Question 1:
This question is rather comprehensive. The overarching point of this question is to provide the students with an opportunity to show that they can read a regression table from a published article and make sense out of this table.
1a
There are three different claims to evaluate:
1. aid-dependent power-sharing elites will hold cleaner elections
2. aid-dependent power-sharing elites will limit judicial independence
3. aid-dependent power-sharing elites will increase particularistic spending
All three lack a specific comparison, as in cleaner than whom, and this is important in evaluating the question. The basic knowledge needed to evaluate this table is to realize that is an interaction term. Students should identify the coefficients involved. The next step is to identify which regression models that corresponds to what claim, and make some kind of a justified claim as to whether Haass is right.
The models behave differently, as some models have all three coefficients pointing in the same direction, and others have coefficients that for a specific set of X and Z values, will produce a 0 effect for either X or Z.
The good students point out that we lack important information in order to justify some of the claims, in particular the variance-covariance matrix and the descriptive statistics. Also, very good students will see that the fraction Aid/GDP is likely to be between 0 and 1, and that the log of this number is negative, but this is not necessary for an A.
Very good students correctly identify the claims that are clearly supported from those that we need more information to evaluate, and identify what information that is needed.
1b
This is a panel structure, or a cross-sectional time series (CTST). Beyond this point, students are rewarded for providing a good discussion of problems related to CTST. We do not ask for all or most of the relevant challenges, so the points that are raised are rewarded based on the quality of the discussion and the justification of the relevance.
Most students are prone to raise issues related to the variance of the error term, autocorrelation and omitted variable bias.
1c
Haass uses OLS with robust standard errors clustered on country and with temporal fixed effects. The latter is used to de-trend the data and the former is usually used to deal with i.i.d.-related challenges. Good students point out that autocorrelation is not mentioned, and that the model might suffer from OVB. Very good students observe that the relatively few observations per country makes these corrections difficult to perform.
Question 2:
Four different terms to be defined. A and B were assumed to be quite easy, and C and B were assumed to be more difficult.
2a
E: Vaguely connect to diagnostics
D: Connect to diagnostics
C: Know that it has to do with regression diagnostics and case-based diagnostics
B. Know the correct definition
A: Know the correct definition and explain why it is important
2b
E: Connect to diagnostics
D: Know that it has to do with regression diagnostics and case-based diagnostics
C: Decompose the term into residuals, leverage and influence
B. Discuss how to handle outliers
A: All of the above, with confidence and precision.
2c
C: Know that this term is connected to the concept of uncertainty.
B: Connect the SE to the uncertainty of the coefficient in a regression model
A: Connect the SE to standard deviation of the sampling distribution of a regression coefficient. Little credit is given for almost reciting the definition.
2d
The link function in a regression setting is the link between the observable part (XB), and the outcome. In logistic regression, the link function is what map a value between plus and minus infinity to a value between 0 and 1. This question is used to separate between A and B students.
Question 3:
The manner in which this question is answered is very important. Both sub-questions asks for deliberation, and we are not as much looking for the right answer as we want to see if the students are able to discuss different challenges up against each other.
3a
Under ideal circumstances, the logit estimator is more efficient when the dependent variable is binary, and a binary dependent variable is likely to provoke a breach of some OLS assumptions, in particular regarding the assumptions. However, logistic regression is more vulnerable to omitted variable bias.
3b
Again, this question is largely used to differentiate between A and B students. Fixed effects is inter alia problematic in conjunction with logistic regression when some units have no variation in the dependent variable. This would lead the model to assign a probability of either 0 or 1 to that unit, which corresponds to a logit value of plus or minus infinity. OLS does not suffer from this problem.