Commentary on Pedersen and Skardhamar (2010): Does cannabis use predict non-drug offending?
Article first published online: 14 DEC 2009
© 2010 The Author. Journal compilation © 2010 Society for the Study of Addiction
Volume 105, Issue 1, pages 119–120, January 2010
How to Cite
FARRINGTON, D. P. (2010), Commentary on Pedersen and Skardhamar (2010): Does cannabis use predict non-drug offending?. Addiction, 105: 119–120. doi: 10.1111/j.1360-0443.2009.02850.x
- Issue published online: 14 DEC 2009
- Article first published online: 14 DEC 2009
- longitudinal survey;
The paper by Pedersen & Skardhamar  is an excellent contribution to knowledge. They use data from the Young in Norway Longitudinal Study, which is an extremely important four-wave population survey. They show that cannabis use at ages 15 and 20 years predicts subsequent offending, but that the predictability is greater for drug crimes than for non-drug crimes. They also show that the predictability for non-drug crimes becomes non-significant after controlling for prior risk factors, alcohol intoxication and use of other illegal drugs. The analyses are neat and clear. While generally welcoming this excellent paper I wish to raise some issues, and I will divide these into ‘methodological’ and ‘substantive’.
The most important methodological issue is attrition. According to Pedersen et al., a nationally representative sample of seventh and eighth grade students was surveyed, using a self-administered questionnaire, in 1992–93. Of the original sample, 1.5% were excluded [at Time 1 (T1)] because of poor reading skills and 3.0% of the remainder were lost because of lack of student or parent consent or prolonged hospitalization. Of those who completed the T1 questionnaire, 91.6% completed the T2 questionnaire. According to Pedersen , of those who completed the T2 questionnaire, 90.4% consented to be followed-up further. Of those consenters, 84% provided data at T3 and 82% at T4 [1,3]. According to my estimate, about 65% of the original sample (98.5% × 97.0% × 91.6% × 90.4% × 82%) provided data at T4. Similarly, we are told that the cumulative response rate over all data collections was 69%, but this could possibly correspond to about 55% of the original sample.
This is a good response rate for a mail follow-up study, but of course the problem is that those who are lost tend to be the high-risk students: male, with poor school grades, urban residence, and so on [1,3]. Commendably, Pedersen & Skarohamar estimate that the serious offending rate in their sample is about one-third less than in the population (7.8% compared to 11.6%, presumably up to age 27 years). However, the high and biased attrition rate may threaten the validity of the conclusions.
In the Cambridge Study in Delinquent Development (CSDD), 411 London males have been followed-up from age 8 to age 48, using repeated personal interviews . The attrition rate in this study was very low; the percentage of those still alive who were interviewed was 95% at age 18, 94% at age 32 and 93% at age 48. Importantly, the males who were the most difficult to interview were significantly more likely to be convicted , showing that conclusions about predictors of offending in studies with high attrition rates might be misleading.
The second important methodological issue, which is related to some extent to the first, is that the number of offenders is low: only 5.1% at ages 15–20 and 3.5% at ages 20–27. This may make it difficult to obtain statistically significant results even if effect sizes are substantial. Although the male : female ratio for offending is high (5.3 : 1 at ages 20–27), males and females are combined in all the analyses, presumably because of the problem of small numbers. It is not clear to me why gender (and age) were not controlled in the Table 3 logistic regression analyses.
The main conclusion of the paper is that cannabis use does not predict non-drug offending significantly after controlling for confounding factors. I am somewhat dubious about controlling for other drug use in model 3, because it could be argued that cannabis use and other drug use are both indicators of the same underlying construct. In the CSDD, there was a considerable overlap between cannabis and other drug users; the percentage of other drug users who were also cannabis users was 83% at age 18, 91% at age 32 and 64% at age 48. I will focus upon model 2. Here we see non-significant odds ratios (ORs) of the order of 2.0 for the relationship between cannabis use and non-drug offending. According to the Presidential Address by Cohen [6, p. 136], ‘the field of epidemiology tends to regard OR of 2.0 or more as fairly large’.
A fairly strong relationship that is non-significant seems somewhat inconclusive to me, and not a very firm basis for the main conclusion of the paper that ‘the use of cannabis does not seem to represent a risk factor for a general criminal involvement’. The systematic review by Bennett et al. was inconclusive in the opposite direction: the weighted mean OR for cannabis was a rather weak 1.5, but it was statistically significant (because it was based upon an accumulation of results obtained in 10 different studies). The main conclusion from the systematic review is that heroin, cocaine and crack were more strongly related to criminal behaviour than was cannabis.
In the CSDD, cannabis use at age 18 significantly predicted non-drug convictions between ages 19 and 32 [39% of users were convicted, compared with 24% of non-users; OR = 2.0, confidence interval (CI) 1.3–3.3]. However, after controlling for prior convictions between ages 10 and 18 in a logistic regression, the prediction became non-significant (partial OR = 1.4), in agreement with the conclusions of Pedersen & Skarohamar . Cannabis use at age 32 predicted non-drug convictions more strongly between ages 33 and 50 (31% of users were convicted, compared with 9% of non-users; OR = 4.5, CI 2.3–8.8). Furthermore, the significant predictability of cannabis use at age 32 held up after controlling for prior convictions at ages 10–18 and 19–32, heavy smoking, heavy drinking and binge drinking up to age 32, and age 8–10 risk factors that were the best predictors of non-drug convictions at ages 33–50. It may be that cannabis use predicts non-drug convictions more strongly at older ages than at younger ages, in agreement with findings in the systematic review by Bennett et al..
Another issue that might be raised centres upon the validity of the self-report information. The fact that cannabis use predicts later drug charges might be cited as evidence of the predictive validity of the self-reports. A final problem is that both cannabis use and non-drug offending vary continuously as people grow older. In the absence of frequent repeated survey information, it is difficult to investigate whether cannabis use has any causal effect on non-drug crime. Ideally, it would be desirable to investigate whether changes in cannabis use within individuals were followed reliably by changes in non-drug offending within individuals, after controlling for confounding factors.
To conclude, the paper by Pedersen & Skardhamar  reports interesting and thought-provoking results from a very important prospective longitudinal survey of drug use and offending. Many more surveys of this kind are needed, with large samples, low attrition and frequent data collection.
Declaration of interests
- 4The development of offending from age 8 to age 50: recent results from the Cambridge Study in Delinquent Development. Monatsschr Kriminol Strafrechtsreform [J Criminol Penal Reform] 2009; 92: 160–73., ,
- 5Minimizing attrition in longitudinal research: methods of tracing and securing cooperation in a 24-year longitudinal study. In: MagnussonD., BergmanL., editors. Data Quality in Longitudinal Research. Cambridge: Cambridge University Press; 1990, p. 122–47., , , ,