Econometrics I Problem Set 03 Guide

Q: Under which assumptions would the OLS population parameter in a regression of Y on T be equal to the average treatment effect, the average treatment effect on treated, and on untreated?

The OLS parameter of the regression of Y on T is equal to the ATE under the assumption of no confounding (unconfoundedness), which means that T is independent of potential outcomes (Y(0), Y(1)). For ATT, the assumption is that the treated group is randomly selected from the population such that E[Y(0) | T=1] = E[Y(0)]. For ATU, a similar assumption holds for the untreated group .

Q: How does the assumption of linear transformations in the OLS model simplify the process of estimating coefficients for non-standard binary variables?

Linear transformations simplify coefficient estimation by allowing binary variables to be standardized, converting non-standard binary variables into a common format (e.g., from {a, b} to {0, 1}). This maintains the information contained while simplifying mathematical operations, crucial for consistent estimates .

Q: When adding gender as a dummy variable in the OLS model for income based on years of education, what restrictions might result on the conditional expectation function?

Introducing a gender dummy variable and interaction terms imposes restrictions that force the expectation of income to be linear in gender and education separately or combined. These restrictions might limit the flexibility of the model capturing nonlinear relationships between the factors .

Q: What are the expressions for the average treatment effect, the average treatment effect for the treated, and the average treatment effect for the untreated in terms of moments of the joint distribution of potential outcomes and treatment indicator?

The Average Treatment Effect (ATE) can be expressed as E[Y(1) - Y(0)]. The Average Treatment Effect for the Treated (ATT) is E[Y(1) - Y(0) | T = 1], and the Average Treatment Effect for the Untreated (ATU) is E[Y(1) - Y(0) | T = 0]. These are derived using the moments of the joint distribution of (Y(0), Y(1), T).

Q: How can you derive the OLS population parameters for a univariate linear regression when the predictor variable is binary?

For a univariate linear regression with a binary predictor X, the OLS parameters can be derived as β0 = E[Y] - E[X]β1 and β1 = Cov(X,Y)/Var(X). This reflects the average difference in outcomes between the two binary states weighted by their variance, capturing the sloped effect of the binary variable .

Q: In the context of years of education affecting income, how do different ways of writing the linear model change the interpretation of OLS population coefficients?

Different ways of specifying the model, such as Y = Σβs1{S = s} or Y = β0 + Σβs1{S = s}, change the interpretation because they reflect different base categories and reference points. The inclusion or exclusion of an intercept term influences how coefficients are interpreted relative to the base group, impacting the analysis of effects .

Q: Why is the full rank condition essential for defining the OLS parameters, and what happens if it is not satisfied? Can the rank ever be greater than k?

The full rank condition, rank(E[xx'])=k, ensures that the predictor matrix is invertible, allowing for unique solutions of the OLS parameters. If this condition is not satisfied, the matrix is singular, and the parameters cannot be uniquely defined. The rank cannot exceed k because the maximum rank of a matrix is limited by its smallest dimension .

Q: Provide an example when the non-invertibility of matrix A in the context of OLS model transformations could occur, and explain its implications.

Non-invertibility occurs when matrix A has less than full rank, such as a matrix of zeros or repeated rows, leading to no unique solution for the parameter estimates. This results in multicollinearity, where predictors are perfectly correlated, hindering distinct contribution identification .

Q: How can introducing interaction terms in an OLS model offer more insights compared to models without them, especially with gender and education data?

Interaction terms in an OLS model allow for the effects of combinations of variables to be studied, offering insights into how the relationship between education level and income changes with gender. This can capture different slopes or effects that are not visible when adding variables independently .

Q: What is the significance of using re-parametrization in an OLS model when considering binary variables, and can it lead to a Bernoulli distribution for the variable?

Re-parametrization allows us to transform variables such that a binary variable can be made to follow a Bernoulli distribution, simplifying the interpretation of the model. For instance, using a transformation matrix to adjust a binary variable from general values {a, b} to {0, 1} can achieve this outcome, helping in standardizing and easing computational interpretations .

The document is a problem set for an Econometrics I course, detailing various tasks related to causal models, OLS parameters, and regression analysis. It includes questions on deriving population parameters, interpreting models with different specifications, and exploring the implications of including additional variables such as gender. The set emphasizes understanding the relationships between treatment effects, education, and income while applying OLS regression techniques.

Uploaded by

Igor Costa Correia

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

0% found this document useful (0 votes)

10 views2 pages

Econometrics I Problem Set 03 Guide

Uploaded by

Igor Costa Correia

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

ECONOMETRICS I

EESP - Graduate Program in Economics | First Quarter 2024

Problem Set 03
Due Mar 05, 23:59

1. Consider a general causal model in which potential outcomes are given by Y (0) and Y (1). Let
T be an indicator variable that is equal to one if treated.
(a) Write the following population parameters as a function of moments of the joint distribu-
tion of (Y (0), Y (1), T ).
1. Average treatment effect;
2. Average treatment effect for the treated;
3. Average treatment effect for the untreated.
(b) Under which assumptions would you have that the OLS population parameter of a regres-
sion of Y on T (including a constant) is equal to each of the parameters derived in item
(a)? You can use your results for the OLS parameters when T is binary, from previous
problem sets.

2. Consider a vector of real random variables x′ = (X1 , X2 , . . . , Xk ) with length k ∈ N, and a

real random variable Y . Assume that rank(E [xx′ ]k×k ) = k.
Define the populational OLS parameters β, that is:

β := argminb∈Rk E (Y − x′ b)2

(a) Derive the expression for β explicitly.

(b) What does the full rank condition mean? Why can’t β be defined when this condition is
not satisfied (that is, rank(E [xx′ ]k×k ) < k)? Can it be that rank(E [xx′ ]k×k ) > k?
(c) (Optional) Consider now the case in which x′ = (1, X). Show that the formula you
found in item (a) recovers the OLS population parameters for the bivariate population
OLS regression (that is, β0 = E[Y ] − E[X]β1 and β1 = Cov(X,Y )
Var(X) ).

3. Let Y be income, S be years of education (which can go from 0 to 8 years) and 1 be the
indicator function. Compute the population OLS coefficients (the βs) in the following models:
(a) Y = 8s=0 βs 1{S = s} + u.
P

(b) Y = β0 + 8s=1 βs 1{S = s} + u.

(c) Can we consider the model Y = γ + 8s=0 βs 1{S = s} + u?

The idea in this question is to note, in items (a) and (b), that there are different ways to
write a linear model that recovers exactly the same approximation to the CEF. However, the
interpretation of the OLS population coefficients will be different, depending on how you write
it. You will use this idea in Q4.

4. Back to Question 3 of Problem Set 2:

(a) Propose an OLS population multivariate regression such that its coefficients identify the
causal effects (Y (1) − Y (0)) and (Y (2) − Y (0)).
ECONOMETRICS I Problem Set 03 - Page 2 of 2 Due Mar 05, 23:59

(b) Propose an OLS population multivariate regression such that its coefficients identify the
causal effects (Y (1) − Y (0)) and (Y (2) − Y (1)).
If your answer is correct, note that both OLS population regressions above recover the CEF
E[Y |T ] (because there is no approximation error of the linear approximation of the CEF
in these models). We just have a re-parametrization that changes the interpretation of the
parameters.

5. Now let’s consider richer models relative to the one in Q3. We still have that Y is income and
S is years of education (from 0 to 8 years). But we also have a dummy variable for gender, F .
(a) Consider the following OLS population regressions, and discuss the restrictions imposed
in the approximation to the conditional expectation function E[Y |F, S].
• Y = β0 + β1 F + β2 S + u
• Y = β0 + β1 F + β2 S + β3 S · F + u
• = β0 + β1 F + 8k=1 γk 1{S = k} + u (what if we had 8k=0 γk 1{S = k}?)
P P
Y
• = β0 + 8k=1 γk 1{S = k & F = 0} + 8k=0 θk 1{S = k & F = 1} + u
P P
Y
(b) In which cases above do we have E[u|F, S] = 0?

1 Bonus problems
1. In addition to being good practice for matrix algebra, this problem is worth 10
points that can replace points that you lost in this or other problem sets.
This question considers a more general setting, relative to Q3, in which we can consider linear
transformations of the random variable x.
Let x′ := (X1 , . . . , XK ) ∈ RK be a random vector such that E [xx′ ] has rank K, and let β be
the population OLS coefficient of Y on x. You may consider that X1 = 1 (the intercept).
(a) Let A be a (non-stochastic) K × K matrix. If A is invertible, then we have that x̃ = Ax
also satisfies rank(E[x̃x̃′ ]) = K. Intuitively, x̃ = Ax is a linear combination of the random
variables x, that provides the same information as x̃ (because we have x = A−1 x̃).
Let β̃ be the population OLS coefficient of Y on x̃. How does β̃ relates to β?
(b) Show that x̃′ β̃ = x′ β. Note that this means that the two model recover exactly the same
linear approximation to the CEF E[Y |x]. We just have a re-parametrization.
(c) Consider now the case in which x′ = (1, X), where X is a binary variable that assumes
values a with probability p, and b with probability 1 − p. Show that, if we use

1 0
A= a 1 ,
− b−a b−a

then x̃ will be a constant and a Bernoulli random variable. This result shows that, in
a model with a constant term in which X is binary (not necessarily {0, 1}, then we can
always consider a re-parametrization of the model so that we have a Bernoulli variable.
Therefore, there is no loss in considering the case X ∈ {0, 1} when we say that X is
binary.
(d) Suppose now that A is not invertible (which is equivalent to say that rank(A) < K).
What would happen in this case? Provide an example.

Common questions