library(pacman)
library(dplyr)
Attaching package: 'dplyr'The following objects are masked from 'package:stats':
filter, lagThe following objects are masked from 'package:base':
intersect, setdiff, setequal, unionlibrary(ggplot2)
p_load("tidyverse")
p_load('fixest')
p_load("lmtest")
p_load("stargazer")
p_load("data.table")
p_load('haven')
p_load('broom')
p_load('modelsummary')
p_load('patchwork')allsites <- read_dta('allsites.dta')
glimpse(allsites)Rows: 42,974
Columns: 44
$ fips <chr> "01001020200", "01001020300", "01001020400", "0100102…
$ state <chr> "", "", "", "", "", "", "", "", "AL", "", "", "", "",…
$ npl2000 <dbl> 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0,…
$ pop_den8 <dbl> 1634.88367, 1497.60767, 1995.14172, 454.03589, 913.84…
$ shrblk8 <dbl> 0.606448, 0.149201, 0.000000, 0.059753, 0.156902, 0.2…
$ shrhsp8 <dbl> 0.012328, 0.016613, 0.016842, 0.014321, 0.003367, 0.0…
$ child8 <dbl> 0.355618, 0.328753, 0.300121, 0.357530, 0.347474, 0.3…
$ old8 <dbl> 0.139876, 0.098083, 0.059050, 0.025185, 0.056565, 0.1…
$ shrfor8 <dbl> 0.010431, 0.021405, 0.009334, 0.014321, 0.000000, 0.0…
$ ffh8 <dbl> 0.249146, 0.166015, 0.097989, 0.066844, 0.175700, 0.1…
$ smhse8 <dbl> 0.517978, 0.404166, 0.488126, 0.408845, 0.422598, 0.6…
$ hsdrop8 <dbl> 0.141361, 0.066079, 0.083333, 0.272222, 0.196428, 0.0…
$ no_hs_diploma8 <dbl> 0.4954874, 0.2879581, 0.2278614, 0.2380952, 0.3936529…
$ ba_or_better8 <dbl> 0.10288808, 0.14077953, 0.20651032, 0.20602527, 0.087…
$ unemprt8 <dbl> 0.108241, 0.035384, 0.033933, 0.073512, 0.079545, 0.0…
$ povrat8 <dbl> 0.302460, 0.077174, 0.057426, 0.070123, 0.084511, 0.2…
$ welfare8 <dbl> 0.203566, 0.091003, 0.033996, 0.076256, 0.070646, 0.1…
$ avhhin8 <dbl> 14465.63, 18397.63, 23987.15, 22034.01, 19911.98, 151…
$ tothsun8 <dbl> 741, 1018, 1633, 641, 1029, 1077, 803, 1013, 708, 113…
$ ownocc8 <dbl> 491, 716, 1293, 487, 834, 826, 590, 715, 617, 658, 67…
$ firestoveheat80 <dbl> 0.05802969, 0.00000000, 0.01959584, 0.02340094, 0.017…
$ noaircond80 <dbl> 0.33063427, 0.11296660, 0.01898347, 0.06240249, 0.165…
$ nofullkitchen80 <dbl> 0.006747638, 0.017681729, 0.004286589, 0.034321371, 0…
$ zerofullbath80 <dbl> 0.03373819, 0.03143419, 0.00000000, 0.01560062, 0.016…
$ statefips <dbl> 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,…
$ bedrms0_80occ <dbl> 0.0000000000, 0.0000000000, 0.0000000000, 0.000000000…
$ bedrms1_80occ <dbl> 0.069246434, 0.008379889, 0.000000000, 0.012170386, 0…
$ bedrms2_80occ <dbl> 0.15274949, 0.03910615, 0.08971384, 0.17241380, 0.232…
$ bedrms3_80occ <dbl> 0.5987780, 0.7416201, 0.6140758, 0.5354970, 0.5455636…
$ bedrms4_80occ <dbl> 0.13849287, 0.20391062, 0.27610210, 0.26977688, 0.157…
$ bedrms5_80occ <dbl> 0.040733196, 0.006983240, 0.020108275, 0.010141988, 0…
$ blt0_1yrs80occ <dbl> 0.03258656, 0.06145252, 0.02552204, 0.09146342, 0.056…
$ blt2_5yrs80occ <dbl> 0.11201629, 0.07541899, 0.10208817, 0.24390244, 0.229…
$ blt6_10yrs80occ <dbl> 0.2749491, 0.3198324, 0.1755607, 0.4857724, 0.2553957…
$ blt10_20yrs80occ <dbl> 0.04276986, 0.33659217, 0.52668214, 0.16869919, 0.326…
$ blt20_30yrs80occ <dbl> 0.08350305, 0.09916201, 0.16550657, 0.00000000, 0.093…
$ blt30_40yrs80occ <dbl> 0.124236256, 0.018156424, 0.004640371, 0.000000000, 0…
$ blt40_yrs80occ <dbl> 0.32993889, 0.08938547, 0.00000000, 0.01016260, 0.022…
$ detach80occ <dbl> 0.9804772, 1.0000000, 0.9718969, 0.8747433, 0.7125457…
$ attach80occ <dbl> 0.008676790, 0.000000000, 0.003903201, 0.000000000, 0…
$ mobile80occ <dbl> 0.01084599, 0.00000000, 0.02419984, 0.12525667, 0.287…
$ occupied80 <dbl> 0.9352227, 0.9715128, 0.9675444, 0.9126365, 0.9475219…
$ lnmeanhs8 <dbl> 10.33014, 10.62553, 10.82902, 10.80235, 10.53244, 10.…
$ lnmdvalhs0 <dbl> 11.21855, 11.29849, 11.40645, 11.66993, 11.35158, 11.…reg2 <- lm(lnmdvalhs0 ~ npl2000, data = allsites)
summary(reg2)
Call:
lm(formula = lnmdvalhs0 ~ npl2000, data = allsites)
Residuals:
Min 1Q Median 3Q Max
-2.50946 -0.40522 -0.02112 0.37935 2.11709
Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 11.719696 0.003052 3840.095 <2e-16 ***
npl2000 -0.021269 0.020159 -1.055 0.291
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Residual standard error: 0.6254 on 42972 degrees of freedom
Multiple R-squared: 2.591e-05, Adjusted R-squared: 2.635e-06
F-statistic: 1.113 on 1 and 42972 DF, p-value: 0.2914Our results suggest that when a Census tract was listed on the NPL, on average, the median house price falls by approximately 2 percent. Although, this coefficient is not statistically significant at any level, which means we likely need to control for something else…
reg3 <- feols(lnmdvalhs0 ~ npl2000, cluster = ~statefips, data=allsites)
list('Heteroskedastic' = reg2,
'Cluster' = reg3) %>% modelsummary(gof_map = 'nobs', output = 'huxtable')| Heteroskedastic | Cluster | |
| (Intercept) | 11.720 | 11.720 |
| (0.003) | (0.092) | |
| npl2000 | -0.021 | -0.021 |
| (0.020) | (0.049) | |
| Num.Obs. | 42974 | 42974 |
Clustering the standard errors increases them from approximately 0.003 to 0.092. Since we have multiple observations in each state that are likely correlated with each other, we should expect clustered standard errors to be different from standard errors.
reg4i <- feols(lnmdvalhs0 ~ npl2000 + lnmeanhs8, cluster = ~statefips, data=allsites)
reg4ii <- feols(lnmdvalhs0 ~ npl2000 + lnmeanhs8 + povrat8 + pop_den8, cluster = ~statefips, data=allsites)
reg4iii <- feols(lnmdvalhs0 ~ npl2000 + lnmeanhs8 + povrat8 + pop_den8| statefips, cluster = ~statefips, data=allsites)
etable(reg3, reg4i, reg4ii, reg4iii) reg3 reg4i reg4ii
Dependent Var.: lnmdvalhs0 lnmdvalhs0 lnmdvalhs0
Constant 11.72*** (0.0924) 2.404*** (0.4442) 2.615*** (0.5827)
npl2000 -0.0213 (0.0488) 0.0400 (0.0247) 0.0972*** (0.0179)
lnmeanhs8 0.8557*** (0.0408) 0.8336*** (0.0529)
povrat8 -0.4645** (0.1483)
pop_den8 1.43e-5*** (1.22e-6)
Fixed-Effects: ----------------- ------------------ --------------------
statefips No No No
_______________ _________________ __________________ ____________________
S.E.: Clustered by: statefips by: statefips by: statefips
Observations 42,974 42,974 42,974
R2 2.59e-5 0.57916 0.62474
Within R2 -- -- --
reg4iii
Dependent Var.: lnmdvalhs0
Constant
npl2000 0.0346 (0.0217)
lnmeanhs8 0.7768*** (0.0412)
povrat8 -0.4046*** (0.1119)
pop_den8 8.4e-6*** (1.44e-6)
Fixed-Effects: -------------------
statefips Yes
_______________ ___________________
S.E.: Clustered by: statefips
Observations 42,974
R2 0.70986
Within R2 0.58936
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1For my economic control variable I used poverty rate in 1980, which I view as a control for socioeconomic status. Furthermore, I used the population density to control for demographics. We see that in all regression done in question 4, the sign of the npl2000 estimate has changed from negative to positive. Where reg4ii, column 3, is statistically significant at all levels. We also see that with state FE, the estimate on npl2000 is no longer statistically significant.
To satisfy the CIA (which, in turn, creates an unbiased estimator) we need the NPL in 2000 to be as good as random conditional on the covariates.
allcovariates <- read_dta('allcovariates.dta')
reg6i <- feols(povrat8 ~ npl2000, data = allcovariates)
reg6ii <- feols(pop_den8 ~ npl2000, data = allcovariates)
reg6iii <- feols(avhhin8 ~ npl2000, data = allcovariates)
reg6iv <- feols(no_hs_diploma8 ~ npl2000, data = allcovariates)
etable(reg6i, reg6ii, reg6iii, reg6iv) reg6i reg6ii reg6iii
Dependent Var.: povrat8 pop_den8 avhhin8
Constant 0.1125*** (0.0005) 5,424.1*** (43.18) 21,510.2*** (39.45)
npl2000 -0.0069* (0.0032) -4,017.1*** (302.2) -1,170.0*** (276.1)
_______________ __________________ ___________________ ___________________
S.E. type IID IID IID
Observations 48,245 48,245 48,245
R2 9.35e-5 0.00365 0.00037
Adj. R2 7.28e-5 0.00363 0.00035
reg6iv
Dependent Var.: no_hs_diploma8
Constant 0.3147*** (0.0008)
npl2000 0.0280*** (0.0054)
_______________ __________________
S.E. type IID
Observations 48,245
R2 0.00056
Adj. R2 0.00054
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1All of our coefficients are statistically significant at some level. Furthermore, all variables with the exception of no high school diploma suggest that it would be less likely that an area would be NPL listed. Meaning that when a higher poverty rate, population density, or average household income are present, a tract is less likely to be NPL listed. Alternatively, when a low high school graduation rate (or high no diploma rate) is present, we see that a tract is more likely to be NPL listed.
sitecovariates <- read_dta('sitecovariates.dta')
reg7povi <- feols(povrat8 ~ npl2000, data = sitecovariates)
reg7povii <- feols(povrat8 ~ I(hrs_82 >28.5), data = sitecovariates)
reg7poviii <- feols(povrat8 ~ I(hrs_82 > 28.5), data = filter(sitecovariates, between(hrs_82, 16.5,40.5)))
etable(reg7povi,reg7povii,reg7poviii) reg7povi reg7povii reg7poviii
Dependent Var.: povrat8 povrat8 povrat8
Constant 0.1169*** (0.0068) 0.1139*** (0.0063) 0.1072*** (0.0093)
npl2000 -0.0168* (0.0083)
I(hrs_82>28.5) -0.0134. (0.0080) 0.0043 (0.0120)
_______________ __________________ __________________ __________________
S.E. type IID IID IID
Observations 487 487 227
R2 0.00842 0.00583 0.00059
Adj. R2 0.00637 0.00378 -0.00385
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1reg7popdeni <- feols(pop_den8 ~ npl2000, data = sitecovariates)
reg7popdenii <- feols(pop_den8 ~ I(hrs_82 >28.5), data = sitecovariates)
reg7popdeniii <- feols(pop_den8 ~ I(hrs_82 > 28.5), data = filter(sitecovariates, between(hrs_82, 16.5,40.5)))
etable(reg7popdeni,reg7popdenii,reg7popdeniii) reg7popdeni reg7popdenii reg7popdeniii
Dependent Var.: pop_den8 pop_den8 pop_den8
Constant 1,770.8*** (205.2) 1,670.2*** (190.2) 1,360.9*** (264.7)
npl2000 -620.1* (248.5)
I(hrs_82>28.5) -512.8* (239.9) -209.9 (340.7)
_______________ __________________ __________________ __________________
S.E. type IID IID IID
Observations 487 487 227
R2 0.01268 0.00933 0.00168
Adj. R2 0.01064 0.00729 -0.00275
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1reg7hhinci <- feols(avhhin8 ~ npl2000, data = sitecovariates)
reg7hhincii <- feols(avhhin8 ~ I(hrs_82 >28.5), data = sitecovariates)
reg7hhinciii <- feols(avhhin8 ~ I(hrs_82 > 28.5), data = filter(sitecovariates, between(hrs_82, 16.5,40.5)))
etable(reg7hhinci,reg7hhincii,reg7hhinciii) reg7hhinci reg7hhincii reg7hhinciii
Dependent Var.: avhhin8 avhhin8 avhhin8
Constant 19,547.5*** (441.5) 19,635.3*** (408.5) 19,812.2*** (576.9)
npl2000 1,265.8* (534.7)
I(hrs_82>28.5) 1,233.5* (515.3) 488.6 (742.5)
_______________ ___________________ ___________________ ___________________
S.E. type IID IID IID
Observations 487 487 227
R2 0.01142 0.01168 0.00192
Adj. R2 0.00939 0.00964 -0.00252
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1reg7nohsi <- feols(no_hs_diploma8 ~ npl2000, data = sitecovariates)
reg7nohsii <- feols(no_hs_diploma8 ~ I(hrs_82 >28.5), data = sitecovariates)
reg7nohsiii <- feols(no_hs_diploma8 ~ I(hrs_82 > 28.5), data = filter(sitecovariates, between(hrs_82, 16.5,40.5)))
etable(reg7nohsi,reg7nohsii,reg7nohsiii) reg7nohsi reg7nohsii reg7nohsiii
Dependent Var.: no_hs_diploma8 no_hs_diploma8 no_hs_diploma8
Constant 0.4175*** (0.0111) 0.4053*** (0.0104) 0.3881*** (0.0144)
npl2000 -0.0755*** (0.0135)
I(hrs_82>28.5) -0.0624*** (0.0131) -0.0348. (0.0185)
_______________ ___________________ ___________________ __________________
S.E. type IID IID IID
Observations 487 487 227
R2 0.06064 0.04468 0.01550
Adj. R2 0.05871 0.04271 0.01112
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1We look at all of the same variables we looked at in question 6, which leads to 12 different regressions.
For poverty rate, we see that the first and second column are almost identical. Although, we notice that with the HRS>28.5 coefficient, it is not statistically significant at any level, while the npl2000 coefficient is significant at the 10% level. When we create a range of HRS, we also see that there is more balance among the covariates in comparison to the first two columns.
For population density we see very similar results for columns 1 and 2. For column 3, we don’t see as much balance as with poverty rate. And we can see that a confidence interval for this would include zero, indicating that the coefficient on the range of HRS is not statistically significant.
For average household income, we see nearly identical results in columns 1 and 2. For column 3 we see a significant dip in the HRS score coefficient and with the standard error, our hypothesis test would consider a negative result for that coefficient. Similar to the poverty rate, we see that there is more balance among covariates in the first two columns, compared to the third column.
For no high school diploma we have the same story as the previous 9 regressions.
Overall, our results suggest that covariates are typically balanced when using NPL prior to 2000, vs not, and on the above and below HRS 28.5. When we considered the range of HRS scores, we lost statistical significance every time, and also saw less balance.
For the IV to be valid, we need the HRS to be exogenous, meaning that the covariance between HRS and the error are zero. We also need the HRS score to be correlated with whether a tract is NPL listed or not. The final assumption we require is monotonicity, meaning that NPL and HRS move in the same direction.
We need HRS to vary smoothly in NPL. And we need NPL to be correlated with HRS.
Because we require exogeneity, the first fact may cause IV to be invalid. Since 28.5 was chosen for manageability reasons, this may suggest that it is actually endogenous, and would fail to meet the assumptions needed. As for the last two facts, one big thing we are concerned with is sorting. Since parties were not aware of the threshold/cutoff, it seems we do not have a sorting issue. So long as the other assumptions are met (which there is no indication that they are not), both of the latter two facts would be a valid IV/RD.
miledata <- read_dta('2miledata.dta')
#First Stage
FirstStage <- feols(npl2000 ~ hrs_82, cluster = miledata$statefips, data = miledata)
#Reduced Form
ReducedForm <- feols(log(mdvalhs0) ~ hrs_82, cluster = miledata$statefips, data = miledata)
#Using the fitted values to estimate in two stages, for fun
miledata1.1 <-miledata %>% mutate(D_hat = FirstStage$fitted.values)
FirstAndReduced <- feols(log(mdvalhs0) ~ D_hat, cluster = miledata1.1$statefips, data = miledata1.1)
#2SLS
TwoSLS <- feols(log(mdvalhs0) ~ 1 | npl2000 ~ hrs_82, cluster= miledata$statefips, data = miledata)
etable(FirstStage, ReducedForm, TwoSLS, FirstAndReduced, digits=6) FirstStage ReducedForm
Dependent Var.: npl2000 log(mdvalhs0)
Constant 0.003736 (0.038960) 11.4298*** (0.107503)
hrs_82 0.020178*** (0.000931) 0.006238*** (0.001724)
npl2000
D_hat
_______________ ______________________ ______________________
S.E.: Clustered by: cluster by: cluster
Observations 483 483
R2 0.55801 0.04640
Adj. R2 0.55709 0.04442
TwoSLS FirstAndReduced
Dependent Var.: log(mdvalhs0) log(mdvalhs0)
Constant 11.4286*** (0.106443) 11.4286*** (0.107732)
hrs_82
npl2000 0.309157*** (0.083740)
D_hat 0.309157*** (0.085426)
_______________ ______________________ ______________________
S.E.: Clustered by: cluster by: cluster
Observations 483 483
R2 0.01276 0.04640
Adj. R2 0.01071 0.04442
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1From using our fitted values from the first stage, we can see that the first stage and reduced form are equal to the 2SLS. (The first and reduced form have different standard errors from the 2SLS, which is expected).
#First Stage
FirstStage12 <- feols(npl2000 ~ hrs_82 | statefips, cluster = miledata$statefips, data = miledata)
#Reduced Form
ReducedForm12 <- feols(log(mdvalhs0) ~ hrs_82 | statefips, cluster = miledata$statefips, data = miledata)
#2SLS
TwoSLS12 <- feols(log(mdvalhs0) ~ 1 | statefips | npl2000 ~ hrs_82, cluster = miledata$statefips, data = miledata)
etable(FirstStage12, ReducedForm12, TwoSLS12, digits=6) FirstStage12 ReducedForm12
Dependent Var.: npl2000 log(mdvalhs0)
hrs_82 0.019981*** (0.001095) 0.003293* (0.001411)
npl2000
Fixed-Effects: ---------------------- --------------------
statefips Yes Yes
_______________ ______________________ ____________________
S.E.: Clustered by: cluster by: cluster
Observations 483 483
R2 0.61462 0.40333
Within R2 0.52153 0.01636
TwoSLS12
Dependent Var.: log(mdvalhs0)
hrs_82
npl2000 0.164785* (0.071730)
Fixed-Effects: --------------------
statefips Yes
_______________ ____________________
S.E.: Clustered by: cluster
Observations 483
R2 0.39709
Within R2 0.00607
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1We see that the reduced form has gone from approximately 0.006 to 0.003 (we also went from statistically significant at all levels, to statistically significant at the 5 percent level), which suggests that we are not controlling for everything, and furthermore, that we may have a weak instrument.
#Q13: Indicator 2SLS
#First Stage
FirstStage13 <- feols(npl2000 ~ I(hrs_82>28.5), cluster = miledata$statefips, data = miledata)
#Reduced Form
ReducedForm13 <- feols(log(mdvalhs0) ~ I(hrs_82>28.5), cluster = miledata$statefips, data = miledata)
#2SLS
TwoSLS13 <- feols(log(mdvalhs0) ~ 1 | npl2000 ~ I(hrs_82>28.5), cluster= miledata$statefips, data = miledata)
etable(FirstStage13, ReducedForm13, TwoSLS13, digits=6) FirstStage13 ReducedForm13
Dependent Var.: npl2000 log(mdvalhs0)
Constant 0.163842*** (0.031526) 11.5279*** (0.090761)
I(hrs_82>28.5) 0.826354*** (0.033228) 0.178662** (0.062024)
npl2000
_______________ ______________________ _____________________
S.E.: Clustered by: cluster by: cluster
Observations 483 483
R2 0.73776 0.03000
Adj. R2 0.73721 0.02799
TwoSLS13
Dependent Var.: log(mdvalhs0)
Constant 11.4925*** (0.097697)
I(hrs_82>28.5)
npl2000 0.216205** (0.073486)
_______________ _____________________
S.E.: Clustered by: cluster
Observations 483
R2 0.02641
Adj. R2 0.02439
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1Both the first stage and reduced form coefficients increased, with the first stage increasing by nearly 600%. Because of this increase, we can conclude that by having the indicator for HRS score, the treatment effect is greater than it previously was.
In this case, we will have a fuzzy RD. This is because when we cross the 28.5 HRS threshold, NPL doesn’t go from 0 to 1.
#Data Set
plot_miledata <- miledata %>%
mutate(run = hrs_82 - 28.5, run_bin = run %/% 1*1) %>%
group_by(run_bin) %>%
transmute(
n = n(),
mean_y = mean((mdvalhs0), na.rm = T),
mean_treatment = mean(npl2000),
mean_hhinc = mean(avhhin8_nbr),
mean_povrate = mean(povrat8_nbr),
mean_popden = mean(pop_den8_nbr),
mean_hsdiploma = mean(no_hs_diploma8_nbr)
) %>% ungroup()
#Outcome with running variable
ggplot(data = plot_miledata, aes(x=run_bin, y = mean_y, group = run_bin<0)) +
geom_vline(xintercept = 0, linewidth=0.2, color = 'blue')+
geom_line()+
geom_smooth()+
geom_point()+
labs(x= "Distance from HRS Threshold", y= "Mean Household Value")+
theme_minimal()`geom_smooth()` using method = 'loess' and formula = 'y ~ x'
#Treatment and running variable
ggplot(data = plot_miledata, aes(x=run_bin, y = mean_treatment, group = run_bin<0)) +
geom_vline(xintercept = 0, linewidth=0.2, color = 'blue')+
geom_line()+
geom_smooth()+
geom_point()+
labs(x= "Distance from HRS Threshold", y= "Ratio of Treated Tracts")+
theme_minimal()`geom_smooth()` using method = 'loess' and formula = 'y ~ x'
#Covariates and Running Variable, 4 separate plots
ggplot(data = plot_miledata, aes(x=run_bin, y = mean_hhinc, group = run_bin<0)) +
geom_vline(xintercept = 0, linewidth=0.2, color = 'blue')+
geom_line()+
geom_smooth()+
geom_point()+
labs(x= "Distance from HRS Threshold", y= "Mean Household Income")+
theme_minimal()`geom_smooth()` using method = 'loess' and formula = 'y ~ x'
ggplot(data = plot_miledata, aes(x=run_bin, y = mean_povrate, group = run_bin<0)) +
geom_vline(xintercept = 0, linewidth=0.2, color = 'blue')+
geom_line()+
geom_smooth()+
geom_point()+
labs(x= "Distance from HRS Threshold", y= "Mean share of Poverty Rate")+
theme_minimal()`geom_smooth()` using method = 'loess' and formula = 'y ~ x'
ggplot(data = plot_miledata, aes(x=run_bin, y = mean_popden, group = run_bin<0)) +
geom_vline(xintercept = 0, linewidth=0.2, color = 'blue')+
geom_line()+
geom_smooth()+
geom_point()+
labs(x= "Distance from HRS Threshold", y= "Mean Population Density")+
theme_minimal()`geom_smooth()` using method = 'loess' and formula = 'y ~ x'
ggplot(data = plot_miledata, aes(x=run_bin, y = mean_hsdiploma, group = run_bin<0)) +
geom_vline(xintercept = 0, linewidth=0.2, color = 'blue')+
geom_line()+
geom_smooth()+
geom_point()+
labs(x= "Distance from HRS Threshold", y= "Mean share without HS Diploma")+
theme_minimal()`geom_smooth()` using method = 'loess' and formula = 'y ~ x'
#Bin Counts and Running Variable
ggplot(data = plot_miledata, aes(x=run_bin, y = n, group = run_bin<0)) +
geom_vline(xintercept = 0, linewidth=0.2, color = 'blue')+
geom_line()+
geom_smooth()+
geom_point()+
labs(x= "Distance from HRS Threshold", y= "Number of Observations")+
theme_minimal()`geom_smooth()` using method = 'loess' and formula = 'y ~ x'
Most of our covariates were pretty smooth across the threshold. Furthermore, with the treated we saw a large discontinuous jump at the threshold. In the mean household value, we saw a large jump in the terms of the numbers, but a relatively small jump in the grand scheme of things, which speaks to the treatment effect on the house value. Finally, the number of observations does jump at the threshold, which is troublesome. Overall, I think this does seem like a plausible RD.
miledata1.2 <- miledata %>% filter(hrs_82 %>% between(16.5,40.5))
#2SLS
FuzzyRD <- feols(log(mdvalhs0) ~ 1 | npl2000 ~ I(hrs_82>28.5), data = miledata1.2)
#State FE
FuzzyRD_StateFE <- feols(log(mdvalhs0) ~ 1 | statefips| npl2000 ~ i(hrs_82>28.5), data = miledata1.2)
etable(FuzzyRD, FuzzyRD_StateFE, digits=6) FuzzyRD FuzzyRD_StateFE
Dependent Var.: log(mdvalhs0) log(mdvalhs0)
Constant 11.5553*** (0.071503)
npl2000 0.056360 (0.091011) 0.092180 (0.072554)
Fixed-Effects: --------------------- -------------------
statefips No Yes
_______________ _____________________ ___________________
S.E. type IID by: statefips
Observations 226 226
R2 -0.00168 0.41230
Within R2 -- -0.00676
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1Without state FE, we estimate that being NPL listed leads to a 5.6% increase in median house value in 2000. With state FE we estimate a 9.2% increase in median house value in 2000 when NPL listed. Neither of these estimates are statistically significant from zero.
I would argue that the model with state FE is more credible. The reason for this is two fold; the first reason being that we see a lower standard error, which hopefully means we are capturing some of the differences between states. The second reason is that we have a much higher R2 value with state fixed effects. Granted, this could be slightly inflated with the inclusion of those fixed effects, but the model does seem to have a better fit than without. One reason for concern would be the low number of observations (which is true for both models).
In this case, we are looking at the Local Average Treatment Effect (LATE). We estimate, as stated earlier, that when NPL is listed, we have a 9.2% increase in median house value.
If I understand the question correctly, I would say yes. When we look at this estimand, we can see whether being an NPL tract area is helpful towards housing prices. Depending on the policy goal, this estimate could be relevant in determining the effect of the policy.