Recidivism, defined as the tendency for an individual to re-offend after a prison sentence, is one of the biggest problems currently facing the criminal justice system. Criminals who are not properly rehabilitated just end up getting arrested again and returning to prison, leading to the justice system becoming one big revolving door. The United States, in particular, deals with a massive recidivism problem, as its recidivism rate of 44% after one year is one of the highest rates in the world. It does not help that the U.S. also has the largest percentage of its population imprisoned among any first-world country as well. It is evident that this country needs to take a step back, find out what is driving these high recidivism rates, and make the appropriate changes.
In an attempt to find out what is causing the U.S.’s recidivism crisis, we are going to critically examine data regarding recidivism rates in Georgia from 2013 to 2015. This dataset is from the Office of Justice Program and includes information regarding age, race, gender, gang affiliation, education, employment status, drug use, prior arrests and convictions, and most importantly, whether the individual was arrested within three years of being released from prison. Each row in the dataset represents an individual that has gone to prison and is in the Georgia Department of Community Supervision. Each column in the dataset represents a variable or feature related to the individual, such as the ones mentioned eariler. We hope to use this data to determine what factors seem to increase the likelihood of committing another crime, and using what we find in order to inform criminal justice policy strategies.
While there is a laundry list of questions one could ask with this dataset, we are going to focus on three in particular. Starting off, we wanted to understand how the predicted risk score for individuals relates to actual recidivism outcomes. Next, we were curious about whether individuals who followed the supervision programs, were employed, and avoided drugs less likely to recidivate. Finally, we studied how recidivism rates vary based on the type and severity of the offense.
Individuals in our dataset have predicted risk scores which indicate how likely they are to recidivate, and we wanted to understand if there was a correlation between predicted risk scores and actual recidivism outcomes and see how accurate the predicted risk scores are.
The graph above depicts the distribution of predicted risk scores by recidivism rates within 3 years. We can see that in the first facet, where individuals did not recidivate within 3 years, the distribution of the data looks more normal. However, in the second facet where individuals did recidivate within 3 years, the distribution is skewed to the right. It looks like higher predicted risk scores have a higher count of individuals that recidivate. We would also expect lower risk scores to have a higher count of individuals that do not recidivate, but the graph shows only a slight increase. However, it appears that overall, predicted risk scores and actual recidivism outcomes are correlated which makes sense.
While analyzing the predicted risk score, we wanted to see what attributes of an individual, such as race, gender, and education level, correspond to higher predicted risk scores. We first analyzed the predicted risk score in terms of race and gender of the individual.
From the faceted graph above, we can see that a majority of the individuals being analyzed are males, in comparison to females. From the graph of females, we can see that the distribution of predicted risk score is unimodal and skewed right, with a majority being individuals of the race White. From the graph of males, we can see that the distribution is skewed to the right, with a majority being individuals of the race Black. We can conclude that more White females have a higher predicted risk score than Black females and that more Black males have a higher predicted risk score than White males. One drawback of this data analysis is that there is limited data on females in comparison to males and there is only data on two races, which could affect the conclusions made above.
We also wanted to see if there was a correlation between education level and predicted risk score. From the box plot above, we can see that individuals with at least some college education have a median predicted risk score of 5.0, while individuals with a high school diploma or less have a predicted risk score greater than 6. Looking at the box plot shape, the plot of individuals with a high school diploma looks to be normally distributed, while individuals with at least some college and less than a high school diploma have a positive skew or skewed to the right. We can conclude that higher level education corresponds to a lower predicted risk score. Because our dataset did not specify the calculation behind the predicted risk score, we wanted to see if attributes of an individual play a role in the numerical value of the score. From our analysis of the graphs above, we can see that lower education levels and Black males have a higher predicted risk score.
To approach this question, we were curious about the relationships between recidivisim rates and individuals drug usage and employment status. Our initial guess was that lower positivtiy rates for drug test and higher employment rates would correspond to lower recidivism rates within 3 years and we tested our hypothesis below. Our dataset gave us information about individual’s drug usage in terms of cocaine, meth, THC, and other types of drugs. We consolidated this data to determine individual’s overall drug usage.
This plot shows us that the far majority of people who do not reoffend have low positive drug test rates and high employment rates. That means that people who are following the guidelines of avoiding drugs and finding work opportunities are typically the ones who have the most success when it comes to avoiding another arrest. This tells us that in order to lower recidivism rates, focusing on helping former criminals gain employment and avoid drugs should be a winning strategy.
We also wanted to gain insights about whether there was a relationship between program attendance and initial supervision level, as well as program violations and recividism outcome rates.
By looking at the plot above, we see that the program attendance of 0 is the most common because the rectangle has the largest area on the mosaic plot. Meanwhile, It seems like program attendance of 1-5, and 7-9 seems to be the least common as it has the smallest area on the mosaic plot. 6 and 10 have decent program attendance which makes sense because most prisoners would either choose to just not attend the programs or attend and actually partake in them for some time. From our data, also we can see that high levels of initial supervision seem to be counterintuitive and have the fewest number of observations than expected. Specialized levels of initial supervision, which in Georgia are given only to people who commit sex crimes, seem to be best as they have the most number of observations than expected throughout the attendance program with only fewer observations than expected for 0 attendance, which is great and highlights that supervision and program attendance could be very helpful. This graph shows that a specialized initial supervision level is correlated with program attendance, which makes sense because sex crimes are generally considered to be the worst type of crime so probation officers may place a greater emphasis on the importance of following the program.
From the mosaic plot above, we see that the prisoners who conduct program violations such as Electronic Monitoring, Instruction, Failure to Report, and Moving without Permission, are the least common because the rectangle has the smallest area on the mosaic plot. Meanwhile, the combination of prisoners who don’t conduct program violations is the most common as its rectangle is the largest in area.
The colors represent the residual of that cell/combination of levels. Blue means there is more observation than expected under the null model and red meaning that there are fewer observations than expected. From our data, we can see that prisoners who conduct program violations but do not recidivate within 3 years have the fewest number of observations than expected. Prisoners who conduct program violations but do recidivate within 3 years have the most number of observations than expected. This shows that program violations could very well be an indication of recidivism in the future.
To gain a higher-level understanding of how the quantitative variables of our dataset are related, we performed principal component analysis.
From the PCA plot, we can see that the data is roughly clustered into bands according to initial supervision level. The variable that is most strongly related to the two principal components is supervision risk score, and the arrow for this variable is almost perpendicular to the bands which makes sense as supervision level is determined by risk score. Variables relating to drug usage are also approximately perpendicular to the bands, implying that people with higher supervision levels tend to have higher rates of positive drug tests. This agrees with our previous graphs which showed that recidivism is correlated with drug usage and risk scores.
Our final research question is to understand the relationship between the 3 year recidivism rates and the type of offense and the severity of the offense. To do so, we first chose to analyze recidivism rates within 3 years based on the length of individuals’ sentences and the type of offense committed.
This stacked bar plot shows that property crimes are by far the most common offenses while sex crimes are comparatively rare. Looking specifically at property crimes, the recidivism rate is the highest out of all crime options, even though the sentence lengths are low. Additionally, the only type of offense not to see a recidivism rate above 50% is sex crimes. This graph seems to imply that crimes that are the most severe/have the highest sentence lengths actually have the lowest recidivism rate. It is possible that the longer people are in prison, the more they are incentivized to take steps to avoid going back.
Finally, we wanted to gain insights on recidivism rates within 3 years by the number of prior misdemeanor arrests.
This bar plot shows that the majority of subjects had six or more prior misdemeanor arrests. Interestingly, the second most common number of prior arrests is zero, and the frequency for the remaining categories decreases as the number of arrests increases. We can also see that people with six or more priors have the highest recidivism rate, while people with zero priors have the lowest rate. In fact, a proportion table reveals that recidivism rate is strictly increasing with number of prior arrests. Based on this analysis we can conclude that first-time offenders are not likely to recidivate, but the more times a person is arrested, the more likely they are to be arrested again in the future.
Looking at the visualizations we made, there are some clear conclusions we can come to about recidivism. Answering our first research question, individuals with a higher predicted risk score do have higher rates of recidivism. Looking at predicted risk scores, race does seem to play a factor, as Black males have higher predicted risk scores than White males, while White females have higher predicted risk scores than Black females. The less education someone has also leads to a higher predicted risk score as well. However, we do need to use risk scores with caution and ensure that the algorithm used to calculate the scores is fair and unbiased.
Now examining our second question, individuals who follow the guidelines post-prison do tend to have lower recidivism rates. That means people who are attending their probation appointments, avoiding drugs, not violating the terms of their release, and gaining employment are the ones who are staying out of prison. This is an unsurprising result and it tells us that more efforts should be made towards getting former prisoners to follow these guidelines, because that is what will lower the recidivism rates.
Finally, we found how recidivism rates vary based on the type and severity of the offense. What we found is that those who were in prison for shorter amounts of time and for lighter (comparatively) offenses such as property crime, were most likely to reoffend. However, individuals who spent a longer time in prison for more serious crimes were less likely to reoffend. We also found the more offenses an individual had committed in the past, the more likely they were to reoffend, which highlights the importance of early intervention.
To sum up, the data mostly wasn’t that surprising. Individuals who had committed crimes in the past and had higher predicted risk scores had higher recidivism rates. However, it was interesting to see that those who had longer sentences for more serious crimes reoffended less. The main takeaway of our analysis is that the most important factors to consider if we wish to reduce recidivism rates are the success of the probation supervision programs and early intervention.
Unfortunately, there are some clear limitations to this data. First of all, the dataset is from 2013 to 2015, so the factors that may have influenced recidivism may be different today than they were years ago. Additionally, the data comes entirely from Georgia, which may not be representative of the United States as a whole. The data also does not include gang affiliation for female prisoners, which could possibly result in some missing context. Also, it stops separating for age after 48 years, so we are unable to tell whether there is any difference in recidivism for those aged 48-60 and 61+ for example. The racial data only includes Black and White, while excluding prisoners who are Asian, Hispanic, or a member of another racial group. The dataset also does not include where in Georgia these arrests are taking place, as recidivism rates might very well be different between downtown Atlanta and more rural parts of Georgia. Finally, the lack of quantitative data made it difficult to provide a more quantitative analysis towards recidivism rates. Despite the problems with this dataset, it is still very detailed and we can still make several conclusions based on it. In future research, we recommend using data from various states of the United States, which is more diverse and inclusive in terms of gender, age, and race and containing more geographic and quantitative data.