Blog

Overview

In recent years, algorithmic-driven policies and recommendations have become a ubiquitous and integral part of our society, from online shopping to job screening. Driven in part by this transformative change, the academic literature on optimal policy learning has flourished. The increasing availability of granular data about individuals at scale means that the use of these new methodologies will only continue to grow.

This investigatory study focuses on safe policy learning for pre-trial risk assessments. The study defines policies as “safe” if they do not lead to worse outcomes than the status quo on average. Pre-trial risk assessment instruments (RAIs) may help judges decide whether to release with different levels of conditions or detain an arrested individual before trial by providing simple scores classifying the risk of their failing to appear at court or committing a crime. Among other reasons, because arrestees are presumed innocent, it is important to avoid unnecessary incarceration. Pre-trial RAIs have become increasingly prominent in criminal proceedings, used in at least 23 states and affecting the lives of thousands every day. 

The study proposes that existing studies of RAIs have been asking the wrong question: they focus on how RAIs have impacted the rate of failure to appear or pre-trial crime, when they should investigate how RAIs improve or worsen judges’ decision-making. Unlike a pill, the intervention – the RAI in this case – is not given to the person whose behavior it is meant to change, but rather a third party, the judge. Thus, the effectiveness of the intervention must be measured by changes in judicial decisions. But to evaluate such changes in a randomized controlled trial would require randomizing RAI scores given to judges and measuring their responses, a prohibited act as it entails deliberately providing judges wrong information. 

How can we learn and create better policies on algorithmic assessment? Unfortunately, prior methods for policy learning are not applicable because they require existing policies to be “stochastic”  (meaning they vary randomly) rather than deterministic (meaning each set of inputs produces only one output value). This study develops a robust optimization approach that partially identifies the expected utility of a policy, and then finds an optimal policy by minimizing the “worst-case regret” (meaning the lowest possible utility). The resulting policy is conservative but has a statistical safety guarantee, allowing the policy-maker to limit the probability of producing a worse outcome than the existing policy. The researchers extend this approach to common and important settings where humans make decisions with the aid of algorithmic recommendations, such as judges using RAIs. We derive new classification and recommendation rules that retain the transparency and interpretability of the existing instrument while potentially leading to better overall outcomes at a lower cost.

Pre-trial Risk Assessment

The study’s methodology is motivated by a popular pre-trial risk assessment instrument called the Public Safety Assessment (PSA), used in Dane County, Wisconsin, along with many other places.

The PSA consists of classification scores based on the risk that each arrestee will engage in three types of risky behavior: (i) failing to appear in court (FTA), (ii) committing a new criminal activity (NCA), and (iii) committing a new violent criminal activity (NVCA). Judges abalance between these risks and the cost of incarceration when making their pre-trial release decisions.

The PSA consists of separate scores for FTA, NCA, and NVCA risks, based on 9 risk factors. Importantly, the only demographic factor used is the age of an arrestee, and other characteristics such as gender and race are not used. The other risk factors include the current offense and pending charges as well as criminal history, which is based on prior convictions and prior FTA. Each of these scores is constructed by taking a linear combination of underlying risk factors and thresholding the integer-weighted sum. For the sake of transparency, policy makers have made these weights and thresholds publicly available (see https://advancingpretrial.org/psa/factors).

Developing a Safe Policy 

The study’s primary goal is to construct new algorithmic scoring and recommendation rules that could lead to a higher overall expected utility than the status quo rules, while retaining a high level of transparency, interpretability, and robustness. While there are many factors besides the risk assessment instruments that affect the judge’s decision and the arrestee’s behavior, the study focuses on changing the existing algorithms rather than the other factors.

The researchers develop their optimal safe policy approach in two parts. To construct a safe policy in the population, i.e., with an infinite number of samples, they analyze the population optimization problem that constructs a safe policy. Then, they give concrete examples to build intuition before describing our methodology in greater generality. Finally, they develop several theoretical properties of our approach. 

Because we do not have access to an infinite amount of data, we cannot compute the population safe policy. Instead, the researchers show how to learn an empirical safe policy from observed data of finite sample size.

Applying the Safe Policy to Pre-trial Risk Assessment

The study applied this optimization methodology to the PSA. Given the highly technical nature of the findings, they will not be summarized in detail here. However, they are relevant for policy makers and technologists as they suggest an opportunity for improving the existing scoring rules.

Future Research

There are several avenues for future research. First, on implementation choices under the proposed approach: while we consider several representative cases, there are many other structural assumptions that would lead to different forms of extrapolation. 

Second, we can use similar statistical tools to create tests of safety for given policies. By creating a worst-case upper bound on the regret of a policy relative to the status quo, we can test for whether the proposed new policy is an improvement over the existing status quo.

Third, there are many ways in which optimal algorithmic recommendations may differ when considering long term societal outcomes rather than short term ones. For example, pre-trial detention driven by a risk assessment recommendation may alter the long term behavior and welfare of an arrestee. Understanding how to design algorithms when they affect long-term future outcomes is key to ensuring that recommendations do not take a myopic view. One potential way to incorporate long term outcomes may be with the use of surrogate measures. 

Finally, within the robust optimization framework, the notion of “safety” can be considerably expanded. In this paper, we consider policies to be safe if they do not lead to worse outcomes on average; however, this does not guarantee that outcomes are not worse for subgroups. 

The rise in court-hosted online dispute resolution (“ODR”) is noticeable. But does it work? It seems intuitive that if courts are widely deploying ODR, touted to improve access to justice by removing the inconvenience of—some argue any necessity to—having to visit a physical courthouse to participate in one’s case, we would find strong evidence to support such wide adoption.

That strong evidence base does not currently exist.[1] In fact, no credible evidence as to ODR’s effectiveness, one way or another, exists. We tried ourselves to investigate whether ODR works. Unfortunately, we still don’t know.

If you are reading this post, you likely know already that at the Access to Justice Lab at Harvard Law School, we focus on credible, meaning almost always randomized, evaluation to understand the direct causal impact of those interventions using the randomized control trial.[2] In this post, we summarize two evaluations completed in partnership with the Iowa Judicial Branch and the 11^th Judicial Circuit of Florida, invaluable partners.

In the coming months, you can watch for a longer publication going into more detail. Here we will proceed as follows: first we will remind readers of why randomization is an important component to a credible evaluation design. Then we will summarize our two studies, which differ in goals and design, but neither of which tested whether the key question of whether ODR (versus no ODR) affects relevant outcomes. Finally, we will end with lessons learned and opportunities for future research/court partnerships.

The Importance of Randomization in Evaluation Designs
Why are we so committed to evaluations incorporating randomization? Randomized studies select groups statistically identical to one another except for that one is not exposed to an intervention or program (here, the availability of an ODR platform), allowing us to know with as much certainty as our current knowledge systems allow that the reason for any observed differences in outcomes is the intervention or program. By contrast, a commonly used methodology that compares outcomes prior to an intervention’s implementation to outcomes after the implementation could be rendered of little value by changes, fast or evolutionary, occurring at about the same time as the intervention. Such factors might include a change in presiding judge, a new crop of mediators or lawyers working on these cases, a change in mechanism to access the court such as by phone or synchronous or synchronous online interaction (uncoincidentally, similar to how ODR works), a change in filing fee amount, a change in way cases process through the court, or change in thinking among members of the bar regarding what is trial worthy and what is better to settle. The gold-standard RCT neutralizes these potentially influencing factors as much as we currently know how to neutralize them.

It was the Best of Times, it was the Worst of Time: A Tale of Two Studies
As suggested above, the threshold question of whether ODR works or not is not yet answered. Before we understand what components of ODR make it better or worse, we need to know if the concept overall works. We went into these evaluations hoping to investigate that important first inquiry. We came closer in Florida.

Florida
In Florida we attempted to use an encouragement design.[3] We sought to answer the question of whether providing encouragement to use an ODR platform to resolve traffic compliance matters results in more use of the platform and, if so, whether those that use the platform experience better outcomes as compared to those who do not. The hope was that people receiving encouragement would do the thing we encouraged them to do at much higher rates than those who did not receive the encouragement. If that hope had been realized, then by randomizing encouragement to use the intervention—giving encouragement to some and not others, randomly selected—we would have effectively been randomizing the intervention.[4]

Encouragement came in the form of a postcard. Individuals with eligible alleged traffic infractions were randomly assigned to receive this encouragement or not. Nothing else about their case changed: law enforcement still issued the same citations, cases were still scheduled as a matter of course with the court, and they proceeded if no other action was taken prior to that scheduled date (such as paying the ticket or using the ODR platform to show remediation of noncompliance). We ended with 289 study participants getting the encouragement and 274 not.

Encouragement designs work only if the encouragements . . . er . . . encourage. Our postcard didn’t. We weren’t terribly optimistic that it would. But we were unable to persuade stakeholders to adopt a stronger design.

In other words, our two groups used the ODR platform at nearly the same rate, meaning we are not able to untangle the effect of ODR itself from other possible outcomes.

What we did see in our data is a possible reminder effect[5] to a large degree. We saw those that received the encouragement postcard were more likely to appear at their subsequent court events at a rate of twelve percentage points. And, perhaps a product of that appearance, those receiving the postcard were more likely to resolve their case at a rate of twenty-five percentage points. Previous researchers have observed reminder effects, usually with postcards or text messages about hearings, but that those have tended to be in the five-eight percent range.[6] The fact that we’re observing larger effects leads us to hypothesize that the combination of a reminder and a convenient method of resolution may be larger than the reminder effect alone.

What we think we see in this data is a lot of people were able to access the ODR platform and did so notwithstanding encouragement. But, subsequently, those that received the encouragement paid more attention to their cases than those who did not. We cannot be definitive about this finding. This is not the outcome for which we were testing. But, it is a hypothesis that emerged and deserves more attention from future research.

Iowa
The Iowa study differed in goals and design. The Iowa Judicial Branch deployed the ODR platform, for the purposes of this evaluation, for a handful of pre-selected traffic infractions. The Iowa Judicial Branch agreed to randomize neither access to the platform nor encouragement to use it. Instead, the Branch agreed to randomize information about payment plans available via the court system as well as information about what prosecutors ordinarily negotiate. Randomization would occur for litigants who created accounts to use the platform.

In the case of information about payment plans, that information was not a secret but also was not openly available, either. Usually it required a litigant to affirmatively ask for a payment plan rather than being an option the litigant could select.

In the case of reasonable expectations, the idea was that giving the litigant some information about what a prosecutor might offer in a plea might help the litigant to decide to pursue negotiation or resolve the matter more quickly (presumable by paying or expediting a not guilty plea).

We were not particularly optimistic that randomization of information of this type would produce much of a treatment contrast. As it turned out, the issue was not the treatment contrast but the platform. During the several months of enrollment, no one used it. Actually, over the course of enrollment, only one participant successfully made it through the platform. Volume did tick up slightly after enrollment closed, but remained too low for an evaluation.

Lessons Learned and Opportunities
Some themes emerged from both evaluation attempts which we think are useful to courts as they move to ODR 2.0. We will use this section to highlight some. Not all applied in each jurisdiction; some applied in both.

Informing the User about the Option
It seemed that courts did not make substantial efforts to inform the user about the option to use ODR. Neither of our evaluation jurisdictions implemented a program mandating the use of ODR to resolve the selected use, which likely would de facto serve to inform the user about the platform. Instead, both made use of ODR opt-in. Litigants cannot opt into something unless they know it exists. In traffic matters, most courts attempt to alert litigants to the existence of ODR by including text about it on citation forms. In most jurisdiction that we have observed, in this study and others, citations are not a good forum for notifying anyone of anything important. They are packed full of dense text with numerous unintelligible statutory citations. The notification of the ODR option amounts to a URL that may not intuitively appear connected to the court. Notwithstanding the URL and the option to resolve one’s case without attending a court hearing, the citation has a court date at which, the citation says, the alleged offender is compelled to appear. It is easy to see how litigants may get confused and/or disregard the ODR option.

Expanding Eligibility Guidelines
In implementing ODR, some courts fear that a deluge of users will flood the platform, making the platform cumbersome for the court to manage. This fear can result in narrowly drawing eligibility guidelines. That could be a method to thin the herd of users. But attempts to clarify who can and cannot use an ODR platform may also confuse potential users, resulting in some eligible users concluding that they cannot use the platform and some ineligible users concluding the opposite. When determining eligibility for ODR, and perhaps any program, we should consider what will make logical sense to the user. Particularly given the Iowa experience, broad rather than narrow eligibility guidelines may help to increase usage.

Clear Timelines
Courts often develop processes around the idea that judicial time should not be wasted. Likely the process discussed above of providing in the citation both information about ODR while also scheduling the next hearing is to preserve judicial time and to keep cases moving. Internal processes around deadlines for users to complete ODR developed with likely the same goal in mind. However, these ODR-specific internal deadlines are not, as far as we were able to observe, communicated to the user, and they differed from the live-appearance deadlines. This results in individuals otherwise eligible to use the platform being unable to do so because an undisclosed deadline for platform use has passed. After the passing of this deadline, the only option is to appear in court to resolve the matter. ODR-specific deadlines should be communicated. Indeed, we wonder what harm there is in either letting someone resolve their case using the platform right up until the day of the hearing, or only scheduling a hearing after a disclosed deadline passes, with the latter option seeming to accommodate both the robust use of the platform and the preservation of judicial time.

Increasing Access to Justice with Simplification and Reminders
Not everything went as we had hoped in these two evaluations. We came away with some thoughts on improvement of court installed ODR platforms. We also came away with a hypothesis that could be a boon to the access to justice community. The idea that combining tools that are separately thought of as improving access to justice (here, reminders and a tool to ease case resolution) might work. The takeaway here is that simplified processes coupled with reminders may create a significant increase in usage of tools designed, we think, to improve access to justice. We are looking forward to testing, and observing others test, this hypothesis in the future.

Support for this project was provided in part by The Pew Charitable Trusts. The views expressed herein are those of the author(s) and do not necessarily reflect the views of The Pew Charitable Trusts.

[1] There are a handful of empirical studies that investigated user experiences in ODR processes. See, Martin Gramatikov & Laura Klaming, Getting Divorced Online: Procedural and Outcome Justice in Online Divorce Mediation, 14 J.L. & FAM. STUD. 97, 117 – 8 (2012) (“finding high levels of satisfaction with online divorce procedures and quality of outcomes of both male and female divorcees in the Netherlands, although the former focused more on monetary and time costs while the latter focused on negative emotions”); See, Katalien Bollen & Martin Euwema, The Role of Hierarchy in Face-to-Face and E-Supported Mediations: The Use of an Online Intake to Balance the Influence of Hierarchy, 6 Negotiation and Conflict Management Research 4:305 – 19, 313 (2013) (“finding that a hybrid process combining online intake with face-to-face mediation had an equalizing effect in hierarchical labor settings on parties’ fairness and satisfaction perceptions”); See, Marc Mason & Avrom H. Sherr, Evaluation of the Small Claims Online Dispute Resolution Pilot, Institute of Advanced Legal Studies, at 19 (Sept. 1, 2008), (finding a lower settlement rate than offline small claims mediations, as well as problems such as the online system timing out, the registration process, spam filtering, a lack of transparency, and digital access and competency, although the study was limited in scope and only had a sample size of 25 cases in the UK); See, Laura Klaming, Jelle van Veenen and Ronald Leenes, I Want the Opposite of What You Want: Reducing Fixed-pie Perceptions in Online Negotiations, 2009 J. DISP. RESOL. 139:85 – 94, 92 – 93 (finding that “providing negotiators with incentives independent from the resources that have to be divided, as well as providing them with information about the opponent’s preferences, led to more agreements”); See, Udechukwu Ojiako et al., An Examination of the ‘Rule of Law’ and ‘Justice’ Implications in Online Dispute Resolution in Construction Projects, 36 International Journal of Project Management 301, 305 , 308(2018) (“finding that the ODR process does not affect parties’ satisfaction with the “rule of law” or “justice” in small claims ODR in construction projects, while suggesting further research on the cultural contexts around these concepts”), Due to the limited scale on which ODR has been implemented in American courts, there are few independent efforts to quantify the outcomes of ODR initiatives in the public sector. “There are some self-reported pre-ODR and post-ODR datasets, mostly compiled by courts and private platforms and unconfirmed by independent research” (See, Joint Technology Committee Resource Bulletin, Case Studies in ODR for courts; A View from the Front Lines, 3 – 18 (2017); Amy Schmitz, Expanding Access to Remedies Through E-Court Initiatives, 67 Buff. L. Rev. 89, 158 (2019); Kevin Bowling, Jennell Challa, & Di Graski, Improving Child Support Enforcement Outcomes with Online Dispute Resolution, Trends in State Courts, 43 – 8, 46 (2019), and Avital Mentovich, J.J. Prescott, & Orna Rabinovich-Einy, Are Litigation Outcome Disparities Inevitable? Courts, Technology, and the Future of Impartiality, 73 Ala. L. Rev. (2020) at 893.

[2] See, Joshua D. Angrist, Instrumental Variables Methods in Experimental Criminological Research: What, Why and How, 2 Journal of Experimental Criminology 23 – 44, 24 (2006) (arguing that randomized studies are considered the gold standard for scientific evidence).

[3] Conner Mullally, Steve Boucher, & Michael Carter, Encouraging Development: Randomized Encouragement Designs in Agriculture, 95 American Journal of Agricultural Economics 5:1352 – 8, 2 (2013) (defining the encouragement design); Paul J. Ferraro, Counterfactual Thinking and Impact Evaluation in Environmental Policy, Environmental Program and Policy Evaluation: Addressing Methodological Challenges. New Directions for Evaluation, 75 – 84, 80 (2009) (suggesting the encouragement design may be appropriate when randomly restricting access to an intervention cannot be done). One assumes Ferraro means literally cannot, or, in other words, ethically impermissible to, rather than a preference not to, randomly restrict access. As an aside, Ferraro’s description of the need for rigorous evaluation in environmental policy is directly analogous to the need for same in the legal sphere. See generally, D. James Greiner & Andrea Matthews, Randomized Control Trials in the United States Legal Profession, Annu. Rev. Law Soc. Sci. 12:295 – 312 (2016).

[4] Id., Mullally et. al., (“Assignment to th ‘encouraged’ group is then used as an instrumental variable in order to estimate the impact of the treatment”).

[5] For more information about reminder effects, See, e.g., Brian H. Bornstein, et. al., Reducing Courts’ Failure to Appear Rate: A Procedural Justice Approach, (May 2011), https://www.ojp.gov/pdffiles1/nij/grants/234370.pdf (evaluating the effectiveness of different messaging approaches in mailed post cards); Timothy R. Schnacke, et. al., Increasing Court Appearance Rates and Other Benefits of Live-Caller Telephone Court-Date Reminders: The Jefferson County, Colorado, FTA Pilot Project and Resulting Court Date Notification Program, 393 Ct. Rev.: J. Am. Judges 86 (2012) (evaluating the effectiveness of providing information about the consequences of failing to appear by live-calling individuals); Christopher T. Lowenkamp, et. al., Assessing the Effects of Court Date Notifications within Pretrial Case Processing, 43(2) Am. J. Crim. Just. 167, 173 (2017) (evaluated the effectiveness of different messaging approaches and different methods of delivering notifications); Brice Cooke, et. al., Using Behavioral Science to Improve Criminal Justice Outcomes: Preventing Failures to Appear in Court, (January 2018), https://www.prisonpolicy.org/scans/Using_Behavioral_Science_to_Improve_Crimina_Justice_Outcomes_Cooke_et_al_2018.pdf (evaluated the effectiveness of the timing and messaging approaches in text notifications); Stephen H. Taplin, et. al., Testing Reminder and Motivational Telephone Calls to Increase Screening Mammography: A Randomized Study, 92(3) J. of the Nat’l Cancer Inst. 233 (2000) (finding reminders to be as efficacious as addressing barriers with phone call reminders performing better than postcards); Susan Maxwell, et. al., Effectiveness of Reminder Systems on Appointment Adherence Rates, 12(4) J. of Health Care for the Poor and Underserved 504, 508 (2001) (finding show rates for appointments for those who received no reminder to be 49.9% as compared to those that received a mailed reminder to be 52.1%); Mary Elaine Koren, et. al., Interventions to Improve Patient Appointments in an Ambulatory Care Facility, 15(4) J. Ambulatory Care Mgmt. 76 (1994) (finding insignificant difference between the type of reminder, when using phone or mailings, but did find some reminder to be more effective than no reminder).

[6] See, Id.