Portfolio item number 1
Short description of portfolio item number 1
Posts by Collection
portfolio
Portfolio item number 2
Short description of portfolio item number 2
publications
The Heterogeneous Effect of Dropping Out for Higher Education Students: The French Case
Working paper on SSRN, 2024
For higher education systems with general and vocational degrees, the question of resource allocation for dropout policies is fundamental. In France, the university concentrates most of the focus and resources compared to the vocational track. I estimate the heterogeneous effect of dropping out on labor market outcomes (employment rate and wages), conditional on the student’s degree, to evaluate the validity of this allocation. I use a causal Random Forest methodology to account for the heterogeneous cohort composition of these degrees, with the distance to the closest higher education institution as an instrument for dropout. Using 2SLS leads to underestimating the overall effect of dropping out by 9 percentage points for the employment rate and by 4 percentage points for the average wage. Vocational degree dropouts are more penalized than university dropouts on their average wage but not on their time in employment. Finally, using a multidimensional categorization of students can be beneficial for creating a targeted dropout policy.
Recommended citation: Tissandier Gaspard. (2024). The Heterogeneous Effect of Dropping Out for Higher Education Students: The French Case
Download Paper
Resources Misallocation and Predictive Policing: A Data Generation Process
Forthcoming working paper, 2024
The use of Machine Learning and Deep Learning methods in data-driven policy making, such as predictive policing, has surged in recent years. This paper explores the connection between misallocation of police resources and the replication in prediction errors due to non-representative data. The study presents a framework where a policy maker allocates police resources to minimize crime in a jurisdiction using predictive models, and addresses two key questions: how misallocation errors in resource allocation can replicate as prediction errors and how to estimate the replication factor between these errors when no closed form expression exists. The paper demonstrates that over-policing in an area amplifies misallocation errors in predictions, creating a feedback loop. In cases involving non-trivial pre-dictive methods, the paper highlights an errors replication pipeline, allowing empirical estimation of the replication factor. The methodology is applied to predict crime levels in New York City.
Recommended citation: Tissandier Gaspard. (2024). Resources Misallocation and Predictive Policing: A Data Generation Process
PAVED With Good Intentions? An Evaluation of a French Police Predictive Policing System
R&R at the International Review of Law and Economics, 2025
From late 2017 to early 2019, one of the two french law enforcement agencies (the Gendarmerie) tested in 11 out of 101 departments a predictive policing system named PAVED. The system designed by the Gendarmerie predicts burglaries and vehicle thefts with the stated objective of better allocating patrols and thus increasing deterrence. We use month-law enforcement jurisdiction area panel data to evaluate whether the system produces the expected reduction in these thefts. Using a TWFE approach and considering several alternative counterfactuals, our results consistently indicate no detectable effect of PAVED on burglaries. With regard to vehicle theft, small variations are observed following the implementation of PAVED, but these variations are not consistent or robust across the different counterfactuals considered.
Recommended citation: Lecorps, Yann and Tissandier Gaspard. (2025). PAVED With Good Intentions? An Evaluation of a French Police Predictive Policing System
Download Paper
Multilevel Modeling
Oxford Research Encyclopedia of Criminology and Criminal Justice, 2025
Multilevel models are statistical methods used to investigate associations between data from different levels of analysis or clusters. Given the inherent relationships between crime and its surrounding social and physical context, multilevel modeling is crucial for criminology and criminal justice research. Typically, research designs often assume observations are independent and internally homogeneous regarding the characteristics of interest. However, the social world consists of different grouping structures, where observations within the same group share more common characteristics than those in different groups. Ignoring these group-level influences can lead to incorrect analysis and interpretations. Multilevel models extend single-level regression models to account for the nested data structures. These models offer several statistical and theoretical advantages. On the one side, they avoid biased estimation of the relationships of interest. On the other side, they prevent interpretative errors, such as ecological and atomistic fallacies, and allow exploration of associations between variables measured at different levels of analysis. Various types of multilevel models exist, including: random intercept models that estimate a baseline level of the outcome of interest for each cluster of observations; random coefficient models that capture how the studied relationship varies across different groups; and fixed effect models that control for all characteristics shared by observations within the same cluster. Additionally, growth model is a specific type of multilevel model used to analyze the effects of time-related variables on an outcome measured consistently over time within the same set of observations. The choice of the most suitable multilevel model depends on statistical, methodological, and theoretical assumptions.
Recommended citation: Dugato, Marco and Tissandier, Gaspard. (2025). Multilevel Modeling; Oxford Research Encyclopedia of Criminology and Criminal. doi:10.1093/acrefore/9780190264079.013.ORE_CRI‑00816 Justice
From Data to Action: Benchmarking the Accuracy-Interpretability Trade-off of Machine Learning Algorithms for Crime Analysis
Working paper, 2026
Over the past two decades, crime modeling has shifted from traditional statistical approaches toward increasingly complex machine learning systems, raising a central question: are gains in predictive accuracy sufficient to justify reduced interpretability? This paper benchmarks a range of models for predicting cell-month crime concentration across five crime types in Newark, New Jersey. We compare traditional approaches (moving averages, kernel density estimation), high-performing machine learning models (XGBoost), and explicitly interpretable methods (FasterRisk, SIRUS, Explainable Boosting Machines), using a common set of predictors including lagged crime, socioeconomic indicators, place-based characteristics, and weather variables. Models are evaluated using area-based classification metrics, with particular attention to sparse and highly concentrated outcomes typical of crime data. We find that the trade-off between accuracy and interpretability is limited. Interpretable models restricted to ten components achieve performance levels only marginally below more complex algorithms. For aggravated assault and homicide, the gap in AUC between the best-performing model and the least accurate interpretable alternative is approximately four percentage points. Sensitivity remains high across models, while traditional approaches show larger performance gaps. We also document the “top-k trap” affecting threshold-dependent metrics and argue for joint use of ROC and precision-recall curves. Overall, interpretable models provide competitive predictive performance while offering greater transparency and actionable insights for policy design.
Recommended citation: Tissandier Gaspard and Gimenez-Santana Alejandro (2026). From Data to Action: Benchmarking the Accuracy-Interpretability Trade-off of Machine Learning Algorithms for Crime Analysis
Download Paper
Widening the Aperture of the Opportunity Lens: Engaging Local Stakeholders in the Interpretation of Crime Concentrations
Crime Science, 2026
This article discusses the value of integrating spatial crime analysis with local stakeholder engagement to fully leverage opportunity theory in community crime prevention. We begin by describing the literature on opportunity theory applications and the limitations of that research, noting both the value and rarity of employing inclusive research practices. To quantify the frequency, types, and trends in stakeholder engagement employed in opportunity theory-based studies, we conducted a content analysis of Crime Science articles published from 2012 through 2025. Of the 212 eligible studies reviewed, 12.3% involved a mix of quantitative and qualitative methodologies, 13.7% incorporated some degree of stakeholder perspectives, yet only 1.9% specifically included community members residing in the geographic areas under study. We then present three case studies on gun violence, auto thefts, and street robberies in Newark, New Jersey through the Newark Public Safety Collaborative, an anchor initiative housed in Rutgers University’s School of Criminal Justice. The case studies illustrate how data analysis alone pointed to opportunity-reducing strategies that would have been less effective or misguided without community member input. This paper concludes by advocating for the field of crime science to employ more comprehensive, triangulated methodologies that combine spatial analysis with the lived experience of community members as well as the professional expertise of other essential stakeholders.
Recommended citation: La Vigne, N., Giménez-Santana, A., Tissandier, G., & Santos, A. (2026). Widening the aperture of the opportunity lens: engaging local stakeholders in the interpretation of crime concentrations. Crime Science. doi:10.1186/s40163-026-00285-x
Evaluating the Effect of Enhanced LED Street Lighting on Nighttime Crime
Journal of Quantitative Criminology, 2026
This study evaluates the deterrent effect of enhanced street lighting on crime in the City of Newark, NJ. Between 2019 and 2021, a prioritization program was launched to replace high-pressure sodium lights (HPS) with light-emitting diodes (LED) fixtures across the City of Newark, NJ. Grounded in environmental criminology, which underscores the influence of the physical environment on criminal behavior, we explore the effect of enhanced street lighting on the distribution of nighttime crime. We hypothesize that enhancing street lighting in crime-prone areas can be an effective crime reduction strategy to improve public safety and community well-being. Leveraging a quasi-experimental design, we analyze a three-year replacement initiative involving approximately 1,500 streetlight poles fitted with new LED fixtures. We evaluate the program’s impact on both violent and property crime, at both the aggregate and disaggregated levels, using a Difference-in-Differences (DiD) methodology and the Gardner estimator Our findings reveal a significant short-term reduction in nighttime and outdoor violent crime of approximately 46% in the first two quarters post-replacement; the effect is mainly driven by a decrease in aggravated assaults and robberies. However, this effect disappears after two quarters, suggesting an adaptation to the new lighting environment. Long-term analysis does not show any significant reduction in violent or property crime, up to two years post-replacement. The results of this study suggest that enhanced street lighting can be an effective strategy for reducing nighttime violent crime in urban areas. However, this deterrent effect diminishes over time, possibly indicating that motivated offenders adapt to these environments. Other factors, such as changes over time in the conditions of streetlighting equipment, may also contribute to suboptimal streetlighting conditions, attracting crime in these areas.
Recommended citation: Gimenez-Santana, Alejandro; Tissandier, Gaspard; Zlaoui, Khalil; Santos, Adriana; Caplan, Joel; Kennedy, Les. (2026). Evaluating the Effect of Enhanced LED Street Lighting on Nighttime Violent Crime in Newark, NJ. Journal of Quantitative Criminology. doi:10.1007/s10940-026-09690-6
talks
Talk 1 on Relevant Topic in Your Field
Published:
This is a description of your talk, which is a markdown file that can be all markdown-ified like any other post. Yay markdown!
Tutorial 1 on Relevant Topic in Your Field
Published:
Talk 2 on Relevant Topic in Your Field
Published:
Conference Proceeding talk 3 on Relevant Topic in Your Field
Published:
This is a description of your conference proceedings talk, note the different field in type. You can put anything in this field.
teaching
Teaching experience 1
Undergraduate course, University 1, Department, 2014
This is a description of a teaching experience. You can use markdown like any other post.
Teaching experience 2
Workshop, University 1, Department, 2015
This is a description of a teaching experience. You can use markdown like any other post.