Are Blue or Red States Worse on Crime?

In a recent interview, California Governor Gavin Newsom emphasized that California’s homicide rate is comparatively low and pointed out that the highest homicide rates are found in Republican-led states. His response followed House Speaker Mike Johnson’s suggestion that Newsom was an accomplice to civil unrest in California, implying that the state generally has high crime. On its face, Newsom’s statement is accurate: when measured by homicides on a per capita basis, California does not rank among the most violent states in the nation. He is also correct that “red states” do more often appear at the top of murder rate lists. But this isn’t the whole picture. Perceptions of crime and criminal justice tend to be heavily related to partisan identity. What does the data say? 

Over the past several months, we have been compiling state-by-state data, including annual crime statistics, for a forthcoming Independent Institute project. Our goal is to standardize and aggregate publicly available data into easy-to-read metrics that paint a holistic picture of different aspects in different states. For crime, we have assembled data not only on homicides, but on all categories of violent and nonviolent crime per capita, recidivism rates, police clearance rates, and even indicators like response times and victimization rates. We prioritized obtaining comprehensive data from various sources instead of relying heavily on the FBI’s data, which depends on voluntary reporting.

The homicide rate is defined as the number of murders per 100,000 residents in a given year. Because it is related to the most severe form of violent crime, it is often used as a shorthand measure of public safety. While tracking homicide rates is certainly important, focusing solely on such crimes distorts other aspects of crime and public safety. A state can have a comparatively low homicide rate but still experience elevated levels of aggravated assault, robbery, or sexual assault. Nonviolent offenses like burglary, theft, and property destruction also impose severe financial and psychological burdens on victims, even if they are short of lethality.

Crime statistics themselves tell only part of the story unless we examine how the justice system reacts. Recidivism rates, for instance, track how frequently convicted criminals resort to crime after having served their sentence. A high recidivism rate may indicate failures in deterrence, prosecution, or rehabilitation. Clearance rates offer another perspective, measuring how frequently reported crimes are “cleared” by the identification and apprehension of a suspect. When clearance rates are low, crimes go unsolved, denying not only justice but also potentially decreasing deterrence by broadcasting a message of impunity. Lastly, benchmarks such as average police response time and victimization surveys, which estimate crimes never even reported to law enforcement, provide additional details.

Together, these variables provide a clearer picture of crime. Homicide is still a significant piece of the puzzle. We don’t believe it is a stretch to view homicides essentially as the worst type of crime, yet it can’t be the sole proxy for public safety.

Normalized Z-Scores

Having assembled the various indicators concerning crime and justice effectiveness, we used z-score normalization. In simple language, normalization converts values to a common measure that expresses the degree to which each observation differs from the average. For example, if this methodology were used for school exam scores, the class average would represent the center, and scores for each test taker would be issued a new metric in conjunction with how far they deviated from the mean, and in which direction. 

The formula is straightforward: subtract the mean of a variable from a given observation X, and then divide the result by the standard deviation. Z = (x-𝜇)/𝜎, where μ mu is the mean, and σ sigma is the standard deviation (dispersion distance from the mean). While not presented here, instead of only comparing it to the literal mean, we also constructed a synthetic hypothetical U.S. state with “true” neutral mean attributes to compare a given state and anticipate counterfactuals. 

This approach has an advantage as it reduces all variables to the same plane, centered around zero. This adjustment is necessary as otherwise integrating homicide rates, property crime, clearance rates, and recidivism into a single index would be challenging, as the raw figures are on varying scales.

Variable Weights

Having converted each measure to normalized scores, the task was then to aggregate them into one index per state. Here we encountered a methodological dilemma. Not all crimes or justice measures are created equal, and one might reasonably argue that homicide is more important than theft of property, or that clearance rates are especially crucial since they reflect directly on the performance of law enforcement. The issue, therefore, is how to determine reasonable weights.

The most straightforward approach would have been to give all measures equal importance. Granted, this avoids subjective judgments about which types of crime are most crucial, but it also waters down the impact of especially severe crimes.

In principle, one might attempt to determine weights statistically by running a regression against some target variable, perhaps some hypothetical measure of safety, and then using backwards induction, or using heteroskedasticity methods from econometrics to determine the weight. We began by considering dollar costs as a proxy for social harm to anchor the weights, including abstracting from the costs of crime, e.g, the market value of a stolen vehicle, and loss of life equating to the statistical value of a life. There also exists a Cambridge Crime Harm Index and a Pinkerton Crime Severity Score that assigns a harm score to crimes by counting the number of days an individual is incarcerated for having committed a given crime. Using the Cambridge methodology as an example, a homicide could be worth 9,000 days (around 25 years) and an assault is worth 510 days (1.5 years). Those weights can then be used to adjust per capita crime rates. However, there are a number of known shortcomings to this type of weighing. There are also possible methods derived from entropy methods, such as Shannon-Entropy weighting. The problem is that crime does not present a clear, singular outcome to optimize against, and no consensus exists about which metric should serve as the gold standard. 

For now, our project remains cautious about the weighting question. We are experimenting with hybrid approaches and different weights and considering variations, but we recognize that the final decision will have normative and statistical implications. In our future project, we plan to describe this problem in detail and present the results under several weighting scenarios. Presently, we are using weights based on average imprisonment days for supercategories, modulated by a few other indicators of criminal justice effectiveness. So far, the largest weights that are most consequential are the per capita categories of violent crime (0.56) and nonviolent crime (0.17), each encompassing their subcomponents made up of individual crimes.

Once the normalized scores were compiled and aggregated, each state could be placed on a single scale that captured its overall crime and justice performance. California was in the bottom quintile of the distribution (in the bottom 20% of states). 

Some patterns in the results were clear. New Mexico emerged as the worst-performing state by a considerable margin. Alaska, Tennessee, Arkansas, and Louisiana followed, rounding out the bottom five states. The best-scoring states were Maine, New Hampshire, Idaho, Rhode Island, and Hawaii.

These results cast doubt on the narrative that one political party’s states are consistently safer than the other’s. While Tennessee, Louisiana, Arkansas, and Alaska—all red-leaning states—performed poorly, several of the strongest states, such as Idaho, Wyoming, Mississippi, and Nebraska, are also red. 

How Our Index Scores 2024 (Filter by checkbox, sort by Y-axis label)

Blue Cities in Red States

There is a common argument suggesting that violent blue cities in red states negatively affect red states. Examples would include Memphis in Tennessee, St. Louis in Missouri, Indianapolis in Indiana, Jackson in Mississippi, and New Orleans in Louisiana. We deemed this “blue cities-red states” objection to be credible and worthy of inquiry. Such was the claim of the widely circulated white paper published by the Heritage Foundation, The Blue City Murder Problem. While the report presents some interesting points and is sufficient for forming a hypothesis, the Heritage paper lacks sufficient scientific rigor and methodological strength to support a conclusion or provide strong evidence, primarily due to sample selection limitations and extrapolation problems. We will explore this topic in greater detail in future reports. For now, we acknowledge this concern, but proceed with our state-level analysis because of the the evidence in favor of this hypothesis is lacking and the following:

1. Excluding the largest blue cities from high-crime red states still results in those red states having higher crime rates compared to other states. This suggests that, when adjusted for population, smaller red cities or rural areas also contribute to the crime statistics of these states at meaningful levels, rather than the crime being solely attributed to the blue cities. For example, you can exclude New Orleans city-parish, a city and metro with a high homicide rate, from Louisiana, and Louisiana is still a state with a high homicide rate.
2. Blue states also have blue cities, and typically more blue cities, which is why the states are blue. It is unclear why the blue cities in red states would be abnormally worse solely because of that fact, and how they would negatively impact red states more than they would negatively impact blue states.
3. The blue cities-red state hypothesis somewhat sidesteps the principal issue in this context, which is the presence of an ineffectual state government. One could also argue, why hasn’t the red state government successfully fixed its blue cities?
4. Local dynamics undoubtedly influence crime rates; however, state-level data offer a more consistent and comprehensive framework for comparison. It allows us to examine broader policy environments, legal structures, and demographic trends that transcend individual municipalities.
5. In the Heritage paper, thirty of the purportedly worst homicide cities were seemingly hand-selected. While the paper states that it chose the thirty worst per capita cities, their selection is wrong prima facie. There seemed to be some population control, which gravely biases the results. For example, cities such as Wichita Falls, TX (11.3) and Hattisburg, MS (10.8) would be in the top half of their list, just as two random examples, but appear definitionally excluded. Also confounding is the fact that smaller high-crime blue cities with lower populations in red states, such as Conway, AR (10.9), are not listed by Heritage. In fact, there are a number of cities that have per capita homicide rates in this approximate range, but are not included. By not having equatable sampling, these examples skew results and do not provide a reliable account.
6. Furthermore, this ignores how other red cities may weigh down blue states. Although some cities vary in population, on a per capita basis, cities such as Alturas, Anderson, Bakersfield, Corning, Redding, Red Bluff, Susanville, and Yreka in California have high violent crime rates relative to the national average. The statistics are even worse if Fresno, Modesto, and Stockton are counted as “red cities.” Additionally, cities like Colorado Springs in Colorado, Spokane in Washington, Mesa in Arizona, and Miami in Florida have notable crime rates. It’s worth mentioning that Arizona still leans more conservative in elections, but has flipped in presidential elections, while Florida has trended increasingly red; both states have been considered swing states in recent decades.
7. Previous literature has investigated the party affiliations of local officials and crime rates, finding effectively no relationship. Other research has found no relationship between the “type” of prosecutor that the city is purview to, e.g., a “progressive” reformer or a tough-on-crime “traditional” prosecutor. This eliminates some confounding variables that would seemingly be associated with a difference among cities with respect to their political makeup.

Given Governor Newsom’s claim and the tendency for crime to become a political talking point, it was worth testing more rigorously whether partisan alignment at the state level has any predictive relationship with crime outcomes. 

Correlation with Partisanship as a Binary Variable

To do this, we used a point-biserial correlation. Point-biserial correlations are used to find relationships for data where one variable is continuous and the other is binary. In this case, the continuous variable was each state’s overall crime index score, while the binary variable was whether the state voted Republican (coded as 1) or Democratic (coded as 0) in the last presidential election. The resulting correlation coefficient measures the strength and direction of any linear relationship between political alignment and crime outcomes.

The point-biserial correlation is a technique variant of the better-known Pearson correlation coefficient r, typically employed for linear regression that enables the use of dummy variables. Ranging from -1 to +1, r shows the direction and magnitude of correlation, where +1 is perfect positive correlation, 0 shows no relationship, and -1 is perfect negative correlation. The coefficient can be expressed as:
$$ r = \frac{\sum (x_i – \bar{x})(y_i – \bar{y})}{\sqrt{\sum (x_i – \bar{x})^2 \sum (y_i – \bar{y})^2}} $$

Where\( X \) represents the dummy partisan variable, and \( \bar{x} \) the average of all \( X \) ’s. Let \( Y \) represent the safety and crime score, and \( \bar{y} \) the average of all \( Y \) ’s and ∑ is the summation of all the X ⊻ Y values.

The calculation produced a coefficient of r = -0.1363977357 for an R² value of 0.01860434231. R² shows the strength of how much of the variation can be explained by the independent variable. Because it is squared, it is always in absolute terms, and therefore does not show the direction of the correlation, positive or inverted. In practical terms, this r value is very low. It is well below what is generally considered the threshold for a weak correlation, meaning there is no meaningful relationship between whether a state is red or blue and where it ranks in crime performance. A low degree of correlation typically begins around 0.29-0.49, depending on specifics not relevant here. 

Secondly, we did a simple t-test to gauge if there was a statistically significant difference between the population mean for red and blue states. Here, we used a two-tailed test because we would like to capture any difference between the populations, with the alternate hypothesis H₁ signifying a statistically significant difference between the red and blue states. 

$$ H_0: \mu_{\text{red}} = \mu_{\text{blue}} $$

$$ H_1: \mu_{\text{red}} \ne \mu_{\text{blue}} $$

Because there is a relatively small number of blue states, instead of using Cohen’s d 0.8 for a large effect size, we applied an effect size sample correction of 0.984. The T distribution was within the 95% region of acceptance. Therefore, at a significance level of α = 0.05, we failed to reject the H₀ because the p-value of 0.3845 is greater than α. 

We used the heuristic Tukey fence method for detecting outliers using k = 1.5. The Tukey fence is a way to flag samples based on the interquartile range and first and third quartiles. K = 1.5 is the go-to value suggested by John Tukey, but it may be a little overaggressive for such a dataset and count too many outliers. However, New Mexico was identified as the sole outlier, having the worst score by a considerable amount. Whether or not New Mexico was included or excluded did not make a difference insofar as accepting the null hypothesis went. The test was run twice. 

Given that this hypothesis testing relies on relatively small samples, the test’s a priori power (probability of detecting an effect) is decent but not incredibly high; there is a risk of not properly rejecting H₀.

Isolating Homicides

Returning to homicides as an isolated variable, we used ten years of homicide and presidential election data to test whether there is a difference between red and blue states. We used a t-test with a significance level of α = 0.05 and effect size of 0.2. With a p-value of 0.302, we do indeed reject H₀ in favor of H_1, meaning that there is nominally a statistically significant difference (that is, there is reason to believe the results are not due to chance) between the blue and red states with respect to homicide. Blue states had an average lower homicide rate of 1.835766 per 100,000 and did better than red states. However, the difference is microscopic in practical terms.

	Blue	Red
Sample average	5.597696	7.433462
n	217	260
S	3.713594	3.989038
Skewness	2.769639	1.200576

Using the Tukey Fence method, k=1.5, we identified eight outliers among red states and five among blue states, and reran the test excluding these samples, reaching the same conclusion with respect to the hypothesis test.

With this in mind, we performed another point biserial regression on the data to see how much of the difference might be attributed to the partisan dummy variable. This regression returned a r = 0.2305581561. Such a low correlation coefficient implies that the political party is a poor forecaster of a state’s homicide rate. Further adjustments for covariates would plausibly tease this effect out. With such a small difference, the result does not seem practically significant and appears to be a semantically moot point.  

Decoupling Crime and Justice

A potential objection could be that crime and measures of law enforcement efficacy should not be combined into a single index. The thought would be that the two are separate domains that would confuse outcomes and inputs and that the effectiveness of law enforcement would already be included in crime statistics. While we hold that our index is proper for our original purpose of providing a holistic picture of states for comparison, for the purpose of interstate comparisons of crime, we recognize this could be an issue.

To address this potential methodological problem, we adjusted the weights to exclude all measures of law enforcement so as to only count per capita crime, both violent and nonviolent. We updated our previous weights to create a new score and reran our tests. Adjusting the index resulted in minimal changes. Oregon and Colorado performed slightly worse when we normalized the scores, and there was some negative skewness (left-skewed, with more upward clustering). However, the overall pattern and rankings of the states were quite similar to our earlier tests. This consistency is to be expected, given that the law enforcement components, such as clearance rates, were minor factors in the previous assessments.

Conclusion

The implication is straightforward: there is a lack of evidence that partisan results are a reliable predictor of safety outcomes. States vary widely within each category, and while certain individual states perform particularly well or poorly, those differences do not map onto red vs blue boundaries.

Governor Newsom is correct that California’s homicide rate is considerably lower than that of several Republican-led states, but homicide is only one measure of public safety. A fuller accounting requires examining violent and nonviolent crime, clearance rates, recidivism, and system performance indicators like response times and victimization surveys. When those dimensions are standardized and combined into a single index, California isn’t the worst but far from exemplary.

While some of the worst-performing states in the index are indeed red, some of the best-performing states are also red. Similarly, the worst-performing state is blue, though the best-performing state is also blue. This failed to demonstrate that political control, by itself, relates to crime outcomes. The point-biserial correlation between political alignment and crime index score is very weak. 

Our findings generally comport with the literature on the subject: Ferreira and Gyourko’s 2007 paper found that the party of a city’s mayor had no relationship to property crime, and Brownstein et al.’s 2021 study found that democratic cities had higher property crime overall, but when adjusted for city size, the relationship between party and property crime disappeared.

In the forthcoming report, we will present the full methodology, including normalization, weighting scenarios, and a more comprehensive discussion of results. Decisions about weighting, for instance, could affect the results, and there may be additional measures worth including to capture public safety more comprehensively. We aim to strengthen our findings by incorporating donut regression to better adjust for how blue or how red a given state is. Previous literature has generally shown that other demographic factors gleaned from multivariate regression are better indicators of crime. Work investigating the results of a party controlling the city or county government while the other party controls the state government has previously been investigated. We welcome constructive feedback, particularly from those with expertise in criminology, statistics, or public policy. Suggestions and critiques offered in the comments will help refine the approach and ensure the eventual index is as rigorous and useful as possible.

For now, the preliminary findings suggest caution against drawing political conclusions from narrow slices of the data. There is an absence of empirical evidence that either blue or red states have a meaningful difference in crime. We find that the inference that partisanship is a predictor of crime is unsubstantiated. 

Postscript

We ran a number of comparisons on a collated dataset. Repeated testing using varied methods risks incorrectly finding a statistically significant relationship. In addition to donut regression, further study could be strengthened by correcting the alpha values based on the number of tests (Bonferroni Correction).

Scores

State	Safety and Crime Score
NM	-2.3317
AK	-1.5820
TN	-1.3828
AR	-1.3538
LA	-1.1625
CO	-0.8443
CA	-0.8346
WA	-0.4399
SC	-0.4652
MO	-0.5448
MD	-0.5440
KS	-0.3995
MI	-0.3761
MT	-0.3658
NC	-0.3050
DE	-0.2964
TX	-0.2998
OK	-0.2856
NY	-0.1387
AL	-0.1863
AZ	-0.1199
GA	-0.0658
IL	-0.0551
OR	-0.0358
SD	0.0095
OH	0.1376
IN	0.0892
ND	0.2274
PA	0.3463
FL	0.3625
MA	0.3792
WI	0.4385
VT	0.4375
KY	0.4595
WV	0.4990
NJ	0.5399
UT	0.5660
IA	0.6649
VA	0.6356
MN	0.6258
CT	0.7152
MS	0.7065
NE	0.7033
WY	0.7554
HI	0.8734
RI	0.9481
ID	0.9822
NH	1.1963
ME	1.4439