Identification

Seemingly Unrelated Regression Models

Reciprocal Causation (Endogeneity)

Multinomial Choice Models

Seemingly Unrelated Regression Models 6 / 23 .

A Multiple Equation Model

• Joint Stochastic Component

𝑌𝑖 ∼ 𝑓 ( 𝜃𝑖 , 𝛼 ) for each 𝑖 (𝑖 = 1, … , 𝑛)
𝑁 ×1 𝑁 ×1 𝑁 ×𝑁

• Systematic components

𝜃1𝑖 = 𝑔1 (𝑥1𝑖 , 𝛽1 )
𝜃2𝑖 = 𝑔2 (𝑥2𝑖 , 𝛽2 )
⋮
𝜃𝑁 𝑖 = 𝑔𝑁 (𝑥𝑁 𝑖 , 𝛽𝑁 )

• Independence Assumption
• 𝑌𝑖 á 𝑌𝑗 ∣ ∀𝑖 ≠ 𝑗
• Not an assumption about relationships among 𝑌𝑖1 , … , 𝑌𝑖𝑁

Seemingly Unrelated Regression Models 7 / 23 .

When is a M.E. Model Better than 𝑁 Separate Models?
When elements of 𝑌𝑖 ∣𝑋 are stochastically or parametrically dependent

• Full Probability, with 𝑁 = 2: 𝑓 (𝑦∣𝜃) = ∏𝑛𝑖=1 𝑓 (𝑦1𝑖 , 𝑦2𝑖 ∣𝜃1𝑖 , 𝜃2𝑖 )

• Assume stochastic independence so we can factor 𝑓
𝑛
𝑃(𝑦∣𝜃) = ∏ 𝑓 (𝑦1𝑖 , 𝑦2𝑖 ∣𝜃1𝑖 , 𝜃2𝑖 )
𝑖=1
𝑛
= ∏ 𝑓 (𝑦1𝑖 ∣𝜃1𝑖 )𝑓 (𝑦2𝑖 ∣𝜃2𝑖 )
𝑖=1

with log-likelihood
𝑛 𝑛
ln 𝐿(𝜃1 , 𝜃2 ∣𝑦) = ∑ ln 𝑓 (𝑦1𝑖 ∣𝜃1𝑖 ) + ∑ ln 𝑓 (𝑦2𝑖 ∣𝜃2𝑖 )
𝑖=1 𝑖=1

• Also assume parametric independence so 𝜃1 and 𝜃2 are

distinct
↝ Equivalent to maximizing equation-by-equation likelihoods
Seemingly Unrelated Regression Models 8 / 23 .
Seemingly Unrelated Regression Model (SURM)

• The Model
1. 𝑌𝑖 ∼ 𝑁 ( 𝜇𝑖 , Σ )
𝑁 ×1 𝑁 ×1 𝑁 ×𝑁
2. 𝜇𝑖𝑗 = 𝑋𝑖𝑗 𝛽𝑗 for equation 𝑗 (𝑗 = 1, … , 𝑁 )
𝑁 ×1 𝑁 ×𝑘𝑗 𝑘𝑗 ×1
3. 𝑌𝑖 á 𝑌𝑗 ∣ ∀𝑖 ≠ 𝑗

• Likelihood
𝑛
𝐿(𝛽, Σ) = ∏ 𝑁 (𝑦𝑖 ∣𝜇𝑖 , Σ)
𝑖=1
𝑛
1
= ∏(2𝜋)−1 ∣Σ∣−1/2 exp [− (𝑦𝑖 − 𝜇𝑖 )′ Σ−1 (𝑦𝑖 − 𝜇𝑖 )]
𝑖=1 2

Seemingly Unrelated Regression Models 9 / 23 .

SURM Notes

• If 𝑌 ∣𝑋 are stochastically independent, and 𝛽’s are

parametrically independent: SURM = equation-by-equation
LS.
• In any PDF
• stochastic independence [P(𝑎, 𝑏) = P(𝑎)P(𝑏)] ⟹
• mean independence [𝐸(𝑎𝑏) = 𝐸(𝑎)𝐸(𝑏)] ⟹
• uncorrelatedness (no linear relationship) [Corr(𝑎, 𝑏) = 0].
• If 𝑌 is multivariate normal
• Uncorrelatedness ⟹ stochastic independence
• Identical 𝑋 ’s: SURM = equation-by-equation LS
• ⟹ identification of extra parameters with identical 𝑋 :
solely from modeling assumptions (i.e., lots of model
dependence)
• Programming is more complicated: Y is 𝑛 × 𝑁 instead of 𝑛 × 1
• Computational tricks: can make estimation a lot faster

Seemingly Unrelated Regression Models 10 / 23 .