This working paper examines the dynamics of offender risk score change across multiple LSI-R reassessments, addressing the question of what actually changes when offender risk scores change. It represents an important empirical and theoretical contribution to the understanding of dynamic risk assessment in criminal justice, connecting criminological risk assessment to broader literatures in psychometric theory, complexity science, cybernetics, and nonlinear dynamics. The prevailing paradigm in corrections treats offender risk as dynamic but implicitly assumes that change is linear — that when criminogenic needs are met, risk will decrease and remain decreased. This paper challenges that assumption directly, demonstrating that the Newtonian worldview underlying current dynamic risk assessment is inadequate for capturing the actual dynamics of offender risk over time. Using multiple LSI-R reassessments from a Central Minnesota offender population, this study finds that the majority of offenders (73.6%) show no significant linear change in risk over time, while a minority show significant linear decreases (18.0%) or increases (8.4%). More importantly, the risk trajectories of most offenders follow a cyclic sinusoidal pattern — fluctuating around a more slowly changing average level of risk — rather than the linear change assumed by current practice. This finding has profound implications for offender treatment research, since single posttest assessments conducted within a one year period may capture a high point, low point, or intermediate point in a naturally fluctuating risk cycle, producing ambiguous or misleading results. The paper addresses several important methodological issues arising from these findings. True score theory is applied to the measurement of change in risk scores, and methods are demonstrated for calculating the Reliable Change Index (RCI) to determine whether observed changes in risk scores represent clinically significant change. The non-ergodic nature of offender risk data is identified as a critical issue — group level results cannot be used to predict individual behavior and individual behavior cannot be used to predict group level results — with important implications for study design and statistical analysis. Practical guidance is provided for both practitioners and researchers. For practitioners, a second LSI-R assessment at approximately six months is recommended, with subsequent reassessments delayed until the 2-3 year period when rank order changes in risk become more meaningful. For researchers, the sinusoidal nature of risk trajectories requires study designs using multiple carefully spaced measurements rather than single posttest periods. Theoretical implications are discussed connecting the nonlinear sinusoidal risk trajectories to the Trans-Theoretical Model of behavior change, Lewin's approach-avoidance conflict framework, cybernetic homeostasis models, and complexity theory. The sinusoidal patterns observed are consistent with the micro/macro chaos framework developed more fully in the broader Physics of Living Systems theoretical framework, in which apparent stability at the macro level reflects micro level chaotic processes operating around propensity selection thresholds. This paper connects directly to Arnold (2007), The Dynamic Predictive Validity of the Level of Service Inventory-Revised, Arnold (2009), The Nonlinear Dynamics of Criminal Behavior, and Arnold (2022), An Examination of the Dynamic Predictive Validity Thesis, forming part of a sustained body of empirical and theoretical work examining the dynamics of offender risk assessment. Licensing Note regarding the Copyright Notice in the Document: This paper was originally copyrighted by the author in 2010. The author is uploading this work to Zenodo and licensing it under Creative Commons Attribution 4.0 International. The copyright notice in the document reflects the original date of authorship.
Thomas K. Arnold (Tue,) studied this question.