Can Punishment Encourage Offending? Investigating The Resetting Effect

Foundations of Deterrence Theory and the Positive Punishment Effect

• Doctrine of Specific Deterrence:

  • Posits that subjecting an individual to criminal sanctions deters them from committing subsequent offenses.
  • Under traditional theory, experiencing punishment transitions the abstract threat of the law into concrete reality, intensifying fear and making future consequences clearer and more compelling than theoretical knowledge alone.
  • A central premise of deterrence theory is that sanction recipients subsequently view future legal threats with a higher perceived threat level than they did prior to being sanctioned.

• Central Perceptual Threat Dimensions:

  • Deterrence theory highlights three distinct sanction threat perceptions:
  1. Sanction Certainty: The perceived probability or likelihood of apprehension and sanction.
  2. Sanction Severity: The perceived harshness or disutility of the sanction.
  3. Sanction Celerity: The perceived swiftness with which the sanction is applied.
  • Empirical deterrence research consistently demonstrates that sanction certainty is the most influential dimension in guiding human conduct. Certainty incorporates the core probabilistic elements of sanction threats.

• Bayesian Updating Framework:

  • Rational choice economics and decision theory utilize Bayesian updating to model how individuals process new information to revise judgments.
  • Formally, experiencing punishment communicates definitive information: the individual exposed themselves to detection, and detection/punishment actually occurred.
  • Under rational actor models, a punished individual should update their prior probability estimate upward, resulting in a higher perceived certainty of future punishment compared to unpunished or less punished individuals, controlling for offending propensity.

• The Positive Punishment Effect and Perceptual Anomalies:

  • Counter to classic deterrence theory, empirical studies across diverse populations consistently document a positive punishment effect: punished individuals exhibit a higher likelihood of future offending and lower perceived sanction certainty than unpunished or less punished individuals.
  • Observed across distinct subject demographics:
  • High school students ($1981$ southeastern U.S. survey).
  • Licensed drivers ($1986$ nationwide survey; telephone survey of $1686$ drivers).
  • University undergraduates ($1999$ southwestern U.S. public university survey).
  • Observed across distinct offense categories:
  • Underage alcohol consumption.
  • Marijuana use.
  • Driving under the influence (DUI) / driving while intoxicated (DWI).
  • Defiance theory suggests that individuals who perceive prohibitions against mala prohibita acts as unjust may re-offend intensely as social protest following punishment. However, defiance fails to explain why punished individuals simultaneously lower their estimates of sanction certainty.

The Selection Hypothesis versus the Resetting Hypothesis

• The Selection Hypothesis:

  • Maintains that punishment coincides with subsequent offending without causing or encouraging it.
  • Punishment serves as a selection mechanism that identifies the most active, determined, and committed offenders.
  • Committed offenders engage in frequent criminal acts, increasing their cumulative exposure to detection. Consequently, they are the most likely to experience punishment.
  • Committed offenders hold deeply entrenched beliefs that punishment is unlikely—views that endure even after experiencing one or more sanctions. Professional shoplifters, for instance, treat arrest and jail time as routine operational hazards.
  • Testable Implications:
  1. Punished individuals possess a higher concentration of offending risk factors (situational and individual correlates) than unpunished counterparts.
  2. Offending risk status, rather than punishment experience itself, drives lower sanction certainty estimates. Controlling for risk status eliminates the negative relationship between punishment experience and perceived certainty ($\pi_{11} < \pi_{21}$ and $\pi_{12} < \pi_{22}$).

• The Resetting Hypothesis:

  • Proposes that experiencing punishment causes individuals to actively lower their sanction certainty estimates due to a systematic decision-making heuristic: the gambler's fallacy.
  • The gambler's fallacy occurs when individuals treat independent chance events as interdependent, expecting short-run outcomes to reflect long-run statistical distributions.
  • Mathematical representation: In a 3-digit lottery where each digit ranges randomly from $0$ to $9$, draws are statistically independent. Over the long run, any specific sequence is rare ($P = 0.1^3 = 0.001$). Laboring under the fallacy, individuals assume that after a sequence hits, its immediate probability drops because "things must even out."
  • Analogous real-world and cultural examples:
  • Card players increasing bets after losing consecutive hands, believing a win is due.
  • Lottery players reducing wagers on numbers that recently won.
  • Real estate decisions based on perceived "disaster-proofing" after an unprecedented accident (e.g., a plane crash into a house).
  • Application to offending decisions: Offenders operate under the assumption that apprehension is rare ("lightning rarely strikes twice"). Experiencing punishment causes offenders to rationalize that their unpleasant experience was due to temporary bad luck and that they are now insulated from immediate re-apprehension. They "reset" their certainty estimate downward, prompting more earnest re-offending ("Next time I won't be caught").
  • Testable Implications:
  1. The negative relationship between punishment and perceived certainty is causal rather than spurious. Controlling for offending risk status does not eliminate the negative relationship between punishment experience and perceived sanction certainty ($\pi_{12} < \pi_{11}$ and $\pi_{22} < \pi_{21}$).
  2. The lower sanction certainty among punished individuals is explicitly driven by a gambler's fallacy cognitive bias.
  • Two Manifestations of Resetting:
  • Cross-sectional / Cross-group: Comparison of mean certainty estimates between punished and unpunished groups, controlling for risk status.
  • Longitudinal / Intraindividual: Assessment of within-person reductions in sanction certainty estimates measured prior to and following a punishment experience.

Methodological Framework and Study Design

• Sample and Survey Administration:

  • Subject pool consisted of $N = 256$ undergraduate students ($N = 253$ complete cases in multivariate regressions) recruited from large classes and posted flyers at a public university in the southwestern United States.
  • Focus offense: Driving under the influence of alcohol.

• Experimental Scenario:

  • Participants read a standardized hypothetical scenario: driving alone to a local bar, consuming enough drinks to exceed the legal blood alcohol limit, residing $10\,\text{miles}$ away with early morning work obligations, and deciding whether to drive home or leave the vehicle (requiring an early morning return to retrieve it).

• Measurement Constructs:

  • Sanction Certainty: Respondent's estimated likelihood ($0$ to $100$ scale) of being pulled over by police if driving home drunk under scenario conditions. Apprehension by police serves as a traumatic, salient punishment event.
  • Offending Likelihood: Respondent's estimated likelihood ($0$ to $100$ scale) of driving while intoxicated under scenario conditions.
  • Prior Punishment Experience: Total number of times operating a vehicle and being pulled over by police when believing BAC exceeded the legal limit ($43$ out of $256$ subjects had prior punishment experience; mean = $0.24$, SD = $0.63$).
  • Offending Risk Status Index: A composite $10$-item index combining key situational and individual correlates of drunk driving, dichotomized at the median ($1 = \text{high risk}$, $0 = \text{low risk}$) and summed ($0$ to $10$ score range; mean = $4.84$, SD = $2.67$):
  1. Drinking Behavior: Days in the past month consuming $\ge 3$ alcoholic drinks (mean = $4.69$, SD = $5.00$).
  2. Prior Drunk Driving: Frequency of driving when suspecting BAC exceeded legal limit.
  3. Shoplifting Frequency: Lifetime shoplifting occurrences (mean = $3.69$, SD = $11.74$).
  4. Vandalism Frequency: Lifetime vandalism occurrences (mean = $3.78$, SD = $4.46$).
  5. Fistfighting Frequency: Lifetime fistfight occurrences (mean = $1.28$, SD = $2.90$).
  6. Sex: Coded $1$ for male, $0$ for female ($53\%$ male). In national arrest data ($1999$), males accounted for $84\%$ of drunk driving arrests ($835000$ total) and nearly $5\times$ more violent offense charges.
  7. Impulsivity: Sum of $6$ 5-point Likert items from the Barratt Impulsivity Index and NEO Personality Inventory (items: acting on impulse, spur of moment, considering consequences [reverse coded], rarely making hasty decisions [reverse coded], regretting impulse acts, acting first and thinking later; score range $5$ to $30$).
  8. Peer Drunk Driving: Number of closest $10$ friends driving intoxicated $> 1$ time (mean = $4.50$, SD = $3.34$).
  9. Passenger Stop Frequency: Times riding as a passenger in a vehicle stopped by police (mean = $3.04$, SD = $4.23$).
  10. Party Police Attendance: Times attending a party where police were called (mean = $12.42$, SD = $16.36$).
  • High Risk Classification: Dichotomized cutoff at score $\ge 7$ (upper $27\text{th}$ percentile of distribution).
  • Gambler's Fallacy Operationalization: A coin-flip probability task. A nickel lands on heads $4$ consecutive times; respondents select whether the $5\text{th}$ flip is more likely heads, tails, or equally likely. A total of $51$ respondents ($20\%$) selected tails, demonstrating evening-out reasoning characteristic of resetting.

• Quasi-Experimental Analytic Matrix:

  • Sample partitioned into four cells based on Offending Risk Status (High vs. Low) and Punishment Experience (Not Punished vs. Punished):
  • Cell $\pi_{11}$: High Offending Risk, Not Punished ($n = 47$).
  • Cell $\pi_{12}$: High Offending Risk, Punished ($n = 23$).
  • Cell $\pi_{21}$: Low Offending Risk, Not Punished ($n = 166$).
  • Cell $\pi_{22}$: Low Offending Risk, Punished ($n = 20$).

Empirical Findings and Quantitative Results

• Replication of Positive Punishment Effect and Deterrence Models:

  • Tobit regression models ($N = 253$, accounting for $22\%$ clustering at $0$ likelihood):
  • Model 1 (Positive Punishment Effect): Prior punishment coefficient = $13.41$ ($p < 0.10$, SE = $7.37$). Effect size calculation: multiplying Tobit coefficient by probability of a non-clustered response ($0.78$) yields $(13.41) \times (0.78) = 10.46$ percentage points higher offending likelihood for punished individuals.
  • Model 2 (Perceived Certainty Deterrence Effect): Sanction certainty coefficient = $-0.52$ ($p < 0.01$, SE = $0.12$), confirming that higher perceived certainty deters offending.
  • Model 3 (Combined Model): Sanction certainty coefficient = $-0.50$ ($p < 0.01$, SE = $0.12$); Prior punishment coefficient = $10.54$ ($p > 0.10$, SE = $7.22$). Bivariate correlation between punishment experience and perceived certainty is negative ($r_{\text{punishment, certainty}} = -0.18$).
  • Control Variable Effects across Models:
    • Age: $0.17$ (Model 1), $0.07$ (Model 2), $-0.08$ (Model 3) [all non-significant].
    • Accident Involvement (Self/Friend/Relative): $-1.76$, $-2.03$, $-2.27$ [all non-significant].
    • Male: $15.10$ ($p < 0.01$), $11.72$ ($p < 0.05$), $10.73$ ($p < 0.05$).
    • Weekly Income: $0.04$ ($p < 0.05$ across all models).
    • Frequency of Alcohol Consumption: $1.58$ ($p < 0.01$), $1.41$ ($p < 0.05$), $1.36$ ($p < 0.05$).

• Selection Hypothesis Threshold Test:

  • Logit regression predicting Punishment Experience ($N = 256$, Log-likelihood = $-104.53$):
  • High Offending Risk Score ($\ge 7$): Coefficient = $1.43$ ($p < 0.01$, SE = $0.36$). High-risk individuals are $23$ percentage points more likely to have experienced punishment. Logit transformation formula:       $P = \frac{1}{1 + e^{-(\alpha + \beta X)}}$
  • Age: Coefficient = $0.08$ ($p < 0.05$, SE = $0.04$).
  • Accident: Coefficient = $0.07$ (SE = $0.13$, non-significant).
  • Weekly Income: Coefficient = $-0.001$ (SE = $0.001$, non-significant).

• Evaluation of the Analytical Matrix Means:

  • Observed mean certainty estimates across groups:
  • $\pi_{11}$ (High Risk, Unpunished) = $18$ (SE = $2.43$, $n = 47$)
  • $\pi_{12}$ (High Risk, Punished) = $19$ (SE = $3.79$, $n = 23$)
  • $\pi_{21}$ (Low Risk, Unpunished) = $33$ (SE = $1.95$, $n = 166$)
  • $\pi_{22}$ (Low Risk, Punished) = $22$ (SE = $4.18$, $n = 20$)
  • Statistical structure of empirical results:     $\pi_{21} > \pi_{11} = \pi_{12} = \pi_{22}$
  • Independent $t$-tests show $\pi_{21}$ ($33\%$) is significantly higher than all other three cell means ($p < 0.05$).
  • Findings partially support selection ($\pi_{21} > \pi_{11}$) but fail on $\pi_{12} = \pi_{22}$.
  • Findings partially support resetting ($\pi_{21} > \pi_{22}$) but fail on $\pi_{11} = \pi_{12}$.

Conditioning Factors and the Role of Offending Naivety and Impulsivity

• Risk-Status Conditioning of the Gambler's Bias:

  • Ordinary Least Squares (OLS) regression of Sanction Certainty on Gambler's Bias stratified by offending risk status:
  • High Offending Risk Subsample: Gambler's bias coefficient = $-4.27$ (SE = $4.71$, non-significant); Raw risk index score coefficient = $-1.47$ (SE = $1.85$, non-significant).
  • Low Offending Risk Subsample: Gambler's bias coefficient = $-7.48$ ($p < 0.10$, SE = $4.60$); Raw risk index score coefficient = $-3.62$ ($p < 0.05$, SE = $0.986$).
  • Interpretation: The resetting effect is operative primarily among low-risk, naïve offenders. Controlling for risk score, low-risk subjects exhibiting the gambler's bias perceive sanction certainty to be $7.5$ percentage points lower than those without the bias.

• Disaggregated Risk Factor Analysis (OLS Regression Models):

  • Comparing Gambler's Bias coefficients across Upper vs. Lower distributions of individual risk factors:
  • Drinking Frequency: Upper = $-4.56$ (SE = $4.90$, ns); Lower = $-9.17$ ($p < 0.10$, SE = $5.20$).
  • Drunk Driving Frequency: Upper = $-1.74$ (SE = $0.67$, ns); Lower = $-8.02$ ($p < 0.10$, SE = $4.69$).
  • Shoplifting Frequency: Upper = $-3.66$ (SE = $4.72$, ns); Lower = $-11.80$ ($p < 0.05$, SE = $5.42$).
  • Vandalism Frequency: Upper = $-4.81$ (SE = $4.90$, ns); Lower = $-10.08$ ($p < 0.05$, SE = $5.01$).
  • Fistfighting Frequency: Upper = $-3.77$ (SE = $4.83$, ns); Lower = $-9.75$ ($p < 0.10$, SE = $5.27$).
  • Sex: Male (Upper) = $-3.78$ (SE = $4.29$, ns); Female (Lower) = $-10.93$ ($p < 0.10$, SE = $6.15$).
  • Impulsivity Score: Upper (High Impulsivity) = $-10.12$ ($p < 0.05$, SE = $4.49$); Lower (Low Impulsivity) = $-5.29$ (SE = $6.01$, ns).
  • Friends' Drunk Driving: Upper = $-5.60$ (SE = $4.20$, ns); Lower = $-8.60$ (SE = $6.06$, ns).
  • Passenger in Vehicle Stopped: Upper = $-0.88$ (SE = $4.39$, ns); Lower = $-16.43$ ($p < 0.01$, SE = $6.16$).
  • Party Police Attendance: Upper = $-3.78$ (SE = $6.44$, ns); Lower = $-9.17$ ($p < 0.05$, SE = $4.10$).

• Psychological Mechanics of the Resetting Effect:

  • Biases operate where there is cognitive "room to operate." Experienced offenders often make crime decisions without considering sanction certainty at all, rendering them immune to probabilistic cognitive biases.
  • Resetting requires a specific confluence of psychological traits: low offending experience (naivety) combined with high impulsivity, where impulsive traits prompt cavalier or immediate cognitive processing of punishment events.

Theoretical Limitations, Future Research Directions, and Policy Implications

• Non-Probabilistic Sanction Threat Dimensions:

  • Severity and Celerity: Differ fundamentally from certainty because they are non-probabilistic.
  • Whether punishment increases or decreases perceived severity and celerity depends on ex ante expectations under a Bayesian framework:
  • If an experienced sanction is less severe or celeritous than anticipated, perception updates downward, emboldening future offending.
  • If a sanction exceeds severity or celerity expectations, perception updates upward, strengthening deterrence.

• Study Limitations:

  1. Single offense domain (drunk driving) examined within a college student sample.
  2. Single-item coin-flip measure assumes the gambler's fallacy is non-domain-specific across chance processes and criminal offending.
  3. Unmeasured psychological constructs such as unrealistic optimism and self-serving bias.
  4. Exclusion of offender learning and expertise acquisition (e.g., active offenders learning specific evasion tactics that reduce actual risk exposure).
  5. Cross-sectional design captures only cross-group comparisons rather than longitudinal intraindividual changes over time.

• Policy Implications for Criminal Justice:

  • Specific deterrence policy efficacy is challenged by the positive punishment effect and resetting phenomenon.
  • Policies should focus formal enforcement efforts on apprehending committed, high-risk offenders and leveraging their apprehension to promote vicarious deterrence among the general public.
  • Concentrating enforcement on general/vicarious deterrence maximizes overall deterrent threat while minimizing the potential for resetting biases among naïve offenders.

• Unexplored Research Avenues:

  • Triggering Mechanisms: Testing whether apprehension for one specific offense triggers resetting across other crime categories (crime-specific vs. crime-general resetting).
  • Nonpunishment Triggering Events: Evaluating whether negative non-sanction events (e.g., traffic accidents) induce resetting based on "unlucky" event rationalizations.
  • Vicarious Resetting: Investigating whether punishment or triggering events occurring to family members or close peers induce resetting in an individual ("it would never happen to both of us").