library(readxl)
library(dplyr)
library(tidyr)
library(ggplot2)
library(haven)
library(nnet)
library(clarify)
library(texreg) DATA <- read_xlsx("hate_crime copy 2.xlsx")
print(colnames(DATA))## [1] "data_year" "pug_agency_name" "state_abbr"
## [4] "state_name" "incident_date" "offender_race"
## [7] "offender_ethnicity" "offense_name" "bias_desc"colnames(DATA) <- c("Year", "Pug Agency Name", "State", "State Name", "Incident Date",
"Offender Race", "Offender Ethnicity", "Offense", "Reason/Bias")
DATA$Year <- as.numeric(DATA$Year)
violent_crimes <- c("Aggravated Assault", "Murder and Nonnegligent Manslaughter",
"Rape", "Robbery")
DATA <- DATA %>%
mutate(violent_crime = ifelse(grepl(paste(violent_crimes, collapse = "|"), Offense), 1, 0),
`Offender Race` = as.factor(`Offender Race`),
`Offender Ethnicity` = as.factor(`Offender Ethnicity`))
model_1 <- glm(violent_crime ~ `Year`, family = binomial(link = "logit"), data = DATA)
model_2 <- glm(violent_crime ~ `Year` + `Offender Race`, family = binomial(link = "logit"), data = DATA)
model_3 <- glm(violent_crime ~ `Year` + `Offender Race` + `Offender Ethnicity`, family = binomial(link = "logit"), data = DATA)model_comparison <- data.frame(
Model = c("Model 1", "Model 2", "Model 3"),
AIC = c(AIC(model_1), AIC(model_2), AIC(model_3)),
BIC = c(BIC(model_1), BIC(model_2), BIC(model_3)),
LogLikelihood = c(logLik(model_1), logLik(model_2), logLik(model_3))
)
print(model_comparison)## Model AIC BIC LogLikelihood
## 1 Model 1 199453.1 199473.9 -99724.53
## 2 Model 2 179915.4 180009.4 -89948.68
## 3 Model 3 179412.7 179548.5 -89693.34anova(model_1, model_2, test = "Chisq")## Analysis of Deviance Table
##
## Model 1: violent_crime ~ Year
## Model 2: violent_crime ~ Year + `Offender Race`
## Resid. Df Resid. Dev Df Deviance Pr(>Chi)
## 1 253774 199449
## 2 253767 179897 7 19552 < 2.2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1anova(model_2, model_3, test = "Chisq")## Analysis of Deviance Table
##
## Model 1: violent_crime ~ Year + `Offender Race`
## Model 2: violent_crime ~ Year + `Offender Race` + `Offender Ethnicity`
## Resid. Df Resid. Dev Df Deviance Pr(>Chi)
## 1 253767 179897
## 2 253763 179387 4 510.69 < 2.2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1set.seed(123)
sim_model_3 <- sim(model_3, n = 1000, vcov = "robust")
# Check simulation object class
print(class(sim_model_3)) ## [1] "clarify_sim"# Compute Average Marginal Effects (AMEs)
variables <- c("Year", "Offender Race", "Offender Ethnicity")
ame_results <- lapply(variables, function(v) {
sim_ame(sim_model_3, var = v)
})
names(ame_results) <- variables
print(ame_results)## $Year
## A `clarify_est` object (from `sim_ame()`)
## - Average marginal effect of `Year`
## - 1000 simulated values
## - 1 quantity estimated:
## E[dY/d(Year)] -0.00145698
##
## $`Offender Race`
## A `clarify_est` object (from `sim_ame()`)
## - Average adjusted predictions for `Offender Race`
## - 1000 simulated values
## - 8 quantities estimated:
## E[Y(American Indian or Alaska Native)] 0.22218936
## E[Y(Asian)] 0.20205824
## E[Y(Black or African American)] 0.27298622
## E[Y(Multiple)] 0.26860282
## E[Y(Native Hawaiian or Other Pacific Islander)] 0.19032922
## E[Y(Not Specified)] 0.05842007
## # ... and 2 more
##
## $`Offender Ethnicity`
## A `clarify_est` object (from `sim_ame()`)
## - Average adjusted predictions for `Offender Ethnicity`
## - 1000 simulated values
## - 5 quantities estimated:
## E[Y(Hispanic or Latino)] 0.2550358
## E[Y(Multiple)] 0.1829382
## E[Y(Not Hispanic or Latino)] 0.1398939
## E[Y(Not Specified)] 0.1291668
## E[Y(Unknown)] 0.1421222print(levels(DATA$`Offender Race`))## [1] "American Indian or Alaska Native"
## [2] "Asian"
## [3] "Black or African American"
## [4] "Multiple"
## [5] "Native Hawaiian or Other Pacific Islander"
## [6] "Not Specified"
## [7] "Unknown"
## [8] "White"print(levels(DATA$`Offender Ethnicity`))## [1] "Hispanic or Latino" "Multiple" "Not Hispanic or Latino"
## [4] "Not Specified" "Unknown"# Set reference values for prediction
set_values <- sim_setx(sim_model_3, x = list(
Year = mean(DATA$Year, na.rm = TRUE),
`Offender Race` = "White",
`Offender Ethnicity` = "Not Hispanic or Latino"
))predictions <- sim_apply(sim_model_3, FUN = predict, setx = set_values)
print(predictions)## A `clarify_est` object (from `sim_apply()`)
## - 1000 simulated values
## - 253776 quantities estimated:
## 1 -0.7904012
## 2 -0.7904012
## 3 -0.7904012
## 4 -0.7904012
## 5 -1.3124653
## 6 -1.3124653
## # ... and 253770 more# Compute first differences
fd_results <- lapply(variables, function(v) {
sim_ame(sim_model_3, var = v, contrast = "difference")
})
names(fd_results) <- variables
print(fd_results)## $Year
## A `clarify_est` object (from `sim_ame()`)
## - Average marginal effect of `Year`
## - 1000 simulated values
## - 1 quantity estimated:
## E[dY/d(Year)] -0.00145698
##
## $`Offender Race`
## A `clarify_est` object (from `sim_ame()`)
## - Average adjusted predictions for `Offender Race`
## - 1000 simulated values
## - 8 quantities estimated:
## E[Y(American Indian or Alaska Native)] 0.22218936
## E[Y(Asian)] 0.20205824
## E[Y(Black or African American)] 0.27298622
## E[Y(Multiple)] 0.26860282
## E[Y(Native Hawaiian or Other Pacific Islander)] 0.19032922
## E[Y(Not Specified)] 0.05842007
## # ... and 2 more
##
## $`Offender Ethnicity`
## A `clarify_est` object (from `sim_ame()`)
## - Average adjusted predictions for `Offender Ethnicity`
## - 1000 simulated values
## - 5 quantities estimated:
## E[Y(Hispanic or Latino)] 0.2550358
## E[Y(Multiple)] 0.1829382
## E[Y(Not Hispanic or Latino)] 0.1398939
## E[Y(Not Specified)] 0.1291668
## E[Y(Unknown)] 0.1421222In this analysis, using the FBI Hate
Crime Dataset, I examined factors that influence whether a hate
crime is violent or non-violent by using logistic regression models. The
dependent variable, violent_crime, was coded as 1 for
violent crimes (aggravated assault, murder and nonnegligent
manslaughter, rape, and robbery) and 0 for non-violent crimes. Each
model added more variables to see how well they explained the likelihood
of a hate crime being violent.
The first model included only the year of the incident, while the second model added the offender’s race. The third model expanded on this by including offender ethnicity. To compare the models, I used the Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC), where lower values indicate better model fit. The results showed that adding race significantly improved the model, with the AIC dropping from 199,453.06 in Model 1 to 179,915.36 in Model 2. A likelihood ratio test confirmed that this change was statistically significant. However, adding offender ethnicity in Model 3 only slightly improved the model, reducing the AIC to 179,412.68. The likelihood ratio test showed this change was also statistically significant, but the small improvement suggests that offender ethnicity contributes less than race in predicting violent hate crimes. Since Model 2 provided the biggest improvement while Model 3 added little additional value, I determined that Model 2 best represents the data.
Using this data, after determining that Model 2 best represents my
findings, I used the clarify package to gain a deeper
understanding of how offender characteristics impact the likelihood of a
hate crime being violent. The clarify package allowed me to
compute average marginal effects (AMEs) and predicted probabilities for
offender characteristics, to help interpret the practical significance
of these factors beyond just statistical fit. The average marginal
effect of year was estimated at -0.00146, meaning that each additional
year is associated with a very slight decrease in the probability of a
violent hate crime. While statistically significant, the effect size is
very small, suggesting that time alone is not a strong predictor of
violent hate crimes. Looking at offender race, the model predicted
different probabilities of a hate crime being violent depending on the
racial category. Offenders identified as Black or African American had
the highest estimated probability at 0.273, followed closely by those
categorized as multiple races at 0.269. White offenders had a lower
estimated probability, while the lowest probability was found among
cases where the race of the offender was “Not Specified” at 0.058, which
may reflect missing or incomplete data. These findings do not
necessarily indicate that certain racial groups commit more violent hate
crimes but could instead reflect disparities in crime reporting, law
enforcement practices, and data collection methods. Systematic biases in
policing, legal definitions of hate crime severity, or broader social
factors may influence these numbers. A similar pattern emerged when
examining offender ethnicity. The results showed that offenders
categorized as Hispanic or Latino had the highest estimated probability
of committing a violent hate crime at 0.255, while those categorized as
Not Hispanic or Latino had a lower probability at 0.140. Cases labeled
“Not Specified” or “Unknown” had even lower probabilities. Like race,
these findings may be influenced by factors beyond offender behavior,
such as differences in reporting, law enforcement classification
methods, and social dynamics surrounding crime categorization.
These findings suggest that offender race has a strong impact on predicting whether a hate crime is violent, while offender ethnicity has a smaller but still significant effect. The impact of year is minimal, suggesting that other factors, such as social dynamics, law enforcement focus, and crime classification policies, play a larger role. Given the possibility of reporting and classification biases, these results should be interpreted with caution, as they do not necessarily reflect inherent differences in criminal behavior across racial or ethnic groups. Instead, they may be shaped by structural factors that influence how hate crimes are reported, investigated, and recorded. This analysis highlights how different offender characteristics correlate with the likelihood of a hate crime being violent. While statistical models suggest that race is a key predictor, it is essential to consider how systematic biases in crime data and law enforcement practices may shape these results. Further research is needed to explore the social and systemic factors driving these differences and to account for potential biases in how hate crimes are reported and recorded.