20240111

Elaborate two advantages of using probit or logit over linear regression models for binary
outcomes. (5p)

Two key advantages of using probit or logit models over linear regression models for binary
outcomes are:

1. Bounded Predictions:
Linear regression can produce predicted probabilities outside the [0, 1] range, which is
not appropriate for binary outcomes. For example, a linear regression model might
predict probabilities less than 0 or greater than 1, which doesn't make sense in the
context of binary outcomes (e.g., the probability of an event occurring or not). In
contrast, both the logit and probit models transform the output through a link function
(logistic or normal cumulative distribution function, respectively), ensuring that the
predicted probabilities are always between 0 and 1, making them more suitable for
binary data.

2. Interpretability in Terms of Odds and Risk:

Logit and probit models offer more straightforward interpretations for binary outcomes
compared to linear regression. In a logit model, the coefficients represent the change in
the log-odds of the outcome, which is directly interpretable in terms of the odds of the
event occurring. This makes it easier to understand the relationship between predictors
and the probability of the binary outcome. Similarly, in a probit model, the coefficients
relate to the z-score in a normal distribution, offering a clear interpretation of the
standard normal change in the probability of the event. Linear regression does not
directly offer this level of interpretability for binary outcomes.
Is the interpretation of pseudo-𝑅𝑅2 of a binary outcome model the same as that of standard
𝑅𝑅2? Why? (5p)
Standard R2R^2R2 in Linear Regression:

In linear regression, R2R^2R2 represents the proportion of the variance in the dependent
variable that is explained by the independent variables. It ranges from 0 to 1, with a value of 1
indicating that the model explains all of the variability in the dependent variable and a value of 0
meaning that the model explains none. This interpretation works well because linear regression
assumes a continuous outcome and measures how well the model fits the data.

Pseudo R2R^2R2 in Binary Outcome Models:

For binary outcomes, such as those modeled with logit or probit regression, the pseudo
R2R^2R2 is an approximation of model fit but does not have the same interpretation as the
standard R2R^2R2. The pseudo R2R^2R2 is often based on different calculations that involve
likelihood functions rather than variance explained. Common pseudo R2R^2R2 measures
include:

• McFadden's R2R^2R2: It is based on the log-likelihood values of the model, comparing

the fit of the model to a baseline (intercept-only) model. Higher values indicate better
model fit, but these values are generally lower than in linear regression models. A
pseudo R2R^2R2 value of 0.2 to 0.4 is often considered a good fit in a binary outcome
model, whereas in linear regression, an R2R^2R2 value much higher than that is
common.

• Cox and Snell's R2R^2R2, Nagelkerke's R2R^2R2, etc.: These are also variations that
aim to measure the goodness of fit, but like McFadden's, they do not have the same
straightforward interpretation as the standard R2R^2R2.

Key Differences in Interpretation:

• Standard R2R^2R2 is directly interpretable as the proportion of variance explained by

the model in linear regression, while pseudo R2R^2R2 in binary models provides an
indirect measure of fit based on likelihood functions, so it doesn’t represent variance
explained in the same way.

• Pseudo R2R^2R2 values are usually smaller and don't offer the same percentage-based
interpretation (e.g., you can't say a pseudo R2R^2R2 of 0.2 explains 20% of the
"variance" in a binary outcome).

Thus, the interpretation of pseudo R2R^2R2 is more about model fit in terms of likelihood rather
than variance explanation, which is why it should not be directly compared to standard
R2R^2R2.

2. Panel data (20 points in total)

a) Elaborate two advantages of using panel data compared to cross-section data or time-series
data. Explain the advantages with an empirical research example. (8p)

Panel data has several advantages over cross-section and time-series data. Two key advantages
are:

1. Better Control for Unobserved Heterogeneity

Panel data allows researchers to control for individual-specific, time-invariant factors that could
bias results in cross-sectional or time-series analyses. Since panel data follows the same units
(e.g., individuals, firms, or countries) over time, it enables the use of fixed-effects or random-
effects models, which help control for omitted variable bias.

2. Enhanced Ability to Study Dynamics and Causal Relationships

Panel data allows researchers to analyze changes within entities over time, making it easier to
study causality and dynamic relationships. It provides the opportunity to assess lagged effects,
path dependency, and state dependence, which are difficult to capture with pure cross-
sectional or time-series data.

b) Write the linear squares dummy variable (LSDV) regression for an individual fixed effect
model. (4p)

c) Describe the purpose of the Hausman test in panel data analysis (2p) What is its null
hypothesis? (2p) Explain verbally how the test helps in choosing between different panel data
models. (4p) (You do not need to write the test statistics of the Hausman test)
3. Consider a financial literacy program on savings behavior across different regions. The
program is implemented in some regions but not others, and within these regions, it targets
households with children. (20 points in total)

a) Design a difference-in-differences (DD) strategy for evaluating the impact of the financial
literacy program on savings behavior. (Note: To get full points, you need to specify the DD
estimator either based on conditional expectations of the outcome variable or in a regression.)
(4p)

To evaluate the impact of the financial literacy program on savings behavior, we can use a
difference-in-differences (DD) approach, which compares changes in savings behavior over
time between treated and untreated groups.

1. Treatment and Control Groups

• Treated group: Households with children in regions where the financial literacy
program was implemented.

• Control group: Households with children in regions where the program was not
implemented.

• The key assumption is that both groups would have followed a similar savings trend in
the absence of the program (parallel trends assumption).

2. Time Periods

• Pre-treatment period: Before the financial literacy program was implemented.

• Post-treatment period: After the financial literacy program was implemented.

b) Elaborate two key assumptions necessary for your DiD method to provide valid causal
inferences. (You should discuss these assumptions in the context of this research problem.)
(6p)

For the DiD estimator to provide valid causal inferences, two key assumptions must hold:

1. Parallel Trends Assumption (4p)

The most critical assumption in DiD is that, in the absence of the financial literacy program,
savings behavior in treated and control groups would have followed similar trends over
time. This means that any observed difference in savings between the groups after the program
must be due to the program itself and not pre-existing differences in savings trajectories.

Context in This Research Problem:

• The assumption requires that, before the program was introduced, savings trends
among households with children in treated regions were evolving similarly to those in
control regions.

• If treated regions had faster-growing savings rates before the program (e.g., due to better
financial infrastructure), then the estimated impact of the program could be overstated.

• A way to check this assumption is to compare pre-treatment trends in savings

behavior between the treated and control groups. If they were parallel, it strengthens
the validity of the DiD approach.

2. No Differential Shocks Affecting Treated and Control Groups (2p)

The DiD method assumes that no other events or policies differentially affected the treated
group compared to the control group during the study period. If external shocks impact only
the treated group, the estimated program effect may be biased.

Context in This Research Problem:

• Suppose the government introduced a separate tax incentive for savings that only
applies to households in treated regions. This policy, rather than the financial literacy
program, could be responsible for increased savings.

• Similarly, if an economic crisis disproportionately affected control regions, then their

savings might drop more than expected, making the program seem more effective than
it actually is.

• To address this, researchers should control for major regional economic shocks or
conduct robustness checks, such as using additional control groups.

Conclusion:

If these assumptions hold, the DiD estimator will capture the true causal effect of the
financial literacy program on savings behavior. However, violations of these assumptions can
lead to biased estimates, and researchers must check them carefully through placebo tests,
pre-trend analysis, and robustness checks.
c) Design a difference-in-differences-in-differences (DDD) strategy for this problem. (Note: To
get full points, you need to specify the DD estimator either based on conditional expectations of
the outcome variable or in a regression. (8p)

The Difference-in-Differences-in-Differences (DDD) approach extends the standard

Difference-in-Differences (DiD) method by introducing a third level of comparison, further
improving causal identification. In this case, we introduce a second control group—households
without children—to account for time trends and regional differences more effectively.

1. Groups for the DDD Approach

We now define four groups:

• Treated group: Households with children in regions where the program was
implemented.

• Control group 1 (DiD control): Households with children in regions without the
program.

• Control group 2 (DDD control): Households without children in both treated and
control regions.

This additional comparison group helps account for other regional factors that might
influence savings behavior, such as regional economic shocks or policy changes affecting all
households in treated regions.
4. Why Use DDD Instead of DiD?

• The DDD approach accounts for regional shocks that might impact savings in both
treated and control regions.

• It also helps control for differences between households with and without children,
isolating the program’s effect more cleanly.

• If a general savings trend affected both households with and without children, DDD
would remove this bias by comparing changes in savings across both groups and
regions.

Conclusion:

• The DDD estimator (β7\beta_7β7) provides a more robust measure of the

program's impact by removing confounding effects that could bias a standard DiD
estimate. This strategy ensures that we isolate the true causal effect of financial
literacy on household savings behavior.

d) What bias can be avoided by the DDD strategy but not by the DD strategy? (2p)

Bias Avoided by the DDD Strategy but Not by the DD Strategy (2p)

The DDD strategy helps eliminate biases arising from time-varying, region-specific shocks
that could affect both treated and control regions differently in a DD design. Specifically, DDD
accounts for:

1. Regional Time Trends:

o In a standard DiD approach, if there are other policies or economic shocks

that impact savings behavior differently across regions, the estimated program
effect may be biased.
 o The DDD strategy controls for these regional trends by including households
without children, who should not be affected by the program but still experience
the same regional influences.

2. General Differences Between Households With and Without Children:

o Households with children might have intrinsically different savings behaviors

compared to those without children (e.g., they might save more for education).

o If this difference changes over time due to external economic factors, a

standard DiD estimate could mistakenly attribute these changes to the
financial literacy program.

o The DDD strategy removes this bias by comparing savings trends among
households with and without children across both treated and control
regions.

Conclusion:

The DDD strategy eliminates biases from confounding time-varying shocks specific to
treated regions and differential trends in household savings behavior that could distort the
true effect of the financial literacy program.
b) What is the null hypothesis implied by the restricted model? (2p) How to test this null
hypothesis? (2p)
C)How to estimate the risk premium for each factor? (4p)
d)Is the risk premium estimated in c) equivalent to the risk premium estimated from the cross-
sectional regression approach? Why? (4p)
Great question! The risk premium estimated in the approach I outlined earlier (based on time-
series regressions for each asset) is not exactly equivalent to the risk premium estimated using
a cross-sectional regression approach. Here’s why:

Time-Series Approach:

In the time-series approach, you're estimating the relationship between the returns of each
asset and the risk factors over time. The process involves:

1. Regression of Asset Returns on Risk Factors: For each asset, you regress its excess
return (return minus the risk-free rate) on the excess returns of the risk factors.

2. Risk Premiums: The risk premiums for the factors are typically estimated as the
average excess returns of the factors over the period you are analyzing.

In this method, you are focusing on the historical time-series data to estimate how much the
factors explain the variation in the asset returns over time.

Key Differences:

1. Time-Series vs. Cross-Sectional Focus:

o In the time-series approach, the risk premium for each factor is an average
excess return over time, considering how the factor returns evolve.

o In the cross-sectional approach, the risk premium represents how much return
is required by investors for taking on exposure to a factor at a given point in
time.

2. Factor Exposure (Betas):

 o In the time-series approach, factor loadings (betas) for each asset are
estimated through regression over time, where the assets' returns are explained
by the factors’ returns.

o In the cross-sectional approach, the betas are determined at a single point in

time for each asset. These betas represent how much each asset is exposed to
the factors at that specific time, not across the entire period.

3. Risk Premium Interpretation:

o The risk premium in the time-series approach is based on the average of

factor returns over time.

o The risk premium in the cross-sectional approach is related to how investors

are willing to pay for exposure to those factors at a specific time.

Equivalence?

The risk premiums estimated in both approaches could be similar if factor returns are stable
over time and if the factor exposures (betas) of the assets are consistent across the sample
period. However, in practice, the two methods can yield different results:

• The cross-sectional regression approach gives a snapshot of the factor prices (risk
premiums) at a single point in time.

• The time-series regression approach gives the historical average of the risk premiums
over the time period considered.

Thus, while both approaches estimate the compensation investors demand for exposure to risk
factors, the time-series method focuses on how the returns of factors correlate with asset
returns over time, whereas the cross-sectional method focuses on the factor's pricing at a
given point in time across assets.

In summary, the risk premium from the time-series regression approach is not identical to that
from the cross-sectional regression approach, because they are estimating risk premiums in
different contexts (historical average vs. cross-sectional pricing). However, under certain
conditions (such as stable factor returns and exposures), the two approaches could yield
similar estimates.
b) Write the formula for GJR-GARCH(1,1). (5p) Which stylized feature of financial time-series
data can be captured by GJR-GARCH in addition to those by GARCH? (3p) How is this feature
parametrized in GJR-GARCH? (3p)