EDA - Corruption and Parking Violations

Cultural norms or legal constraints?

We were hired by the World Bank to take a study the parking behavior data of United Nations diplomats as a way to analyze the effect of cultural norms and legal constraints on controlling corruption. Diplomatic immunity protected UN officials from parking enforcement reprecussions until 2002, when enforcement authorities were given the ability to confiscate diplomatic plates of violators. Therefore, before 2002 behavior was constrained by cultural norms alone, and after 2002 both cultural norms and threat of legal action could have had an impact on behavior.

Data

We have been given a dataset for a selection of UN diplomatic missions, Corrupt.R. The dependent (or target) variable in this data is named violations. We are given the labels of some of the variables that are listed below. We are told that the rest of the variables should be self-explanatory. • corruption: Country corruption index, 1998 • violations: Unpaid New York City parking violations • trade: total trade with the United States (1998 US$)

Objective

The World Bank would like to know what if any relationship there is between corruption and parking violations both pre and post 2002 and if there are any other relevant explanatory variables.

Description of the Data

The data set we are working with includes 364 observations of 28 variables. There are two observations for each unique value in the wbcode and country variable. Within each of these observation pairs there are only 3 variables that differ within the pairs. These are prepost (a binary categorical variable), violations, and fines variables. It is important to note that violations (dependent variable) is almost colinear with fines (as discussed in the following investigation), such that the relevant changes in the dataset between the observation pairs are narrowed to prepost and violations. For the sake of the assigned task, we assume that the prepost variable relates to the pre and post 2002 periods when the enforcement action took place.

Scope of Analysis

Given that only the dependent variable changes (and none of the other variables change with the exception of fines as discussed), our analysis of the effects of the enforcement action on behavior is essentially limited to analyzing (1) changes in the violations variable relative to the key independent variable corruption (in both pre and post enforcement periods), (2) correlations between the dependent variable and other independent variables (in both pre and post enforcement periods), besides corruption, and (3) correlations between corruption and the independent variables themselves (regardless of pre or post period) to identify potential confounding effects.

A full summary of the variables in the dataset are below:

Changes to the Original Dataset

1. Since violations is the target variable, we chose to exclude 66 records with an NA value in the violations column.This brings the total number of records to 298 (i.e. 149 observation pairs). NA values remain in the gov_wage_gdp, pctmuslim, majoritymuslim, trade, cars_total, cars_personal, cars_mission, ecaid, milaid, region, totaid, and distUNplz variables. As these are not the the main variables of investigation, and because R handles NA values in its analysis pretty well, these records will remain in the dataset. However, it is worth noting special attention is required for the gov_wage_gdp variable, which has 114 NAs after the NAs are removed from violations.

#1. 
df <- FMcorrupt[complete.cases(FMcorrupt[, 3]), ]

Table: NAs in Changed Dataset

# returns a single vector with violations = 0 NAs and then counts the NAs in the other variables.
na_count <-sapply(FMcorrupt[!is.na(FMcorrupt$violations),], function(y) sum(length(which(is.na(y)))))
na_count <- data.frame(na_count)
na_count <- data.frame(cbind(row.names(na_count),na_count$na_count))
na_count <- na_count[na_count$X2 != 0,] 
colnames(na_count) <- c("Variable","NAs")
rownames(na_count) <- NULL
na_count

##          Variable NAs
## 1    gov_wage_gdp 114
## 2       pctmuslim   4
## 3  majoritymuslim   4
## 4           trade   4
## 5      cars_total  20
## 6   cars_personal  20
## 7    cars_mission  20
## 8           ecaid   4
## 9          milaid   4
## 10         region   2
## 11         totaid   4
## 12      distUNplz   6

2. Looking at the data, the majoritymuslim values that are -1 line up with a pctmuslim value of 0. Therefore, we assigned those values a 0 instead of a -1.

test1 <- df[df$majoritymuslim < 0, c(9,10)] 
df$majoritymuslim[df$majoritymuslim == -1] <- 0

3. Records with no violation values included those that pertained to nominal economic groupings as coded by World Bank. For example, observation #337 in wbcode = ‘WLD’ which is “World” according to World Bank coding. As a result of ommiting observations that a did not carry violation values (as they have no diplomatic representative standing as countries), these economic groupings were omitted from the analysis.

#1. 
df <- FMcorrupt[complete.cases(FMcorrupt[, 3]), ]    # Moved up under 1. as a suggestion
summary(df)
#2. 
test1 <- df[df$majoritymuslim < 0, c(9,10)]      # moved up under 2. as a suggestion
df$majoritymuslim[df$majoritymuslim == -1] <- 0
#3. 
test2 <- df[df$pop1998 ==max(df$pop1998), ]  # left as is under #3

4. Besides cleaning up records we didn’t want to include in our analysis/felt needed adjustment, we also wanted to restructure the table to have one row for each country, and have individual columns to represent the pre 2002 and post 2002 fines and violations. We also wanted to add a new column for the difference in violations and fines, more easily calculated in this new structure. We used the merge function to ensure that the post data was reassigned to the correct wbcode.

#4.
df_pre <- df[df$prepost == "pre", ]
df_post <- df[df$prepost == "pos", c(1,3,4)]
colnames(df_pre)[3:4] <- c("pre_violations", "pre_fines")
colnames(df_post) <- c("wbcode", "post_violations", "post_fines")

FMcorrupt_new <- merge(df_pre, df_post, by = "wbcode")

FMcorrupt_new$dif_violations = FMcorrupt_new$pre_violations-FMcorrupt_new$post_violations
FMcorrupt_new$dif_fines = FMcorrupt_new$pre_fines-FMcorrupt_new$post_fines
FMcorrupt_new = FMcorrupt_new[, c(1,20,3,4,29,30,31,32,5:19,21:28)]          #dropped prepost here, put ordered corruption 2nd and reordered pre/post violations to follow

Univariate Analysis

A simple plot of the total amount of fines against the number of violations shows a fairly strong linear relationship, suggesting that either parameter could be chosen as the dependent variable in the analysis. However, if the average fine is calculated, it varies by country and ranges from roughly $48 to $56. While this difference doesn’t appear to be particularly large, the amount of an individual fine could vary by severity of the offense or change over time, and so for the purposes of this analysis we will choose to focus on the number of violations as the dependent variable.

# plot(FMcorrupt_new$pre_violations,FMcorrupt_new$pre_fines,xlab = 'Number of violations',
#     ylab = 'Total fines', main = 'Relationship between fines and violations')
b = FMcorrupt_new$pre_fines/FMcorrupt_new$pre_violations
min(b, na.rm = T)
max(b, na.rm = T)

Our research question is focused on the relationship between violations and corruption, so these are clearly key variables. Summary statistics for the number of violations in the pre 2002 period are:

summary(FMcorrupt_new$pre_violations)

##    Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
##    0.00   17.22   51.65  198.07  189.59 3392.96

There is a positive skew, with a few countries whose number of violations is much higher than the average. This is skew is visible before and after 2002.

par(mfrow = c(1,2))
hist(FMcorrupt_new$pre_violations, breaks = "FD", xlab = "Violations", main = "Histogram of violations pre 2002")

hist(FMcorrupt_new$post_violations, breaks = "FD", xlab = "Violations", main = "Histogram of violations post 2002")

Compared to pre 2002, there is a significant decrease in violations overall post 2002. When examining the change in violations, we use the convention that a positive number indicates an overall decrease in the number of violations post 2002. It was noted that 10 countries experienced an increase in violations after 2002, though very modest. For example, Azerbaijan, which showed the largest increase in violations, had no violations before 2002 and only 5 afterwards. It is worth noting that countries that had an increase in violations post 2002, typically had a low number of pre 2002 violations to begin with.

summary(FMcorrupt_new$dif_violations)

##     Min.  1st Qu.   Median     Mean  3rd Qu.     Max. 
##   -4.906   14.481   51.324  194.383  185.588 3385.111

FMcorrupt_new[FMcorrupt_new$dif_violations < 0, c("country","pre_violations","post_violations")]

##          country pre_violations post_violations
## 6      AUSTRALIA      0.0000000       0.3270609
## 8     AZERBAIJAN      0.0000000       4.9059138
## 12  BURKINA-FASO      0.0000000       0.9811828
## 37       DENMARK      0.0000000       0.3270609
## 56        GREECE      0.0000000       2.2894266
## 65       IRELAND      0.0000000       0.6541219
## 67        ISRAEL      0.0000000       1.3082438
## 71         JAPAN      0.0000000       0.6541219
## 102  NETHERLANDS      0.4051054       1.6353047
## 106         OMAN      0.0000000       1.3082438

If we examine “corruption”, we find it is left skewed as per its histogram:

hist(FMcorrupt_new$corruption, breaks = "FD", xlab = "Corruption Index", main = "Histogram of Corruption Index")

summary(FMcorrupt_new$corruption)

##     Min.  1st Qu.   Median     Mean  3rd Qu.     Max. 
## -2.58299 -0.41515  0.32696  0.01364  0.72025  1.58281

As an index value we cannot assume that the corruption index criteria yields meets the requirements for an iid variable. In order to get a closer approximation of the true relationship between violations and corruption, further transformation of the corruption variable is likely warranted. However, any such transformation warrant domain knowledge of the methodology for deriving the corruption index.

unique(FMcorrupt_new$region)

## [1]  6  3  7  2  4  5  1 NA

FMcorrupt_new[FMcorrupt_new$region == 7, 'country']

##  [1] ""             "BAHRAIN"      "CYPRUS"       "EGYPT"       
##  [5] "IRAN"         "ISRAEL"       "JORDAN"       "KUWAIT"      
##  [9] "LEBANON"      "OMAN"         "PAKISTAN"     "SAUDI ARABIA"
## [13] "SYRIA"        ""             "YEMEN"        NA

options(scipen = 3)
boxplot(milaid~region, data = FMcorrupt_new, main = 'Milaid by Region', xlab = 'Region', ylab = 'milaid (log scale)', ylim = c(0.01,10000), log = "y")

Analysis of Key Relationships

In this section we will be looking for any potential relationships between parking violations and corruption. Part I looks at the changes in the violations variable relative to the key independent variable corruption (in both pre and post enforcement periods). Part II examines correlations between the dependent variable and other independent variables (in both pre and post enforcement periods), besides corruption. Part III looks at correlations between corruption and the independent variables themselves (regardless of pre or post period) to identify potential confounding effects.

First, we start from the general, overall correlation between each of the variables using corrplot package in R:

Chart: Overall Correlation between Variables

M <- cor(FMcorrupt_new[,c(2:8,10:29,31)], use = "complete.obs")
corrplot(M, order = "hclust", method = "square", tl.col ='black')

Several key features that emerge and analyzed in depth below. These are summarized as the following: * Violations are only mildly correlated with corruption overall, * Violations are more strongly correlated with aid (i.e. totaid, ecaid), particularly in the pre 2002 period, * Violations are modestly correlated with other variables (i.e. region, r_middleeast, pctmuslim) in both pre post periods, * Corruption is strongly correlated with gdppcus1998 (i.e gdp per capita), * Corruption is strongly correlated with variables that are modestly correlated with violations.

scatterplotMatrix( ~ corruption + pre_violations + post_violations, data = FMcorrupt_new, diagonal = 'histogram')

par(mfrow = c(1,2))
binned_corruption = seq(min(FMcorrupt_new$corruption), max(FMcorrupt_new$corruption), length.out = 11)
sliced = cut(FMcorrupt_new$corruption, breaks = binned_corruption, labels = round(binned_corruption[1:10] + ((max(binned_corruption) - min(binned_corruption))/20), digits = 2))
plot(sliced, FMcorrupt_new$pre_violations, main = 'Violations\n by Corruption (pre 2002)', xlab = 'Corruption (bins = 10)', ylab = 'Number of Parking Violations')
plot(sliced, FMcorrupt_new$post_violations, main = 'Violations\n by Corruption (post 2002)', xlab = 'Corruption (bins = 10)', ylab = 'Number of Violations' )

violations_means_pre = by(FMcorrupt_new$pre_violations, sliced, mean,  na.rm = T)
violations_means_post = by(FMcorrupt_new$post_violations, sliced, mean, na.rm = T)

par(mfrow = c(1,2))
plot(y = as.single(violations_means_pre), x = sort(unique(sliced)), type = "p", main = 'Mean violations\n by corruption pre 2002', xlab = 'Corruption', ylab = 'Mean Violations')
plot(y = as.single(violations_means_post), x = sort(unique(sliced)), type = "p", main = 'Mean violations\n by corruption post 2002', xlab = 'Corruption (bins = 10)', ylab = 'Mean Violations')

par(mfrow = c(1,2))
violations_median_pre = by(FMcorrupt_new$pre_violations, sliced, median, na.rm = T)
violations_median_post = by(FMcorrupt_new$post_violations, sliced, median, na.rm = T)

plot(y = as.single(violations_median_pre), x = sort(unique(sliced)), type = "p", main = 'Median violations \nby corruption pre 2002', xlab = 'Corruption', ylab = 'Median Parking Violations')
plot(y = as.single(violations_median_post), x = sort(unique(sliced)), type = "p", main = 'Median violations \nby corruption post 2002', xlab = 'Corruption', ylab = 'Median parking Violations')

vcorr<- data.frame(cor(FMcorrupt_new[,c(2:8,10:29,31)], use = 'complete.obs')[c(2,4),])
vcorr <- data.frame(t(vcorr))
vcorr <- vcorr[order(-abs(vcorr$pre_violations)),][c(5,6,9:28),]
kable(vcorr)

scatterplotMatrix( ~ pre_violations + post_violations + milaid + totaid + ecaid + region + r_middleeast + pctmuslim + majoritymuslim, data = FMcorrupt_new, diagonal = 'histogram', smoother = F)

par(mfrow = c(1,2))
mjr_mus = subset(FMcorrupt_new, majoritymuslim = 1)
min_mus = subset(FMcorrupt_new, majoritymuslim < 1)
boxplot(min_mus$pre_violations, mjr_mus$pre_violations, names = c("Minority", "Majority"), main = "Violations by Muslim Minority\n and Majority Countries (pre 2002)", ylab ='Number of Violations')
boxplot(min_mus$post_violations, mjr_mus$post_violations, names = c("Minority", "Majority"), main = "Violations by Muslim Minority\n and Majority Countries (pos 2002)", ylab ='Number of Violations')

crptcorr<- data.frame(cor(FMcorrupt_new[,c(2:8,10:29,31)], use = 'complete.obs')[1,])
colnames(crptcorr) <- "corr"
crptcorr <- data.frame(row.names(crptcorr), crptcorr$corr)
colnames(crptcorr) <- c("Variable", "Corr")
crptcorr <- crptcorr[2:28,]
crptcorr <- crptcorr[order(-abs(crptcorr$Corr)),]
kable(crptcorr, row.names = F)

scatterplotMatrix( ~ corruption + gdppcus1998 + r_europe + trade + r_africa +  region + pctmuslim, data = FMcorrupt_new, diagonal = 'histogram', smoother = F)

scatterplot(jitter(FMcorrupt_new$corruption,2),jitter(FMcorrupt_new$gdppcus1998,2), pch = 17, cex = 1,
            lty = "solid", ylab = 'GDP Per Capita (USD)', xlab = 'Corruption Index',
            main = 'Correlation between GDP per Capita and Corruption', smoother = FALSE,
            boxplot = FALSE)
text(FMcorrupt_new$corruption,FMcorrupt_new$gdppcus1998, labels=FMcorrupt_new$wbcode, cex= 0.7, lwd = 3, pos = 4)

Confounding Variables

There are two types of confounding variables we are concerned with. The first type in our dataset, and the second type are those that are not in the dataset, but those that we would want to consider if we had the data. To address the first type, we saw that the relationship between violations and corruption was relatively modest. Instead, we observed that violations variable (pre 2002) is most strongly correlated with aid variables (i.e. milaid, totaid, ecaid), and that corruption was most strongly correlated with gdppcus1998. We also saw that aid variables were weakly associated with corruption and gdppcus1998 was weakly associated with violations. But we have to ask ourselves, is it possible that an other variable in our data has a significant effect on both aid and gdppcus1998 at the same time? However, a quick glance at Chart: Overall Correlation between Variables suggests that the relationship between milaid and gdppcus1998 is relatively weak. We confirm this by calculation, cor = 0.1030945.

Among the variables that are not in our dataset, other confounding variables could possibly include: * Diplomatic staff. Since violations are committed by individual staff it makes sense to have a stronger understanding about staff as the main subject of observation. It is plausible that staff pick up certain behaviors while abroad, rather than from their home countries. We would want data for example on how long a staff person has they been in NY. We’d want to ask how long have the staff been abroad? How long have the staff been in other countries with low or high corruption scores? How much are staff concerned with the implications of parking violations via their home offices? What is the rank of staff? A more granual picture would also serve to strengthen the unit of observation which is currently spread over countries, diplomatic missions, and diplomatic staff. * Parking enforcement - While the corruption variable may be an accurate measure for abuse of power by public official, we do not know what if the index directly relates to the actual case of parking violations. It could be that in countries with high levels of corruption, that enforcement of parking violations is high as in the case of security states. Conversely, countries with low levels of corruption may have lax enforcement of parking regulations whereas they may have stricter regulations of things like grand corruption. Therefore, it could be the case that the corruption effect on violations is actually lower in our dataset than it might be otherwise. Therefore, data parking enforcement would be important to have to ensure it was not a confounding variable. * Exchange relationship with U.S. - As it is possible that treatment of U.S. laws is a reflection of respect or disrespect for the U.S., it would be important to understand the type of relationship that the U.S. has. For example, does the U.S. have a mutually beneficial trade relationship with the other country? What is the balance of trade in terms of types of goods and services? Is the relationship primarily defined by trade or security? What is the number of diaspora per capita by country residing in the U.S. and vice versa?

Conclusion

Our initial task was to look at the relationship between corruption and parking violations of UN diplomats, but the analysis we have done shows nothing clear or strong. There is clearly some relationship, but it is more complicated than a linear correlation. The increase in correlation after the legal enforcement began seem to indicate that for some countries the cultural norms as operationally constructed by the corruption index could in fact be stronger than threat of legal penalties. However, analysis of some of our compounding variables, such as percent of the country that is Muslim and region, showed to be even stronger relationships to both the primary and independent variables. Also, several other factors in the data set showed skews and outliers that could be explored further. A larger multivariate analysis is needed before any concrete conclusions can be drawn about the impact of cultural norms vs. legal penalties as pertaining to UN parking tickets.