arXiv subjects

Kristian Lum

Publications and source records attributed to Kristian Lum.

29 records · Page 2Linked to original sources

arXiv2020

Prediction-Based Decisions and Fairness: A Catalogue of Choices, Assumptions, and Definitions

A recent flurry of research activity has attempted to quantitatively define "fairness" for decisions based on statistical and machine learning (ML) predictions. The rapid growth of this new field has led to wildly inconsistent terminology and notation, presenting a serious challenge for cataloguing and comparing definitions. This paper attempts to bring much-needed order. First, we explicate the various choices and assumptions made---often implicitly---to justify the use of prediction-based decisions. Next, we show how such choices and assumptions can raise concerns about fairness and we present a notationally consistent catalogue of fairness definitions from the ML literature. In doing so, we offer a concise reference for thinking through the choices, assumptions, and fairness considerations of prediction-based decision systems.

stat.AP↗

arXiv2020

Estimating the number of SARS-CoV-2 infections and the impact of social distancing in the United States

Understanding the number of individuals who have been infected with the novel coronavirus SARS-CoV-2, and the extent to which social distancing policies have been effective at limiting its spread, are critical for effective policy going forward. Here we present estimates of the extent to which confirmed cases in the United States undercount the true number of infections, and analyze how effective social distancing measures have been at mitigating or suppressing the virus. Our analysis uses a Bayesian model of COVID-19 fatalities with a likelihood based on an underlying differential equation model of the epidemic. We provide analysis for four states with significant epidemics: California, Florida, New York, and Washington. Our short-term forecasts suggest that these states may be following somewhat different trajectories for growth of the number of cases and fatalities.

stat.AP↗

arXiv2020

The impact of overbooking on a pre-trial risk assessment tool

Pre-trial risk assessment tools are used to make recommendations to judges about appropriate conditions of pre-trial supervision for people who have been arrested. Increasingly, there is concern about whether these models are operating fairly, including concerns about whether the models' input factors are fair measures of one's criminal activity. In this paper, we assess the impact of booking charges that do not result in a conviction on a popular risk assessment tool, the Arnold Public Safety Assessment. Using data from a pilot run of the tool in San Francisco, CA, we find that booking charges that do not result in a conviction (i.e. charges that are dropped or end in an acquittal) increased the recommended level of pre-trial supervision in around 27% of cases evaluated by the tool

stat.AP↗

arXiv2019

Removing the influence of a group variable in high-dimensional predictive modelling

In many application areas, predictive models are used to support or make important decisions. There is increasing awareness that these models may contain spurious or otherwise undesirable correlations. Such correlations may arise from a variety of sources, including batch effects, systematic measurement errors, or sampling bias. Without explicit adjustment, machine learning algorithms trained using these data can produce poor out-of-sample predictions which propagate these undesirable correlations. We propose a method to pre-process the training data, producing an adjusted dataset that is statistically independent of the nuisance variables with minimum information loss. We develop a conceptually simple approach for creating an adjusted dataset in high-dimensional settings based on a constrained form of matrix decomposition. The resulting dataset can then be used in any predictive algorithm with the guarantee that predictions will be statistically independent of the group variable. We develop a scalable algorithm for implementing the method, along with theory support in the form of independence guarantees and optimality. The method is illustrated on some simulation examples and applied to two case studies: removing machine-specific correlations from brain scan data, and removing race and ethnicity information from a dataset used to predict recidivism. That the motivation for removing undesirable correlations is quite different in the two applications illustrates the broad applicability of our approach.

stat.AP↗

arXiv2017

The causal impact of bail on case outcomes for indigent defendants

We use near-far matching, a technique for estimating causal relationships, to explore whether bail causes a higher likelihood of conviction. We find evidence of a strong causal impact. This paper was compiled as a submission to the 2017 Fairness, Accountability, and Transparency in Machine Learning (FAT ML) workshop.

stat.AP↗

arXiv2017

Theoretical Limits of Record Linkage and Microclustering

There has been substantial recent interest in record linkage, attempting to group the records pertaining to the same entities from a large database lacking unique identifiers. This can be viewed as a type of "microclustering," with few observations per cluster and a very large number of clusters. A variety of methods have been proposed, but there is a lack of literature providing theoretical guarantees on performance. We show that the problem is fundamentally hard from a theoretical perspective, and even in idealized cases, accurate entity resolution is effectively impossible when the number of entities is small relative to the number of records and/or the separation among records from different entities is not extremely large. To characterize the fundamental difficulty, we focus on entity resolution based on multivariate Gaussian mixture models, but our conclusions apply broadly and are supported by simulation studies inspired by human rights applications. These results suggest conservatism in interpretation of the results of record linkage, support collection of additional data to more accurately disambiguate the entities, and motivate a focus on coarser inference. For example, results from a simulation study suggest that sometimes one may obtain accurate results for population size estimation even when fine scale entity resolution is inaccurate.

math.ST↗

arXiv2017

An algorithm for removing sensitive information: application to race-independent recidivism prediction

Predictive modeling is increasingly being employed to assist human decision-makers. One purported advantage of replacing or augmenting human judgment with computer models in high stakes settings-- such as sentencing, hiring, policing, college admissions, and parole decisions-- is the perceived "neutrality" of computers. It is argued that because computer models do not hold personal prejudice, the predictions they produce will be equally free from prejudice. There is growing recognition that employing algorithms does not remove the potential for bias, and can even amplify it if the training data were generated by a process that is itself biased. In this paper, we provide a probabilistic notion of algorithmic bias. We propose a method to eliminate bias from predictive models by removing all information regarding protected variables from the data to which the models will ultimately be trained. Unlike previous work in this area, our framework is general enough to accommodate data on any measurement scale. Motivated by models currently in use in the criminal justice system that inform decisions on pre-trial release and parole, we apply our proposed method to a dataset on the criminal histories of individuals at the time of sentencing to produce "race-neutral" predictions of re-arrest. In the process, we demonstrate that a common approach to creating "race-neutral" models-- omitting race as a covariate-- still results in racially disparate predictions. We then demonstrate that the application of our proposed method to these data removes racial disparities from predictions with minimal impact on predictive accuracy.

stat.AP↗

arXiv2016

A statistical framework for fair predictive algorithms

Predictive modeling is increasingly being employed to assist human decision-makers. One purported advantage of replacing human judgment with computer models in high stakes settings-- such as sentencing, hiring, policing, college admissions, and parole decisions-- is the perceived "neutrality" of computers. It is argued that because computer models do not hold personal prejudice, the predictions they produce will be equally free from prejudice. There is growing recognition that employing algorithms does not remove the potential for bias, and can even amplify it, since training data were inevitably generated by a process that is itself biased. In this paper, we provide a probabilistic definition of algorithmic bias. We propose a method to remove bias from predictive models by removing all information regarding protected variables from the permitted training data. Unlike previous work in this area, our framework is general enough to accommodate arbitrary data types, e.g. binary, continuous, etc. Motivated by models currently in use in the criminal justice system that inform decisions on pre-trial release and paroling, we apply our proposed method to a dataset on the criminal histories of individuals at the time of sentencing to produce "race-neutral" predictions of re-arrest. In the process, we demonstrate that the most common approach to creating "race-neutral" models-- omitting race as a covariate-- still results in racially disparate predictions. We then demonstrate that the application of our proposed method to these data removes racial disparities from predictions with minimal impact on predictive accuracy.

stat.ML↗

arXiv2016

Estimating the observable population size from biased samples: a new approach to population estimation with capture heterogeneity

Capture-recapture methods aim to estimate the size of a closed population on the basis of multiple incomplete enumerations of individuals. In many applications, the individual probability of being recorded is heterogeneous in the population. Previous studies have suggested that it is not possible to reliably estimate the total population size when capture heterogeneity exists. Here we approach population estimation in the presence of capture heterogeneity as a latent length biased nonparametric density estimation problem on the unit interval. We show that in this setting it is generally impossible to estimate the density on the entire unit interval in finite samples, and that estimators of the population size have high and sometimes unbounded risk when the density has significant mass near zero. As an alternative, we propose estimating the population of individuals with capture probability exceeding some threshold. We provide methods for selecting an appropriate threshold, and show that this approach results in estimators with substantially lower risk than estimators of the total population size, with correspondingly smaller uncertainty, even when the parameter of interest is the total population. The alternative paradigm is demonstrated in extensive simulation studies and an application to snowshoe hare multiple recapture data.

stat.ME↗

arXiv2013

An agent-based epidemiological model of incarceration

We build an agent-based model of incarceration based on the SIS model of infectious disease propagation. Our central hypothesis is that the observed racial disparities in incarceration rates between Black and White Americans can be explained as the result of differential sentencing between the two demographic groups. We demonstrate that if incarceration can be spread through a social influence network, then even relatively small differences in sentencing can result in the large disparities in incarceration rates. Controlling for effects of transmissibility, susceptibility, and influence network structure, our model reproduces the observed large disparities in incarceration rates given the differences in sentence lengths for White and Black drug offenders in the United States without extensive parameter tuning. We further establish the suitability of the SIS model as applied to incarceration, as the observed structural patterns of recidivism are an emergent property of the model. In fact, our model shows a remarkably close correspondence with California incarceration data, without requiring any parameter tuning. This work advances efforts to combine the theories and methods of epidemiology and criminology.

stat.AP↗

arXiv2012

Bayesian variable selection for spatially dependent generalized linear models

Despite the abundance of methods for variable selection and accommodating spatial structure in regression models, there is little precedent for incorporating spatial dependence in covariate inclusion probabilities for regionally varying regression models. The lone existing approach is limited by difficult computation and the requirement that the spatial dependence be represented on a lattice, making this method inappropriate for areal models with irregular structures that often arise in ecology, epidemiology, and the social sciences. Here we present a novel method for spatial variable selection in areal generalized linear models that can accommodate arbitrary spatial structures and works with a broad subset of GLM likelihoods. The method uses a latent probit model with a spatial dependence structure where the binary response is taken as a covariate inclusion indicator for area-specific GLMs. The covariate inclusion indicators arise via thresholding of latent standard normals on which we place a conditionally autoregressive prior. We propose an efficient MCMC algorithm for computation that is entirely conjugate in any model with a conditionally Gaussian representation of the likelihood, thereby encompassing logistic, probit, multinomial probit and logit, Gaussian, and negative binomial regressions through the use of existing data augmentation methods. We demonstrate superior parameter recovery and prediction in simulation studies as well as in applications to geographic voting patterns and population estimation. Though the method is very broadly applicable, we note in particular that prior to this work, spatial population estimation/capture-recapture models allowing for varying list dependence structures has not been possible.

stat.ME↗