Asymmetrical causation and criminal desistance.

• Author: Uggen, Christopher

Although criminologists have long been concerned with desistance or cessation from crime,(1) tests of theory are typically based on etiological invesugations.(2) Desistance studies, in contrast, have historically been used for program evaluations, undertaken by professionals in social work, clinical psychology, and corrections. We argue that theory and research on desistance is absolutely critical to advancing scientific and policy goals. We do not attempt to break new theoretical or empirical ground in this paper, but instead present a systematic explication of the argument for desistance research.

1. CAUSAL ASYMMETRY, CRIME, AND DESISTANCE

   1. THE GROWING UNIVERSE OF POTENTIAL DESISTERS

      Though ex-offenders are socially marginalized in America, they are no longer a statistically marginal group. As of December 31, 1997, American state and federal prisons held 1,244,554 prisoners, with over 500,000 additional inmates held in local jails.(3) Despite recent increases in mandatory minimum and mandatory life sentences, almost all of these prisoners will eventually rejoin civil society.(4) Each year, several hundred thousand releasees pour into the general population, with over 400,000 entering parole in 1997.(5) In fact, over the past twenty-five years the trends in prison release closely mirror rising incarceration rates. Figure 1 plots the number of U.S. prisoners incarcerated and the number of prisoners released each year.(6) As the figure indicates, the two data series are correlated quite closely (r = .98). In fact, more prisoners were released in 1996 than were incarcerated in 1986 and more than twice as many were released in 1996 as were incarcerated in 1976. Some of these released prisoners will resume crime and others will desist from crime--they will temporarily or permanently cease offending--yet little is known about the desistance process. This paper argues that a research program to identify the causes of desistance will advance both scientific and policy concerns.

      [Figure 1 ILLUSTRATION OMITTED]

      Despite a longstanding concern within the discipline, criminologists today devote relatively little attention to deriving theoretical understanding of the desistance process. This is because criminological theory and research are primarily concerned with questions of etiology, or the causes of crime. Discipline-based criminologists such as sociologists, psychologists, and economists have focused their attention on etiological research addressing individual involvement in crime or community crime rates. Unfortunately, the study of these phenomena is conceptually complex, fraught with daunting methodological barriers, and in many (though not all) ways, without policy relevance. Desistance research may prove more rewarding for theory and policy, in part because it is more manageable conceptually and methodologically.

      In this paper, we are less concerned with why people commit crime than with the conditions promoting social reintegration and desistance from crime. We make the following assertions: (1) that the causes of desistance likely differ from the causes of crime; (2) that knowledge of the true causes of desistance will be easier to obtain than knowledge of the true causes of crime; and (3) that it will be possible to translate scientific knowledge about desistance into specific policy interventions. We begin with a general discussion of crime and causality, then present the case for and against desistance research.

   2. CRIMINOLOGICAL PROBLEMS POSED BY THE RUBIN/HOLLAND CAUSAL MODEL

      As social scientists researching crime and conformity, we often set out to make causal inferences. Temporal order, statistical association, and lack of spuriousness are generally accepted as minimal criteria for establishing causality in the social sciences,(7) though these standards are rarely approached in practice. In the study of delinquency, for example, we continue to debate the putative causes: are delinquent friends causes or consequences of involvement in crime? Does family alienation precede or follow from one's involvement with delinquent peers? Definitive answers to such questions are elusive because of both conceptual confusion over the meaning of causality and operational difficulties implementing critical tests of theoretical propositions.

      Rubin(8) and Holland(9) offer a statistical model of causal inference that lays bare the obstacles to establishing causation in criminology. The Rubin/Holland model highlights the difficulties in establishing the causes of crime and illustrates the potential for desistance research. The central prescription of the model is to seek the effects of manipulable causes rather than to trace the causes of observed effects. They arrive at this conclusion in the following manner:

      Holland defines a true causal effect of some factor T on response variable Y for individual unit u as

      [Y.sub.T](U) - [Y.sub.C)(U)(10)

      This difference implies that the effect of any cause T must always be assessed in relation to some other cause, the counterfactual condition not-T or C. The fundamental problem of causal inference, in criminological research as elsewhere, is that it is impossible to observe the value of [Y.sub.T](U) and [Y.sub.C](U) on the same individual person or "unit." If a respondent is employed at age fifteen and commits delinquency at age sixteen, for example, we cannot determine whether she would have committed delinquency had she not been working. Because we can only observe one condition per unit, we face a missing data problem for the counter-factual condition. Holland distinguishes between scientific and statistical solutions to this problem. Both have been applied in criminological research.

      1. Scientific Solutions

         Among the scientific solutions, one can attempt to assure (or simply assume) that each individual unit is identical (assuming unit homogeneity) and submit unit one to treatment T and unit two to treatment C. If so, then the true causal effect is easily obtained as:

         [Y.sub.T]([U.sub.1) - [Y.sub.C]([U.sub.2])

         Alternatively, if it is reasonable to believe that prior exposure does not affect subsequent response, the scientist could expose the same unit to each treatment in succession (assuming temporal stability and causal transience).

         [Y.sub.T(TIME 2)] ([U.sub.1]) - [Y.sub.c(TIME 1)] ([U.sub.1])

         Although such assumptions may be reasonable in laboratory work, they are rarely justified in criminological research: humans are not identical in all relevant respects, and human responses are neither constant over time nor unaffected by previous exposures. To continue the previous example, we cannot assume that a working teen is identical in all relevant respects to a non-working teen, for workers may be more ambitious, less impulsive, or more opportunistic than non-workers. Nor can we assume that exposure to work in eighth grade will have the same effects as exposure to work in tenth grade, or, for that matter, that working in eighth grade will not affect one's response to non-work in tenth grade. Therefore, as social scientists we must often rely on a statistical solution to the inference problem.

      2. Statistical Solutions

         The Rubin/Holland statistical solution is to find the expected value (E) of the average causal effect T, of W (relative to C) over a population U, or

         E([Y.sub.T] - [Y.sub.C]) = T

         which can be expressed as

         T = E([Y.sub.T]) - E([Y.sub.C])

         To continue the earlier example, one could deviate the average number of crimes (or the proportion committing a crime) among workers from the average number of crimes among non-workers. This replaces the unobservable causal effect of T on a specific unit with an estimate of the average causal effect of T over a population of units. Unfortunately, this approach breaks down in practice because an important assumption is unlikely to hold.

      3. The Mean Independence Assumption

         For the statistical solution to hold, we must assume mean independence the mean values on Y for the T group and the C group must be independent of the selection or assignment mechanism (S) that determines whether [Y.sub.T] or [Y.sub.C] is observed for any given unit. In general, mean independence fails in criminology and other social sciences and the observed average treatment effect [T.sub.(OBS)] is not equivalent to the true treatment effect T.(11)

         [T.sub.(OBS)] = E([Y.sub.T]|S=T) - E([Y.sub.C]|S=C)

         The mean independence assumption for groups is thus analogous to the unit homogeneity assumption for units. Both allow the comparison of observed quantifies with latent or unobserved quantities. Only when [Y.sub.T] and [Y.sub.C] are both mean independent of S, however, does

         E([Y.sub.T]) - E([Y.sub.T]|S=T) and E([Y.sub.C]) - E([Y.sub.C]|S=C)

         So that

         T = [T.sub.(OBS)] = E([Y.sub.T]) - E([Y.sub.C])

         In short, the two conditional means E([Y.sub.T]|S=T) and E ([Y.sub.T]|[S=C) must be independent for them both to equal the unconditional mean E([Y.sub.T]).(12) If the selection mechanism S is randomized assignment to employment, this is a reasonable assumption for the work and crime example above. If S is self selection, however, this assumption is invalid: those that self-select into employment are likely to have lower crime means than those that self-select out of employment. If selection is partially determined by "ambition" and ambition is associated with crime, for example, the conditional distributions are unequal and the mean independence assumption breaks down. This is simply one important variant of the more general omitted variable problem that biases parameter estimates. Criminology is particularly vulnerable to violations of mean independence, however, because the selection processes into levels of our independent and dependent variables are so poorly understood.

      4. Selection Mechanisms and Strong Ignorability

         When researchers cannot control the assignment mechanism (S), as in observational studies of delinquency, causal inference requires "strong...