Week 7 · Dr. Aluma Dembo

Multiple Hypothesis Tests

Multiple Hypothesis Tests

Part of: Intro Econometrics | Key concepts: F-test, Joint Hypothesis Test, Restricted Model, Unrestricted Model, Residual Sum of Squares, t-test, Returns to Scale


Where this lecture sits

Quick recap (week 3 material)
  • The multivariate model is yi=β0+β1x1i+⋯+βKxKi+uiy_i=\beta_0+\beta_1x_{1i}+\dots+\beta_Kx_{Ki}+u_i; each βk\beta_k is the effect of xkx_k holding all other variables fixed.
  • OLS Estimation minimises ∑u^i2\sum\hat u_i^2; the partial derivatives give K+1K+1 first-order conditions ∑u^i=0\sum\hat u_i=0, ∑u^ixki=0\sum\hat u_ix_{ki}=0.
  • In R: summary(lm(y ~ x1 + x2 + x3 + x4, mydata)).
  • A single coefficient is tested with a t-test or a Confidence Interval.

New this lecture: how to test several restrictions at once with an F-test, and how to build the restricted model for any null hypothesis. The later lectures reuse this machinery: Lec_08-Goodness of Fit & Multicollinearity (joint significance when t-stats are all small) and Lec_11-Interaction Terms (Chow-type joint tests on dummies and interactions). The whole course is summarised in Lec_13-Exam Review; practice is in Assign_04-Omitted Variables, F-tests & Multicollinearity (Birthweight & Soda Prices).

Title slide vs content

The title slide also lists "regression through the origin" and "change of units". The 37 slides cover only multiple hypothesis tests, so those two topics are not in this note.


1. Warm-up: reading one coefficient (Activity 7, part 1)

Model on 540 US workers (2002):

log⁡(wagei)=β0+β1Si+β2Expi+β3IQtesti+β4Agei+β5Hoursi+ui\log(wage_i)=\beta_0+\beta_1S_i+\beta_2Exp_i+\beta_3IQtest_i+\beta_4Age_i+\beta_5Hours_i+u_i
Variable Estimate Std. error t p
(Intercept) 1.385217 0.4218969 3.28 0.001
s 0.0981245 0.0108345 9.06 0.000
exp 0.0402309 0.0056855 7.08 0.000
IQ_test 0.0112204 0.0029018 3.87 0.000
age -0.0336168 0.0101734 -3.30 0.001
hours 0.0041505 0.0024687 1.68 0.093

Residual SE 0.49889 on 534 df; R2=0.3011R^2=0.3011, adjusted R2=0.2946R^2=0.2946; F=46.02F=46.02 on 5 and 534 df.

Effect of age: β^4=−0.0336\hat\beta_4=-0.0336, so one more year of age lowers wage by about 3.36% (log-level model, so multiply by 100). It is significant at 10%, 5% and 1% (p=0.001p=0.001), so we reject H0:β4=0H_0:\beta_4=0. The 95% interval is

−0.034±0.0101734⋅t0.025[534]⇒[−0.0536, −0.0136]-0.034\pm 0.0101734\cdot t_{0.025}[534]\quad\Rightarrow\quad[-0.0536,\,-0.0136]
Why this matters for what follows

A t-test answers "is this one coefficient zero?". It cannot answer "are x2x_2 and x3x_3 both irrelevant?" or "does β3=β4\beta_3=\beta_4?". Those are joint hypotheses, so we need the F-test.


2. Joint hypotheses

Take y=β0+β1x1+β2x2+β3x3+β4x4+uy=\beta_0+\beta_1x_1+\beta_2x_2+\beta_3x_3+\beta_4x_4+u and test that x2x_2 and x3x_3 both have no effect:

H0: β2=0  and  β3=0H1: β2≠0  or  β3≠0H_0:\ \beta_2=0\ \text{ and }\ \beta_3=0 \qquad H_1:\ \beta_2\neq0\ \text{ or }\ \beta_3\neq0
The alternative is "everything else"

H1H_1 is true if at least one restriction fails: (β2≠0,β3≠0)(\beta_2\neq0,\beta_3\neq0), (β2=0,β3≠0)(\beta_2=0,\beta_3\neq0) or (β2≠0,β3=0)(\beta_2\neq0,\beta_3=0). It is not "both are non-zero".

An F-test compares two models:

Unrestricted model (UR) Restricted model (R)
What it is the original model we care about the original model with the null imposed
Example y=β0+β1x1+β2x2+β3x3+β4x4+uy=\beta_0+\beta_1x_1+\beta_2x_2+\beta_3x_3+\beta_4x_4+u y=β0+β1x1+β4x4+uy=\beta_0+\beta_1x_1+\beta_4x_4+u
Parameters free forced by H0H_0

See Unrestricted Model and Restricted Model.


3. Building the restricted model

Substitute the null into the model, then simplify. Always start from y=β0+β1x1+β2x2+β3x3+β4x4+uy=\beta_0+\beta_1x_1+\beta_2x_2+\beta_3x_3+\beta_4x_4+u.

Null hypothesis Restricted model What is new
β2=0\beta_2=0 and β3=0\beta_3=0 y=β0+β1x1+β4x4+uy=\beta_0+\beta_1x_1+\beta_4x_4+u variables dropped
β0=0\beta_0=0, β2=0\beta_2=0, β3=0\beta_3=0 y=β1x1+β4x4+uy=\beta_1x_1+\beta_4x_4+u (no intercept) intercept and variables dropped
β2=0\beta_2=0 and β3=β4\beta_3=\beta_4 y=β0+β1x1+β4(x3+x4)+uy=\beta_0+\beta_1x_1+\beta_4(x_3+x_4)+u new variable x3+x4x_3+x_4
β2=0\beta_2=0 and β3=2β4\beta_3=2\beta_4 y=β0+β1x1+β4(2x3+x4)+uy=\beta_0+\beta_1x_1+\beta_4(2x_3+x_4)+u new explanatory variable 2x3+x42x_3+x_4
β2=1\beta_2=1 and β3=0\beta_3=0 y−x2=β0+β1x1+β4x4+uy-x_2=\beta_0+\beta_1x_1+\beta_4x_4+u new dependent variable y−x2y-x_2
Rule of thumb (slide 20)
  • If a restriction fixes a coefficient to a non-zero constant (βk=b, b≠0\beta_k=b,\ b\neq0), the restricted model has a new outcome variable (y−bxky-bx_k).
  • If a restriction ties coefficients together (βj=kβm\beta_j=k\beta_m), the restricted model has a new explanatory variable (combine xjx_j and xmx_m).
  • If it sets a coefficient to zero, just drop the variable.
Activity 7, part 2

H0: β2=3, β1=−β4, β3=0H_0:\ \beta_2=3,\ \beta_1=-\beta_4,\ \beta_3=0.

  • H1: β2≠3H_1:\ \beta_2\neq3 or β1≠−β4\beta_1\neq-\beta_4 or β3≠0\beta_3\neq0.
  • Unrestricted: the full four-regressor model.
  • Restricted: y=β0−β4x1+3x2+0x3+β4x4+u=β0+3x2+β4(x4−x1)+uy=\beta_0-\beta_4x_1+3x_2+0x_3+\beta_4x_4+u=\beta_0+3x_2+\beta_4(x_4-x_1)+u.
  • Rearranged: y−3x2=β0+β4(x4−x1)+uy-3x_2=\beta_0+\beta_4(x_4-x_1)+u.
  • New variables: dependent variable y−3x2y-3x_2 and explanatory variable x4−x1x_4-x_1.

4. The F-statistic

Residual sum of squares

RSS=∑i=1Nu^i2RSS=\sum_{i=1}^N\hat u_i^2

See Residual Sum of Squares (written SSR in some notes; same thing).

Key fact:

RSSUR≤RSSR\boxed{RSS_{UR}\le RSS_R}
Why dropping variables can never lower the RSS (slide 22)

OLS picks the coefficients that minimise RSS. The unrestricted model has all the freedom of the restricted one plus more: it could choose β^2=0\hat\beta_2=0 and copy the restricted estimates exactly, giving the same RSS. Since it can always do at least that well, it never does worse. So RSSR<RSSURRSS_R<RSS_{UR} unless the unrestricted estimates happen to satisfy the null exactly, in which case they are equal.

Idea of the test: if the null is (nearly) true, imposing it barely raises the RSS, so the gap RSSR−RSSURRSS_R-RSS_{UR} is small. If the null is false, forcing it costs a lot of fit and the gap is large.

The statistic

F=(RSSR−RSSUR)/qRSSUR/(N−M) ∼ F(q, N−M)\boxed{F=\frac{(RSS_R-RSS_{UR})/q}{RSS_{UR}/(N-M)}\ \sim\ F(q,\,N-M)}
Symbol Meaning
RSSURRSS_{UR} RSS of the unrestricted model
RSSRRSS_R RSS of the restricted model
qq number of restrictions in the null
N−MN-M degrees of freedom of the unrestricted model
NN number of observations
MM number of first-order conditions (estimated parameters) in the unrestricted model, i.e. K+1K+1
The denominator is σ^2\hat\sigma^2

RSSUR/(N−M)RSS_{UR}/(N-M) is the unbiased estimator of the error variance in the unrestricted model (the textbook remark on slide 34). So FF is "extra RSS per restriction" measured in units of the noise level.

Computing qq

q=dfR−dfURq=df_R-df_{UR}: the difference in degrees of freedom, which equals the number of parameters removed. With NN observations and y=β0+β1x1+β2x2+uy=\beta_0+\beta_1x_1+\beta_2x_2+u (UR has N−3N-3 df):

Null Restricted model dfRdf_R q=dfR−dfURq=df_R-df_{UR}
β1=β2\beta_1=\beta_2 y=β0+β1(x1+x2)+uy=\beta_0+\beta_1(x_1+x_2)+u N−2N-2 (N−2)−(N−3)=1(N-2)-(N-3)=\mathbf{1}
β1=β2=0\beta_1=\beta_2=0 y=β0+uy=\beta_0+u N−1N-1 (N−1)−(N−3)=2(N-1)-(N-3)=\mathbf{2}
β1=β2=5\beta_1=\beta_2=5 y−5(x1+x2)=β0+uy-5(x_1+x_2)=\beta_0+u N−1N-1 (N−1)−(N−3)=2(N-1)-(N-3)=\mathbf{2}
Shortcut

qq is just the number of equals signs you are testing (count independent restrictions). Use the df difference to double check.

Decision rule

Lec07_f_distribution_rejection

  • The FF distribution takes only positive values, so the rejection region is a single right tail: reject H0H_0 if F>Fcrit(q,N−M)F>F_{crit}(q,N-M).
  • If F≤FcritF\le F_{crit}, we fail to reject the null.
  • Critical values come from the table in the formula sheet (slide 26 shows the 5% table: columns are qq, rows are N−MN-M; the "infinity" row for q=4q=4 gives 2.37) or from R.

5. Relation between the F-test and the t-test

When there is a single restriction (q=1q=1) the two tests are the same test:

F[1, N−M]=(t[N−M])2\boxed{F[1,\,N-M]=\big(t[N-M]\big)^2}

Slide 27 illustrates with an R regression of consumption on GDP: t=42.6t=42.6 and the ANOVA F value is 1814.9=42.621814.9=42.6^2.

Only for q=1q=1

For two or more restrictions there is no t-test equivalent. A joint test can reject even when each individual t-test fails to reject, and vice versa. This is exactly the situation with highly correlated regressors (see Lec_08-Goodness of Fit & Multicollinearity).


6. Worked example: electricity production cost (Activity 7, part 3)

Setup

Cost is a Cobb-Douglas function (Cobb-Douglas Function) of output YY and the prices of fuel PFP_F, labour PLP_L and capital PKP_K:

Cost=B YβyPLβ1PKβ2PFβ3euCost=B\,Y^{\beta_y}P_L^{\beta_1}P_K^{\beta_2}P_F^{\beta_3}e^{u}

Data: 99 monthly observations (cost in millions of dollars, output in billions of kWh). Taking logs makes it linear (ln⁡B\ln B becomes the intercept):

ln⁡Cost=ln⁡B+βyln⁡Y+β1ln⁡PL+β2ln⁡PK+β3ln⁡PF+u\ln Cost=\ln B+\beta_y\ln Y+\beta_1\ln P_L+\beta_2\ln P_K+\beta_3\ln P_F+u

OLS output (unrestricted model)

Variable Estimate Std. error t p
(Intercept) -8.548757 1.394584 -6.13 0.000
ln_y 0.824102 0.013085 62.98 0.000
ln_pk 0.2028625 0.1445721 1.40 0.164
ln_pl 0.1778888 0.142561 1.25 0.215
ln_pf 0.6932505 0.0807074 8.59 0.000

Residual SE 0.21197 on 94 df; R2=0.9798R^2=0.9798, adjusted R2=0.9790R^2=0.9790; F=1140.68F=1140.68 on 4 and 94 df.

Interpretation

The coefficient on ln⁡Y\ln Y is an Elasticity: a 1% rise in output raises cost by about 0.8%, less than 1%. This means Returns to Scale are increasing: doubling output less than doubles cost, which is the signature of a natural monopoly (the slide's own conclusion).

Test of constant returns to scale

H0: βy=1H1: βy≠1H_0:\ \beta_y=1\qquad H_1:\ \beta_y\neq1

Constant returns means cost grows one-for-one with output. The null fixes a coefficient to a non-zero constant, so (by the rule in section 3) the restricted model has a new dependent variable:

ln⁡Cost−ln⁡Y=ln⁡B+β1ln⁡PL+β2ln⁡PK+β3ln⁡PF+u\ln Cost-\ln Y=\ln B+\beta_1\ln P_L+\beta_2\ln P_K+\beta_3\ln P_F+u

From R (anova(...)["Residuals",2]): RSSUR=4.2234RSS_{UR}=4.2234 and RSSR=12.3426RSS_R=12.3426. Here q=1q=1, N=99N=99, M=5M=5, so N−M=94N-M=94, and Fcrit(1,94)=3.92F_{crit}(1,94)=3.92.

F=(12.3426−4.2234)/14.2234/(99−5)=8.11920.04493=180.71F=\frac{(12.3426-4.2234)/1}{4.2234/(99-5)}=\frac{8.1192}{0.04493}=180.71

Since 180.71>3.92180.71>3.92 we reject constant returns to scale.

Same answer from the t-test

t=β^y−1se(β^y)=0.824102−10.013085=−13.443,t0.025[94]≈−1.98t=\frac{\hat\beta_y-1}{se(\hat\beta_y)}=\frac{0.824102-1}{0.013085}=-13.443,\qquad t_{0.025}[94]\approx-1.98

−13.443<−1.98-13.443<-1.98, so reject. And 13.4432=180.7113.443^2=180.71, matching FF exactly, as the q=1q=1 relation predicts.

Typo on the slides

Slide 33 writes "13.433^2 = 180.71". The t-statistic on the same slide is 13.443 and 13.4432≈180.713.443^2\approx180.7, so 13.433 is a typo.


Summary

  • A t-test handles one restriction; an F-test handles qq restrictions at once (Joint Hypothesis Test).
  • Write H0H_0, H1H_1 (any restriction fails), the unrestricted model, and the restricted model (impose H0H_0 and simplify).
  • Restricted-model recipe: zero restrictions drop variables; βk=b≠0\beta_k=b\neq0 creates a new dependent variable y−bxky-bx_k; βj=kβm\beta_j=k\beta_m creates a new explanatory variable.
  • RSSUR≤RSSRRSS_{UR}\le RSS_R always, because OLS minimises RSS and the unrestricted model has more freedom.
  • F=(RSSR−RSSUR)/qRSSUR/(N−M)∼F(q,N−M)F=\dfrac{(RSS_R-RSS_{UR})/q}{RSS_{UR}/(N-M)}\sim F(q,N-M); q=dfR−dfURq=df_R-df_{UR}; reject if F>FcritF>F_{crit} (right tail only).
  • For q=1q=1: F=t2F=t^2 and the two tests agree.
  • Electricity example: β^y=0.82\hat\beta_y=0.82 (elasticity below 1, increasing returns); F=180.71>3.92F=180.71>3.92, so constant returns to scale is rejected.