Risk Assessment Approaches

Violence risk assessment is relevant to the field of law and psychology because it occurs at numerous junctures in the legal system, and it is one of the key areas of research and clinical practice in forensic psychology. This entry reviews two primary approaches to risk assessment: unstructured and structured. The former approach, sometimes also called clinical prediction or judgment, imposes no rules on the decision-making process, whereas the latter approach does. Two primary approaches to structured risk assessment include (1) actuarial and (2) structured professional judgment (SPJ). Although both structured approaches impose rules in terms of which risk factors are considered and defined, actuarial risk assessment uses an algorithm to combine risk factors into a final decision, whereas SPJ does not. This entry will describe these approaches, along with their attendant strengths and weaknesses.

The Relevance and Context of Risk Assessment

Risk assessment informs decisions about future violence. In numerous legal and clinical practice areas, such decisions are required by statute, professional ethics, or common law. For instance, in most jurisdictions, a person must pose, inter alia, a risk of harm to others (or to self) to be involuntarily civilly committed. The release of prisoners from institutional to community placement is typically contingent on whether they constitute an undue risk to public safety. Risk assessment is the basis for decision making in these situations. Depending on jurisdiction in the United States, psychologists and other mental health professionals have common law and ethical duties to protect third parties from the violence posed by their patients. Correspondingly, risk assessment is used to determine whether a sexual offender will be subjected to postsentence involuntary commitment under sexual predator laws. Analogous “indeterminate sentencing” provisions exist in non-U.S. jurisdictions as well, for example, Canada and the United Kingdom. As such, risk assessment plays a pivotal role in balancing public safety with constitutionally protected rights and freedoms such as liberty.

Approaches to Risk Assessment

Contemporary approaches to risk assessment have been heavily influenced by Paul Meehl’s distinction between actuarial and clinical prediction. The former refers to decision-making procedures that involve the formal combination of variables or pieces of information, by way of equations or other algorithmic processes, to reach a decision. The latter is defined by a lack of such rules. In contemporary risk assessment, approaches may be generally classified as structured and unstructured. As described below, unstructured risk assessment is, in essence, the clinical prediction to which Paul Meehl referred. Structured risk assessment includes actuarial prediction as well as a more recent approach termed structured professional judgment.

Unstructured Risk Assessment

Conventionally, unstructured clinical judgment is the most common approach to appraising an individual’s risk for violence. By definition, it is based primarily on professional opinion, intuition, and clinical experience. Assessors have absolute discretion in terms of selecting risk factors to consider, how to conceptualize them, how to synthesize case material, and how to interpret this information to render decisions. As such, this method is inherently informal and subjective. Although clinical judgment is a routine and necessary component within many clinical decision-making contexts, the defining feature of clinical judgment in terms of prediction is the lack of rules to integrate case information. Although this permits flexibility, ostensible widespread applicability, and relevance to the individual patient, there are numerous problems with this approach.

First, because of the lack of rules, critics contend that the technique generally lacks consistency because independent clinicians may focus on dissimilar sources of information and subsequently form disparate conclusions (low interrater reliability). Second, clinicians may or may not attend to variables that actually relate to violent behavior (low content validity). Third, either failing to attend to important risk factors, attending to irrelevant variables, or giving improper weight to risk factors, will inevitably decrease the accuracy of decisions (low predictive validity). Fourth, detractors argue that unaided clinical decision making precludes transparency of decision making, which is essential in a legal forum (low legal helpfulness). Other factors leading to low (or at least inconsistent) accuracy include susceptibility to decisional biases and heuristics, failure to consider base rate information, failure to integrate situational information, and a lack of specificity about the criterion variable. Research bears these weaknesses out: The accuracy of unstructured risk assessment has been shown (a) to vary considerably across different clinicians and (b) though predictive of violence, to be less strongly related to violence than more systematic approaches.

Structured Risk Assessment

In response to the shortcomings of the unstructured clinical approach and the disquieting implications these held for important legal decisions, researchers started to investigate structured approaches. Contemporary structured risk assessment approaches share common features such as (a) inclusion of a fixed set of risk factors, (b) operational definitions of risk factors, (c) scoring or coding procedures for risk factors, and (d) direction for how to integrate risk factors to reach a final decision about risk. As described below, however, there are important differences between the two primary approaches to structured risk assessment.

Actuarial Prediction

The first structured approach that was investigated was actuarial prediction. Technically, a prediction approach is said to be actuarial when it uses formal rules to combine variables or risk factors to make a decision. This process, therefore, involves the formal application of a predetermined set of explicit and formulaic decision rules to make a decision about the likelihood of violence. The actuarial approach has been described as algorithmic, mechanical, well specified, and completely reproducible. An associated, though not defining, feature of actuarial prediction is the use of empirical item selection; that is, the variables that comprise risk factors on an actuarial risk assessment measure are often selected because they demonstrated statistical associations with violence in one (or, more rarely, two or more) specific construction or calibration sample. Another associated feature of actuarial prediction is that the risk factors that are derived empirically are typically weighted according to the strength of association with violence observed in the construction sample(s).

The primary argument in support of actuarial prediction techniques is that they facilitate interrater reliability and predictive validity, especially in comparison with unstructured approaches. Because actuarial procedures use explicit rules for combining risk factors, they yield the same decision regardless of who uses them (high interrater reliability), and given the presence of the same risk factors across cases, they yield the same outcome. Furthermore, they are transparent (reviewable and accountable). Many actuarial prediction techniques are statistically optimized because they weigh variables according to their relationship with violence. Hence, at least in the samples in which they were developed, they tend to have high predictive validity in comparison with unstructured approaches.

There is general agreement that the actuarial approach to risk assessment yields higher predictive accuracy than does the unstructured approach when the two are compared for group-based (nomothetic) predictions within the same sample. Perhaps the best evidence of this stems from a meta-analysis of 136 studies conducted by William Grove and colleagues that directly compared actuarial prediction with unstructured clinical prediction. Actuarial prediction was more accurate than clinical prediction in approximately one-third to half of the studies. In approximately half the studies, there was no difference in predictive accuracy. In a small minority of studies, unstructured clinical prediction was more accurate. On average, actuarial prediction was more accurate than clinical prediction by an approximately 10% increase in hit rate.

Despite the important advantages of enhanced inter-rater reliability and predictive validity that actuarial prediction possesses, commentators have noted several weaknesses. Perhaps most important, the predictive properties of actuarial models tend to be optimized within the sample of development, with no guarantee that these properties will apply to novel settings or samples (generalizability). For this reason, the precise numerical probability estimates, or bright-line classification cut scores, that tend to be used in actuarial prediction are in crucial need of cross-validation and replication prior to use.

Second, some actuarial techniques may have limited clinical applicability, in that decision makers may be concerned about violence in a context (e.g., imminent violence) that is incongruent with existing actuarial protocols constructed with a specific set of conditions (e.g., a long-term follow-up period). Third, actuarial approaches tend to ignore low base rate factors that failed to enter nomothetically derived statistical equations because of their rarity or their case-specific nature, even if they may be important in individual cases. Under the strictest actuarial approaches, any extraneous information not contained on the instrument cannot be considered. Fourth, some actuarial models have been criticized for not being helpful in terms of risk management, treatment, or risk reduction more broadly because they tend to focus on static risk factors as opposed to dynamic (changeable) risk factors that may be better suited to treatment efforts.

Structured Professional Judgment

To contend with these weaknesses, a more recently developed risk assessment approach, termed structured professional judgment, has been forwarded. Like most actuarial approaches, the SPJ approach specifies a fixed set of operationally defined risk factors with explicit coding procedures. The purpose of this structure is to facilitate both interrater reliability and comprehensive domain coverage, or content validity. It has three primary differences compared with most actuarial approaches. First, SPJ approaches use logical or rational item selection as opposed to empirical item selection procedures to select risk factors. This process involves extensive consultation of the scientific and professional literature to select risk factors with broad support across contexts. In theory, this approach fosters generalizability as well as comprehensiveness of the set of risk factors.

Second, SPJ approaches do not require algorithmic combinations of risk factors to derive risk estimates, and hence they are not actuarial. There are four primary reasons why SPJ approaches do not adopt algorithmic item combinatory procedures. (1) Such procedures are susceptible to degradation of predictive accuracy across contexts, meaning that a cutoff score in one sample cannot be assumed to apply to another context. (2) While combination rules promote consistency, they may do so at the expense of individual relevance; that is, certain risk factors will be more relevant for one person’s violent risk than for another’s risk, and a risk assessment process should be able to account for this differential individual rele-vance. (3) Decisions based on fixed algorithmic procedures presume that the future is fixed as well; if circumstances change, the actuarial estimate may be invalid. (4) There may be cases with only a few risk factors present, but their salience compels a conclusion of high risk.

SPJ approaches attempt to optimize the relevance of nomothetically derived risk factors to the individual— which, whether for legal or clinical purposes, is the level of decision making. Final decisions of low, moderate, or high risk are formed by decision makers after consideration of the number and relevance of risk factors present in the case and the intensity and urgency of any necessary intervention or management strategies to mitigate risk. The SPJ model does not provide estimated numerical probability levels of future vio-lence for the individual case, because it is assumed that it is not actually possible to do so given the problems with lack of stability of such procedures, as reviewed above. Furthermore, actuarially derived numerical probability estimates are group-based estimates (i.e., 53 of 100 persons in X risk group were violent); their applicability to what an individual who was not in this group might do in future is tenuous.

Critics of the SPJ approach have argued that it lowers reliability and validity through the allowance of discretion at the variable integration phase of decision making. Though this is a controversial aspect of SPJ, research to date suggests that the reliability and predictive validity of the SPJ approach are at least comparable with the reliability and predictive validity of the actuarial approach—and in some studies, exceed them.

Researchers continue to study the strengths and limits of both actuarial and SPJ approaches to risk assessment. Both have promise, and both have limitations. The field would benefit from research on how to increase individual relevance, treatment relevance, and cross-validated generalizability of actuarial procedures. In terms of SPJ research, questions in need of research include whether additional structure can be added to the final decision without introducing the problems associated with actuarial decision making.

References:

  1. Douglas, K. S., & Kropp, P. R. (2002). A prevention-based paradigm for violence risk assessment: Clinical and research applications. Criminal Justice and Behavior, 29, 617-658.
  2. Grove, W. M., Zald, D. H., Lebow, B. S., Snitz, B. E., & Nelson, C. (2000). Clinical versus mechanical prediction: A meta-analysis. Psychological Assessment, 12, 19-30.
  3. Meehl, P. E. (1954). Clinical versus statistical prediction. Minneapolis: University of Minnesota Press.
  4. Monahan, J., Steadman, H. J., Silver, E., Appelbaum, P. S., Robbins, P. C., Mulvey, E. P., et al. (2001). Rethinking risk assessment: The MacArthur study of mental disorder and violence. New York: Oxford University Press.

See also: