“Of all the gin joints, in all the towns, in all the world, she walks into mine.”
—Rick Blaine, Casablanca
In studying a cohort of males born in Berkeley, California, in the late 1920s, Elder (1986) observed entry into military service, either late in WWII or the Korean War, led to certain continuities and discontinuities in the individuals’ lives, including obtaining greater educational opportunities and later transitions to certain adult roles, such as starting a family, compared to nonveterans. Importantly, Elder (1986) also noticed that the timing of entry into the military was important, as those who entered into military service at a younger age displayed larger gains in psychological health and social competence between adolescence and adulthood compared to nonveterans. This led Elder (1986, p. 238) to conclude that military service could essentially serve to reshape the life course for those with a disadvantaged past by providing individuals with a more auspicious future, or, as he described it, leading to “a pronounced break in the lifetime, a discontinuity between past and future.”
Not long afterward, Sampson & Laub (1993) observed that in a historical sample of delinquent males from Boston born in the late 1920s, marriage (among other things) was prominently implicated in the process of criminal desistance (the cessation of criminal activity), as it provided an opportunity for individuals to knife off, or sever themselves from harmful environments. Similarly, Uggen (2000) analyzed data from the National Supported Work Demonstration Project commissioned in the mid-1970s. He found that, at least for certain individuals, being provided with employment opportunities was linked to a reduction in self-reported offending and arrest compared to those who were not, suggesting that entry to employment was linked to the process of desistance. Uggen again emphasized the importance of timing in the process, noting that work tended to be more influential for older rather than younger offenders.
Military service, marriage, and employment thus have become three of the most common types of salient life events, or turning points, that are theorized to bring about opportunities for change and are capable of altering an individual's trajectories of behavior. These events have become of great theoretical importance since being introduced to criminology by Sampson & Laub (1993, p. 233), who left little room for ambiguity in their position: “[o]ur major thesis is that…turning points are important concepts in understanding processes of change in the adult life course.” This point is underscored by Elder (1986, p. 233), who argued that “[u]nderstanding such change presents a major challenge to social science, especially in identifying mechanisms that enable men and women to turn their lives around.”
In this review, we reflect on how criminologists have sought to identify both turning points in the criminal career and the process of desistance from crime by directly considering the measurement and identification of turning points. There has been increasing attention paid to integrating a causal framework into the study of turning points. A primary consideration, however, is that, by nature, life events such as marriage and employment are endogenous. Thus, efforts to identify the effect of certain key life events on future behavioral trajectories are closely linked to the formal concept of econometric identification (Manski 1995). We present a formal way to think about the causal identification of turning points in the face of endogeneity, paying particular attention to threats to identification created by not only selection bias but also simultaneity bias, which we argue is a key concept implicated within the life-course framework. We also consider several conceptual issues relating to the measurement and definition of turning points that we hope may stimulate creative advances to existing empirical methods. Overall, our aim is to provide an outline for future directions in which the literature can move.
The remainder of this review is organized as follows. First, we provide a brief overview of the concept of turning points within the broader context of the life-course perspective and provide a summary of how scholars have conceptualized turning points. Second, we review the empirical evidence of the role of certain turning points in criminal trajectories. Our goal is not to provide an exhaustive review of the empirical literature on turning points in sociological and criminological literatures but rather to pay special focus to the issues and limitations associated with the statistical identification. Third, we consider the limitations of the current methodological approaches, specifically as they relate to the connection between empirical models and theoretical conceptualizations. We conclude with a research agenda by outlining ways to investigate the role of turning points in the desistance process.
TURNING POINTS IN THE LIFE-COURSE FRAMEWORK
Turning points are a key concept in the life-course approach and have garnered substantial interest in criminology. The idea of turning points emerged from two prominent themes in life-course processes: trajectories and transitions. Trajectories are long-term patterns of behaviors and embedded in trajectories are transitions, which span shorter periods of time, are usually age-graded, and slightly modify trajectories (Elder 1985). Sometimes, transitions can drastically alter trajectories and have been referred to as turning points. In addition to trajectories and transitions, Elder (1994) recognized four factors associated with determining the shape and patterns in the life course: (a) location in time and place, which recognizes the importance of understanding lives within their larger dynamic and complex environment, particularly social, cultural, and historical contexts; (b) linked lives, through which humans are tied both to one another and to social institutions that can promote or inhibit behavior; (c) human agency, the idea that the life course is constructed by the choices and actions individuals take within the opportunities and constraints of history and social circumstances; and (d) timing of lives, which concerns individual strategic adaptation of events. Other major themes in the life-course approach include the concept of cumulative disadvantage, which is the interdependency of earlier transitions on later transitions or events (Elder 1985, Rutter 1996) and diversity in trajectories due to an accumulation of historical, cultural, and social factors, individual characteristics, and individual agency.
In the original conceptualization of the life-course approach, human agency is a central aspect not only of constructing the life course but also of turning points embedded within the life course. Hareven (1982, 2000) points out that human agency is perhaps one of the most important aspects of the life-course perspective but individuals’ choices are constrained by the structural and cultural arrangements and opportunities available. Indeed, according to Hitlin & Elder (2007, p. 183), “Transition is closest to what we term life course agency, the ability of individuals to make choices at turning points in the life course.” Thus, agency is a core concept that relates to issues of endogeneity with which we are concerned.
In addition to the primacy of human agency, the idea of turning points as processes is dominant. Pickles & Rutter (1991) underscore the need to consider the processual nature of turning points rather than viewing them as a singular event. Similarly, Vaughan (1986) describes a long process of “uncoupling” in the dissolution of marriage. Hareven & Masaoka (1988) write about turning points as a process that alters life paths and requires certain strategies and choices (see also Abbott 1997, Sampson & Laub 2005). Pickles & Rutter (1991, p. 134) also describe a dynamic branching process incorporating agency, process, and cumulative disadvantage: “At each point some individuals choose one way; others choose another, and it is necessary to examine both the consequences of their choices and the mechanisms resulting in one choice rather than the other” [see also Hofferth's (1987) description of motherhood].
Popularization of the concept of turning points in criminology can be attributed to the influential work of Sampson & Laub (1993). Sampson & Laub (1993) put forward the idea that important life events such as marriage and stable employment can serve as turning points away from a criminal trajectory. Original conceptualization of turning points in criminological discourse is largely consistent with the broader life-course treatment. For example, Laub & Sampson (1993, p. 306) wrote about the interdependency of earlier events and transitions on later ones. They stated that “delinquency incrementally mortgages the future by generating negative consequences for the life chances,” suggesting a dynamic, endogenous process among earlier life events, turning points, and subsequent events. Laub & Sampson (1993) also highlighted the role of human agency and choice in acknowledging that self-selection into turning points and adult roles is a function of individual differences and prior experiences. They also note that the possibility for change is likely mediated by individual contingencies. Even in later extensions of their theory, Laub & Sampson (2003, p. 9) echo the original conceptualization of turning points by stating that desistance results in “the interplay of human agency and choice, situational influences, routine activities, local culture, and historical context.”
In sum, the theoretical foundations of the concept of turning points were built on several key life-course assumptions, including the important role of choice and agency in structuring events into turning points, that the processes that define turning points have a duration (Abbott 1997), and the interdependent nature of earlier life events on later ones, or the notion of cumulative disadvantage. Unfortunately, however, theory and research have largely occurred in isolation. Instead of attempts at modeling turning points to incorporate their fundamental characteristics, empirical investigations on turning points have been focused on identifying exogenous turning points, such as the causal effect of marriage and employment, in dramatically shifting an individual's offending trajectory.
COMPLEXITIES WITH MEASURING TURNING POINTS
In the past, scholars conceptualized turning points as single rare events; however, most scholars now agree that turning points are process oriented (Abbott 1997, Sampson & Laub 2005). For example, Hareven & Masaoka (1988, p. 274) assert that “a turning point is not an isolated event of short duration…[but rather] a process involving the alteration of life path.” Similarly, Abbott (1997) argued that turning points are, in fact, not points per se but rather a process that has a duration.
The nature of turning points presents a quantitative challenge to the understanding of how turning points are related to offending trajectories as well as the desistance process, which itself is somewhat difficult to cleanly operationalize. First, the true form of turning points is unclear. If turning points reflect a gradual process, they should resemble transitions away from offending rather than an abrupt shift in offending trajectory. However, both processes can have a lasting effect “on a person's self-concept or the views and expectations of other people, one of the fundamental characteristics of turning points” (Rutter 1996, p. 614).
Second, and perhaps more importantly, most contemporary methods used to examine the desistance process retrospectively identify offending groups by their offending rates, hindering identification of more subtle or extended turning point processes. Specifically, to examine the desistance process, researchers have recently relied on semiparametric group-based mixture models (Nagin & Land 1993), which identify population heterogeneity in offending over time. Bushway et al. (2003) argue that because of differential rates of offending, these models are the optimal way to capture the dynamic nature of desistance (see also Bushway et al. 2001). However, by defining offending groups based on individuals’ longitudinal offending vectors over the entire observed period, retroactively considering the role of life events within this time frame produces a fundamental identification problem. Abbott (1997, p. 249) anticipated problems with these contemporary methods and observed that “turning points, precisely because they are the more causally central shifts in the regime, will not be uncovered by methods aiming at uncovering regimes.”
Third, scholars have discussed the complex nature of analyzing turning points. Issues related to time dependencies, omitted variables, interdependency of turning points with other life-course transitions, and individual differences are all tenuous issues with which to methodologically contend (Pickles & Rutter 1996). Rather than attempting to consider some of these issues, the trend in criminological discourse is to focus on identifying the causal effect of a particular life event, such as marriage or employment, on offending. We echo Abbott's (1997) argument that the concept of turning points is unique because the nature of the process is extremely gradual rather than being one of simply change, causality, or succession. Therefore, the complexities at the core of turning points are obfuscated with much of the current treatments of turning points. We expand on this last point in detail below.
EMPIRICAL EVIDENCE FOR TURNING POINTS IN CRIMINOLOGY
Since Sampson & Laub's (1993) initial conceptualization of turning points in criminology, there has been considerable interest in analyzing the effects of turning points on offending. The majority of the efforts have focused on uncovering marriage (e.g., Farrington & West 1995, Horney et al. 1995, Warr 1998) or employment effects (e.g., Uggen 2000, Uggen & Staff 2001, Uggen & Shannon 2014), and as such our discussion also focuses largely on marriage and employment. It is important to note, however, that even though marriage and employment take the lion's share of empirical investigation, scholars have also recognized other important life transitions, such as military service (Bouffard 2003, Elder 1986, Sampson & Laub 1996, Teachman & Tedrow 2016), parenthood (Kreager et al. 2010, Edin et al. 2004, Landers et al. 2015, Michalsen 2011), gang membership (Haviland & Nagin 2005, Melde & Esbensen 2011), residential change (Kirk 2012), and high school dropout (Bersani & Chappie 2007, Dupéré et al. 2015), as potential turning points that can dramatically alter a criminal trajectory.1
There have been many studies that have successfully demonstrated an empirical association between marriage and/or employment and a reduction in later offending. The concept of marriage is thought to provide benefits of increased social bonding and informal social control (Laub & Sampson 2003), and, as a turning point capable of deflecting criminal trajectories, has been an attractive subject of inquiry for scholars studying the desistance process. Besides operating through a mechanism of informal social control, competing theories of desistance offer that marriage can promote reduction in offending through identity change or cognitive transformation (Giordano et al. 2002, Maruna 2001, Paternoster & Bushway 2009), reduction in exposure to delinquent peers (Warr 1998), and changes in routine activities (Osgood & Lee 1993, Osgood et al. 1996). As such, there has been considerable attention devoted to uncovering a marriage effect on criminality across multiple longitudinal samples and countries.
Craig et al. (2014) review 85 studies intended to study the relationship between marriage and future criminality. A negative association between marriage and future offending has been documented in nearly every major longitudinal study used by criminologists, including the AddHealth (Barnes et al. 2014), National Longitudinal Survey of Youth (NLSY) (Forrest & Hay 2011), National Youth Survey (King et al. 2007), Cambridge Study in Delinquent Development (Farrington & West 1995), California Youth Authority (Piquero et al. 2002), Dutch Criminal Career and Life-Course Study (Bersani et al. 2009), and Glueck data (Sampson & Laub 1993). Marriage has also shown to be associated with a reduction in other harmful activities, including substance use and binge drinking (Duncan et al. 2006, Teruya & Hser 2010). Moreover, this association has been shown to be robust to race and gender, although perhaps varying in magnitude for the latter (Doherty & Ensminger 2011).
Similar developments have been done examining the relationship between adult employment and crime using observational data. Sampson & Laub (1993) argued that stable employment can have an effect on the desistance process by serving as a turning point in a criminal career. Accordingly, the stronger the ties to work, the less likely an individual is to engage in criminal activities. Legal employment can structure one's daily activities and reduce time spent partying and hanging out with friends as well as serve as a platform to create or strengthen bonds to conventional individuals. Other scholars also underscore the importance of employment in the desistance process but assert that cognitive and identity changes must occur prior to employment for work to facilitate, sustain, and reinforce the desistance process (i.e., Giordano et al. 2002, Maruna 2001, Paternoster & Bushway 2009). Nonetheless, theory and observational data support the positive role of employment in the desistance process. Similar to marriage and crime, the relationship between employment and crime has also been demonstrated with a number of samples, including youth samples (Apel et al. 2006, Staff et al. 2010), adult offending samples (Apel & Horney 2017, Horney et al. 1995), and supported work samples (Uggen & Shannon 2014).
However, an association between, for instance, marriage and desistance or employment and desistance is a necessary, but not sufficient, condition to demonstrate that such an event represents a turning point from crime. In fact, many earlier studies are plagued by coarse measurement observations, which limit temporal sequencing of marriage and crime cessation. More importantly, earlier studies often utilize regression-based methods with observational data, which limits the potential for causal identification. In other words, the association of marriage with a reduction in offending could be completely spurious and therefore would not demonstrate that marriage is a turning point. Instead, to qualify as a turning point an event must be demonstrated to deflect a trajectory of behavior. The latter issue is the point of departure for Sampson et al. (2006, p. 466), who note that “the literature is replete with findings suggesting that marriage is linked with well-being. The meaning of all these associations is another matter altogether. Questions of selection and confounding are paramount.” Although Sampson and colleagues were speaking specifically about marriage, the generality of their argument extends to nearly all other kinds of life events, including getting a job or joining the military.2
Dealing with selection and causality naturally implicates the important concept of econometric identification. Identification can be thought of as the set of assumptions that are necessary to draw causal inference from observational data (Angrist & Krueger 1999). Our discussion of identification is heavily influenced by Manski (1995), who makes a distinction between problems of statistical inference, or what can be learned from a finite sample (i.e., precision), and problems of identification, or what could be learned (or not) from an infinite sample. Complications from endogeneity are generally threats to identification.
TURNING POINTS IN AN IDENTIFICATION FRAMEWORK
Important life events such as marriage or (often) employment are not randomly assigned. That is, historical factors, cultural factors, social factors, individual characteristics, and individual agency all play a pivotal role in important life events. Because individuals often nonrandomly select into life events, this presents an inherent problem: Those who get married (or secure employment or join the military) and those who do not are not appropriate comparison groups. The two groups likely differ in important ways before the event and may have future outcomes related to these differences and not necessarily the event itself. Importantly, we can think of a life event such as marriage as being endogenous, that is, the event is related to other factors in the process.3 One important type of endogeneity is selection bias—that is, people who select into marriage are simply different than those who do not.
The influence of self-selection into turning points is evident in the early work of Elder (1986). The Berkeley men were born right before the Great Depression and early entry into military service was favored by those who came from disadvantaged families, had low high school grades, and felt a sense of inadequacy as teenagers, with the latter differentiating between those who joined and nonveterans. Events like getting married and gaining employment are not purely exogenous occurrences but rather involve some degree of choice on the part of the individual. This is consistent with one of the key principles of the life-course framework, namely the role of human agency. In fact, it is the very notion of selection that life-course scholars specifically invoke when suggesting that agency or choice has a pivotal role in the life course and life-course transitions. According to Elder (1994, p. 6): “Within the constraints of their world, people are planful and make choices among options that construct their life course…selection processes have become increasingly important in understanding life course development.”
Although scholars have largely attended to issues related to selection when studying turning points, there are other types of endogeneity—specifically, simultaneity bias and reverse causality, such as when offending alters marriage or employment. Just as agency and selection are central themes in the life-course framework, interdependency is also a key theme; it may, in fact, be more complicated to address and less examined compared to selection bias.
IDENTIFYING TURNING POINTS IN THE PRESENCE OF SELECTION BIAS
Consider the following model, which specifies offending as a function of age and marriage:4
where individuals and age are subscripted by i and t, respectively, ai represents fixed, individual differences, and εit is a random error. In this model, we wish to identify the key parameter β, or the effect of marriage on subsequent offending.5 To make a causal interpretation of β in a normal regression, we must be willing to assume that cov(marriageit, ai)=0. If there are unobserved differences between those who select into marriage and those who do not that are also related to offending, then this assumption fails, meaning, to identify β, we must devise a more plausible identification strategy.
Marriage and Crime
Disentangling selection from causation is a key theme of research on the marriage effect, starting with Sampson et al. (2006, p. 469), who refer to it as a first-order issue that must be addressed before attempting to study any theorized mechanisms through which it operates. In general, most studies seek to employ some sort of counterfactual design, where individuals who marry are compared to others who do not but are otherwise comparable on observable dimensions. One common strategy is to create a suitable counterfactual through the use of a balancing score, such as a propensity score (Rosenbaum 2002, Rosenbaum & Rubin 1984). Here, a counterfactual can be thought of as an individual who, conditional on all observable characteristics, has the same probability of entering marriage (perhaps even at age t) as the treated subject but does not get married. Thus, a reasonable control group can be constructed, and, under appropriate assumptions, a causal effect of the treatment can be identified.
For instance, King et al. (2007) used propensity scores constructed from 16 covariates to match married individuals in the National Youth Survey to similar nonmarried individuals and observed a reduction effect of marriage on subsequent criminality. As hypothesized, the authors also observed an important effect of selection into marriage among this sample. Theobald & Farrington (2009) tested for a marriage effect in the Cambridge data by matching criminal histories of 162 males who entered marriage. They observed a reduction in offending for those who were married between the ages 18–24 (who they termed early or mid-range) but not for those who married after age 25. Using the Glueck data, Sampson et al. (2006) employed a counterfactual approach using inverse probability-of-treatment weighting, which allows for the time-varying nature of the treatment. For the same man, they found that being married reduced the odds of crime, on average, approximately 35% compared to being unmarried.
Of course, a key assumption of propensity score-based approaches necessary for a causal interpretation is that treatment (in this case entry into the turning point) is random, conditional on observable characteristics (i.e., selection on observables) (Heckman & Robb 1985). Put differently, this means we must be willing to assume that those who marry and those who do not differ only by a set of observable factors for which we can directly control.6 If there are key factors that are not directly observed, residual bias due to selection may still be a problem. To deal with concern about fixed, unobserved heterogeneity, another analytic strategy is to directly eliminate ai as a confounder by using some type of within-individual estimator such as a first-difference or fixed-effects regression (Wooldridge 2002). In this case, the individual represents his or her own counterfactual.
For instance, Bersani & Doherty (2013) conduct multilevel, within-individual analysis on subjects in the NLSY and observe a reduction in criminality that they attribute to the marriage effect. However, they also observe that the marriage effect disappears after marriage dissolution, suggesting that the effect may not be permanent. Similarly, van Schellen et al. (2012a) conducted fixed-effects Poisson models using data from the Criminal Career and Life-Course Study in the Netherlands and found marriage reduces the frequency of conviction if the marriage is to a spouse with no conviction history.
Although panel fixed-effect regression models are a powerful analytic strategy, the method only eliminates fixed unobserved heterogeneity. If there are important unobserved confounders that are time varying, then bias is still likely. Blokland & Nieuwbeerta (2005, p. 1,229) offer the following caution of within-person analysis on their own results: “It must be noted that we cannot entirely rule out the possibility that our results are due to some unmeasured time-varying variable that influences both life circumstances and crime.” As we develop below, there are reasons to suspect that certain characteristics, such as attitudes or propensity, that are important to explaining both selection into the turning point and the outcome may, in fact, be time varying.
Employment and Crime
Lageson & Uggen (2013) reviewed studies that examine how crime affects work and work affects crime. Overall, there is convincing evidence to suggest that both pathways are important (see also Uggen & Wakefield 2008). Early studies such as West & Farrington (1977) using the Cambridge cohort and Wolfgang et al. (1972) using a Philadelphia cohort found a positive relationship between unemployment and criminal activity. Similar to marriage and crime, more recent developments in the employment and crime literature have increasingly been associated with empirical strategies that attempt to rule out selection bias. For example, Aaltonen and colleagues (2013) estimate fixed-effects regression models to measure the association between changes in unemployment and changes in violent crime, property crime, and driving under the influence. The researchers found a positive relationship between unemployment and property crime but not violent crime or driving under the influence. More recently, Apel & Horney (2017) use the Second Nebraska Inmate Study to estimate fixed-effects multilevel models and found that employment reduces offending but only for men who were committed to their jobs. Verbruggen et al. (2015) used random and fixed-effect models on a sample of Dutch men and women who were incarcerated in adolescence and found employment was associated with a reduction in offending for both men and women. Based on findings from a Norwegian sample, Skardhamar & Savolainen (2014) used spline-regression techniques and found little support for the causal effect of employment on crime.7
A key difference between employment and marriage is the possibility of experimental manipulation of employment. Overall, results of the evaluations demonstrate that the beneficial effects of employment programs are mixed (Sherman et al. 1998, Visher et al. 2005). These studies are concerned with assessing the effectiveness of in-prison vocational training programs and work-release programs to reduce recidivism rates for ex-prisoners. For example, in their meta-analysis of eight studies using random assignment experimental designs, Visher et al. (2005) examined the effects of noncustodial employment services for ex-offenders in employment programs. The results revealed that the eight interventions had no significant effect on the probability that participants would be rearrested. The mixed effects of employment programs have been echoed in more recent reviews (i.e., Bushway & Apel 2012, Crutchfield 2014). Cook et al. (2015) present results of a randomized control trial among high-risk offenders in Milwaukee. Participants who received subsidized work were significantly less likely to be rearrested after the first year of release, but there was no difference in the rates of reincarceration. Zweign et al. (2011) evaluated a Center for Employment Opportunities transitional job program and found long-term effects for a reduction in recidivism, especially for those at highest risk of recidivating.
In sum, scholars who study turning points have employed advanced empirical strategies, such as propensity score methods, fixed-effect regression models, and randomized control trials, to address selection biases. These strategies have been used with increasing sophistication and regularity, greatly improving our understanding of how important life events can potentially have causal effects on offending. However, as we have argued throughout this review, a close reader of the seminal life-course literature might caution that explaining away selection bias departs from the original conceptualization of turning points, where choice, human agency, and cumulative effects are integral to the life-course framework.
IDENTIFYING TURNING POINTS IN THE PRESENCE OF SIMULTANEITY BIAS
To be clear, we laud the efforts of those who attempt to move past merely demonstrating associations between life events and subsequent offending and deal more explicitly with issues of disentangling selection bias. However, as we noted, selection bias is only one way to think of endogeneity in the problem context. A second and, in our reading of life-course theory, more critical way in which endogeneity manifests itself into turning points is through simultaneity bias.
To illustrate, consider the following system of equations:
In this case, β is still the key parameter we wish to identify (i.e., the marriage effect), bi and ci denote fixed, unobserved individual differences, but now marriage can be thought to depend on offending. In other words, the key explanatory term (marriage) is, in fact, being jointly determined with the independent variable (offending). First, merely establishing clear time order between offending and the event if there is unobserved, latent variation that is driving both factors is alone not a sufficient condition to identify β (or π). More importantly, merely controlling for observable factors (e.g., through matching) to compare those who marry and those who do not is also not sufficient to identify β.
Life-course scholars have described turning points and the measurement of turning points as recursive processes, where the measures and processes of change are interrelated (Pickles & Rutter 1991, Winship & Mare 1983). In fact, scholars have employed empirical strategies such as simultaneous equation models to capture the recursive nature of turning points (e.g., Chamberlain & Griliches 1975, Winship & Mare 1983). LaBel et al. (2008, p. 132) astutely observed that disentangling the differential impacts of external and internal processes is extremely difficult and that “longitudinal studies can statistically control for unmeasured, stable differences across individuals, but fail to take more dynamic individual-level factors, including personal goals and motivation, into account.”8
Scholars have argued that cognitive changes occur prior to entering into turning points, and turning points are important to the desistance process but are not necessary for desistance (e.g., Giordano et al. 2002, Maruna 2001, Paternoster & Bushway 2009). Giordano et al. (2002) argue that individuals must be aware and be willing to leave behind a life of crime for positive life events to serve as hooks for change. Offenders must not only be cognizant of their willingness to change but must also take advantage of opportunities for change. Giordano et al.’s (2002) depiction of the desistance process incorporates a complex relationship between human agency, cognitive transformations, and structural changes.
Paternoster & Bushway's (2009) Identity Theory of Desistance (ITD) emphasizes human agency and choice in the desistance process and asserts that other theories present an oversocialized view of human nature. Paternoster & Bushway (2009) argue that offenders are at the brink of entering into the desistance process when they view the future of their continued criminality with fear and apprehension. It is this cluster of internal changes in identity and preferences and the motivations to become the person that one wants to be that facilitate prosocial opportunities like marriage and employment. Paternoster et al. (2015, p. 216) note that ITD can articulate how “a change in a former offender's identity that both explains the movement into conventional roles or ‘hooks’ and explains why those who had previously been involved in crime would ever be receptive to these prosocial influences.”9 The importance of considering other notable theories of desistance for the purposes of our argument is that they clearly portray an intricate process of desistance that includes both selection processes and simultaneity in important life events and desistance.
Criminologists also specifically point to covariates such as changes in social bonds, informal social control, routine activities, and time spent with peers in relation to both entry into marriage (and/or work) and criminal activity (Sampson & Laub 1993, Warr 1998). Early on, Thornberry & Christenson (1984) demonstrated the simultaneity of employment and crime using a subsample of boys from the Philadelphia birth cohort of 1945 by showing that unemployment and crime have reciprocal causal effects. Barnes et al. (2014, p. 234) explicitly argue that the “marriage-crime relationship may be the result of an endogenous relationship between marital status and criminality such that offenders are less likely to marry and married individuals are less likely to offend.” Using the NLSY, Barnes et al. (2014) used cross-lagged models to estimate the reciprocal relationship between marriage and crime and results largely supported their hypothesis that marriage and crime operate in tandem. There is established literature on the simultaneity of marriage decisions and other outcomes such as those in the labor market (e.g., Choo & Siow 2006). If there is, in fact, a reciprocal relationship between turning points and crime, then merely treating differences in those who marry and those who do not as selection bias is not sufficient for identification.
Although not a turning point, it is possible under this specification to identify the π in Equation 2b. van Schellen et al. (2012b) conversely considered the impact of criminal history on marriage outcomes. Using a large registry sample of Dutch offenders, they found that having a criminal history lowered the unconditional probability of getting married and, conditional on getting married, individuals were likely to select a more criminal partner. Again suggesting that timing matters, the longer in the past a conviction occurred, the less it matters in the process. Huebner (2005) examined the effect of incarceration on the probability of marriage and full-time employment and found that incarceration was associated with a lower probability of both marriage and gaining employment. The growing literature investigating the mark of a criminal record on gaining employment also overwhelmingly corroborates with the notion that the relationship between turning points and criminal activity is complex and bidirectional (i.e., Pager 2003, Uggen et al. 2014).
Title: Large-Scale Multi-Label Learning with Incomplete Label Assignments
Authors:Xiangnan Kong, Zhaoming Wu, Li-Jia Li, Ruofei Zhang, Philip S. Yu, Hang Wu, Wei Fan
(Submitted on 6 Jul 2014)
Abstract: Multi-label learning deals with the classification problems where each instance can be assigned with multiple labels simultaneously. Conventional multi-label learning approaches mainly focus on exploiting label correlations. It is usually assumed, explicitly or implicitly, that the label sets for training instances are fully labeled without any missing labels. However, in many real-world multi-label datasets, the label assignments for training instances can be incomplete. Some ground-truth labels can be missed by the labeler from the label set. This problem is especially typical when the number instances is very large, and the labeling cost is very high, which makes it almost impossible to get a fully labeled training set. In this paper, we study the problem of large-scale multi-label learning with incomplete label assignments. We propose an approach, called MPU, based upon positive and unlabeled stochastic gradient descent and stacked models. Unlike prior works, our method can effectively and efficiently consider missing labels and label correlations simultaneously, and is very scalable, that has linear time complexities over the size of the data. Extensive experiments on two real-world multi-label datasets show that our MPU model consistently outperform other commonly-used baselines.
Submission historyFrom: Xiangnan Kong [view email]
[v1] Sun, 6 Jul 2014 20:13:48 GMT (154kb,D)
Which authors of this paper are endorsers? | Disable MathJax (What is MathJax?)