Homework 5 11/30/25, 9:41 PM
Homework 5
AUTHOR
Logan Kenyon
Chapter 18: Coding Question 1
a.
Yes, if you were to run a regression of total_med_expenditure on free,
the coefficient on the free term would equal the Average Treatment
Effect on the Treated (ATT). Given that this data comes from an
experiment involving random assignment, we can assume that the
treatment is independent of potential outcomes, by construction. This
allows us to determine that the ATT is just the difference in outcomes
among the treated group relative to average outcomes amount the
untreated group. This is exactly what the coefficient in the regression
represents.
b.
RAND_reg <- lm(total_med_expenditure ~ free, rand_hie_subset)
summary(RAND_reg)
The estimated ATT is given by the coefficient on free, $140.12. The
result is statistically significant at the 5% level. This means that
individuals who were assigned to the “free” insurance plan incurred, on
average, $140.12 more in medical expenses than those assigned to the
“Catastrophic” plan.
c.
RAND_reg_face <- lm(face_to_face_visits ~ free, rand_hie_subset)
summary(RAND_reg_face)
The estimated ATT is given by the coefficient on free, 1.7361. The result
is statistically significant at the 5% level. This means that individuals
who were assigned to the “free” insurance plan had, on average, 1.7361
more face-to-face doctor visits than those assigned to the
“Catastrophic” plan.
d.
[Link] Page 1 of 4
Homework 5 11/30/25, 9:41 PM
RAND_reg_index <- lm(health_index ~ free, rand_hie_subset)
summary(RAND_reg_index)
The estimated ATT is given by the coefficient on free, -0.7407. The result
is not statistically significant at the 5% level. This means that individuals
who were assigned to the “free” insurance plan were, on average, .74
points lower on the overall health index compared to those assigned to
the “Catastrophic” plan. However, given the large p-value, we cannot
reject that the ATT of the insurance plans on the health index is equal to
0.
e.
The results in parts b-d demonstrate that when individuals are on a
“free” health insurance plan they are likely to use their insurance more,
1.74 more visits, which in turn increases the overall costs of healthcare.
However, even though participants visited the doctor more, in person,
there was no statistically significant improvement in the health of the
average person involved in this experiment.
Question 2
library(AER)
data("Fertility")
a.
fertility_reg <- lm(work ~ morekids, Fertility)
summary(fertility_reg)
The estimated coefficient on morekids of -5.387 should not be
interpreted as the causal effect of having more than two children. There
are likely more variables that can influence a woman’s amount of work
per week besides their number of children.
b.
Relevance: The relevance assumption which assumes that the
instrument actually affects whether units get treatment likely holds in
this context. The instrument, which is the sex of the first two children,
influences the treatment, morekids. When women have the same sex
children, they are more likely to have more than 2 kids, meaning this
increases the probability of the treatment.
[Link] Page 2 of 4
Homework 5 11/30/25, 9:41 PM
Independence: This assumption likely holds in this scenario because
the sex of the first two children of a woman is random, essentially a
coin flip. This means that the instrument is “as-good-as randomly
assigned”.
Exclusion Restriction: I believe this assumption mostly holds because
it is unlikely that the combination of genders of the woman’s children
affects how much she works. This instrument only affects how many
total children she has, which in turn affects her outcome, work hours.
Monotinicity: I also believe this assumption could hold in this context.
This assumption says that woman would not systematically oppose
the common mechanism of having another child when the first two
are the same gender. If we follow this mechanism, than we can say
that the monotonicity assumption holds in this scenario. We can say
that there is not a distinct group of people who defy the norms.
c.
Fertility$samesex <- 1 * (Fertility$gender1 == Fertility$gender2)
library(estimatr)
fertility_iv_model <- iv_robust(work ~ morekids | samesex, Fertility)
summary(fertility_iv_model)
The estimate of -6.314 represents the Local Average Treatment Effect
(LATE). This results means that, for the specific group of woman
affected by the instrument, having more than 2 children causes a woman
to work 6.314 fewer hours per week.
Question 4
a.
The main limitation from this design is that the treatment, Incumbency,
is not randomly assigned. There are many different factors that likely
affected how these candidates became incumbents or not, and these
factors are not controlled for without random assignment.
b.
Using a regression discontinuity design, we would assign the margin of
victory in the past election as the running variable in our experiment.
This value would have a cutoff at 0. This would make the treatement
indicator whether or not the candidate is an incumbent. If they are, they
would have a margin greater than 0 in the last election. Our outcome
[Link] Page 3 of 4
Homework 5 11/30/25, 9:41 PM
variable would be the margin of victory in the current election. We would
only analyze the candidates just above and below the cutoff in order to
determine whether the jump in the data at the cutoff was attributable to
their incumbency status.
c.
In this context, the continuity assumption declares that candidates who
barely won, meaning they had a margin barely above 0, have the same
outcomes are candidates that barely lost. We are assuming the
candidates who barely won and lost are very similar in their
characteristiscs, therefore, we would expect their results in the current
election to be similar absent incumbency status. This assumption might
not hold in this example because incumbency status likely comes with
additional resources such as greater fundraising ability and access to
greater media attention. These factors may aid candidates in their
efforts to win their next election.
d.
library(tidyverse)
load("[Link]")
h <- .05
house$Incumbent <- 1 * (house$Dmargin1 >= 0)
rd_data_bw <- house %>%
filter(abs(Dmargin1 - 0) <= h)
rd_reg <- lm(Dmargin2 ~ Incumbent + I(Dmargin1-0) + Incumbent:I(Dmargin1-0), rd_d
summary(rd_reg)
The estimated casual effect of incumbency is .0487. This result is
statistically significant at the 5% level. We estimate that being an
incumbent in the previous election, on average, increases the
candidates’ margin of victoryvby 4.87 percentage points in the next
election.
[Link] Page 4 of 4