Abstract
Mass incarceration is commonly understood as a sweeping national policy development, which has obscured remarkable local variation at the policy implementation stage. California’s “Realignment” (Assembly Bill [AB] 109; 2011) is a reform that exploits this variation by design. Research consistently finds that, net of crime, demographic, political, and system capacity characteristics explain the variation in incarceration across local jurisdictions. Do these characteristics also explain decarceration? This study uses group-based trajectory modeling and logistic regression to examine the association of such characteristics with California county trajectories of state prison use in the decade preceding Realignment. County imprisonment trajectories and their related characteristics are then assessed as explanations for decarceration under AB 109. Distinct “risk” factors for high and/or increasing imprisonment trajectories are identified, as well as apparent protective factors. A clear association was found between previous trajectories and decarceration, but county-level characteristics did not demonstrate the predicted effects. Results indicate that decarceration cannot be explained as merely the mirror image of incarceration and should be examined as a distinct phenomenon. Implications for future research and policymaking are discussed.
The nationalization of crime policy is a key theme in the origin story of mass incarceration. Until the 1940s, criminal justice had historically been a federalist policy ideal, remaining almost exclusively the domain of state and local government. However, postwar episodes of racial violence and political unrest invoked calls for an expanded role by the federal government. The swiftly rising crime rates of the 1960s further solidified the legitimacy of federal expansion into criminal justice and cemented crime’s place on the national policy agenda (Gottschalk 2006; Miller 2008; Murakawa 2014; Simon 2007). The result was an expansion of the governmental capacity to pursue, prosecute, and punish lawbreakers unprecedented in U.S. history (see National Research Council 2014).
In tracing nationalization to the dramatic growth of incarceration, the story has been largely focused on the front-end policy stages of agenda setting and policy formulation—stages in which Miller (2008) has highlighted the “perils” of federalism for poor and minority group representation in the political process. This article pivots attention to the implementation stage, where national policy filters to lower levels of government and diffuses into local organizational practices. Recent scholarship demonstrates that even against the backdrop of nationalized crime policy, remarkable variation exists in local penal policy and practice, where the incarceration process begins for the vast majority of inmates (e.g., Barker 2009; Campbell and Schoenfeld 2013; Goodman, Page, and Phelps 2015; Lynch 2011). This variegated implementation has revealed a related “paradox” of federalism: nationalization at the policymaking stage goes hand-in-hand with localization at the implementation stage. Under America’s tripartite federalist structure and traditions, national crime policy has been overwhelmingly carried out by state and county governments with the practical effect of localizing the implementation of what has been otherwise conceived (and perhaps misspecified) as the “nationalized” policies that fueled mass incarceration.
As the nation reaches a potential turning point in the growth of incarceration, what are the implications for the reform goal of decarceration? California’s “Realignment” (Assembly Bill [AB] 109 [2011]) represents a distinctive response to federal judicial intervention in that the state localized the onus of legal compliance to counties. Rather than expanding prison capacity through new construction in response to a population cap ordered by the U.S. Supreme Court in Brown v. Plata (2011), the state opted to comply by enacting AB 109, which devolves the supervision of most nonviolent offenders to the county level and delegates remarkable discretion to local practitioners to decarcerate those previously sent to state prison (Schlanger 2013). The decentralization of policy and “grassroots” governance strategies are by no means historically unprecedented in the U.S. (e.g., Selznick 1965), nor in California’s history across multiple policy domains (California Legislative Analyst’s Office 1991). Therefore, understanding the complexities of local implementation and how underlying local variation shapes the present-day enactment of AB 109 has broader implications for assessing the prospects of regulatory models premised on localization well beyond this state and the domain of criminal justice.
Crime and recidivism have so far been the main outcomes of interest in gauging the success of Realignment’s “great experiment” in localization (The Economist 2012). However, this study is one of the first to analyze the outcome of decarceration: the system-wide reduction of incarceration in state prisons and local jails. Decarceration is an outcome of particular interest in addressing the prison overcrowding crisis because it aims to achieve deinstitutionalization rather than trans-institutionalization to jails or other locked institutions. As measured in this study, decarceration is distinctly revealing as a metric of institutional change in that it accounts for the possible displacement of incarceration to local jails—the concern that AB 109 will become a “shell game,” having “simply changed the address where offenders live and report” (Petersilia and Snyder 2013, 304). Unlike the outcomes of crime and recidivism, which aim to assess the effects of Realignment on changes in individual behavior (but see Bird and Grattet, this volume), decarceration aims to measure changes in institutional behavior. For this reason, I explore previous organizational practices and their enduring effects on decarceration in present-day implementation.
I approach local variation in counties’ historical rates of state prison use from a life course perspective, which typologizes developmental processes by classifying variation into empirically derived and theoretically meaningful categories (e.g., Nagin and Tremblay 2005). Through group-based trajectory modeling, I first demonstrate that the local imprisonment variation among California counties is historically patterned and can be grouped into five distinct trajectories. Second, I begin to explain these patterns by presenting results from multinomial logistic regression analyses of the relationship between a range of local characteristics and each trajectory type. Third, I relate these historical trajectories to present-day decarceration outcomes. I conclude with a discussion of implications for future research and policymaking.
Local Variation and the Implementation of Realignment
Wide variation in county responses to AB 109 has been documented at the planning stage of implementation (Abarbanel et al. 2013; Bird and Grattet 2014; Verma 2015), as well as remarkable preexisting variation in counties’ reliance on state prisons prior to Realignment (Ball 2012; Males and Goldstein 2014). The underlying sources of these variations, and the relationship between them, have yet to be explained.
After standardizing for population size, and despite the fact that all counties adjudicate the same statewide penal code, data demonstrate remarkable intercounty variation in state prison admission rates in the decade preceding Realignment. Figure 1 depicts this variation by county in each of the years from 2000 to 2009.
Annual State Prison Use Rate, by County
These data clearly show heterogeneity in carceral behavior, but they do not reveal whether there are patterns and meaningful groupings among counties. Ball’s (2012) analysis controls for local crime levels and finds that the same “high prison use” counties repeatedly fell into the top quartile of prison use for nearly all years, while the same “low prison use” counties repeatedly fell into the lowest quartile for nearly all years. These findings suggest that, despite variation, there are underlying county groupings to be identified.
Explaining local imprisonment variation
When incarceration is conceptualized as a straightforward tool for crime control, incarceration levels may logically appear as responses to crime levels. However, research roundly rejects this view (see National Research Council 2014, 44–47). Studies of interjurisdictional penal variation demonstrate that, net of crime rates, sociodemographic, political, and system capacity characteristics exert causal effects on incarceration levels even when jurisdictions operate under identical criminal codes and sentencing statutes. High degrees of poverty and income inequality are associated with high incarceration rates, as are the percentage of the nonwhite population and overall racial/ethnic heterogeneity (e.g., Bridges and Crutchfield 1988; Greenberg and West 2001; Kane 2003; Percival 2010); these characteristics show especially strong effects in urban areas (e.g., Beckett and Western 2001). Conservative partisanship, often approximated through measures of Republican Party affiliation, also consistently predicts high incarceration rates (e.g., Jacobs and Helms 1996; Stucky, Heimer, and Lang 2005).
According to group threat theories (e.g., Blalock 1960), rather than functioning straightforwardly as a tool for crime control, incarceration is more adequately explained as a tool for controlling groups that threaten the dominant social order (the wealthy and the white). As the proportion of poor and minority groups grows, incarceration rates are hypothesized to rise in response to the dominant population’s perception of social threat (see Liska 1992). As a political manifestation, Republican partisanship is understood to represent party alignment with punitive “law and order politics” (Jacobs and Carmichael 2001) mobilized in response to group threats (e.g., Percival 2010). Outside of the threat literature, organizational capacity theories of “prisons as self-regulating systems” (Berk et al. 1983) suggest that high prison and/or jail occupancy rates impose structural limits on the institutional capacity to incarcerate and should be controlled for in analyses.
While these explanations may apply to the observed variation in California counties’ historical use of state prison, do they apply to the outcome of decarceration? Lofstrom and Raphael (2013a, 23) found significant variation in county “jail-use responses” to AB 109, measured in the form of a ratio that reflects the change in a county’s jail incarceration rate before and after AB 109 relative to the change in its state prison incarceration rate before and after AB 109. Ratios indicate “prison-to-jail transfer” (p. 19), effects in thirty-nine counties, and overall decarceration effects in eighteen counties. Lofstrom and Raphael (2013a, 2) conclude that, “on average, a county’s jail population increases by one for every three felons no longer assigned to state prison. However, the effect of realignment on jail populations differs across counties, with some counties incarcerating much higher percentages of their caseloads.”
AB 109 has the potential to catalyze a genuine cultural shift in corrections that other states may emulate. However, without a theoretical framework for explaining why local systems react differently to the same legal intervention, as well as the processes that diverse institutional actors use to translate these interventions on the ground, the lessons of AB 109 may be missed in future sites.
Local Variation and the Life Course of U.S. Incarceration
The historical processes of policy development and, more specifically, the build-up of mass incarceration (e.g., Gottschalk 2006; Schoenfeld 2010), have been conceptualized in terms of path dependence or policy feedback, in which the past imposes constraints on what becomes understood as “possible” in the future, but does not determine what is actually made possible (e.g., Pierson 1994; Skocpol 1994). Legacy is also a heuristic for explaining the continuation of policy and practice over time (e.g., King, Messner, and Baller 2009; Verma 2015). Campbell and Schoenfeld (2013, 1377) synthesize these concepts into a “historicized political sociology of punishment” to “explain both the consistent, persistent punitive shift and the variation in the extent and timing of that shift” across states and regions (p. 1383). California’s Realignment exposes intrastate variation as a more fine-grained dilemma of federalism, which in turn raises new questions about the county-level dynamics that drove the punitive turn to mass incarceration in the first place and how they might shape prospects for present-day decarceration.
Life course research on human development investigates patterns and subgroups of variation to reveal the distinct etiologies that underlie a given developmental process (Nagin 2005), including the contingent factors, or “turning points,” that create opportunities for change (Laub and Sampson 1993). A life course perspective of county-level variation suggests the presence of underlying patterns and that, in different “types” of places, policy followed distinct developmental pathways, leading to the observed variation in imprisonment. This study adds to historical explanations by systematically classifying “types” of local variation through the use of life course trajectory modeling methods and by examining the proposition that these local variants and etiologies of imprisonment, in turn, shape how counties respond to reform in the present tense. It also extends life course research, in which the predominant unit of analysis has been the individual. Following studies that apply life course methods to examine developmental trajectories of crime among geographical and jurisdictional units of analysis (e.g., Groff, Weisburd, and Yang 2010; Schupp and Rivera 2010), this article examines imprisonment trajectories among counties as a way of understanding local variation in the “life course” of U.S. incarceration and decarceration as the potential “turning point.”
Research questions
I examine the following questions: (1) Can counties be grouped according to distinct trajectories of state prison use over time? (2) If distinct trajectories are found, what explains them? and (3) How do these trajectories shape the local implementation of AB 109 with respect to decarceration?
Hypotheses
Ball (2012, 1082) identified eighteen “high prison use” and fifteen “low prison use” counties and considers the remaining twenty-five counties “middle use.” I examine the proposition that there are also meaningful groups within this middle range that reveal imprisonment trajectories distinct from those of outlier groups. I expect to find trajectories that differ by level, shape, and direction, with some groups steadily increasing or decreasing state prison use throughout the decade and others changing course. I proceed on the theoretical basis that the characteristics explaining local imprisonment variation in cross-sectional studies also explain variation longitudinally. Therefore, I hypothesize these characteristics’ same direction of effects on county membership in a particular trajectory group over time—specifically, that high levels of poverty, income inequality, the proportion of the nonwhite population, racial/ethnic heterogeneity, and conservative political partisanship will distinguish counties with high and/or increasing imprisonment trajectories from those with low and/or decreasing trajectories, controlling for crime, urbanization, and jail occupancy.
The path dependence and policy feedback literature suggests that counties will largely follow their previous trajectories despite the reform intervention of AB 109, leading to the proposition that decarceration will be more likely in low and/or decreasing trajectory group counties, while high and/or increasing trajectory group counties will be more likely to have displaced incarceration to local jails. Accordingly, I hypothesize that the characteristics found to explain local imprisonment variation will also explain the variation in decarceration (by mirroring effects in the opposite direction).
Data and Methodology
Data
Analyses deploy publicly available federal and state administrative data on all county-level outcomes and characteristics. I developed a 2000 to 2009 panel data set, where “county” is the panel variable i(=1 . . . 58) and “year” is the time variable t(=1 . . . 10), yielding a total of 580 observations. The panel combines a subset of the “Tough on Crime” data compiled by Ball (2012), which includes annual county-level data reported by the California Department of Corrections and Rehabilitation (CDCR), the Department of Justice, Department of Finance, and Secretary of State, with additional data from the State Elections Board, Board of State and Community Corrections (BSCC), the U.S. Decennial Census, 2000 and the U.S. Department of Agriculture (USDA). I use 2000 census data because the possible predictors of trajectory group membership should generally be established by the time of the initial period of trajectories (see Nagin 2005, 96).
Dependent variables
I analyze two county-level outcomes: state prison use in the decade preceding AB 109’s enactment and decarceration under AB 109.
State prison use is measured by the annual rate of “new felony admissions” from 2000 to 2009, as reported by CDCR; “new felony admissions” is the annual number of individuals sent by each county to state prison for a new crime. Ball (2012) calculated the rate per 100,000 of each county’s annual population. I use this measure for two reasons. First, in line with previous research on interjurisdictional variation in penal practices, I measure the “flow” of people from county jurisdictions into state prison rather than the county origin of the “stock” of people in the state prison population; this isolates my main construct of interest (the propensity for a county to use state prison) without the confounds contained in “stock” measures, which reflect sentence lengths and state administrative release practices (see, e.g., Schupp and Riviera 2010, 58). Second, I do not include CDCR figures on “new parole violations leading to a new prison term” because, prior to Realignment, parole determinations were made by state officials with less input from county officials.
Decarceration is measured as a dichotomous variable, where 1 indicates that both jail and prison incarceration rates declined under AB 109; all other scenarios are coded as 0. I coded counties according to a range of pre-to-post AB 109 response scenarios. Because the legislation’s central mandate rendered a sizable class of offenses no longer eligible for incarceration in state prisons, scenarios in which a county’s use of state prison increases are theoretically possible but virtually impossible in practice. One such scenario is an increase in both jail and prison incarceration; another is that jail incarceration decreases while prison incarceration increases, and yet another is that jail incarceration remains unchanged even as prison incarceration rises. The more likely scenarios entail the mandated reduction in county use of state prison, where jail incarceration could increase, decrease, or remain the same. In a “displacement” scenario, a decrease in prison incarceration is accompanied by an increase in jail incarceration. In a “transfer” scenario, increases in the jail incarceration rate exceed, or offset, declines in the prison incarceration rate. In a “decarceration” scenario, both jail and prison incarceration rates decline. Overall decarceration could also occur, where jail incarceration remained unchanged while prison incarceration decreased, or where jail incarceration decreased while prison incarceration did not change. A final scenario is that incarceration in neither jails nor prisons changed from the pre- to post-Realignment periods.
I assess these scenarios using Lofstrom and Raphael’s (2013b) calculation of the pre-post AB 109 change in county jail and prison incarceration rates for the pre-Realignment time period of September 2010 to June 2011 and the post-Realignment time period of October 2011 to June 2012: DJail Incarceration Rateit and DState Prison Incarceration Rateit, where i=(1, . . . 57) indexes counties 1 and t=(1, . . . 9) indexes the first nine post-Realignment months (October 2011–June 2012). DJail Incarceration Rateit is the pre-post Realignment change in the jail incarceration rate (per 100,000 county residents) in county i in month t, and DState Prison Incarceration Rateit is the pre-post Realignment change in the rate of county residents incarcerated in state prison. To isolate changes in county jail and prison incarceration rates net of any seasonal variations, Lofstrom and Raphael calculate the pre-post Realignment changes relative to September 2011 for each post-Realignment month and then subtract corresponding changes that occurred in the previous year. The jail incarceration rate is measured using monthly BSCC data on the average daily jail population (ADP) in county i for month t. The prison incarceration rate is based on cumulative weekly CDCR counts of county-level prison admissions and releases, from which Lofstrom and Raphael calculate the monthly prison ADP for each county. 2
I found only four scenarios present in these data: system-wide decarceration, where both jail and prison incarceration rates declined; displacement; transfer; and an increase in both jail and prison incarceration rates.
Independent variables
My selection of crime, demographic, political, and system capacity variables follows previous research on interjurisdictional penal variation (see generally Liska 1992):
Crime
I measure the Type 1 crime rate based on the FBI UCR’s index of serious, reported violent and property crimes, which includes aggravated assault, forcible rape, murder, robbery, arson, burglary, larceny-theft, and motor vehicle theft.
Demographics
I include percentage in poverty and an income inequality index as indicators of economic threat. I use the county poverty levels reported in the 2000 census; the income inequality index is based on a calculation of the Gini coefficient from the same census. The index scale is 0 to 100, in which 0 represents complete income equality and 100 indicates maximum inequality. I also include the population percentage black and percentage Latino and racial/ethnic heterogeneity from the 2000 census as indicators of racial threat. I measure racial/ethnic heterogeneity with a Herfindahl index based on five groupings (white, African American, Latino, Asian, and other races). The index scale is 0 to 100, in which 0 represents a racially/ethnically homogenous population and 100 indicates maximum heterogeneity (i.e., all groups represent equal proportions of the population). Urbanization is a continuous measure that applies a more nuanced indicator than the dichotomous “metropolitan/nonmetropolitan” variable. The USDA’s rural-urban continuum code ranks counties on a scale from 1 to 9 based on population data and adjacency to metropolitan areas. Higher continuum codes indicate more rural counties with lower degrees of urbanization (see Lee, Maume and Ousey 2003, 117).
Politics
I measure the annual percentage of voters registered Republican as an indicator of conservativism. I add to previous studies that examine the effects of Republicanism by measuring the annual percentage registered to vote as a gauge of political participation regardless of party affiliation.
In addition, California’s voter initiative process enables a more specific measurement of “penal punitiveness” through voter support for particular penal policies on the ballot during this time period that reflect the “propensity, extensiveness and intensity” of support for incarceration (see Schupp and Rivera 2010, 58). Therefore, I measure the percent voting no on relevant ballot initiatives in the decade preceding Realignment that entail reducing incarceration (Proposition 66 [2004] to reduce penalties under California’s “Three Strikes” law, Proposition 5 [2008] to rehabilitate nonviolent drug offenders, and Proposition 19 [2010] to legalize marijuana) and the percent voting yes on initiatives to enhance incarceration (Proposition 9 [2008] to establish a crime victims’ “bill of rights” to participate in criminal sentencing and prison release decisions). 3 To address multicollinearity among these measures, I performed a principal components factor analysis with promax rotation, which demonstrated that they load onto a single component, which I label penal punitiveness. I performed a second factor analysis that included percent Republican with the proposition measures, finding that they also load onto one component. I label this factor conservative punitiveness. While the inclusion of factors occurring during the time period analyzed may appear to undermine the time-order requirement for establishing causality, I treat measures of voter support for ballot initiatives at time points throughout the decade as reflecting already-established levels of punitiveness. In other words, I analyze punitiveness as a time-fixed rather than time-varying characteristic in this study.
Capacity constraints
I measure the annual percentage of jail occupancy to account for constraints on the local capacity to incarcerate as well as county propensity to use incarceration as a response to crime. I calculated the jail occupancy percentage by first determining the yearly average of each county’s ADP, which is reported monthly by the BSCC. I then divided the yearly ADP by the rated capacity of all jail facilities within each county, as reported in December of each year.
Table 1 summarizes all variables.
Summary Statistics for California Counties, 2000–2009
The proportion is reported for “decarceration under AB 109” (0/1). Alpine County not included.
Methods and analytic strategy
I use group-based trajectory modeling (GTM) to identify statistically meaningful relationships among counties based on their use of state prison over time. GTM situates each county within one of several distinct classes characterized by group homogeneity but heterogeneity between classes. GTM is an application of the finite mixture modeling framework, which relies on the modeling assumption that the population comprises a mixture of a finite number of unobserved groups (see Nagin 2005). The Bayesian Information Criterion (BIC) is one of several measures used to guide model specification. The model with the smallest absolute BIC value is conventionally selected as best fitting. Once specified, GTM yields a “posterior probability” for each case (county), which reflects the certainty with which any given case is classified within a particular trajectory group. Cases are classified into the trajectory group for which they have the maximum posterior probability. GTM models require high probabilities (>.7) for classification of cases (see Nagin 2005, 88), but many have imperfect classification certainty, which results in statistically derived approximations of how many cases belong within a given class rather than providing clear cut-offs for cases with lower posterior probabilities. Therefore, the proper interpretation of trajectory groups is as a “statistical approximation to a more complex underlying reality” rather than as “literally distinct entities” (Nagin and Tremblay 2005, 84).
This analysis proceeds in three steps. First, after describing the univariate growth curve of state prison use, I present GTM results from a five-class linear solution. Second, I examine the association of county characteristics with membership in a given trajectory group by cross tabulating county trajectory group assignment with explanatory variables to generate a basic profile of the groups. I then specify multinomial logistic regression models to test the prevalence and intensity of associations. Third, I relate counties’ classification within distinct trajectory groups to decarceration under AB 109 using binomial logistic regression.
Results
Trajectories of state prison use
Figure 2 depicts the univariate (or single-class) model by plotting the mean state prison use rate for all counties for each of the years examined. State prison use rates averaged from a low of 117 in 2000 to a high of 148 in 2005, and the overall county imprisonment trajectory rose steadily throughout the first half of the decade before declining steadily after 2005.
State Prison Use by California Counties, 2000–2009
I specified a group trajectory model using the MPlus statistical package to test the hypothesis that the single-class model masks significant heterogeneity. GTM yielded a five-class solution, which indicates five distinct imprisonment trajectories underlying the statewide trend. Table 2 reports the mean of each variable by trajectory group, the number and proportion of counties in each group, and the average assignment probability for counties classified within a given group. In an ideal model, the posterior probability for each county would equal 1, resulting in an average class probability of 1. The average class probabilities shown in Table 2 nearly approach 1 and remain well above the minimum cut-off (.7) for all groups. With the exception of five counties, all maximum posterior probabilities were well over .7, with many at or approaching 1. 4
Trajectory Group Profiles
NOTE: Except for “decarceration under AB 109,” for which the proportion is reported, entries for dependent and independent variables report mean values by group.
Figure 3 depicts the trajectories by plotting the mean state prison use rate for each group in each year. All but one group (Group 3) tracks the shape of the univariate trajectory, but the groups notably diverge in their levels of state prison use. Group 1 tracks the shape of the overall imprisonment trajectory shown in the univariate model but at exceptionally high levels and with a steeper increase in the first half of the decade, followed by a shallower decline in the last half of the decade. Groups 2, 4, and 5 show the same overall shape but at significantly lower levels, with Group 4 being the lowest, Group 2 in the middle, and Group 5 being the highest of the three. Under GTM, Group 2 is considered normative because it contains the largest number of cases and tracks the overall trend shown in the univariate model; thus, Group 2 serves as the reference group for further analyses (see Nagin 2005). I summarize these groups as:
Middle Increasing (Group 2, the normative reference group)
Middle Decreasing/Stable (Group 5, conforming somewhat to the norm)
Low Increasing/Stable (Group 4, deviating)
High Increasing (Group 1, deviating)
High Decreasing (Group 3, deviating)
Historical Trajectories of State Prison Use, Plotted by Mean, 2000–2009
Explaining imprisonment trajectories
Life course studies position the characteristics that predict membership within a certain trajectory group as “risk” factors, and each trajectory group is understood to have a distinct “risk profile” (Nagin 2005). The group profiles presented in Table 2 offer initial insight into how group membership probability varies with each county characteristic; however, they do not convey the statistical significance of observed differences, nor do they specify the mathematical relationship between individual county characteristics and group membership.
Table 3 reports results from four multinomial logistic regression models that assess the effects of each county characteristic on the probability of group membership across the five groups. The first model includes the percentage of voters registered Republican and the penal punitiveness factor. To address multicollinearity, the second and third models include one or the other of these measures. The fourth model uses the conservative punitiveness factor instead, which captures both Republican partisanship and support for punitive ballot propositions.
Multinomial Logistic Regression Results for Trajectory Group Membership
NOTE: β / relative risk ratio (RRR) / (SE). Standard errors are reported for coefficients.
p < .05. **p < .01. ***p < .001.
The coefficients estimate how each characteristic influences the probability of membership in the particular trajectory group relative to the Middle Increasing reference group (Group 2). Because the parameter estimates are relative to the referent, the standard interpretation of coefficients is that for a unit change in the predictor variable, the logit of falling into the comparison group versus the reference group is expected to change by its respective parameter estimate, holding all other variables in the model constant. Table 3 also reports the relative risk ratio (RRR), which offers another interpretation. The RRR indicates how the “risk” of falling into the comparison group versus the “risk” of falling into the reference group changes with the variable in question. An RRR of greater than 1 indicates that the comparison outcome is more likely, while an RRR of less than 1 indicates that the outcome is more likely to be in the reference group. While the coefficients most directly estimate the direction of effects, the RRRs help to communicate the intensity of effects. 5
Across all groups in all models, the Type 1 crime rate demonstrates minimal effects on assignment relative to the Middle Increasing group, holding all other factors constant. These results support previous findings that crime rates alone fail to adequately explain the variation in local penal practice. Similarly, the percentage of jail occupancy just slightly reduced the relative risk of assignment to a given group relative to the reference group, except for the Low Decreasing/stable group, in which jail occupancy lacked statistical significance in all models. Counties might have been expected to increase reliance on the state prison system if their local jail capacities were highly strained; however, based on results from this analysis, I conclude that jail occupancy was not a significant factor in shaping different trajectories of state prison use in the decade prior to Realignment. Rather, individual differences in county economic and racial demography, urbanization and politics appear to have distinguished high, middle, and low imprisonment trajectory groups from the statewide norm.
“Risk” factors for high incarceration
Across all models, high poverty rates, income inequality, and the percentage of the black population are associated with assignment to the High Increasing and High Decreasing groups relative to the reference group (Middle Increasing). High degrees of urbanization also increase the likelihood of assignment to a high-level group, with a one-unit increase along the rural-urban continuum (indicating greater rurality) reducing the likelihood of assignment to either group. Across models 2 through 4, the risk of assignment to one of the high groups is anywhere from 4 to 11 times that of the referent for every 1 percent increase in poverty, holding the other variables constant. A one-unit increase in the degree of income inequality increases these group assignments’ likelihood by as much as 3.217 that of the reference group. The percentage of the black population made assignment to the High Increasing group approximately twice as likely as assignment to the referent and assignment to the High Decreasing group slightly greater than to the referent.
On the other hand, an increase in the percentage of the Latino population consistently reduces the likelihood of these group assignments to roughly half that of the referent, with the likelihood of assignment to the High Increasing group ranging from .438 to .666 across all models and, for the High Decreasing group, .352 to .778. Racial/ethnic heterogeneity was not statistically significant with respect to the High Decreasing group in any model. While greater racial/ethnic heterogeneity increased the relative risk of assignment to the High Increasing group in model 2, when the factors that capture penal punitiveness were included in the other models, its effects failed to reach statistical significance.
Regarding politics, the construct of “penal punitiveness” appears more clearly indicative of high incarceration group assignment than Republican party affiliation. Model 4 demonstrates that the combination of both Republican affiliation and penal proposition voting contained in the conservative punitiveness factor offers the most predictive value for assignment to each of the groups relative to the reference group. However conflating these constructs may mask important differences. In model 1, which attempts to isolate the effects of each separately, the percentage of voters registered Republican is not statistically significant for the High Increasing group, while the penal punitiveness factor appears to increase the likelihood of assignment exponentially. For the High Decreasing group, both measures are statistically significant but exert effects in opposing directions, and in model 2, the Republican coefficient changes signs when penal punitiveness is not controlled.
Like the proportion of the Latino population, political participation regardless of political affiliation may serve as a “protective” rather than “risk” factor for high state prison use relative to the norm. An increase in the percentage of the population registered to vote made assignment to the High Increasing group less likely at statistically significant levels across models, and less likely to the High Decreasing group at statistically significant levels in model 2.
Factors associated with midrange incarceration
Many of the same demographic characteristics found to be risk factors for high incarceration also predicted membership in the Middle Decreasing/Stable group relative to the reference group. However, the intensity of economic risk factor effects (increased poverty and income inequality) was notably less for the midrange group, as was the intensity of “protective” factor effects. The percentage of the black population exerted similarly intense positive effects on group assignment, but the negative effects of the percentage Latino were less drastic in reducing the likelihood of midrange group assignment (with RRRs from .856 to .880 across all models, compared to RRRs falling below 4 in some high incarceration group models). Similarly, while increased rurality and the percentage registered to vote consistently reduced the likelihood of midrange group assignment, the intensity of each variable’s effects was less than for the high-level groups.
Unlike model results for the high incarceration groups, the Republican, penal punitiveness and conservative punitiveness measures appear to tell the same story across all models for the Middle Decreasing/Stable group. In model 1, neither of the estimates for the percentage of voters registered Republican nor the penal punitiveness factor rose to statistically significant levels, yet they each exerted significant, positive effects in models 2 and 3. The similarly significant, positive effects for the conservative punitiveness factor shown in model 4, which captures both measures, indicate that they approximate the same general construct rather than meaningfully distinct constructs. Thus, while I found penal proposition voting behavior more dispositive in predicting high incarceration group membership—and Republican affiliation to be an insufficient predictor on its own—such disaggregation is not necessary in models predicting membership in the midrange group.
Factors that distinguish relatively low incarceration
As expected, results for the models predicting membership in the Low Increasing/Stable group (the only group that deviates downward from the reference group) show effects in the opposite direction. The same characteristics that serve as “risk” factors for high incarceration decrease the likelihood of membership in the low-level group. Likewise those factors that “protect” against high incarceration increase the likelihood of membership in this group.
Only the political characteristics appear to take on different meanings. The percentage of Republican voters is a negative predictor of low group assignment, while punitive voting patterns on penal ballot initiatives appear to positively predict low group assignment, net of Republican affiliation. However, when each measure is included separately, or combined into one factor, the results replicate the significant, negative effects of the Republican measure. When the Republican and penal punitiveness measures are combined via the conservative punitive factor in model 4, the RRR falls to .484—making it less than half as likely that a county with relatively high conservative punitiveness would fall within the Low Increasing/Stable group. This suggests that, for relatively low-imprisonment groups, while Republicanism and penal punitiveness are distinct constructs with divergent effects, levels of Republicanism in general tend to overshadow the effects of penal punitiveness in particular. Therefore, Republican measures may be more reliable for predicting low-imprisonment trajectories; however, such measures may mask an important nuance of the relationship between penal punitiveness and low-imprisonment trajectories: even in counties with relatively low Republican partisanship, support for punitive penal policies at the ballot box may be strong—in other words, party affiliation does not necessarily track “law and order” politics in liberal counties.
Trajectories of state prison use and decarceration under AB 109
Figure 4 depicts the distribution of the four observed AB 109 responses across groups (system-wide decarceration, prison-to-jail displacement, prison-to-jail transfer, or an increase in system-wide incarceration). The seventeen counties in which decarceration was observed were distributed as follows: the Middle Decreasing/Stable (Group 5) and Low Increasing/Stable group (Group 4) each contained six decarcerating counties; the High Increasing (Group 1) and Middle Increasing group (Group 2) each contained two decarcerating counties, and the High Decreasing group (Group 3) contained one county that decarcerated.
Distribution of County AB 109 Responses by Trajectory Group, 2000–2009
Results from a bivariate analysis of the relationship between trajectory group membership and decarceration show that differences in decarceration across the groups, as of June 2012, are statistically significant. 6 However, a bivariate approach does not control for alternative explanations and, in particular, the possibility that the same characteristics found to explain trajectory group membership adequately explain decarceration under AB 109, net of the different developmental paths counties took with respect to state prison use in the decade prior to Realignment.
Table 4 reports results from four sets of binomial logistic regression models. Within each set, a model was estimated with and without the inclusion of dummy (0/1) variables for membership within each of the trajectory groups 1, 3, 4, and 5. Group 2 (Middle Increasing) is once again used as the reference group, and the parameter estimates for each trajectory variable are interpreted as comparing a given group to the referent on the outcome of decarceration. As with the previous analysis, each set of models controls for the same crime, demographic and capacity variables but differs in the political variables included.
Binomial Logistic Regression Models Predicting Decarceration under AB 109
NOTE: β / odds ratio (OR) / (SE).
p < .05. **p < .01. ***p < .001.
In addition to coefficient estimates, which are interpreted as in the previous multinomial models, Table 4 also reports the odds ratio (OR) for each variable. An OR of greater than 1 indicates that the decarceration outcome is more likely than all other observed AB 109 responses with a one-unit increase in the variable, while a ratio of less than 1 indicates that decarceration is less likely. 7
The results of models 1a through 4a, which do not include the trajectory variables, diverge somewhat from previous analyses. Across these models, increased poverty and rurality are consistently associated with decarceration under AB 109 at statistically significant levels, while an increase in the crime rate shows a slightly negative effect across models. The percentage of voters registered Republican, the penal punitiveness factor, and the conservative punitiveness factor achieved statistical significance in models 2a, 3a, and 4a, respectively, each predicting less decarceration. While the penal punitiveness and conservative punitiveness factors appear to make decarceration approximately half as likely as the other scenarios, holding all other variables constant, an increase in the percentage of voters registered Republican exerted weaker effects, reducing the odds of decarceration to slightly below 1. In my previous analysis, I found high poverty rates and degrees of urbanization to predict high and/or increasing imprisonment trajectories in the decade preceding Realignment; paradoxically, here, high poverty increases the likelihood of decarceration as well. I did not find a clear association between crime rates and trajectories of state prison use; nor do these models indicate a clear directional association with decarceration under AB 109. Additionally, my previous analyses found a positive association between the percentage registered to vote and lower and/or decreasing trajectories of state prison use; however, this variable was not significant in explaining decarceration. While the effects of the percentage of the population black and the percentage Latino were significant in explaining previous trajectories of state prison use (with the percentage black predicting high or increasing trajectories and the percentage Latino predicting low and/or decreasing trajectories), neither variable exerted statistically significant effects in models 1a through 4a.
When the trajectory variables are added in models 1b through 4b, the percentage in poverty and the degree of rurality remain positively associated with decarceration at statistically significant levels, and the direction and significance of effects for the Republican, penal punitiveness, and conservative punitiveness variables remain unchanged. However, the percentage of the black population demonstrates significant negative effects, reducing the ratio of the odds of decarceration versus all other observed AB 109 responses to between .844 and .871 across models. In comparison to the Middle Increasing reference group, counties falling within the High Increasing and Low Increasing/Stable groups demonstrated the greatest odds of decarcerating across all models, regardless of differences in crime, demographics, politics and jail occupancy. While counties falling within the High Decreasing and Middle Decreasing/Stable groups were also more likely than the reference group to have decarcerated under AB 109, membership within one of these groups appears to make decarceration only slightly more likely in comparison to the High Increasing and Low Increasing/Stable groups.
These results suggest that previous trajectories of state prison use that were already at relatively low levels compared to the norm (the Low Increasing/Stable group) made decarceration the most likely response to AB 109, regardless of differences in individual county characteristics. At the same time, exceedingly high levels of state prison use that had been on an increasing trajectory (the High Increasing group) also appear to have the greatest odds of decarcerating. Counties with trajectories that had already declined in the decade prior to Realignment (the High Decreasing and Middle Decreasing/Stable groups) were less likely than the rest to respond to the reform by decarcerating, net of local differences. These results provide evidence that historical imprisonment trajectories exert significant, independent effects on the outcome of decarceration under AB 109. While certain county demographic and political characteristics remain important factors, the distinct developmental pathways that counties followed with respect to state prison use in the 2000s are just as—perhaps more—important in explaining decarceration in the wake of Realignment.
Conclusion: Toward a Typology of Local Variation
The National Research Council’s (2014, 430) proposed agenda on incarceration calls for “studying the cluster of conditions and variations in the cluster across time and space” and describing “variation in the pattern of correlations … in a way that would be useful for analysis.” The present study begins to answer this call.
Results from the multivariate analysis of California counties’ reliance on state prisons in the decade leading up to Realignment may build toward a more general typology that associates local characteristics with comparatively high, middle and low imprisonment trajectories. In this case, I found poverty, income inequality, the percentage of the black population, urbanization, and the percentage of voters registered Republican to be strongly associated with high prison use. Alternatively, the percentage of the Latino population and the percentage registered to vote were strongly associated with comparatively low prison use. Crime did not figure into this typology as clearly associated with any particular trajectory, providing additional evidence that imprisonment does not function as a straightforward tool for crime control. I also found that, while jail capacity constraints should be controlled for in assessing the effects of other predictors, they explained relatively little about the distinct trajectories of state prison use from 2000 to 2009.
Results further indicate that explanations for the phenomenon of decarceration do not precisely mirror those of incarceration. For example, high poverty rates are at once distinguishing characteristics of counties with high and/or increasing imprisonment trajectories as well as those most likely to decarcerate in response to AB 109. I also found many of the variables significant for explaining incarceration to appear irrelevant as explanations for decarceration. An increase in the percentage of African Americans notably distinguished high imprisonment trajectory group counties but appeared to play a role in distinguishing counties on the basis of decarceration only when previous imprisonment trajectories were controlled for. Conversely, while high Latino percentages were among the most consistent distinguishing characteristics of counties with relatively low imprisonment trajectories, they did not demonstrate an association with decarceration. While the political variables diverged across trajectory groups with respect to explaining state prison use, whether measured in terms of Republican partisanship or, more specifically, penal punitiveness or conservative punitiveness, these variables were consistently significant negative predictors of decarceration.
Current theories do not explain these divergent findings or the characteristics that emerged as “protective” factors against high imprisonment trajectories and the displacement of incarceration to local jails. In particular, theoretical emphases on conservative political partisanship and “law and order” politics may have overlooked the local dynamics of political participation in general and, perhaps relatedly, voter engagement with penal policy in particular (as reflected in the respective measures of the percentage registered to vote and patterns of support for relevant voter propositions). I also found models that included developmental variables for previous imprisonment trajectories to better explain decarceration under AB 109 than the time-fixed county characteristic variables. Such developmental variables that capture the potential path dependencies of organizational patterns of past practice may enhance existing explanations of incarceration as well.
This study’s findings are limited to demonstrating associations between historical imprisonment, local characteristics and decarceration under AB 109. Future research should examine the causal connections and mechanisms that explain these relationships. This will necessarily entail accounting for other plausible predictors of carceral behavior and legal implementation, including policy and fiscal developments in domains seemingly unrelated to criminal justice (see Campbell, this volume), as well as the human agency displayed by individual practitioners and politicians who make decisions and take concrete action under idiosyncratic local conditions and contingencies in the penal field (see Goodman, Page, and Phelps 2015). At the same time, these limitations should be considered in light of the National Research Council’s (2014, 430–31) rationale for promoting such research: [I]t is not a score on a scale but the strength of association of incarceration with other variables that may be consequential for social science and for policy. Motivation for examining the pattern of correlation—rather than trying to isolate the effects of individual factors—might derive from both a high level of interaction operating with incarceration and its correlates and a high level or feedback of endogeneity operating among the factors. In this context, efforts to assess individual causal effects will result in misspecification. Studying the cluster of conditions and variations in the cluster across time and space emerges as an important research priority.
The developmental process, or life course, of U.S. incarceration provides active context legal interventions, whether by courts, as in Brown v. Plata (2011), or legislatures, as with AB 109 (2011). While the path dependencies and legacies of past policy and practice may constrain future reform implementation, the life course perspective draws attention to the possible turning points and opportunities for change that the contingencies of history and local context create. Realignment has presented one such turning point in California. The federalist paradox is that even if a national policy shift to decarceration comes to fruition, it will diffuse and filter in distinct ways to the local level, where its implementation remains contingent on a range of local conditions, including the local legacies of organizational practice left behind by mass incarceration.
Footnotes
NOTE:
The author extends special thanks to Magnus Lofstrom and Brandon Martin at the Public Policy Institute of California, where the author began data collection for this study as a 2014 Richard J. Riordan research intern. The author is also grateful to Charis Kubrin, Carroll Seron, John Hipp, and Bryan Sykes for their support and guidance on this project and beyond.
Notes
Anjuli Verma is a doctoral candidate in the Department of Criminology, Law and Society at the University of California, Irvine. Her research examines organizational regulation and compliance, legal reform and transformations in the governance of crime and punishment. She is a 2015 National Institute of Justice Graduate Research Fellow and recipient of the National Science Foundation’s Doctoral Dissertation Research Improvement Grant. Her work was recently published in Law & Society Review.