The liberal dream of smart detention? Algorithms and the politics of pretrial detention in the US states

In the 2000s and 2010s, US states have seen an important wave of change in criminal justice policies toward a “ smart on crime ” approach. In this context, several states have rolled out algorithmic risk assessment tools for statewide use in pretrial decisions, whereas some others have not, and still others are moving back from using such tools again. The present article examines the explanations for this variance. To this end, it tests competing expectations about the role of functional pressures, including fiscal strain and the party-political balance of power. The findings show that functional pressures, policy diffusion, and politics affect the likelihood that algorithmic tools will be used in criminal justice. Democratic control of both the state executive and legislative branches increases the likelihood that a state will use these tools, indicating that Republicans are reluctant to leave the “ tough on crime ” paradigm behind and to advance the “ smart on crime ” approach.


Abstract
In the 2000s and 2010s, US states have seen an important wave of change in criminal justice policies toward a "smart on crime" approach. In this context, several states have rolled out algorithmic risk assessment tools for statewide use in pretrial decisions, whereas some others have not, and still others are moving back from using such tools again. The present article examines the explanations for this variance. To this end, it tests competing expectations about the role of functional pressures, including fiscal strain and the party-political balance of power. The findings show that functional pressures, policy diffusion, and politics affect the likelihood that algorithmic tools will be used in criminal justice. Democratic control of both the state executive and legislative branches increases the likelihood that a state will use these tools, indicating that Republicans are reluctant to leave the "tough on crime" paradigm behind and to advance the "smart on crime" approach.

| INTRODUCTION
The United States has a long tradition of using risk assessments in criminal justice (Taxman & Dezember, 2016) in order to inform decision-making at the pretrial stage, in corrections, and in sentencing. Although the adoption of actuarial tools for assessing risk in the US criminal justice system began in the 1930s (Harcourt, 2007), there have been intensified efforts to advance the use of computerized tools since the early 2000s as part of a broader trend toward justice reform and "smart on crime" strategies (König & Krafft, 2021;Mayson, 2018;Percival, 2016). This has been observable especially with regard to pretrial release decisions. Not only have many counties adopted or revised pretrial risk assessment instruments, but some states have moved toward a statewide and standardized implementation of such tools-and some much earlier than others (for the situation in 2019, see Figure 1). More recently, however, several states have halted the roll-out of tools (e.g., Iowa and Idaho) or have even repealed the use of pretrial assessment tools (PRATs) that had already been introduced (e.g., in Alaska in 2020 and in Utah in 2021) 1 (Hutson, 2021;König & Wenzelburger, 2021).
Given the variance over time across US states, the present article addresses the question of how these differences in criminal justice policies can be explained. Hence, we work to shed light on the drivers of a major trend in US criminal justice policy during the 2000s and 2010s. Our article contributes to a broader literature on the ebbs and flows of mass imprisonment in the United States (Enns, 2016;Garland, 2002;Gottfredson & Moriarty, 2006;Gottschalk, 2016). Law and order policy in the United States has been marked by a "tough on crime" stance at least since the late 1960s, with both Democrats and Republicans campaigning on this issue in almost every election campaign since 1968 (Enns, 2016;Miller, 2016;Simon, 2009;Tonry, 2007). However, this policy of "mass incarceration" (Gottschalk, 2006) has led to skyrocketing costs in the criminal justice system and increasingly louder calls for a change of course in criminal justice policy (Dagan & Teles, 2014;Gottschalk, 2010;Green, 2015;Schoenfeld, 2016). These calls have been closely connected to a new paradigm in criminal justice policy. Seemingly transcending the opposition of "soft on crime" and "tough on crime," the more recent "smart on crime" strategy 2 has served as a buzzword for penal reform that enables policy-makers to achieve de-carceration without being seen as lenient on crime and endangering public safety (Altheide & Coyle, 2006;Dagan & Teles, 2016;Fairfax, 2010;Green, 2015;Mayson, 2018;Percival, 2016).
An important prerequisite for adopting the "smart on crime" strategy has been the possibility of relying on more advanced risk-assessment tools as a cornerstone in a broader move toward greater use of actuarial methods in public administration and criminal justice (Coglianese & Ben Dor, 2020;Hannah-Moffat, 2019;Rothschild-Elyassi et al., 2019;Tonry, 2019). The use of PRATs is presumed to enhance evidence-based decision-making because these tools are created via statistical analyses of empirical data intended to identify predictors of pretrial outcomes. Using statistical evidence, PRATs help decision-makers in the justice system to differentiate between offenders with higher or lower risks of re-offending or F I G U R E 1 States that have implemented PRATs for statewide use (2000)(2001)(2002)(2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015)(2016)(2017)(2018)(2019). Alaska repealed the legislation for the consistent state-wide use of a PRAT in 2019. Darker shading represents earlier introduction, with dates of introduction ranging from 2003 (Virginia) to 2018 (e.g., Rhode Island). See Supporting Information Appendix Table A.1 for details. missing their court dates (Stevenson, 2018). To the extent that these tools help to accurately identify low-risk cases, they could contribute to lowering incarceration numbers without jeopardizing public safety. 3 However, while greater use of risk assessment tools may help to alleviate functional and financial pressures on the criminal justice system, it also implies a paradigm shift: Whereas the "tough on crime" paradigm puts hard sanctions and security first, "smart on crime" implies a more fine-grained and technocratic approach that includes public safety and efficiency considerations as well as (disputable) promises to achieve greater objectivity and fairness in decision-making (Fairfax, 2010, p. 597).
Building on these insights, we argue theoretically and explore empirically whether functional pressures and ideational differences in the policy approaches to public safety can explain the variance in the use of PRATs in criminal justice. Regarding functional pressures, we test whether a high public debt burden or a high imprisonment rate are correlated with the decision to use PRATs. We also account for diffusion effects and investigate whether the introduction of PRATs spreads from one state to its neighbors. Regarding policy approaches toward safety, we claim that party politics may make a difference. As Republican politicians and their supporters more strongly endorse retribution and tough sanctions (Jacobs & Helms, 1996;Sutton, 2000), we hold that they will be comparatively less open to the "smart on crime" approach. This expectation contradicts parts of the literature that emphasize the bipartisan stance of the "smart on crime" movement (e.g., Dagan & Teles, 2016), but it does resonate with findings according to which Republican calls for criminal justice reforms have mostly been justified by fiscal considerations and less by concerns about human rights or rehabilitation (Beckett et al., 2016).
We test these possible explanations, drawing on a dataset that covers all 50 US states over the last two decades. While algorithmic risk assessments have also been implemented for posttrial decisions, here we focus on pretrial decisions, for three reasons. First, the pretrial stage has seen a wave of change toward the use of algorithmic tools. Second, risk assessments in the pretrial stage come at a particularly sensitive point in the criminal justice process, as they can lead to detention decisions before a suspect has even been judged in a criminal process (Baughman & McIntyre, 2012). Third, data availability is much better for the pretrial stage, which allowed us to construct an encompassing dataset with statewide PRATs.
Our findings suggest that both functional pressures (such as the number of prisoners or a high public debt burden in a state) and ideological factors drive the adoption of PRATs for statewide use. We observe a higher likelihood of a state implementing a PRAT at a certain point in time if the state is governed by a Democratic executive and when Democrats dominate the legislature. We also find that the partisanship of the governor only matters in the absence of divided government.
The remainder of this article is structured as follows: In Section 2, we will briefly summarize the state of the art and discuss theoretical assumptions that will guide our empirical analysis. Section 3 discusses data and methods, and in Section 4, we present our empirical findings. A final section concludes.
2 | "SMART ON CRIME," CRIMINAL JUSTICE REFORM, AND RISK ASSESSMENTS TOOLS 2.1 | From "tough on crime" to "smart on crime": The state of the art After decades of unhalted increase of incarceration rates in the US, recent years have seen a "climate change" (Karstedt et al., 2019, p. 59), with imprisonment rates gradually declining (for a more critical view, see Gottschalk, 2016;Seeds, 2017). Criminologists have started to investigate the reasons for this change and have emphasized the influence of financial pressures (Gottschalk, 2010), while also noting a general change in ideology that emphasized the need to rely on evidence-based practices. These changes have put the reform of criminal justice systems on the political agenda in the United States since the mid-2000s (Pew Charitable Trusts, 2018Taxman et al., 2014). However, in most of the studies that look into these dynamics, imprisonment rates are at the center of researchers' interest and serve as the dependent variable (Green, 2015;Karstedt et al., 2019;Petersilia & Cullen, 2015). While understanding the forces that may have reduced incarceration is certainly important for understanding the general trend, incarceration is a very broad measure of punitiveness (Frost, 2008) and does not tell us much about the political dynamics that have brought these changes about.
In fact, the empirical finding of a general trend toward de-carceration in the United States becomes fuzzier when looking at legislative action and the discursive justifications for the relevant policy changes (Beckett et al., 2016;Gottschalk, 2016;Seeds, 2017). Regarding policy changes, observers have pointed out that the decrease in imprisonment rates is mainly caused by a bifurcation of criminal justice policies in the United States, with a significant difference between violent and nonviolent offenses (Gottschalk, 2016, pp. 165-195). In a similar vein, in their study of legislative action and discourse around penal reform, Beckett et al. (2016) emphasize the partly contradictory patterns in reforms and highlight the fact that most reform initiatives have been directed at drug offenders (Beckett et al., 2016). These findings call for a more nuanced analysis of the elements linked to the policies of criminal justice reform, and especially those that may differentiate a broader strategy of penal reform from a bifurcation strategy, which does reduce incarceration rates for some groups but not for all of them (Takei, 2017).
Our study of the adoption of statewide algorithm-based PRATs contributes to such a nuanced understanding in three respects. First, we do not focus on broad measures such as imprisonment rates but analyze an actual policy change: the state-wide 4 roll out (or repeal) of PRATs. The introduction of such tools is mostly linked to legislative bills, but they are sometimes introduced by the executive. Second, our focus on actuarial risk-assessment tools zeros in on reforms that are not directly related to specific offender groups. This is true to the extent that the recommendations of data-driven actuarial tools are rooted in criteria that are not directly visible to policy actors. With algorithmic tools, the relative importance of such criteria or risk factors, such as criminal history or substance abuse, are first estimated by statistically analyzing a sample of previously released defendants with known pretrial outcomes (usually, being charged with a new crime or to appear in court). One can then calculate the risk of pretrial failure for new defendants by adding the weighted risk factors obtained from the statistical analysis. Compared to legislation that changes an individual measure for a clearly defined group, such as drug addicts, the use of risk assessment tools therefore amounts to a more general change in criminal justice policy.
Third, we focus on pretrial risk assessment tools, as they are a highly relevant example when studying the dynamics around the deployment of "smart justice" policies in the United States. On the one hand, pretrial decisions about detention of suspects are very clearly linked to public safety considerations, because danger to public safety (together with flight risk) is a crucial criterion that can justify detention of suspects, and algorithmic tools are used to identify such highrisk cases. This also means that politically, any "false decision" based on a PRAT may lead to a highly visible crime, which could then be used in political attacks against those lawmakers that have opted for the introduction of the tool (König & Wenzelburger, 2021). On the other hand, the central promise of PRAT is to reduce pretrial detention rates through the identification and release of defendants who pose a low risk of failure to appear and low danger to public safety. Importantly, PRATs are concerned with a specific decision, namely about whether to detain someone who is yet to be judged. This differs from parole decisions, for example, in which a person has already been sentenced. In terms of criminal law concepts, this is a cardinal difference, since basic civil rights such as the presumption of innocence weigh heavily during the pretrial stage (Baradaran, 2011). This is also one reason why pretrial detention and bail reform have been so hotly debated by legal scholars since dangerousness became an important criterion in judicial decision-making (Baughman, 2019;Baughman & McIntyre, 2012;Howard, 1989;van Brunt & Bowman, 2018).
Altogether, risk assessments are not only about public safety, but also about reducing pretrial detention, which has been rising rapidly since the 1990s (by 72%; Baughman, 2018, p. 4). They aim to keep people from going to jail simply because they cannot post bail, or to avoid detaining people for whom detention is an unnecessary measure-a measure that potentially reduces people's integration in society and even increases their risk of engaging in criminal activity. This may make risk assessments appealing to political actors who favor rehabilitative policy, and it very directly links them to the "smart on crime" approach as a way to avoid overcriminalization (Fairfax, 2010). The adoption of PRATs is thus ambiguous with regard to the reasons with which one might justify it, and the preceding considerations make the adoption of PRATs in the US states a particularly relevant object of study, specifically in the context of the "smart on crime" paradigm, which entails a more differentiated approach.

| Theoretical assumptions
To explain what may account for the variance in the use PRATs in the US states, we build on three strands of the literature: theoretical assumptions of policy studies, the more recent literature on criminal justice reform policies, and the insights of research on mass incarceration in the United States. We discuss three main theoretical factors from this literature: functional pressures, policy diffusion, and ideas.
Functional pressures have long been regarded as major drivers of policy change, as they open windows for substantial departures from the existing trajectories. Such pressures can arise from structural transformations of societies or economies (Knill & Tosun, 2012, pp. 71-73). They can result from focusing events or the gradual deterioration of a situation (Kingdon, 2003), and they can increase the likelihood of a policy change in various policy areas (Aldrich et al., 2019;Kuipers, 2006;Vis, 2009). For penal policies, the increase in the violent crime rates in the 1970s, for instance, contributed to the punitive turn in the United States (Enns, 2016;Garland, 2002;Miller, 2016).
When trying to explain the more recent turn toward a "smart on crime" approach, and specifically the use of PRATs, an important strand of the literature argues that prison overcrowding, the related financial burden, and the aim to cut budget deficits were important forces spurring the debate about penal policy reform (Beckett et al., 2016;Brown, 2013;Gottschalk, 2010). Although it is disputed whether budgetary constraints have actually served the liberal ideas of ending mass incarceration or whether they have only generated new fractures and inequalities (Aviram, 2015;Takei, 2017), it is nevertheless clear that budgetary concerns count as one of the major drivers of the "smart on crime" movement (e.g., Wroblewski & Hoffman, 2015). Accordingly, we expect that the use of PRATs is partly driven by the hope of achieving more cost-efficient pretrial decisions, and that the use of PRATs is therefore more likely if functional pressures on the criminal justice system are high. Accordingly, we formulate the following hypothesis.
Hypothesis H1. The stronger the functional pressures on a state's criminal justice system, the more likely that state will be to use a statewide PRAT.
The public policy literature -and particularly the literature studying policy variance across US states-has described policy diffusion as a key driver of policy change (e.g., Berry & Berry, 1990, 1999Matisoff, 2008;Nicholson-Crotty, 2009;Volden, 2016). Numerous studies have shown that different mechanisms of diffusion-such as inter-state competition, learning, or coercion-affect policymakers' choices, and especially so in federal states such as the United States (Gilardi, 2015;Shipan & Volden, 2008). Research specifically on criminal justice policy has also recurrently demonstrated the relevance of policy diffusion between states (Bergin, 2010, p. 408;Nicholson-Crotty, 2004), with geographic proximity being one of the main drivers of policy diffusion. Theoretically, the introduction of PRATs as a new instrument within the criminal justice system is a good candidate for the study of policy diffusion, since policy innovation is a core element of diffusion processes (Berry & Berry, 1990, 1992Rogers, 2010). Given that PRATs are widely considered to be innovative, we therefore expect the following: Hypothesis H2. The likelihood that a state will introduce a PRAT will increase if a neighboring state is already using one.
While functional pressures and diffusion are clearly important drivers of policy change, it is evident that the use of risk assessment instruments is not simply a neutral technical solution to the financial burden of overcrowded prisons. Criminal justice policies always carry ideas concerning how to react to crime and deviant behavior and how to balance rehabilitative goals with public safety concerns (Loader & Sparks, 2016). From the literature about the ideas behind recent "smart on crime" policies, one can distill several arguments to generate expectations about how ideology and partisan politics may affect the decision to introduce a PRAT.
The first important argument sees "smart on crime" reforms following from a bipartisan consensus between Republicans and Democrats (Dagan & Teles, 2016;Petersilia & Cullen, 2015). Percival (2016) has convincingly shown that the coalition of political actors advocating for the smart-on-crime reforms was made up of people from both the Republican and Democratic parties. As Petersilia and Cullen note, "[w]hat had changed, then, is … a way of thinking about incarceration. For so long, mass imprisonment had been the governing policy of corrections-as book after book detailed. But seemingly overnight, its hegemony was shattered, and downsizing quickly emerged as its replacement" (Petersilia & Cullen, 2015, pp. 6-7). Thus, this perspective emphasizes the bipartisan character of the "smart on crime" movement, arguing that both Republicans and Democrats pushed similarly for penal reform.
However, in-depth studies of the political dynamics that have preceded the fundamental shift to the "smart on crime" approach indicate that the forces that led to the turn were more complex. As Percival's analysis shows, support for criminal justice reform came from very different ideological directions, including moral arguments put forward by evangelical Christians in the GOP, efficiency-based arguments on how to spend taxpayer money in a better way, and liberal ideas emphasizing human rights and the need for rehabilitation (Percival, 2016;similarly: Petersilia & Cullen, 2015, p. 9). In a similar vein, the analysis of the discourse around reform by Beckett et al. (2016, p. 250) also illustrates that the preponderance of conservative commentators in newspaper stories about criminal justice reform has not led to a very significant broadening of the conversation about the problems associated with mass incarceration: the fiscal costs associated with mass incarceration were routinely emphasized over and above other considerations.… Interestingly, our findings do not support the speculation that conservatives are raising fundamental questions about sentence length and rehabilitation by introducing quasi-religious concepts such as forgiveness and redemption, as these concepts were quite rarely discussed in newspaper stories about reform.… These findings provide additional reason to suspect that the involvement of conservatives in conversations about criminal justice reform has not led to recognition of the need for comprehensive criminal justice reform. In fact, discussions of more far-reaching reform strategies that would modify punishments for offenses that trigger long prison sentences are exceedingly rare. This assessment, according to which the bipartisan consensus about reform has been built on rather different ideological perspectives, is corroborated by further studies such as Aviram's (2015, p. 73) analysis of penal reform policies, which illustrates that fiscal considerations are often more relevant than humanitarian concerns. Importantly, these studies point to a continuing difference between Republicans and Democrats on the question of criminal justice reform. Republicans appear to see the shift toward less incarceration mostly as a strategy to save money and to more efficiently allocate resources (Barkow, 2005;Fan, 2012), which is why critics question whether the politics of mass incarceration have truly seen a qualitative change (Gottschalk, 2016;Takei, 2017, p. 170).
This discussion suggests that while reforms may at first glance appear to build on a bipartisan consensus, the ideas that animate political actors to engage in these reform efforts differ. Whereas Democrats are more likely to see these reforms as pushing forward more liberal and progressive ideals of rehabilitation and civil rights, while acknowledging budgetary benefits, Republicans are mainly swayed by the latter. In this sense, the question of whether to use risk assessment tools to assist pretrial decisions can be linked to broader ideological questions regarding penal policy rooted in liberal versus conservative values (Wenzelburger, 2020), which come to the fore in the very contested issue of pretrial detention decisions. 5 Liberal and rehabilitative views are especially likely to resonate with the use of PRATs because, as mentioned above, pretrial decisions concern the detention of people who have not even been to trial yet, which makes possible deprivation of liberty a sensitive issue. Accordingly, we would expect that beyond the role of functional pressures, ideology also plays a role in PRAT adoption. The observation that major differences between liberals and conservatives remain (Eaglin, 2013;Levin, 2018) is also corroborated by systematic analyses of the content of legislation. Brown's analysis of correctional reform initiatives in state legislatures finds, for instance, that a higher number of Republican seats in state legislatures significantly reduced the probability of reforms directed at decarceration (Brown, 2013). Based on these considerations, our third hypothesis proposes that the statewide use of PRATs depends on the party that holds executive power in a given state: Hypothesis H3. The greater the executive power of Democrats in a state, the more likely is the statewide use of a pretrial risk assessment tool.
One qualification of this hypothesis is in order, however. While it seems straightforward that executive power matters most for statewide use of PRATs-with some states rolling out tools without or before specific legislation was enacted-a divided government with House and/or Senate opposing the executive could hinder policy change. Hence, if there are too many actors who want to block such reforms or are not willing to carry them through, policy change may not come to pass. We therefore qualify our third hypothesis with regard to such legislative veto powers and formulate H4.
Hypothesis H4. The influence of Democratic dominance of the executive on the probability of statewide adoption of PRATs will be weaker if the legislature is dominated by Republicans.

| DATA, MEASURES, AND METHODS
To test our theoretical claims, we draw on a dataset that covers all 50 US states over a period from 2000 to the end of 2020. We start in 2000, the year in which the first experimental phase for the introduction of a computerized PRAT, in the Commonwealth of Virginia, ended, and a state-wide rollout would have been enviseageable. The PRAT was eventually put in place in 2003, making Virginia the first US state using a statewide tool. In legal terms, the Pretrial Services Act of 1995 was the legislative foundation for the use of risk assessments in pretrial decisions. With Virginia as an anchoring point for the observation period in which the implementation of statewide PRATs can meaningfully be studied, we choose 2000 as the start; however, we also perform a robustness check with 1995 as the start date. 6 Starting in 2000, our dependent variable therefore takes a value of 0 in all years in which no state-wide PRAT was in use in a given state, and a value of 1 in all years in which a PRAT was in use. The resulting data matrix therefore has a panel structure with 1050 state-years (21 years Â 50 states) as units of analysis, and a binary dependent variable that is coded 0 in 93% of the cases (N = 1445), with 7% (N = 105) being coded as 1. Coding all years in which an implemented statewide risk assessment tool is present as 1-and not just the year of introduction-is theoretically sensible given that in some cases these tools have also been subsequently abolished.
We focus on the date of the statewide roll-out of PRATs since this reflects that a crucial policy change has been put into effect, with or without explicit backing through legislation (for an overview, see Supporting Information Appendix Table A.1). Some states have passed bills that foresee the adoption of risk assessments in the pretrial stage and have seen the implementation of a statewide tool in the same year or the year after. However, the presence of a legislative basis and/or the piloting of a tool do not necessarily mean that statewide implementation is inevitable. In the case of Iowa, for instance, a piloted statewide tool was halted by the governor after political resistance emerged. Iowa had already been piloting the Public Safety Assessment, which was intended for statewide use. After strong political opposition to the project and a budget bill in 2018 stipulating its termination, the governor chose to discontinue the pilot at the end of the year. In other cases (e.g., Ohio and Utah), a tool was implemented statewide before legislation was passed. Our dependent variable therefore captures both instances: cases in which a state passed legislation for the adoption of a PRAT and then rolled it out statewide, and cases in which a PRAT was implemented statewide without such legislation. The roll-out dates for statewide tools were compiled through a search of publicly available reports and documents (further details on the coding and the list of sources can be found in Supporting Information Appendix Tables A.1 and A.2).
In order to analyze our time-series-cross-sectional data structure with a binary dependent variable, we estimate panel logit (binary time-series cross-section, BTSCS) regressions. This is a common technique for this kind of data and has been discussed widely in the methodological literature in conflict studies, where binary coded dependent variables (war and peacedyads) are also common (Beck et al., 1998;Beck & Katz, 2001;Carter & Signorino, 2010;Green et al., 2001). Logistic regression analysis with pooled data faces the same challenges as linear TCSC-regression, including heteroscedasticity, serial correlation, and heterogeneity. The econometric solutions for these problems, however, differ from the ordinary-least-squares (OLS)-variant due to the binary nature of the outcome.
We deal with heteroscedasticity by estimating robust standard errors (Huber-White). To correct for serial correlation, Beck et al. (1998) suggest introducing a series of dummy variables. We follow this recommendation. A possible way of dealing with cross-sectional heterogeneity would be to include unit fixed effects and estimate a conditional logistic regression model (Chamberlain, 1980;Verbeek, 2017). However, one major drawback of the inclusion of state dummies is that time-invariant units cannot be included in the model. Thus, all states that had not adopted any tool in 2020 would be dropped from the analysis, which would mean losing relevant information. Following the suggestion by Beck and Katz that "the use of fixed effects is clearly a bad idea for the binary dependent variable case" (Beck & Katz, 2001, p. 488), we therefore do not estimate a conditional logit model. However, to check the robustness of our results, we ran additional analyses, where we account for regional disparities based on the idea that institutional and historical legacies that have shaped certain regional penal policies, such as the "Sunbelt" penal system with its retributive mass imprisonment (Barker, 2009;Lynch, 2020), may influence bail reform dynamics. Hence, to control for such regional differences, we include dummy variables that account for regional specificities, using the nine geographical census divisions employed by the US Census Bureau (2022). As a result, we lose only a few Southwestern states due to perfect correlation with the outcome (no rollout of a PRAT) and therefore perform a robustness test without these four states (see Supporting Information Appendix), which indicates that our main results hold even in this case.
To test our hypotheses, we introduce three sets of variables into our logistic regression. First, to measure functional pressure on the criminal justice system (H1), we draw on two indicators. The first indicator, public spending (per capita) in the area of corrections and police, with the prison population being used as an alternative indicator, aims to directly measure pressure on the criminal justice system. The second, the public debt burden of each state (per capita), taps into more general fiscal pressures that may lead to fiscal adjustments. These indicators are all lagged by one year to rule out endogeneity. Second, to test policy diffusion (H2), we use two different indicators interchangeably: a distance variable that takes the mean value of the distances between the geographic center of a state and the centers of all other states that, in a given year, had previously introduced a PRAT. 7 As an alternative, we created a variable similarly adopted in other analyses (Berry & Baybeck, 2005;Mooney, 2001) that counts the number of neighboring states that had implemented a PRAT in the preceding year relative to the total number of neighbors a state has. Third, to capture the executive power of political parties (H3), we include a binary indicator that takes the value of 1 if the governor is a Democrat. Finally, we examine divided government (H4) by drawing on the seat shares of Republicans and Democrats in the chambers of the state legislatures (in order to include Nebraska, with its unicameral and nonpartisan legislature, we always code the seat shares as 0.5). Data for the 2000 to 2011 period comes from the Correlates of State Policy Project (Jordan & Grossmann, 2020); data for the years 2012-2020 was collected by the authors. Alternativly, we also use a dummy variable indicating whether Democrats control both chambers.
To account for possible confounders, we also include several control variables in the analysis. We include the violent crime rate as one important control variable. The reason is straightforward: If violent crime is high, this may reduce the likelihood that a risk assessment tool will be implemented because high crime rates tend to encourage a tough stance on crime, making it difficult to justify decarceration via the introduction of PRATs (König & Wenzelburger, 2021). To measure the general punitive climate in a state, we use a general state ideology measure of liberalism (Jordan & Grossmann, 2020) as well as several punitiveness variables aggregated to an index, including the existence of three-strikes or truthin-sentencing laws (based on Karstedt et al., 2019). We also include a measure for the popular election of judges, since the introduction of a PRAT gives judges the opportunity to avoid blame for risky decisions and should therefore appeal particularly to elected judges (Eckhouse et al., 2019, p. 203). In addition, following Beck et al. (1998), we always include a count variable in our regression equations to account for temporal dependency (see below). This is important, because it is unrealistic to expect that a state would change back and forth every year. Instead, once a PRAT has been implemented, it is very likely to remain for some time before potentially being abolished (as in Alaska). An overview of the measures used in the analysis and their data sources is presented in Table 1.

| EMPIRICAL ANALYSIS
To obtain a first impression of the variance with regard to the adoption of PRATs in our dataset, Figure 2 illustrates the mean values of four selected variables across two groups: those states and years with a PRAT implemented, and those without one. It is evident that there are remarkable differences between the two groups for all four variables considered here. In the PRAT group, violent crime is lower, while budgetary pressure on the criminal system and overall indebtedness is higher. Moreover, in around 63% of state-years in this group, Democrats hold the executive, compared to 42% in the No-PRAT group. Similarly, the geographic neighborhood of a state that has implemented a PRAT seems to be important as well. States in which a PRAT is in place often also tend to have neighboring states that also use a PRAT-a finding that points to a possible spillover effect. Running t-tests on differences of means reveals significant differences for all the reported comparisons. While these descriptive findings indicate a pattern that seems to support our hypotheses about fiscal pressures and partisan effects as well as the need to control for the spatial proximity of states using a PRAT, only a multiple regression can answer the question of whether the partial effects of these variables are significant while controlling for other covariates. To test our hypotheses, we therefore estimated pooled logistic regressions including indicators for the respective hypothesized independent variables as well as the control variables in the same regression equations. We present the results of these analyses in the next step. The models predict the probability that a PRAT will exist in a given year and state in relation to different combinations of our independent variables. We also include multiplicative terms in the model to estimate the interaction effect posited in H4. To ease interpretation, we will also visualize the estimated relationships. Table 2 presents the results of our analysis. Models 1 and 2 represent the baseline estimations, in which we control for geographical and time dependence as well as the violent crime rate as a demand factor. Models 3 and 4 add indicators for fiscal pressure, the liberalism of a state and punitiveness as well as an indicator measuring the direct election of judges. The three subsequent models account for the direct effect of party influence and include variables for Democratic dominance of the executive or the state Congress. Finally, the last three models display the estimation of interaction effects.
Looking first at the fiscal pressure variables (H1), it is evident that high spending for police and corrections, high public indebtedness, and a high incarceration rate significantly increase the probability that a state will have a PRAT implemented in a given year. Hence, our data indeed indicate that pressure on the criminal justice system pushes state governments to adopt algorithmic tools-a finding that lends support to the idea that the introduction of these tools is driven by efficiency and cost-saving considerations.
As hypothesized, policy diffusion also seems to matter. The multiple regression analysis thus corroborates the descriptive findings from Figure 1. Both indicators measuring geographic proximity as a facilitator of diffusion are significantly related to the likelihood that a state adopts a PRAT: The closer a state is to another state with a PRAT, or the more neighbors with PRATs a state has, the more likely it is that the state will also adopt algorithmic tools in the pretrial stage. Hence, policy diffusion as expected in H2 seems to be relevant, although we must acknowledge that the actual mechanisms can hardly be detected causally via aggregated quantitative data.
Beyond these effects, the coefficients for our party politics variables are also in line with our expectations (H3): A Democratic governor as well as a higher seat share of Democrats in the state House significantly increase the likelihood of the implementation of a PRAT. In contrast, partisan control of the Senate fails to reach conventional levels of significance (but becomes significant in the models using the region dummies, as shown in the Supporting Information Appendix). The fact that both fiscal pressures and ideology emerge as relevant in the analysis implies that fiscal pressures have additional explanatory power on top of Democratic political dominance and can explain a part of the variation regarding the dependent variable that Democratic dominance does not explain. This is in line with the idea that there may be different motives for adopting PRATs. Further, among the control variables, the violent crime rate is negatively associated with the likelihood that a state will adopt a PRAT, suggesting that politicians are reluctant to adopt "smart on crime" policies if crime rates are particularly high. All other control variables are not robustly significant at conventional levels. To gauge the size of the main effects, we present the predicted probabilities for the significant party control variables, the three indicators of functional pressure, and the variable of geographic proximity to other states that have implemented PRATs (Figure 3). We always present predictions for the empirical range of the respective variables, holding all other covariates constant. The graphs clearly indicate that while the political variables do significantly increase the probability that a state will make use of a PRAT, the effect of proximity to states with PRATs and the fiscal variables (especially public debt and spending on criminal justice) are much more substantial. In sum, the results lend strong support to our theoretical expectations, with fiscal pressures and geographic proximity having more substantial effects.
So far, we have looked at the effects of the executive and legislative seat shares of the two main parties separately. However, as we have argued in H4, the conditions for implementing (and continuously using) a PRAT should be particularly favorable where Democrats control not only the executive, but also the legislative branch. To assess whether such a conditional effect emerges from the data, we have included a corresponding interaction term in the regression equations presented in Table 2 (only including the significant terms-see Models 8-10). As the interaction coefficients are not easy to interpret due to the inclusion of the multiplicative term, we follow Brambor et al. (2006) and illustrate the interaction models graphically (see Figure 4). The two graphs in the first row of Figure 4 depict the marginal effect of the presence of a Democratic governor (vs. a Republican governor) on the likelihood of implementing a PRAT conditioned by the partisan composition of the respective state House or state Senate. The estimated marginal effects reveal a clear and consistent pattern over the three interaction models estimated for divided government. A Democratic state executive is more likely to introduce a PRAT if the Democrats also hold a majority in the state legislature. However, with increasing seat shares of the GOP in the state House and Senate, the positive effect of Democratic governorship decreases substantially, becoming no longer significant at higher Republican seat shares in the two chambers of the state legislatures.
The graph in the second row of Figure 4 replicates these findings with a simple dummy variable that takes the value of 1 if the Democrats control both chambers of the state legislature. Again, and as hypothesized in H4, the positive effect of a Democratic executive on the likelihood of PRAT implementation depends on the partisan composition of the state House and only materializes in the case of united government. Hence, as proposed in H4, the likelihood of the introduction of algorithmic tools in the criminal justice system depends on the interplay between governorship and the majority situation in the legislature.
In order to check our results for robustness, we ran several additional regressions (see Tables A .3-A.5 in the Supporting Information Appendix). Importantly, we extended our observation period to the start date of 1995, as the development of the first computerized statewide PRAT in our sample was commissioned after legislation passed in that year. However, including five more years and increasing the number of observations to 1200 does not alter the results. Moreover, the results also do not change if we exclude all cases in which a PRAT has de facto been implemented, but without an explicit legislative basis, as in Nevada. 8 Adding regional dummies to the analysis also leads to substantively identical results. We are therefore confident that our results are not driven by specific cases.

| CONCLUSION
In the recent past, the question of how to decrease skyrocketing incarceration rates has dominated the debate about criminal justice policies in the United States. The existing literature emphasizes several reasons for policymakers' preoccupation with bringing down imprisonment rates. First, fiscal pressures have been identified as drivers of penal policy change (Green, 2015;Karstedt et al., 2019). In this view, prison downsizing is mainly a function of austerity policies (Gottschalk, 2010). Second, new technologies of risk assessment that are based on big datasets and evidence generated by algorithms may also have contributed to the downturn. These technologies enabled criminal justice administrations to differentiate between risk categories and to reduce detention for those persons identified as having a low risk of re-offending (Duwe & Kim, 2017;Hartmann & Wenzelburger, 2021;Stevenson, 2018). Finally, several studies have emphasized a general ideological change in the political climate driving criminal justice reform (Dagan & Teles, 2016). According to this view, a bipartisan strategy of being "smart on crime" has replaced the "tough on crime" paradigm of the old days and paved the way for less punitive policies.
Against this background, and based on a novel dataset built around the implementation of algorithmic PRATs for statewide use in the 50 US states, our article confirms that fiscal pressures and policy diffusion have both played important roles when it comes to the adoption of PRATs as part of "smart on crime" policies. Indeed, high spending on the criminal justice system as well as the existence of neighboring states that have adopted PRATs increase the likelihood that a state will introduce PRATs in its own criminal justice system. Additionally, and adding nuance to the existing literature that emphasizes the bipartisan nature of "smart on crime" policies, our results suggests that the bipartisan consensus on criminal justice reform might well be more fragile than acknowledged in parts of the literature (for a similar view, see Beckett et al., 2016;Gottschalk, 2016). We find that partisan differences still exist, with Democrats being more willing than Republicans to adopt and maintain such tools. These relationships are clearly discernible if we control for functional, including fiscal, pressures, which increase the likelihood of PRAT implementation.
According to our theoretical argument, the effect of partisan differences can be explained by the fact that reliance on risk assessment tools has been discussed politically as a way to release people where possible rather than continuing to detain them. At least in political discourse, this is a break with the "tough on crime" approach that was prominent in the United States for decades. Advancing PRATs as a new penal policy strategy is therefore a tall order for Republicans and would mean changing core ideas in their penal policy program. Our results and arguments fit well with recent studies that question the alleged new bipartisan consensus on criminal justice and call for a more nuanced understanding (Beckett et al., 2016;Gottschalk, 2016;Seeds, 2017). This research indicates that the GOP supports criminal justice reform not so much to achieve rehabilitative ideals but rather out of an interest in fiscal conservatism (Beckett et al., 2016), and that it is key to look at the politico-economic context of the reform attempts (Gottschalk, 2016).
While the analysis above uncovers larger trends, state-specific developments and idiosyncratic political developments are also likely to play a role in criminal justice and PRAT adoption. Dynamics of political competition, political actors vying for public support, and other factors could sideline efforts to introduce PRATs. In California, for instance, this happened when pretrial reform became publicly salient in 2020, with Californians voting to repeal a law that would have introduced a state-wide PRAT to replace the practice of cash bail. In addition, the politicization of PRATs may lead to these tools being repealed due to public security concerns (arguments that should be most salient for the Republican Party and its electorate), as qualitative evidence regarding attempts to stop the introduction or repeal the adoption of PRATs in some states suggests (König & Wenzelburger, 2021). Another potentially relevant channel of influence that this study's large-N design cannot adequately probe is forms of diffusion and cooperation that cross state borders. Professional associations (Douglas et al., 2015;Scott, 2008) and state partnerships such as the Justice Reinvestment Initiative (Harvell et al., 2017) may well have been conducive to the diffusion of PRATs. Shedding light on these factors is a task that remains for future research.
It is also an open question at this stage whether the patterns found in the United States also apply to other countries, as the "tough on crime" policies did (Jones & Newburn, 2002). Several countries, including the United Kingdom and Australia, have also adopted algorithmic systems to inform decision-making in criminal justice, and they may similarly have experienced functional pressures to use such tools to bring down incarceration rates. It is furthermore unclear whether other US states will follow the lead of Alaska and Utah and terminate the "smart on crime" approach in criminal justice policies. If the use of algorithms in pretrial decision-making becomes politicized such that the tools are framed as representative of a "weak on crime" policy that tends to erroneously release dangerous people, this indeed could quickly lead politicians to abandon evidence-based practices that involve the use of PRATs. Hence, the "tough on crime" paradigm seems to be resilient. And despite bipartisan initiatives for criminal justice reform, the results of our study suggest that whether a definitive move toward a "smart on crime" approach occurs will depend at least in part on the balance of power between the parties.

ACKNOWLEDGMENTS
We would like to thank the anonymous reviewers for their constructive comments. The manuscript also owes to discussions with Anja Achtziger, Julia Felfeli, Adam Harkens, Tobias Krafft, Johannes Schmees, Wolfgang Schulz, Karen Yeung, and Katharina Zweig. We also acknowledge funding by the Volkswagenstiftung, Az. 95 438. Open Access funding enabled and organized by Projekt DEAL.
ORCID Georg Wenzelburger https://orcid.org/0000-0001-6195-7664 Pascal D. König https://orcid.org/0000-0001-9466-4024 ENDNOTES 1 The Utah case is somewhat special, as a PRAT was introduced in 2020 only to be repealed in early 2021. A compromise legislation was passed in late 2021 with some changes to the rules but ensuring the use of the PRAT. 2 Some authors differentiate between a more liberal "smart on crime" strategy and a more conservative "right on crime" policy (Seeds, 2017). However, as the general idea of using risk-assessment tools to identify different groups of offenders based on their risk is similar in both approaches, we do not differentiate them further. 3 Empirical evidence on whether this is actually the case is rather mixed, however (Stevenson, 2018;Viljoen et al., 2019; for a more optimistic take, see Desmarais et al., 2022). 4 In fact, several counties have been using PRATs earlier than entire states. However, it seems sensible to focus on statewide adoption, since our main theoretical interest centers on the political dynamics that underpin these decisions, and those should be most visible on the state level (and less so on the county level). 5 In addition, the use of PRATs also entails the promise of realizing greater fairness in criminal justice decisions and thus reducing possible discrimination in the system. Indeed, the American Civil Liberties Union (ACLU, 2011) advocated for the use of risk assessment instruments as way to reduce particularly racially biased decision-making, before changing its stance in light of reports on possible unfair discriminatory impacts of risk assessment tools. 6 One could argue that the beginning of the test period (1996)(1997)(1998)(1999) in Virginia that followed the adoption of the Pretrial Services Act in 1995 could also serve as starting point. However, even for Virginia, statewide implementation was only thinkable after the end of this test period (and was finally accomplished in 2003). This also justifies starting the observation period in 2000. At any rate, extending it to 1995 does not change our results. 7 For all those state-years in which no tool had yet been implemented, we use the maximum value of the calculated average distance for a conservative coding and test of the variable. 8 In Nevada, rollout of a statewide tool began without legislation but was then mandated by the state supreme court.