US homicide rates increase when resources are scarce and unequally distributed

As homicide rates spike across the United States, researchers nominate diverse causes such as temperature, city greenness, structural racism, inequality, poverty and more. While variation in homicide rates clearly results from multiple causes, many correlation studies lack the systematic theory needed to identify the underlying factors that structure individual motivations. Building on pioneering work in evolutionary human sciences, we propose that when resources are unequally distributed, individuals may have incentives to undertake high-risk activities, including lethal violence, in order to secure material and social capital. Here we evaluate this theory by analysing federal data on homicide rates, poverty and income inequality across all 50 US states for the years 1990, 2000 and 2005-2020. Supporting predictions derived from evolutionary social sciences, we find that the interaction of poverty (scarcity) and inequality (unequal distribution) best explains variation in US homicide rates. Results suggest that the increase in homicide rates during the height of the COVID-19 pandemic are driven in part by these same underlying causes that structure homicide rates across the US over the last 30 years. We suggest that these results provide compelling evidence to expand strategies for reducing homicide rates by dismantling structures that generate and concentrate sustained poverty and economic inequality.


Introduction
This document is the supporting information (SI) appendix for the manuscript "Homicide rates in the United States are highest where resources are scarce and unequally distributed" by Weston C. McCool and Brian F. Codding.The document includes additional methods text and all the code required to reproduce the results and figures using the case data.All analyses are run in the R Environment for Statistical Computing (1).

Data
Homicide rate data come from the FBI's Uniform Crime Reporting Program, Crime Data Explorer, Expanded Homicide Data: https://crime-dataexplorer.app.cloud.gov/pages/explorer/crime/shr The proportion of people estimated to be under the poverty level data come from the American Community Survey (ACS, https://www.census.gov/programs-surveys/acs/)variable B17001 "Poverty Status in the Past 12 Months by Sex by Age" of the Census Bureau.
The Gini coefficient is derived from the Lorenz curve of household income within a state to measure income distribution from 0-1.The coefficient for each of the fifty states was calculated from data in the United States census population.As reported in the main text, the response variable is highly skewed non-integer.We fit models specifying a Poisson family with quasi-likelihood estimation ("quasipoisson").## [1] "AK" "ME" "NH" "NM" How many states showed an increase in poverty from 2019 to 2020?

Multimodel comparison
Following theory, we propose that absolute and relative income should influence decisions that may lead to homicide.The unit of analysis is the state-year.Here we construct seven increasingly complex models (below) using generalized additive models from the mgcv library (2-6) which allows us to fit the predictor variables as linear parametric terms while also including random effects and non-linear terms.All models include year as a factorlevel random intercept.This is because we have a "random" sample of years for which data are reported, out of the population of all years of potentially available information on homicide rates, inequality, and poverty.While we also have repeated observations per state, including both year and state as factor level random effects would essentially account for all variation as the unique combination of each factor (state and year).As such, our data retain some non-independence, however, this is unavoidable.Our attempt to partially account for this non-independence by fitting the trend for each year with random intercepts, and subsequently with random slopes and non-linear smooths.We include smooths because the response may be non-linear as pay-offs to risky behavior should plateau at some threshold.Given the nature and distribution of the response variable, we specify a Poisson family and log link with quasi-likelihood estimation to relax assumptions and reduce potential overdispersion.Negative binomial models would produce comparable results.After fitting each model, we compare them in order of complexity using approximate hypothesis tests with the anova.gamfunction.After selecting the "best" model, we run diagnostics, including assessing standardized model residuals by evaluating their distribution, checking for overdispersion, assessing temporal autocorrelation averaged by year, and comparing residuals by state.We then evaluate the results by examining model coefficients and plotting the partial response of predicted homicide rates as a function of poverty and inequality for the years 1990, 2000, 2010, and 2020.
Models: As reported in the main text, the interaction model with smoothed terms by year is a significant improvement on all less complex models.As reported in the main text, standardized residuals are centered on zero, but the model under-predicts several outlier cases (defined as scaled residual greater or less than 3).

Temporal autocorrelation
As there are multiple annual observations per state, we examine mean residual temporal autocorrelation per year: As reported in the main text, there is only meaningful averaged autocorrelation up to one year.

State-level variation in residuals
Examine median residuals by state:

4.3 Variation by state Which states had the highest homicide rates in 2020?
Plot variation over time from 2000 to 2020 by state (grey) and US median (red):

Check for multicollinearity in predictors
Check for multicollinearity in predictor variables using a linear model and correlation coefficient.lm.pov.ine<-lm(poverty_prop ~ gini_index, data = df)