Identification Problems in Probabilistic Measures of Perceived Arrest Risk: Estimating a Partially-Identified Certainty Effect.

Objectives: Test the effect of perceived likelihood of arrest on criminal behavior under a relaxed set of measurement assumptions. Specifically, responses that are commonly associated with inaccurate reporting practices–particularly, the 0%, 50%, and 100% categories–can be treated as partially identified. By doing so, scholars are able to bound the effect of perceived arrest risk on criminality, which provides more credible, although less precise, estimates of β 1 . Scholars can use this approach to not only produce more defensible findings, on the whole, but also gain insight into the possible threat posed by measurement misspecification. Methods: Point estimates of a perceived certainty effect were elicited via Quasi-Poisson regression using data derived from the Pathways to Desistance study. These estimates were subsequently bounded under progressively weaker measurement assumptions by a series of hill-climb algorithms. Results: In nine out of seventeen total algorithms, the worst-case bound remained in the expected direction and was statistically significant. For as long as a relatively minor level of response inaccuracy is assumed, supportive conclusions can be drawn. Conclusions: Support for a certainty effect can be found under relaxed measurement assumptions, up to a point. This not only provides further support for the deterrence hypothesis, but also implies the effect might be somewhat resilient to measurement error. [ABSTRACT FROM AUTHOR]