The rate of drug-related deaths in Connecticut is currently higher than the national average and has continued to steadily increase. While there have been a significant number of initiatives implemented to reduce drug misuse in recent years, it is important to understand the demographic and geographic attributes underlying this public health issue to effectively reduce drug-related deaths in vulnerable populations.
To address our initial research questions, we were interested in specific variables within the Accidental Drug Related Deaths dataset after tidying and cleaning. Variables used included attributes and characteristics of individuals at the time of death, details of the location of death, and type of drug used before death.
Below is a list of variables used in alphabetical order:
age
: Age of individual at death
death_city_long
and death_city_lat
: Location of death by longitude and latitude
drug
: Type of drug which includes heroin, fentanyl, cocaine, ethanol, benzodiazepine, and other
location
: Location of death categorized by location type including nursing home, hospice, or convalescent home and residence, and hospital, and others.
month
and year
: Details of the date of death
race
: Race of individual which include White, Black, Hispanic White, Asian, Hispanic Black, and other
sex
: Sex of individual which include male, female, or unknown
The spider plot above shows the different components of drug use in different race groups. We can see: * Hispanic-Black has a higher proportion of Fentanyl-related death, as compared to other race groups. * Both Asian and Hispanic-White have a high proportion of Heroin-related death, as compared to other race groups. * Black and Other race groups have a high proportion of Cocaine-related death, as compared to other race groups. * White has a high proportion of Heroin-related and Other-drug-related death, as compared to other race groups.
When examining the distribution of deaths at different location types using map visualization, the large majority of death counts are found within personal residences and locations labeled as “other”. Other locations range from public spaces such as restaurants, hotels, and outside land plots, as well as residences of parents or friends. Across all location types, the highest distribution of deaths is found in Hartford, CT. The highest death count, 607 deaths, is found among residences in Hartford, CT.
In addition, the highest distribution of residence deaths are found along major highways near or within the following additional cities: New Haven, Waterbury, Norwich, and Bridgeport. Similarly, deaths at other locations are mainly concentrated within Hartford, New Haven, Waterbury, Norwich, and Bridgeport. The highest number of deaths at other locations are found near Hartford with a death count of 238 and near New Haven with a death count of 195. For the distribution of deaths at hospice, nursing homes, and convalescent homes, only five death counts were recorded. Two of these deaths were in Hartford, one death in Waterbury, one death in Wallingford, and one death in New Haven. Lastly, 34 deaths were reported at hospitals in CT. Single death counts at hospitals are distributed across the state, with the highest death count of 4 deaths occurring at hospitals near Hartford, CT.
# specify predictor variable types and set reference categories by frequency
drug_death_bin =
drug_death_bin %>%
mutate(
age = as.numeric(age),
sex = fct_infreq(sex),
race = fct_infreq(race),
new_county = fct_infreq(new_county)
)
# logistic regression model
model=
glm(death_location ~ age + sex + race + new_county, data = drug_death_bin, family = "binomial")
# all ORs and betas from the model
model %>%
broom::tidy() %>%
mutate(OR = exp(estimate)) %>%
select(term, log_OR = estimate, OR, p.value) %>%
knitr::kable(digits = 3)
term | log_OR | OR | p.value |
---|---|---|---|
(Intercept) | 0.954 | 2.595 | 0.000 |
age | 0.006 | 1.006 | 0.049 |
sexMale | -0.188 | 0.829 | 0.028 |
sexUnknown | 10.217 | 27351.427 | 0.959 |
raceAsian | -0.600 | 0.549 | 0.000 |
raceBlack | -0.655 | 0.519 | 0.000 |
raceWhite | -0.419 | 0.658 | 0.286 |
raceHispanic, White | -0.304 | 0.738 | 0.548 |
raceOther | -0.464 | 0.629 | 0.529 |
new_countymiddlesex | -0.312 | 0.732 | 0.002 |
new_countylitchfield | -0.404 | 0.668 | 0.000 |
new_countynew haven | -0.217 | 0.805 | 0.133 |
new_countytolland | -0.323 | 0.724 | 0.058 |
new_countynew london | 0.167 | 1.182 | 0.412 |
new_countyout of state | 0.151 | 1.163 | 0.497 |
new_countyfairfield | 0.034 | 1.035 | 0.882 |
new_countywindham | -0.476 | 0.621 | 0.026 |
In order to understand the likelihood of a drug related death occuring in the hospital versus not in the hospital, we binarized the variable death_location
as the outcome (1 = In the hospital, 0 = Not in the hospital). For this logistic regression analysis we used a generlized logistic model with the predictors age, sex, race and county of residence. Age
was treated as a continuous variable. Sex
was categorical (Male, Female, and Unknown) with Male as the reference category based on frequency. Race
was categorical with White, Black, Hispanic White, Asian, Other, Hispanic Black; White was the reference category based on frequency and NAs were dropped pre-analysis (n= 16). Finally the 8 Counties
of CT were converted to factor variables with all other counties being converted to “out of state” with Litchfield Country as the reference variable based on frequency, NAs were dropped (n= 7).
The significant variables in the model (p-value < 0.05) were age, Female, Hartford and New London Counties and the “Out of State” category.The odds of dying in the hospital from an overdose decrease by 0.05 for every 1-year increase in age. The odds of dying in the hospital as a female are 1.2 times the odds for men controlling for other covariates. Based on location the odds of dying in the hospital are 1.5 in Hartford, 1.8 in New London and 1.7 for out of state residentess compared to Litchfield county and accounting for other covariates.
Through this assesment of accidental drug-related deaths within the state of Connecticut, we highlighted multiple points of interest that could then be utilized for potential intervention or additional analyses. Significant findings to further consider as research questions include but are not limited to: