ = 0 AND β₃ = 0) by running two separate t-
tests. Joint tests require the F-test because the two estimates may be correlated in the same sample.
The joint probability of both individual CIs containing the true β is not (1 − α)².

Testing Individual Coefficients (t-test)

8.3

t-test H₀: βⱼ=0 p-value confidence interval

KEY CONCEPT

To test H₀: βⱼ = 0 vs H₁: βⱼ ≠ 0, compute the t-ratio. If |t| exceeds the critical value (or p-value
< α), reject H₀.

T-STATISTIC FOR INDIVIDUAL COEFFICIENT

t = (β̂ⱼ − 0) / se(β̂ⱼ) ~ t(n−k)

where k = total number of estimated parameters (including intercept)

df = n − k

Decision via p-value

The p-value is the probability of obtaining a t-value as extreme or more extreme than the one
computed, assuming H₀ is true. A small p-value (e.g., < 0.05) provides evidence against H₀.

Confidence Interval Approach (equivalent)

95% CI FOR Β₂

β̂₂ − t(α/2) · se(β̂₂) ≤ β₂ ≤ β̂₂ + t(α/2) · se(β̂₂)

If the CI does not include 0 → reject H₀ at significance level α.

EXAMPLE 8.1 CONTINUED — Testing PGNP Coefficient

 TEST: H₀: Β₂ = 0 (PGNP HAS NO EFFECT ON CM)

t = −0.0056 / 0.0020 = −2.8187

df = 64 − 3 = 61 (≈ use 60 df from tables)

Critical t (two-tail, 5%): ±2.00
|−2.8187| > 2.00 → Reject H₀ ✓

p-value = 0.0065 (very small → strong evidence against H₀)

95% CI for β₂:

−0.0056 ± 2.00 × 0.0020 = [−0.0096, −0.0016]
This interval excludes 0 → same conclusion.

Interpretation: Holding FLR constant, a $1 increase in per capita GNP reduces child
mortality by 0.0056 deaths per 1,000 live births, on average, and this effect is statistically
significant.

For FLR: t = −10.63, p ≈ 0 → reject H₀ decisively. Female literacy is highly significant.

One-tail vs Two-tail Tests

Two-tail test

H₁: βⱼ ≠ 0
Use when direction of effect is unknown a priori.
Critical region: both tails.

One-tail test

H₁: βⱼ < 0 (or > 0)

Use when economic theory specifies direction.
More powerful for detecting that specific direction.

⚠ Rule: If you reject H₀ in a two-tail test with the correct sign, you will always reject in a one-tail
test. Choosing a one-tail test post hoc (after seeing the sign) is statistically invalid.

Testing Overall Significance: The F-test & ANOVA

8.4

F-test ANOVA joint significance ESS RSS TSS

 KEY CONCEPT

The F-test tests whether all slope coefficients are simultaneously zero: H₀: β₂ = β₃ = ··· = βₖ
= 0. Cannot be replaced by multiple t-tests on the same sample.

ANOVA Decomposition

TSS = ESS + RSS

Σyᵢ² = (β̂₂Σyᵢx₂ᵢ + β̂₃Σyᵢx₃ᵢ) + Σûᵢ²

df(TSS) = n−1
df(ESS) = k−1 (number of slope coefficients)
df(RSS) = n−k

The F-statistic

DE C I S I O N R U L E — O V E R A L L F - T E S T (K - V A R I A B L E M O D E L )

F-STATISTIC

F = [ESS/(k−1)] / [RSS/(n−k)] ~ F(k−1, n−k)

Reject H₀ if: F > Fα(k−1, n−k)

OR equivalently: p-value of F < α

k = total parameters (intercept + slopes)

k − 1 = numerator df (number of slope coefficients)

n − k = denominator df

EXAMPLE 8.1 CONTINUED — Overall F-test

ANOVA TABLE FOR CHILD MORTALITY REGRESSION

Source SS df MSS
─────────────────────────────────────────────
Regression 257,362.4 2 128,681.2
Residual 106,315.6 61 1,742.9
Total 363,678 63
─────────────────────────────────────────────
 F = 128,681.2 / 1,742.9 = 73.83

Critical F(2, 60) at 5% ≈ 3.15 | at 1% ≈ 4.98

73.83 >> 4.98 → Reject H₀: all slope coefficients are NOT zero ✓
p-value ≈ 0 → PGNP and FLR jointly explain CM significantly.

Why Individual t-tests Cannot Replace the F-test

If you test β₂ = 0 and β₃ = 0 separately at 5% each, the joint probability of both tests being correct
is not 95% — it depends on the covariance between β̂₂ and β̂₃. The joint test properly handles this
interdependence.

Result: You can sometimes accept each individual hypothesis but reject the joint hypothesis, and
vice versa.

The Link Between R² and F

8.4 cont.

R² and F equivalent forms

KEY CONCEPT

F and R² are monotonically related. Testing overall significance via F is equivalent to testing
whether the population R² = 0.

F IN TERMS OF R² (K-VARIABLE MODEL)

F = [R²/(k−1)] / [(1−R²)/(n−k)] ~ F(k−1, n−k)

Special case (3-variable model):

F = [R²/2] / [(1−R²)/(n−3)]

When R² = 0 → F = 0
When R² = 1 → F = ∞

EXAMPLE — Low R² but Significant F

GDP GROWTH REGRESSION: R² = 0.1092, N = 190

F = [0.1092/2] / [(1−0.1092)/187]
= 0.0546 / 0.004763
 = 11.46

F(2, 187) critical at 5% ≈ 3.04 → Reject H₀!

p-value = 0.00002

Conclusion: Even a "low" R² of 10.9% can be statistically significant

with a large enough sample. In cross-section data, low R² is common
due to unit diversity — what matters is correct signs and significant t-
values.

Incremental Contribution of a Variable

8.4 cont.

marginal F sequential entry adjusted R²

When adding a new variable to an existing model, the incremental F-test assesses whether the
new variable significantly raises ESS (and R²).

INCREMENTAL F-TEST

F = [(ESS_new − ESS_old) / m] / [RSS_new / (n − k_new)]

Equivalently:
F = [(R²_new − R²_old) / m] / [(1 − R²_new) / (n − k_new)]

m = number of new regressors added

Follows F(m, n−k_new)

Note: R̄² (adjusted R²) increases if and only if |t| of the new variable > 1 (equivalently, F > 1
for that variable).

EXAMPLE 8.1 CONTINUED — Adding FLR to the Model

STEP 1: REGRESS CM ON PGNP ONLY

CMᵢ = 157.42 − 0.0114 PGNPᵢ

R²_old = 0.1662 RSS_old = 303,228.5 ESS_old = 60,449.5

STEP 2: REGRESS CM ON PGNP AND FLR

 R²_new = 0.7077 RSS_new = 106,315.6 ESS_new = 257,362.4

INCREMENTAL F FOR ADDING FLR:

F = [(257,362.4 − 60,449.5)/1] / [106,315.6/61]

= 196,912.9 / 1742.9
= 112.98

F(1, 61) critical at 1% ≈ 7.08 → p ≈ 0 → FLR is highly significant ✓

Note: (−10.6293)² = 112.98 = F value → t² = F for a single added variable.

R̄² rose from 0.1528 to 0.6981 → FLR greatly improves the model.

Testing Equality of Two Coefficients

8.5

H₀: β₃ = β₄ t-test on difference covariance formula

KEY CONCEPT

To test H₀: β₃ = β₄, form a t-statistic using the difference of estimates and its standard
error. This requires the covariance between β̂₃ and β̂₄.

T-TEST FOR EQUALITY OF TWO COEFFICIENTS

t = (β̂₃ − β̂₄) / se(β̂₃ − β̂₄) ~ t(n−k)

where:
se(β̂₃ − β̂₄) = √[var(β̂₃) + var(β̂₄) − 2·cov(β̂₃, β̂₄)]

Note: The covariance term is crucial and available from regression software output.

Economic rationale: E.g., if Y = demand for a good, X₃ = income, X₄ = wealth, the test H₀: β₃ = β₄
tests whether income and wealth elasticities are equal.

EXAMPLE 8.2 — Cubic Cost Function (H₀: β₃ = β₄)

REGRESSION: TOTAL COST ON OUTPUT, OUTPUT², OUTPUT³

 Ŷᵢ = 141.77 + 63.48Xᵢ − 12.96X²ᵢ + 0.9396X³ᵢ
se: (6.38) (4.78) (0.986) (0.0591)
cov(β̂₃, β̂₄) = −0.0576

Test: H₀: β₃ = β₄ (i.e., coefficients of X² and X³ are equal)

t = (−12.9615 − 0.9396) / √[(0.9867)² + (0.0591)² − 2(−0.0576)]

= −13.9011 / √[0.9736 + 0.003494 + 0.1152]
= −13.9011 / √1.0923
= −13.9011 / 1.0451
= −13.31

df = 10 − 4 = 6
Critical t(6): ±3.707 at 1% → |−13.31| >> 3.707
p-value ≈ 0.000006 → Reject H₀ decisively.

Conclusion: The X² and X³ coefficients are statistically different from each

other.

Restricted Least Squares: Linear Equality Restrictions

8.6

RLS Cobb-Douglas constant returns F-test

KEY CONCEPT

When economic theory imposes a linear equality restriction on coefficients (e.g., β₂ + β₃ =

1), test it by comparing the unrestricted and restricted regressions using the F-test.

Two Approaches

① t-test approach

Run the unrestricted regression. Test the restriction post hoc using a t-statistic on the relevant
linear combination.

E.g., for β₂ + β₃ = 1:
t = (β̂₂ + β̂₃ − 1) / se(β̂₂ + β̂₃)

② F-test (RLS) approach

Incorporate restriction into the model, run both restricted and unrestricted regressions,
compare RSS values using the F-test.
General F-test Formula (Restricted vs Unrestricted)

GE N E R A L R E S T R I C T E D L S F - T E S T

F VIA RSS

F = [(RSS_R − RSS_UR)/m] / [RSS_UR/(n−k)] ~ F(m, n−k)

F VIA R² (ONLY WHEN DEPENDENT VARIABLE IS SAME IN BOTH MODELS)

F = [(R²_UR − R²_R)/m] / [(1 − R²_UR)/(n−k)] ~ F(m, n−k)

m = number of restrictions imposed

k = parameters in the unrestricted model
RSS_R ≥ RSS_UR always (restrictions increase RSS)
R²_UR ≥ R²_R always

Large F → restrictions are incompatible with data → Reject H₀

Small F (or large p-value) → restrictions are plausible → do not reject

EXAMPLE 8.3 — Cobb-Douglas Production Function, Mexico 1955–1974

UNRESTRICTED REGRESSION (NO RESTRICTION)

ln GDP = −1.6524 + 0.3397 ln Labor + 0.8460 ln Capital

t values: (−2.73) (1.83) (9.06)
R²_UR = 0.9951 RSS_UR = 0.0136

β̂₂ + β̂₃ = 0.3397 + 0.8460 = 1.1857 (suggests increasing returns?)

RESTRICTED REGRESSION (IMPOSE Β₂ + Β₃ = 1)

Substitute β₂ = 1 − β₃:
ln(GDP/Labor) = β₀ + β₃ ln(Capital/Labor) + u

ln(GDP/Labor) = −0.4947 + 1.0153 ln(Capital/Labor)

t values: (−4.06) (28.11)
R²_R = 0.9777 RSS_R = 0.0166
 F-TEST (M=1 RESTRICTION, N−K=17)

F = [(0.0166 − 0.0136)/1] / [0.0136/17]

= 0.003 / 0.0008
= 3.75

Critical F(1,17) at 5% ≈ 4.45 → 3.75 < 4.45 → Do NOT reject H₀ ✓

Conclusion: Cannot reject constant returns to scale (β₂ + β₃ = 1).

The Mexican economy (1955–1974) was likely characterised by CRS.

EXAMPLE 8.4 — Demand for Chicken (H₀: β₄ = β₅ = 0)

UNRESTRICTED: LN Y = F(INCOME, OWN-PRICE, PORK PRICE, BEEF PRICE)

R²_UR = 0.9823 (n=23, k=5)

H₀: PORK AND BEEF PRICES HAVE NO EFFECT ON CHICKEN DEMAND (Β₄ = Β₅ = 0)

Constrained: ln Y = f(income, own-price)

R²_R = 0.9801

F-TEST (M=2 RESTRICTIONS, N−K=18)

F = [(0.9823 − 0.9801)/2] / [(1 − 0.9823)/18]

= [0.0022/2] / [0.0177/18]
= 0.0011 / 0.000983
= 1.12

Critical F(2,18) at 5% = 3.55 → 1.12 < 3.55 → Do NOT reject H₀ ✓

p-value = 0.3472

Conclusion: Chicken demand is NOT significantly affected by pork or beef

prices.
The restricted model is preferred.

Economic interpretation: Own-price elasticity ≈ −0.38 (inelastic). Income elasticity ≈ 0.45

(not a luxury good). Both have expected signs.

General F-testing Strategy

▸ Start with the larger (unrestricted) model.

▸ Form the smaller (restricted) model by deleting variables or imposing linear equality constraints.

▸ Compute F = [(R²_UR − R²_R)/m] / [(1−R²_UR)/(n−k)].

▸ Reject H₀ (the restriction) if F > critical value.

▸ Important: Use R²-version only when the dependent variable is the same in both regressions.
Otherwise use the RSS-version.

8.7 Structural Stability: The Chow Test

Chow test structural break parameter stability subgroup regression

KEY CONCEPT

The Chow test checks whether the regression coefficients (intercept and slopes) are the
same across two (or more) subperiods or groups. It uses an F-test comparing pooled vs.
separate regressions.

Setup: Three Regressions

Subperiod 1 (n₁ obs): Y = λ₁ + λ₂X + u₁ → RSS₁ (df = n₁−k)

Subperiod 2 (n₂ obs): Y = γ₁ + γ₂X + u₂ → RSS₂ (df = n₂−k)
Pooled (n₁+n₂ obs): Y = α₁ + α₂X + u → RSS₃ (df = n₁+n₂−k)

H₀: λ₁ = γ₁ AND λ₂ = γ₂ (no structural change)

H₁: at least one parameter differs across periods

Chow Test Mechanics (5 steps)

1Run the pooled regression (8.7.3) on all n₁ + n₂ observations. Obtain RSS₃ = RSSR (restricted). df
= n₁ + n₂ − k.

2Run the subperiod 1 regression. Obtain RSS₁. df = n₁ − k.

3Run the subperiod 2 regression. Obtain RSS₂. df = n₂ − k.

4Compute RSSUR = RSS₁ + RSS₂ (unrestricted — allows different parameters). df = n₁ + n₂ − 2k.

5Compute the F-ratio and compare with critical value.

CH O W T E S T F - S T A T I S T I C

CHOW F-STATISTIC
 F = [(RSS_R − RSS_UR)/k] / [RSS_UR/(n₁+n₂−2k)] ~ F(k, n₁+n₂−2k)

RSS_R = RSS₃ (pooled regression)

RSS_UR = RSS₁ + RSS₂ (sum of subperiod RSSs)
k = number of parameters (e.g., 2 for simple regression)

Reject H₀ (structural stability) if F > Fα(k, n₁+n₂−2k)

EXAMPLE — US Savings-Income Regression 1970–1995 (break at 1982)

SUBPERIOD 1: 1970–1981 (N₁=12)

Ŷ = 1.016 + 0.0803 X RSS₁ = 1785.032 MPS ≈ 0.08

SUBPERIOD 2: 1982–1995 (N₂=14)

Ŷ = 153.49 + 0.0148 X RSS₂ = 10,005.22 MPS ≈ 0.015

POOLED 1970–1995 (N=26)

Ŷ = 62.42 + 0.0376 X RSS₃ = 23,248.30

CHOW F-TEST (K=2)

RSS_UR = 1785.032 + 10,005.22 = 11,790.252 (df = 26−4 = 22)

RSS_R = 23,248.30 (df = 26−2 = 24)

F = [(23,248.30 − 11,790.252)/2] / [11,790.252/22]

= (11,458.048/2) / 535.92
= 5729.02 / 535.92
= 10.69

Critical F(2,22) at 1% = 5.72 → 10.69 > 5.72 → Reject H₀ ✓

p-value = 0.00057

Conclusion: Structural break exists — the savings-income relationship

changed significantly around 1982 (the Reagan recession year).
The MPS dropped from ~8% to ~1.5%.
Caveats of the Chow Test

▸ Equal variance assumption: The test requires homoscedastic errors across subperiods (σ₁² = σ₂²).
Check this with an F-test on the two subperiod RSS values before running the Chow test.

▸ Must know the break point: The Chow test requires you to pre-specify the date of the break. If
unknown, use alternative methods (e.g., recursive residuals, CUSUM tests — covered in Ch. 13).

▸ Tells you a break exists, not what broke: Whether it is the intercept, the slope, or both requires
further investigation (addressed using dummy variables in Ch. 9).

▸ Generalises to multiple breaks: For m breakpoints, divide sample into m+1 subperiods and apply the
same logic with multiple RSS terms.

Checking the Equal Variance Assumption (pre-Chow test)

ESTIMATED SUBPERIOD VARIANCES

σ̂₁² = RSS₁/(n₁−k) = 1785.032/10 = 178.50

σ̂₂² = RSS₂/(n₂−k) = 10005.22/12 = 833.77

F = σ̂₂²/σ̂₁² = 833.77/178.50 = 4.67 (larger in numerator by convention)

~ F(12, 10)

Critical F(12,10) at 5% = 2.91 | at 1% = 4.71

4.67 > 4.71 at 1% → variances are significantly different.

PROBLEM: Equal variance assumption is violated here!

Strictly, Chow test should not be applied. Modifications exist but are beyond
scope.
(This is a cautionary finding — the example is used to illustrate mechanics.)

8.10 Testing Functional Form: MWD Test (Linear vs Log-

Linear)

MWD test functional form linear vs log-linear MacKinnon-White-Davidson

KEY CONCEPT

The MacKinnon-White-Davidson (MWD) test formally chooses between a linear model (Y

as function of X's) and a log-linear model (ln Y as function of ln X's) using two t-tests on
artificially constructed variables.

Setup
 H₀: Linear model: Y = α₁ + α₂X₂ + α₃X₃ + u
H₁: Log-linear model: ln Y = β₁ + β₂ ln X₂ + β₃ ln X₃ + u

MWD Test Procedure (6 Steps)

1Estimate the linear model. Obtain fitted values Yf.

2Estimate the log-linear model. Obtain fitted values ln̂ Y.

3Construct Z₁ = ln(Yf) − ln̂ Y.

4Regress Y on X's and Z₁. If Z₁'s coefficient is significant → reject H₀ (linear model is wrong).

5Construct Z₂ = antilog(ln̂ Y) − Yf.

6Regress ln Y on ln X's and Z₂. If Z₂'s coefficient is significant → reject H₁ (log-linear model is
wrong).

⚠ Possible test outcomes: (1) Reject neither → both are compatible; (2) Reject H₀ only → prefer
log-linear; (3) Reject H₁ only → prefer linear; (4) Reject both → neither is adequate; consider other
forms.

EXAMPLE 8.5 — Demand for Roses (Linear vs Log-Linear)

LINEAR: Ŷ = 9734.22 − 3782.20 X₂ + 2815.25 X₃

t: (3.37) (−6.61) (2.97) R² = 0.771

LOG-LINEAR: LN̂Y = 9.228 − 1.761 LN X₂ + 1.340 LN X₃

t: (16.23) (−5.90) (2.54) R² = 0.729

Both fit the data reasonably. How to choose?

STEP 4 — TEST H₀ (LINEAR IS CORRECT): REGRESS Y ON X₂, X₃, Z₁

Ŷ = 9727.57 − 3783.06 X₂ + 2817.72 X₃ + 85.23 Z₁

t of Z₁ = 0.021 p = 0.98 → Z₁ is NOT significant

→ Do NOT reject H₀: Linear model is acceptable ✓

STEP 6 — TEST H₁ (LOG-LINEAR IS CORRECT): REGRESS LN Y ON LN X₂, LN X₃, Z₂

 t of Z₂ is significant → Reject H₁: Log-linear model is NOT adequate.

Overall conclusion: Prefer the linear model for demand for roses.

8.9 The Troika: LR, Wald, and LM Tests

LR test Wald test LM test large sample

KEY CONCEPT

The Likelihood Ratio (LR), Wald (W), and Lagrange Multiplier (LM) tests are three
asymptotically equivalent hypothesis tests that follow the χ² distribution in large samples. For
linear regression models, they reduce to the F-test, making separate treatment unnecessary
at this level.

WHAT'S
TEST INTUITION
ESTIMATED

Likelihood Ratio (LR) Difference in log-likelihoods: restricted vs unrestricted Both models

Wald (W) How far estimated parameters are from restriction Unrestricted
only

LM (Lagrange Multiplier / Whether residuals from restricted model are correlated with Restricted
Score) regressors only

In large samples: LR ≈ W ≈ LM (all ~ χ²). In small samples with linear models, the F-test is
preferred. As Davidson & MacKinnon note, for linear regression "no information is gained from LM,
W, and LR over and above what is already contained in F."

8.8 Prediction with Multiple Regression

mean prediction individual prediction forecast

Prediction in multiple regression is a direct extension of the two-variable case:

▸ Mean prediction: Predict the point on the population regression function (conditional mean of Y given
X values).

▸ Individual prediction: Predict a single new observation of Y given X values — always has wider
confidence interval than mean prediction.

The formulas for prediction variances involve matrix algebra (Appendix C). In practice, regression
software computes these automatically. The key warning: prediction intervals assume the new X
values fall within the range of the original data (extrapolation carries more risk).
Chapter 8 — Revision Summary

▸ Under normality (uᵢ ~ N(0,σ²)), OLS estimators follow t-distributions with n−k df. Check
residuals with Jarque-Bera test.

▸ Six types of tests: individual t, joint F, equality of coefficients, linear restrictions (RLS),
structural stability (Chow), and functional form (MWD).

▸ Individual t-test: t = β̂ⱼ/se(β̂ⱼ). Reject H₀ if |t| > t_critical or p-value < α. Equivalent to checking
whether CI excludes 0.

▸ Joint F-test: F = [ESS/(k−1)]/[RSS/(n−k)]. Cannot substitute multiple t-tests — they are not
independent on the same sample.

▸ F and R² are linked: F = [R²/(k−1)]/[(1−R²)/(n−k)]. Low R² can still be highly significant with
large n.

▸ Incremental F: adding a variable is worthwhile if its t² (=F) > 1 (raises R̄²). Use F = [(R²_new −
R²_old)/m]/[(1−R²_new)/(n−k_new)].

▸ Testing β₃ = β₄: t = (β̂₃ − β̂₄)/√[var(β̂₃)+var(β̂₄)−2cov(β̂₃,β̂₄)].

▸ Restricted LS: impose restriction, run restricted regression, compare RSS (or R²) to
unrestricted. F = [(RSS_R − RSS_UR)/m]/[RSS_UR/(n−k)].

▸ Chow test: split sample, get RSS₁ + RSS₂ = RSS_UR, compare with pooled RSS₃ = RSS_R. F =
[(RSS_R − RSS_UR)/k]/[RSS_UR/(n₁+n₂−2k)]. Pre-check equal variances.

▸ MWD test: construct Z₁/Z₂ from residuals of the two models and add to each regression.
Significant Z₁ → reject linear. Significant Z₂ → reject log-linear.

▸ LR/Wald/LM tests are asymptotically equivalent to F-test in linear models. F-test is preferred in
finite samples.