Advanced Econometrics Team Homework

Problem 1
a. Investment decisions are frequently affected by broader economic conditions. To properly account for
these macroeconomic influences, it’s necessary to include time-specific intercepts in the model. This can be
accomplished by adding T − 1 dummy variables representing each time period.
b. The inclusion of ci in the equation serves as an effective way to control for county-specific character-
istics that remain constant over time but may impact both investment levels and tax policies. For example,
the overall economic climate of a county could influence investment decisions while also being correlated with
how local officials set tax rates. In a purely cross-sectional analysis, we would need to find an instrumental
variable that satisfies two conditions: being uncorrelated with ci while still correlated with tax rates - a
challenging requirement to meet in practice.
c. Standard economic theories of investment suggest that, all else being equal, higher marginal tax rates
tend to discourage investment activity.
d. I would recommend beginning with a fixed effects analysis as it permits arbitrary correlation between
the time-varying explanatory variables and the unobserved effect ci . As a preliminary step, running pooled
OLS can provide useful baseline results for comparison with the FE estimates. The fixed effects approach
relies on the strict exogeneity assumption, meaning that zit , taxit , and disasterit must be uncorrelated with
the error terms uis across all time periods.
Regarding the serial correlation properties of {uit }, there are two possibilities:
• If ci captures most of the persistence, the uit may show little serial correlation, making standard fixed
effects appropriate
• If the uit exhibit positive autocorrelation, first differencing might be preferable

In either case, it’s advisable to compute both conventional and robust standard errors.
e. Assuming taxit and disasterit don’t have lagged effects on investment, the strict exogeneity assumption
could only be violated if future values of these variables are correlated with current errors. While it’s
reasonable to assume future natural disasters aren’t influenced by past investment, tax policies might be
adjusted in response to historical investment levels - particularly if officials have specific revenue targets.
The fixed effects approach, by allowing taxit to correlate with ci , helps mitigate this concern, though the
possibility of feedback effects can’t be completely eliminated.

Problem 2
a. The model allows for consistent estimation of θ2 , δ2 , and γ (assuming all elements of zit vary over
time). However, the first period intercept θ1 and the gender coefficient δ1 cannot be estimated from this
specification.
b.
• θ2 measures the underlying wage growth for men when controlling for other factors. If we set f emalei =
0 and hold zi1 = zi2 , the average change in log wages simplifies to θ2 , representing the growth at-
tributable to economy-wide factors.
• δ2 captures differences in wage growth patterns between women and men. When δ2 = 0, it implies
that men and women with identical characteristics experience the same average wage growth.

1
 c. The wage equations for the two periods are:

log(wagei1 ) = θ1 + zi1 γ + δ1 f emalei + ci + ui1

Taking the difference between the two equations yields:

∆ log(wagei ) = θ2 + ∆zi γ + δ2 f emalei + ∆ui

This formulation shows that wage growth depends on changes in the control variables and gender. When
zi1 = zi2 , wage growth is θ2 for men and θ2 + δ2 for women. While this allows testing for gender differences
in wage growth, it doesn’t provide information about wage differentials in any specific year.

Problem 3
a. For the case with T = 2 time periods, the fixed effects estimator can be expressed in terms of the
transformed variables: !−1 N !
XN X
β̂F E = ∆x′i ∆xi ∆x′i ∆yi = β̂F D
i=1 i=1

This demonstrates that the fixed effects and first difference estimators produce identical results when there
are only two time periods.
b. The fixed effects residuals for the two periods relate to the first difference residuals as:

ûi1 = −ϵi /2 and ûi2 = ϵi /2

where ϵi = ∆yi − ∆xi β̂F D are the first difference residuals. This leads to the following relationship between
the sum of squared residuals:
N N
X 1X 2
(û2i1 + û2i2 ) = ϵ
i=1
2 i=1 i
Consequently, the estimated error variance from fixed effects is exactly half that from first differences
(σ̂u2 = σ̂e2 /2). However, the variance matrices of the coefficient estimates are identical, resulting in the same
standard errors and test statistics for both approaches.

Problem 4
a. Including aggregate time effects (like d2t ) is crucial for distinguishing genuine program effects from
broader temporal trends. For example, if studying a job training program’s impact on unemployment rates,
omitting time effects might lead to attributing general economic improvements to the program.
b. The ci term accounts for time-invariant individual characteristics that might influence both program
participation and outcomes. This is particularly important in non-experimental settings where participants
aren’t randomly assigned.
c. After first-differencing the model, we obtain:

∆yi = θ2 + δ1 progi2 + ∆ui

The OLS estimates from this equation have clear interpretations:

2
 Without the time effect d2t , δ̂1 would simply reflect the average change for the treated group, potentially
confounding program effects with general time trends.
d. For T time periods, the comprehensive model specification includes:

yit = θ1 + θ2 d2t + · · · + θT dTt + δ1 progit + ci + uit

This includes separate intercepts for each period, the unobserved effect, and the program indicator.
e. The general model offers several benefits compared to simplified versions like equation (10.89):
• Accommodates arbitrary patterns of program participation
• Separates program effects from general time trends

• Provides more flexible estimation of aggregate time effects

Team Members
• Mohamed Youssef Azouzi
• Nada Baccouchi
• Ahmed Kahlaoui

• Chedli Ben Amara