EC393: Fall 2021 Final Exam

Instructions:

- This exam is open note, book, and computer but the work and explanations needs to be your own!

- It is out of 80 points.

- You have 3 hours to complete the exam – Moodle will track the time between when you downloaded the
data, .do, and this is file and when it gets submitted. It must be submitted by 9pm on Monday the 20th.

- For time sake, you do not have to write in Greek – i.e. you can say b1hat instead of !"!

- You will need to submit your answers (this document) as well as your .do file. Please follow the naming
conventions we’ve been using on problem sets:
lastname_final.docx for your written answers and table
lastname_final.do for your Stata .do file

- Big thanks for your efforts this term – you’re very close to a well-deserved break!

Please also fill in the following:

Time I began this exam: _______________________

Time I ended: _____________________

I neither gave to nor received assistance from another person during this exam:

______________________
(type your name here)
The Impact of Environmental Events on Risk Aversion

You’ll recall from your micro theory experience that economists think a lot about preferences when
trying to model and understand how people make decisions. Risk aversion is one key aspect of
preferences – someone who is risk averse is less likely to take gambles like switching to a new job,
adopting the latest production technology for their business, migrating, or taking on debt. Economists
increasingly care about how someone’s risk taking attitudes can change over time when good and bad
things happen to them. For example, winning a lottery may make me less risk averse in the future, but
watching my stock portfolio plummet during the Great Recession may make me more risk averse in
the future. I’ve recently been exploring this topic in a project with a few coauthors (including a former
ec393 ‘metrics master). Let’s look at what we’re doing.

We take public longitudinal survey data from Indonesia that measures risk attitudes of approximately
20,000 people in both 2007 and 2014 – these are predominantly small-scale agricultural workers and
rice farmers. We then match that data with satellite-based rainfall data we can use to determine if the
person had experienced droughts over the prior five years. Droughts are particularly bad in this
context – they hurt crop yields and agricultural income for farmers.1

This data is in final21_data.dta and looks like this:

id year riskaverse ndroughts

1 2007 1 3
1 2014 0 1
2 2007 0 0
2 2014 1 4

where risk aversion is coded as a binary variable 1 if you’re risk averse (don’t like risk) and 0 if not
based on preference surveys. The drought information is a count of how many years of the last 5 did
your community experience a drought. We observe each person twice – once in 2007 and again in
2014.

Droughts can matter for lots of potential reasons, but let’s start by seeing if they’re related to
income. Estimate the following bivariate regression of log income on the drought variable with
OLS.

ln($%&'())!" = ,# + ,$ %./'01ℎ3!" + )!" (1)

1. (2 pts) Report your results in column 1 of your table.

Please paste your completed table at the bottom of your exam document.

2. (2 pts) How much does predicted income change with each additional year of
drought?

3. (2 pts) Is 4566(78(9:45;<) , :>65?@ABC) = D

E % ? Why or why not?

1 Not key to this exam, but FYI a “drought year” here is a year where rainfall is below the 15th percentile of the historical
rainfall distribution of the person’s location.
 4. (2 pts) When I estimate equation (1) with LAD, the coefficient on ndrought is -0.129.
What does this tell you?
EC393: Fall 2021 Final Exam

Instructions:

- This exam is open note, book, and computer but the work and explanations needs to be your own!

- It is out of 80 points.

- You have 3 hours to complete the exam – Moodle will track the time between when you downloaded the
data, .do, and this is file and when it gets submitted. It must be submitted by 9pm on Monday the 20th.

- For time sake, you do not have to write in Greek – i.e. you can say b1hat instead of !"!

- You will need to submit your answers (this document) as well as your .do file. Please follow the naming
conventions we’ve been using on problem sets:
lastname_final.docx for your written answers and table
lastname_final.do for your Stata .do file

- Big thanks for your efforts this term – you’re very close to a well-deserved break!

Please also fill in the following:

Time I began this exam: _______________________

Time I ended: _____________________

I neither gave to nor received assistance from another person during this exam:

______________________
(type your name here)

Our primary aim is to identify the impact of droughts on risk aversion. As a good econometrician,
you know visualization is an important first step.

5. (1 pts) Graph the nonparametric regression of the risk aversion indicator on the
number of droughts. Paste the figure here.
(You can omit the scatter plot if you’d like with the ,noscatter option.)
 6. (2 pts) Briefly discuss the shape of the predictions and any patterns you see in the
figure.

Estimate the following regression where /F!" is the risk aversion indicator for individual i in year t:

/F!" = G# + G$ %./'01ℎ3!" + G& F1)!" + G' F1)!"& + G( ℎℎI$J)!" + G) ).0&!" + G* (F//$).!" + ," + 0!" (2)

and ," is an indicator for which year of the survey the observation comes from.

7. (2 pts) Report your results in column 2.

8. (3 pts) How do you interpret the coefficient on the drought variable? Is it statistically
significant?

Now re-estimate the column 2 regression but including an individual fixed effect:

/F!" = G# + G$ %./'01ℎ3!" + KL!" + ,! + ," + 0!" (3)

where !!" represents the controls for age, age2, household size, years of education, and marital status
and "! is an individual fixed effect.

9. (2 pts) Report your results in column 3.

Column 3 is our preferred specification – i.e. the ## we like the most – so let’s dig in a bit.

10. (3 pts) What counterfactual question is $$ trying to answer?

11. (3 pts) Who/what are you comparing to estimate $$ in the fixed effects regression?

12. (2 pts) What was the point of including an individual fixed effect? Why weren’t the
controls for age, household size, education, marital status, and survey-year
sufficient?

13. (3 pts) What does the coefficient on ndrought in the regression in equation (3)
suggest about OVB on the ndrought variable in results from equation (2)? Briefly
explain why this may be the case.

14. (2 pts) What are the appropriate standard errors to use in these regressions. Why?

15. (3 pts) How did the t-stat for %% : $$ = ( from the model in column 2 change when
you added individual fixed effects in column 3. Why?
 16. (2 pts) Why did I ask you to include both age and age2? Would it have been fine to
just use age and omit age2?

17. (3 pts) Test the joint significance of age and age2 and report the p-value. Do you
conclude the same about the joint hypothesis that you do about statistical
significance for each of age and age2? Why?

18. (2 pts) If you wanted to interpret the ndrought coefficient in column 3 in percent
terms, how would you phrase your interpretation?

Suppose you decided to also include a variable that measured the number of years that were not
droughts. As we use a 5-year window to look back, this variable would be (5 – ndrought).

19. (2 pts) What would happen to your estimate of $$ if column 3 if you included this
new variable in the regression?

Suppose you defined the outcome as an indicator for risk-loving that was equal to 1 when riskaverse
= 0 and equal to 0 when riskaverse = 1.

20. (3 pts) If you estimated the same righthand side version as equation (3), what would
the coefficient on ndrought be in this new regression? Why?
(You do not have to actually estimate this to answer the question).

Prior evidence on risk aversion suggests that certain groups may have more flexible preferences than
others. To test this hypothesis, we want to focus on gender differences in the impact of droughts.

21. (2 pts) Why can’t we just add an additional control for male or female to the column
3 regression to answer this question?
(The survey only has male/female as possible categories)

Instead we need to create and include an interaction between droughts and reported gender and
estimate the following:

/F!" = G# + G$ %./'01ℎ3!" + G& %./'01ℎ3LM)(FN)!" + G' L!" + ,! + ," + 0!" (4)

where X are the controls for household size, education, and marital status.
 22. (2 pts) Report your results in column 4.

23. (2 pts) What’s the impact of droughts on risk aversion for males?

24. (3 pts) What’s the impact of droughts on risk aversion for females? Is it statistically
significant at the 1% level?

25. (3 pts) How do these two compare to your estimate of b1 in column (3)? Why?

26. (2 pts) Are the two statistically different from each other?

In week 9 I stressed you needed both variables by themselves if you were going to include an
interaction – i.e. interaction regressions should look like this:
O = G# + G$ P$ + G& P& + G' P$ ⋅ P& + 0

27. (3 pts) The regression you just ran didn’t include the female indicator as a control by
itself. Why is it okay to not do that here?

My coauthors and I like this project, but to keep up with the LaFave-girl demands for Sunrise bagels,
Dairy Cone, and a seemingly endless want for stuffed animals, I need peer-reviewed publications to
my name, not just fun projects I can turn into exams! When we submit this paper for peer-review,
here’s what some of the feedback could be about the results in column (3) – say these in your
grumpiest economist voice for full-effect:

“I worry the authors are p-hacking with the control variables. I doubt the finding on the link
between droughts and risk aversion holds-up after using PDS LASSO to select among the controls
in X in equation (3).”

28. (3 pts) How do you respond?

“The outcome is binary. This analysis is fundamentally flawed due to the use of a LPM.”

29. (2 pts) Explain if you agree or disagree.

“I will only believe this if the logit marginal effects are similar magnitude to these OLS coefficients.”

30. (3 pts) Estimate the logit version of equation (3) and report the ndrought marginal
effect here (you don’t need to add the full column to your table). How do you
respond to this complaint?
“I would find this more convincing if this were a randomized controlled trial instead.”

31. (3 pts) How do you respond?

Let’s take a bit of a step back now to wrap up.

The satellite rainfall data is quite fun – look at this great map I made showing areas with more and
fewer droughts:

but there is a straightforward question already in the survey that says, “Did you experience any crop-
loss due to drought in the last 5 years?”

32. (2 pts) That question would have been much easier to use than all the coding of the
rainfall analysis – why didn’t we just use that question instead as our key righthand
side variable?

33. (4 pts) Focusing on internal validity for this sample, what must be true for your fixed
effects estimates in columns 3 and 4 to be given a causal interpretation? Do you
believe it’s true?

Please paste your results table here