Time Delay on CRIMINAL sentencing
Tian Tong, McCourt School of Public Policy
Georgetown University
December 4, 2024
Evidence from Cook CountyData Post-2020
Time Delays Exacerbate the Severity of Third-Party Punishment
Kundro et al., Psychological Science, 2023.
The researchers analyzed two large-scale datasets of 6 controlled experiments:
-
Civilian Crime Data: 150,392 sentencing decisions from Cook County, Illinois
-
Police Misconduct Data: 10,380 punishment recommendations for NYPD officers from the Civilian Complaint Review Board
Participants were presented with scenarios where they judged the severity of punishment for transgressions with varying time delays between crime and arrestment.
Results consistently showed that perceptions of unfairness due to time delays led participants to recommend harsher punishments.
Normative perspectives on punishment suggest that time delays should be irrelevant when determining punishment severity. This is because fundamental principles of penal justice suggest that punishments should vary only with regard to the 'moral gravity' of the alleged wrongdoing, and morally similar cases should be treated alike, regardless of other factors….
Differences / Enhancements
Temporal Extension and Contextual Relevance
Assessing the Established Relationships
By extending the analysis to include data from 2020 to 2024, this study captures the impact of significant societal events, the COVID-19 pandemic, which may have disrupted judicial processes and sentencing patterns. This extension ensures the findings remain relevant to contemporary judicial practices.
This hierarchical ranking of features adds interpretive depth to the results, offering actionable insights into which variables require attention in policy discussions.
By replicating the original analysis with updated datasets, this section examines whether the previously established relationship still holds true in the contemporary context.
Unveiling Feature Importance
Our Analysis
1
DATA
Using the Cook County State Attorney's Office dataset, we analyzed criminal cases from January 2020 to October 2024, a critical period encompassing three distinct phases of pre-COVID, Peak-COVID and Post-COVID.
​
To enhance our analysis of punishment severity, we created a weighted punishment measure that combines standardized sentence lengths with commitment types (scaled 1-3), generating a new target variable 'Weighted_Punishment'. This composite measure provides a more comprehensive view of punishment severity by accounting for both duration and type of sentence.
​
​
Additionally, while our primary focus was on examining time delays, we recognized the significance of immediate arrests. Therefore, unlike the original experiment, our analysis specifically focuses on delayed arrests rather than all cases.
2
METHODOLOGY
Our analysis employed complementary techniques to understand patterns in criminal justice during the COVID era:
1. Repetitive OLS Model
2. Bisecting K-means Clustering
3. Random Forest Model
By employing bisecting k-means clustering, we can effectively identify these distinct groups in a hierarchical manner. This approach combines the simplicity of traditional k-means with the ability to iteratively refine clusters, making it particularly advantageous for datasets with complex structures.
​
​​​
To complement the OLS models, a Random Forest model was introduced as a novel analytical tool in this context. Random Forest not only provides predictive capabilities but also offers a hierarchical ranking of feature importance, which was not emphasized in the original study.
3
INSIGHTS
​​
Evolving Political Implications from Punishment Severity Analysis:
- Reducing bias in Sentencing
- Providing Sentencing Guideline
- Court Transparency and Accountability
- Advocacy for Human Rights in Justice Systems
- Potential Policing and Sentencing Reform
​
The results of analysis offer political and policy implications that are relevant to criminal justice reform, civil liberties, health care for incarcerated individuals, and demographic equity in sentencing.
It highlights the need for systematic changes to reduce bias and delay in the justice system, improve transparency, and ensure fair treatment across all demographic groups.
-
The COVID-19 pandemic created unprecedented disruptions in the American criminal justice system, fundamentally altering how courts operated and cases were processed. During this extraordinary period (2020-2024), the Cook County State Attorney's Office continued to handle criminal cases, creating a unique dataset that captures how crime, arrests, and sentencing patterns evolved through pre-pandemic, peak-COVID, and recovery phases.
This comprehensive criminal justice data offers a valuable window into examining how time delays between crimes and arrests affected punishment severity during a period of systemic disruption that fundamentally changed how justice was administered. By analyzing patterns in over 42,000 cases, we can understand how extraordinary circumstances influenced the relationship between procedural delays and sentencing outcomes.
Linear Regression Model:
Repetitive Experiment with controlled variables
Original Approach
The study used OLS regression to quantify the effect of time delays on sentencing severity while accounting for potential confounders.
The study found that longer time delays between the crime and the arrest were associated with harsher sentences. This relationship persisted even after controlling for confounding variables such as crime type, judge, demographic factors, and year of sentencing​
OLS with controlled variables:
-
​Interaction term between length of case in days and counts of charge
-
Demographic factors such as race, age and gender
-
Fixed court and time effect
-
Target factor: Weighted punishment severity
Conclusions:
​​
​Statistical summary of independent variable
==============================================================================
Coefficient Std.Error t-statisticst P>|t|
----------------------------------------------------------------------------------------------------------------------------------
Intercept 0.2578 0.019 13.589 0.000
delay in days 0.0196 0.0070 2.675 0.0070
==============================================================================
The transgression was met with harsh penalty when...
Higher Case Complexity
Being Male
Post-COVID
Age at Incident
Aggravated Robbery
Attempt Armed Robbery/Armed Robbery
Attempt Homicide/Homicide
​
Sex Crimes
Human Trafficking​
​
Home Invasion
Attempt Vehicular Hijacking
Violent Crimes
Sexual or Exploitation Crimes
Property-Related Crimes
Limitations:
​
The model demonstrates moderate explanatory power, with an R-squared of 0.360, indicating that 36% of the variance in sentencing severity is explained by the included predictors.
The inclusion of 200 predictors reflects the complexity of sentencing decisions but also increases the risk of overfitting, particularly for variables with many categories.
Also, the relatively high AIC and BIC indicate the trade-off between model complexity and explanatory power.
A key limitation of the model lies in its assumption of a linear relationship between arrest delays and punishment severity, which may not fully align with real-world dynamics. While the findings are consistent with previous literature—showing that longer delays are associated with harsher punishments—the observed effect size is relatively small (0.0070 units), underscoring the need for further exploration of alternative specifications, such as non-linear or interaction effects, to better capture the nature of this relationship.
Key Metrics and Results:
R-squared: 0.3595
Adjusted R-squared: 0.3448
AIC: 26437.68
BIC: 27863.71
Bisecting K-Means Clustering
​
Reasons to use Bisecting K-Means
​
-
Hierarchical Approach:
The hierarchical splitting process ensures more cohesive and well-separated clusters, especially in datasets with complex or nested structures.
-
Adaptability
Bisecting k-means improves cluster quality by better handling irregular or uneven cluster shapes through localized optimization during each split.
-
Balance of Simplicity and Efficiency:
It retains the simplicity and computational efficiency of k-means while introducing hierarchical refinement, making it scalable and practical for large datasets.
​
Deciding the Value of k​
​​
​
​Silhouette Plot:
There's no significant improvement when increasing k from 4 to 7
The stability of the score around k=4 suggests it's a reasonable choice without overfitting
Elbow Plot:
After k=4, the decrease in WCSS becomes more gradual
The rate of WCSS reduction slows significantly after k=4, indicating diminishing returns.
Complex sexual or exploitation related cases with relatively higher resolution time and punishment
Low-severity ,less complex cases involving property crimes
​
Serious cases involving violent crimes, with higher punishment and longer resolution time
​
Very specific outlier group with highly complex cases and extremely high charges per person, resulting in long delay
Cluster 0: 528 cases
Cluster 1: 4531 cases
Cluster 2: 2594 cases
Cluster 3: 51 cases
Random Forest Model
​
​
Hyperparamter Combination
​​​
​
The best hyperparameter combination for the Random Forest model, as determined by the grid search, is as follows: the model uses bootstrap sampling, has a maximum tree depth of 34, considers log2 of the total features at each split for the best feature selection, requires a minimum of 1 sample per leaf node, uses a minimum of 4 samples to split an internal node, and is constructed with 151 decision trees (n_estimators).
Overall Performance:
​
-
Accuracy (0.71): The model correctly classified 71% of all cases in the dataset.
-
Macro Average (0.70): This is the unweighted average of precision, recall, and F1-score across both classes. It treats each class equally, regardless of class size.
-
Weighted Average (0.71): This is the weighted average of precision, recall, and F1-score, accounting for the number of instances in each class. It provides an overall measure that reflects the class imbalance.
Feature Importance Plot
The Random Forest model indicates that log time delay is among the most influential features contributing to the prediction of sentencing severity, corroborating earlier findings from both the OLS regression and the original literature. Despite the model's moderate accuracy (66%), the hierarchical ranking of feature importance aligns with prior evidence that longer time delays are positively associated with harsher punishments.
From the clustering analysis, we observed distinct groups of offenders with varying levels of sentencing severity, where clusters with higher log time delays tended to show patterns consistent with harsher punishments. This further supports the narrative that time delays are not only significant predictors of punishment but also key variables in identifying latent patterns in the justice system.
The original literature also emphasized the role of time delays in exacerbating sentencing severity, a relationship validated across all methods in this study. This consistency across methodologies highlights the robustness of the finding, underscoring the need to address time delays as a critical issue in judicial reform. While the Random Forest model added depth by ranking feature importance, the overall conclusion remains aligned with prior studies: time delays do increase the severity of punishment.
Our
REFLECTION
Our experiment did not yield results as robust as those reported in the existing literature, which is based on multiple experiments (six in total), while our study focuses solely on Cook County data. This limited scope makes our findings less generalizable and may weaken the validity of the conclusions. Additionally, the relatively small number of numerical variables in our dataset, combined with its high-dimensional nature, hindered the effectiveness of Bisecting K-means clustering, which is better suited for lower-dimensional data.
Another factor to consider is that approximately 80% of the dataset consists of one-day arrests, which limits the reliability of conclusions about time delays and punishment severity. To make more conclusive statements in this area, a larger, more diverse dataset or additional personal surveys would be required, as longer delays exacerbate punishment severity due to perceptions of unfairness rather than other mechanisms like moral outrage or deterrence​.
​
The analysis of punishment severity reveals several evolving political implications that could shape justice systems. Reducing bias in sentencing is a critical priority, as disparities often undermine public trust in the legal system to ensure court transparency and accountability. Advocacy for human rights within justice systems is another important area, emphasizing the need for equitable treatment and protection of vulnerable groups.
Citations
Cook County Government. (n.d.). Sentencing [Data set]. Cook County Open Data. Retrieved October 27, 2024, from
https://datacatalog.cookcountyil.gov/Legal-Judicial/Sentencing/tg8v-tm6u
​
Batson, C. D., Kennedy, C. L., Nord, L. A., Stocks, E. L., Fleming, D. Y. A., Marzette, C. M., & Zerger, T. (2007). Anger at unfairness: Is it moral outrage? European Journal of Social Psychology, 37(6), 1272–1285. https://doi.org/10.1002/ejsp.434
​
Kundro, T. G., Nurmohamed, S., Kakkar, H., & Affinito, S. J. (2023). Time and punishment: Time delays exacerbate the severity of third-party punishment. Psychological Science, 34(8), 914–931. https://doi.org/10.1177/09567976231173900​
​
Witte, G., & Berman, M. (2021, December 19). Long after the courts shut down for COVID, the pain of delayed justice lingers. The Washington Post. https://www.washingtonpost.com/national/covid-court-backlog-justicedelayed/2021/12/18/212c1 6bc-5948-11ec-a219-9b4ae96da3b7_story.html
==============================================================================
Coefficient Std.Error t-statisticst P>|t|
----------------------------------------------------------------------------------------------------------------------------------
Intercept 0.2578 0.019 13.589 0.0000
delay in days 0.0196 0.0070 2.675 0.0070
Charge counts -0.0462 0.0090 -4.926 0.0000
Length of cases 0.0007 0.0001 10.118 0.0000
Case complexity 0.0001 0.0000 4.265 0.0000
Age at incident 0.0030 0.0010 2.946 0.0030
COVID Period 0.1613 0.0550 2.945 0.0030
Race: Black 0.1885 0.1030 1.834 0.0670
Gender: Male 0.2485 0.0340 7.244 0.0000
==============================================================================
Full Statistical Summary:
​