A Critique of the Self-Control Theory of Crime: Controversies and Methodological Issues, Study Guides, Projects, Research of Criminology

Download A Critique of the Self-Control Theory of Crime: Controversies and Methodological Issues and more Study Guides, Projects, Research Criminology in PDF only on Docsity!

M A X P L A N C K S O C I E T Y

Preprints of the Max Planck Institute for Research on Collective Goods Bonn 2012/

Low Self-Control

As a Source of Crime

A Meta-Study

Christoph Engel

Preprints of the Max Planck Institute for Research on Collective Goods Bonn 2012/

Low Self-Control As a Source of Crime

A Meta-Study

Christoph Engel

February 2012

Max Planck Institute for Research on Collective Goods, Kurt-Schumacher-Str. 10, D-53113 Bonn http://www.coll.mpg.de

I. Control Theory

Few criminological theories have had such an impact as the “General Theory of Crime” by Gott- fredson and Hirschi (Gottfredson and Hirschi 1990). The theory is not only pervasively cited,^1 and has triggered a lively theoretical debate. It has also been tested empirically hundreds of times.^2 This meta-study organizes the empirical evidence. While a predecessor study 10 years ago had covered 19 papers (Pratt and Cullen 2000), this paper covers 102 publications. It devel- ops a new methodology to make the effect of low self-control on crime and deviant behavior comparable across studies, including the competing operationalizations of self-control. It uses the resulting dataset to test the effect of multiple explanatory variables on the sensitivity of devi- ance to a lack in self-control.

The general theory has become a classic of criminological theory. Suffice it in this introduction to recall its key components and claims. The theory sets out to explain the stability of crime over the lifecourse of the individual (Gottfredson and Hirschi 1990 : 107 and passim), and the versatil- ity of crime, i.e. a lack of specialization in committing certain types of crimes (Gottfredson and Hirschi 1990 : 91-94 and passim). The theory posits the following explanation: Resulting from deficiencies in child rearing (Gottfredson and Hirschi 1990 : 97-107), the inability to resist the drive for immediate gratification is likely to stay with this person for the rest of her life. And a person having a hard time controlling her urge to commit one crime is likely to do no better if the opportunity to commit another crime presents itself (Gottfredson and Hirschi 1990 : chapter 5).

Bold claims provoke critique (for summary accounts see Brannigan 1997 ; Schulz 2006 : chapter 6). Critics have wondered whether low self-control and the propensity to commit crime collapse, which would make the theory tautological (Akers 1991 ; Marcus 2004), and whether self-control is not better analyzed as a feature of the situation, not as a personality trait (Wikström and Treiber 2007). They have pointed to the fact that, at least in the original version of the theory (but see Hirschi 2004 : 543; Piquero and Bouffard 2007), criminal opportunity got short shrift (Barlow 1991). They have wondered whether it is justified to even extend the self-control expla- nation to white collar crime (Tittle 1991 ; Reed and Yeager 1996 ; Herbert, Green et al. 1998), whether it is legitimate to explain the well-known decline of crime with age by a distinction be- tween a stable trait (“criminality”) and its time-variant expression (“crime”) (Tittle 1991 ; Geis 2000 ; Geis 2008), and whether social norms (Taylor 2001) and culture are not a more important determinant of the crime rate than the General Theory admits (Komiya 1999). From the opposite angle, it has been said that the key role of parental management is inconsistent with a purely in- dividualistic explanation of crime (Kissner 2008). Critics have also missed an explicit treatment of the link between crime and power, in particular to the detriment of women (Miller and Burack 1993), they have argued that disruptions of parental attachment in early childhood should be acknowledged as an additional cause of crime (Hayslett-McCall and Bernard 2002), and that the theory should account for the fact that most antisocial children do not turn into adult criminals

(^1) Google Scholar lists 4696 citations, as of Feb 19, 2012. (^2) For detail, see section 0 below.

(Cohen and Vila 1996). In a series of articles, Gottfredson and Hirschi have responded to their critics, have further explicated and slightly modified their theory (Hirschi and Gottfredson 1987 ; Hirschi and Gottfredson 1989 ; Hirschi and Gottfredson 1993 ; Hirschi and Gottfredson 1995 ; Hirschi and Gottfredson 2000 ; Hirschi and Gottfredson 2001 ; Gottfredson and Hirschi 2003 ; Hirschi 2004 ; Gottfredson 2008 ; Hirschi and Gottfredson 2008 ; Gottfredson 2011).

II. Present Research Focus

The General Theory has thus triggered a lively theoretical debate. Yet in comparison the reaction of empirical criminologists has been just overwhelming. Since the book has come out in 1990, hundreds of empirical tests have been performed (see next section for details). It is the purpose of this meta-study to make this huge body of evidence accessible, to make individual contribu- tions comparable, and to derive insights for the relationship between self-control and crime that do only follow from all these studies jointly, not from individual contributions in isolation. Of course, not all of these studies have defined deviance the same way, nor have they operational- ized and measured self-control by the same construct. To make results comparable, I therefore construct an index that is relative to the dependent and the independent variable used by the re- spective study. Conditional on the specifics of the study, the index measures by how many per- cent antisocial behavior increases if self-control drops by 10 percent. I register the specifics of each study in multiple dimensions, and use these indicators as control variables.

Overall, the General Theory has performed very well empirically. In all these empirical tests, low self-control and crime, or deviant behavior more generally, have hardly ever been negatively correlated. Insignificant findings are also rare. In most studies, low self-control has the expected positive effect on the frequency or the severity of crime. Yet contextual factors matter to a re- markable degree. To preview only a few findings: self-control is much more important for actual delinquency than for “analogous behavior”. It is even more important for actual convicts. It is most important for adolescents, but only if analogous behavior is at issue. If one does not control for race, one overestimates the effect of low self-control. If one does not control for family sup- port and for opportunity, one underestimates the effect.

Methodological issues do also become visible: Attitudinal measures, like the popular scale by Grasmick, Tittle et al. (1993), heavily underestimate the effect of low self-control. Several ex- planatory variables significantly interact with the average level of deviance in the sample. If re- searchers have used a non-linear statistical model, they estimate a much smaller effect. The later the study, in historical time, the bigger the estimated effect.

There is one major predecessor study, by Pratt and Cullen (2000). It covered 19 papers, while my meta-study covers 109 (102), with 826 (717) separate observations (see Figure 1). I also develop a new dependent variable, which is better suited to organize this evidence. I explain the meth- odological differences between both meta-studies in section 0 below. Other summary reports and meta-studies are less close. Riddle and Roberts (1977) is a meta-study of psychological research

ance in the first place. 57 papers do not report the range of the dependent variable. I then cannot say by how many percent the dependent variable changes if self-control is reduced by 10%. 66 papers do not report the range of the independent variable. I then cannot say which is the effect of a 10% reduction in self-control. That is, without knowing both ranges, I cannot normalize the findings across studies. I explain this is more detail in sections 0 and 0 below. If the respective paper has used a linear statistical model, I do not necessarily need the mean of the dependent variable. In a linear model, the marginal effect of a one unit change of an independent variable on the dependent variable is the same whichever the level of the dependent variable. This is dif- ferent with non-linear models, like logit or a negative binomial regression. In order to exclude as little papers as possible, I then calculate the marginal effect at the mean of the dependent varia- ble. I must exclude 43 papers since they use a non-linear model, but do not report the mean of the dependent variable. If the range of the dependent or the independent variable is not reported, but if I know their means and standard deviations, I construct the range, assuming that either var- iable is normally distributed. In 49 papers, I cannot use this proxy either since the mean of the independent variable is not reported either. In further 38 papers, information about the standard deviation is missing. Two papers estimate an ordered logit model, but do not report the estimated cut off points. That makes it impossible to calculate the marginal effect of a change in self- control at the mean of the dependent variable. 54 papers do not provide any measure of effect size. One paper suffers from an obvious mistake which I cannot correct. One works with simula- tion. Three papers only report qualitative findings. Three more papers report the same data as an earlier paper published in a different journal.

th th 43 meth^ wr_dv^ wr_iv^ no_n^ m_ra_dv^ m_ra_iv^ m_me_dv^ m_me_iv^ m_sd^ m_cut^ m_eff^ mist^ sim^ no_q^ rep methwr_dv (^543 1 1 1 1 ) wr_ivno_n 37 143 1 11 11 1 1 m_ra_dvm_ra_iv 57 5366 3535 3642 2632 22 m_me_dvme_me_iv (^43 4149 3035 ) me_sdm_cut 38 2 1 m_effmist 54 1 simno_q (^13 ) rep 3

Table 1 Reasons for Excluding Papers from Meta-Analysis th: theory only; meth: empirical methodology only; wr_dv: wrong dependent variable; wr_iv: wrong independent variable; no_n: paper establishes no nexus between low self-control and deviance; m_ra_dv: range of dependent variable not reported; m_ra_iv: range of independent variable not reported; m_me_dv: mean of dependent varia- ble not reported; m_sd: standard deviation of dependent variable not reported; m_cut: coefficients of cut-offs not reported (for ordinary logit); m_eff: no measure of effect size; mist: mistake in data; sim: simulation; no_q: no quantitative information; rep: repeats earlier publication

The remaining dataset covers 1,074,895 observations.

The left panel of Figure 1 shows that interest in self-control theory has peaked after the turn of the century, and has somewhat been on the decline in recent years. This might imply that the

empirical body of evidence is maturing, which would make this a particularly suitable point in time for taking stock with a meta-study. The right panel demonstrates that journals have covered self-control theory unequally, with Criminology standing out.

0

10

20

30

40

(^1920 1940 1960) year 1980 2000 2020 total included in meta-study

development of self-control literature

(^0 10) number of papers 20 30

JustQ

JResCrimDel

JQuantCrim

JCrimJust

IntJOffThCompCrim

EurJCrim

DevPsy

DevBeh

CrimJustBeh

CrimDel

Crim

main journals

excluded included

Figure 1 Features of the Data-Set

2. Dependent Variable

The dependent variable of the meta-study is a composite construct. In the original papers, some measure of deviance is explained by some measure of self-control. Yet for the purposes of the meta-study, this approach would not be meaningful. The measure of interest is not unconditional deviance, but deviance conditional on varying degrees of self-control. If self-control theory has it right, the lower self-control of an individual, or of a defined population for that matter, the higher the predicted level of deviance. Put differently, the explanatory power of self-control theory is the higher, the more the level of deviance is sensitive to changes in self-control. The dependent variable of the meta-study is precisely this. Separately for each paper, and if the paper had cov- ered several populations, measurements, dependent or independent variables, then separately for each of those, it indicates how strongly predicted deviance increases if self-control goes down by 10%. Economists would call this an elasticity. From a policy perspective, the higher this depend- ent variable, the more important self-control is for crime control.

This procedure deviates from the research strategy of the major predecessor paper. Pratt and Cullen (2000) also estimate effect sizes, but they use a standardized correlation coefficient for the purpose. I take an alternative approach because otherwise I would have lost many more ob- servations. In recent years, beta-coefficients have been reported less and less frequently. I can only reconstruct them from the unstandardized regression coefficients if the standard deviations of both the dependent and the independent variables are reported. This is often not the case.

3. Analytic Strategy

For the ease of reading, independent variables are introduced together with the findings from the meta-study. At this point I have to discuss my analytic strategy. My dependent variable is con- tinuous and uncensored. It is somewhat skewed. Yet if I repeat the analysis with a log transform or with a square root transform, coefficients are of course different, but results look qualitatively very similar. Significance levels are usually not affected. Since coefficients can then be inter- preted directly, I prefer (untransformed) ordinary least squares.

I have, however, two complications. First from many papers I have more than one data point. Authors have compared different populations. Longitudinal studies have measured the same population repeatedly. Other authors have used multiple dependent or independent variables. Finally many authors have estimated several statistical models, usually in the interest of adding moderating or mediating factors. In all but the first case, observations are even dependent on ob- servables. And if the same author has tested different populations on the same design, I cannot exclude that these observations depend on unobservables. With such a data generating process, a random effects model would be most efficient. Yet since most of my explanatory variables differ within, not between papers, these explanations would get lost if the Hausman test forced me to use a fixed effects model. More importantly, if I use a random effects model, I cannot weight the data. This, however, is mandatory in a meta-study. Although this involves a slight less in statisti- cal power, I therefore revert to clustering standard errors for publications. To guard against het- eroskedasticity, I also estimate robust standard errors.

The second complication is the mainstay of meta-study. While each statistical model in a paper covered by the meta-study generates but one data point, from the statistical tests performed by the authors of the original papers I have information about the reliability of these findings. Metastudy does not just aggregate this information (e.g. by calculating means). It measures what the entire community of researchers undertaking self-control studies has found, through weigh- ing the contribution of each individual study by its precision. Since I must simultaneously solve the dependence problem, I cannot use ready-made tools like the metareg command of Stata (Harbord and Higgins 2008) for this purpose, though. Yet my solution keeps the essence of me- ta-regression. I weigh each data point by the inverse of the square of the standard error, and then cluster for publications.^4 Recall that my dependent variable is the sensitivity of deviance to a decrease in self-control. Consequently the standard error in question is the standard error of the regression coefficient for self-control in the original paper. Unfortunately some papers neither report standard errors nor p-values, from which standard errors could be reconstructed. Although I otherwise have full data, I have to exclude another 112 data points since I cannot weigh them. This leaves me with a final sample of 717 data points. The final sample still covers 966,364 orig- inal data points.^5

(^4) The Stata code is reg dv ivs [aw=se^(-2)], robust cluster(study_id). (^5) For the unconditional effect of low self-control, I also report unweighted results that use all data, see below 0.

IV. Analysis

1. Descriptives and Unconditional Effect

Figure 2 summarizes the distribution of the dependent variable. One directly sees that it is very rare for the effect of the decrease in self-control on the level of deviance to be negative. This is only observed in 42 of 826, or in 5.08% of all cases. Most frequently, the effect is positive, but small. The median is at .02268. If a person has 10% less self-control than another, this person is 2.27% more likely to behave antisocially, or to deviate that much more from social expectations. Yet quite a few studies have found stronger effects, occasionally even above .1. Such a study predicts that a person with 10% less self-control than another is even more than 10% more likely to behave antisocially. Yet such an extreme effect of self-control is only observed in 11 of 826, or in 1.33% of all cases.

0

50

100

150

200

Frequency

-.1 (^0) effectsize .1.

sensitivity of deviance to decrease in self-control

Figure 2 Distribution of Dependent Variable

In 633 of 717 studies, i.e. in 88.28 % of all studies that report significance, the degree of self- control significantly explains the frequency or intensity of deviance. Table 2 presents two regres- sions with just a constant, as statistical tests for the grand mean. Both have a positive, significant result. Overall, low self-control undoubtedly increases crime. Yet if I weigh observations by the inverse of the standard error, and thereby also exclude observations where the standard error had not been reported, the size of the effect goes down dramatically, and the standard error goes up substantially. The standard error in model 2 is about twice as large, and the effect is only some 13% of the effect measured with the raw data. In absolute terms, the effect is very small. The model only predicts an increase in crime by less than half a percent if self-control reduces by 10%. This of course is only the unconditional effect. If I condition the effect on appropriate con- trol variables, it becomes bigger. This first finding should therefore not be misread. It does not show that the overall effect of self-control on deviance is negligible. It only shows that the un- conditional effect is not the best object of observation.

Hirschi 1990 : 91-94 and passim). This claim has attracted considerable interest among empiri- cists. Studies widely vary with respect to the operationalization of deviance. If I separately ana- lyze those studies that have measured delinquency, I find a significant effect of low self-control (N = 424, cons .00836*), as well as when I confine the analysis to studies with some form of non-criminalized analogous behavior (N = 217, cons .00167). Model 1 of

Table 3 instead uses the type of deviant behavior as a control variable. One may read this result in two ways. The fact that the constant remains positive and significant if one controls for those studies that have measured delinquency shows that low self-control also predicts analogous be- havior (which implicitly becomes the reference category). This supports the General Theory. Yet the coefficient for delinquency is more than twice as large as the constant. If one wants the mod- el prediction for the effect on delinquency, one must add up the constant and the coefficient of the regressor. One then sees that the effect of low self-control on delinquency is about three times as strong as on analogous behavior.

The General Theory has also been influential with coining two broad categories for crime: “force or fraud” (Gottfredson and Hirschi 1990 : 16). Some later contributions have explicitly used these categories. Other studies I have classified analogously, using “force” for all acts directed against the physical integrity of another person, and “fraud” for all acts directed against other individu- al’s property or fortune. As model 4 shows, “fraud” (for which I have 94 observations) has a di- rect and strong effect. The likelihood to commit “fraud” is even more strongly affected by the individual level of self-control than the general willingness to engage in delinquent acts. By con- trast, as models 2 and 3 show, with “force” (for which I have 170 observations), the story is more complicated. If one only controls for this class of deviant behavior, one finds no significant dif- ference from the grand mean. Yet if one controls for delinquency and the interaction term, the main effect of delinquency jumps up, but it is reversed by the interaction term. This implies that low self-control (only) matters disproportionately for delinquent acts that do not involve physical attacks.

model 1 model 2 model 3 model 4 delinquency .00592* (.00268)

.00955*** (.00145) force. (.00303)

. (.00462) delinquencyforce -.0117 (.00551) fraud .00809** (.00248) cons .00244* (.00101)

.00330* (.00133)

.00189** (.00068)

.00349** (.00125) p model .0296 .4652 <.001. R^2 .1105 .0139 .1926.

Table 3 Various Definitions of Delinquency N= 717, OLS, robust standard errors, clustered at the level of publications standard errors in parenthesis, *** p < .001, ** p < .01, * p < .05, +^ p <.

Using the same statistical model, one finds that the propensity to commit specific types of crime is very differently sensitive to more specific types of crime. Naturally, the number of observa- tions on each of these classes of crime is relatively small. This explains why these regressors explain little variance, even if they turn out significant. To save space, I only report the number of observations falling into the respective class, and the coefficient in a regression explaining sensitivity to low self-control with this one dummy. That way I find a strong differential effect of low self-control on dishonest behavior (.01619, 30 cases), on sexual offences (.02451, 21 cases) and on traffic violations (.01870, 18 cases), and a smaller but still sizeable effect on technical violations, for instance by parolees (.00708*, 17 cases). For the propensity to commit all these crimes, the effect of low self-control is particularly pronounced. The opposite is true for gang-related crime (-.00277+, 22 cases).

Table 4 further investigates the effect of low self-control on analogous behavior (for which I have 217 observations). As one should have expected from Model 1 of Table 3, the effect is sig- nificant and negative. For analogous behavior, low self-control is less important than for actual delinquency. Yet this negative main effect disappears once one controls for substance abuse (model 2 of Table 4, 143 cases). This shows that the difference between criminal and non- criminal behavior is not a general one, but is specific to one form of analogous behavior, the abusive consumption of drugs, alcohol and tobacco. Model 3 further differentiates between sub- stances more generally (like, in particular, alcohol) and shows that the effect of low self-control is more similar to delinquency if the consumed substance are drugs (68 cases).

model 1 model 2 model 3 analogous -.00566** (.00194)

. (.00285)

. (.00286) substance -.00932*** (.00222)

-.01105*** (.00217) drug .00516* (.00242) alcohol .00207** (.00066) cons .00733*** (.00190)

.00733*** (.00190)

.00733*** (.00190) p model .0044 <.001 <. R^2 .1379 .1701.

Table 4 Analogous Behavior N= 717, OLS, robust standard errors, clustered at the level of publications standard errors in parenthesis, *** p < .001, ** p < .01, * p < .05, +^ p <.

b) Population Characteristics

If the General Theory has it right, those who have actually been convicted for committing a crime should exhibit particularly pronounced self-control problems. This explains why quite a number of studies have been conducted with convicted criminals. If one confines the sample to these 49 cases, one finds a strong significant effect of low self-control in the expected direction (cons .01633***). Model 1 of Table 5 is even more revealing. It uses the fact that the sample

effects indeed show an independent effect of age on the sensitivity to self-control which is not explained by the age specific level of deviance. Adolescents are not only much more likely to exhibit crime and analogous behavior. Their level of deviance is also more sensitive to the de- gree of self-control than in elder individuals.

model 1 model 2 student age -. (.00249)

.00681 + (.00373) adult -. (.00291)

. (.00594) mean deviance .02732* (.01041) studentmean deviance -.03769* (.01177) adultmean deviance -.05013* (.01780) cons .00638** (.00196)

.00374 + (.00203) p model .3327. N 717 694 R^2 .0371.

Table 6 Age OLS, robust standard errors, clustered at the level of publications standard errors in parenthesis, *** p < .001, ** p < .01, * p < .05, +^ p <.

In principle, the General Theory treats the well documented gender effect on crime the same way as the age effect. It does not aim at explaining the effect. But it claims that, taken the gender dif- ference into account, there should be a differential effect of low self-control (Gottfredson and Hirschi 1990 : 144-149). I have 201 observations with gender-homogenous populations, 98 of them with males. In the male only studies, the effect of a 10% reduction in self-control on devi- ance is even more pronounced, and has a lower p-value (cons .016628**). But in the female only studies, it is significant as well (cons .01274). The evidence thus supports the main claim of the General Theory. Actually if I only control for gender, I do not even find a significant gen- der effect (model 1 of Table 7). This changes if I also control for age brackets and their interac- tion with gender (model 2). I then find that male subjects are much more sensitive to the level of self-control than females. Yet the main effect of gender and the two interaction effects essential- ly cancel out. This implies that the gender effect is confined to adolescents, who are the refer- ence category.

Gottfredson and Hirschi are open to the possibility of an additional difference between males and females in criminality, i.e. in their level of self-control (Gottfredson and Hirschi 1990 : 147). If I try to explain the mean level of self-control in each study by the gender of participants, I do not find a significant result. If I add this level as a control variable (Table 7 model 3), and interact it with gender, the interaction effect is also insignificant. All I find is a significant three-way inter- action. Male students are the more sensitive to self-control the higher the level of self-control in the sample.

model 1 model 2 model 3 male. (.00365)

.01241** (.00390)

.05138 + (.02923) student .03291*** (.00303)

.08570*** (.01068) adult .00977** (.00335)

.04817** (01291) malestudent -.01246 (.00554)

-.17002*** (.03747) maleadult -.01431* (.00459)

-.05720 + (.03019) mean self-control .07966* (.02965) malemean self-control -. (.06491) studentmean self-control -.11724** (.02972) adultmean self-control -.10013 (.03522) malestudentmean self-control .26966** (.07180) maleadultmean self-control. (.06731) cons .01217** (.00374)

.00631** (.00217)

-.02219 + (.01059) p model .2369 <.001 <. N 201 201 200 R^2 .0307 .2140.

Table 7 Gender OLS, robust standard errors, clustered at the level of publications standard errors in parenthesis, *** p < .001, ** p < .01, * p < .05, +^ p <.

The General Theory treats race much the same way as it treats gender. Taking the well- documented race differences in criminality into account, it nonetheless expects a differential ef- fect of low self-control. Yet the theory is also open to the possibility that racial background in- duces differences in criminality, i.e. in self-control (Gottfredson and Hirschi 1990 : 149-153, 179). I have 48 one-race studies in my dataset, of which 21 test African-Americans. I do find a significant sensitivity to self-control in African-Americans (N = 21, cons = .04050**), but I do not find it for Caucasians (N = 27, cons = .01756, p = .125). The deviance of African-Americans is significantly more sensitive to self-control (N = 48, coef = .02294).

“From our theoretical perspective, there is little reason to expect employment to be related to crime independent of the character of the offender” (Gottfredson and Hirschi 1990 : 164). Admit- tedly this statement was made with respect to the unemployment-crime relationship. I have no data on participants selected for being unemployed. But I do know whether most or all of a sam- ple consisted of participants who hold an employment position. This holds for 185 of 717 data points. As Gottfredson and Hirschi expect, this variable does not significantly explain sensitivity to self-control. If I analyze those whom I know to be employed in isolation, I do not find a sig- nificant effect of low self-control on deviance (cons .00299, p = .173).

It is one of the main purposes of the General Theory to get beyond “cultural” explanations. But the authors explicitly call for cross national research, hoping that it will help uncover the causes and consequences of low self-control. For “self-control is presumably a product of socialization”

survey evidence strongly underestimates the sensitivity of deviance to self-control problems. Model 2 demonstrates that asking for attitudes, not behavior, further reduces the estimated sensi- tivity of deviance to self-control. Finally model 3 shows yet another dampening effect of using the Grasmick Tittle scale. These findings suggest that the skepticism of Gottfredson and Hirschi with respect to just asking subjects for their self-control attitudes is well founded.

model 1 model 2 model 3 survey -.00838*** (.00177)

-.00678*** (.00174)

-.00539** (.00187) attitudinal -.00968** (.00313)

-.00580 + (.00339) Grasmick Tittle scale -.00489* (.00196) cons .01168*** (.00128)

.01940*** (.00308)

.01831*** (.00306) p model <.001 <.001 <. R^2 .0610 .1341.

Table 9 Empirical Methodology N = 717, OLS, robust standard errors, clustered at the level of publications standard errors in parenthesis, *** p < .001, ** p < .01, * p < .05, +^ p <.

In the General Theory, self-control is conceptualized as a complex construct (Gottfredson and Hirschi 1990 : 89 f.). Most empirical researchers have used the labels Grasmick, Tittle et al. (1993) have given to the elements: impulsivity, simple task, risk seeking, physical activities, self- centered, temper. Those using attitudinal questionnaires have aimed at first measuring these di- mensions independently. But, again following Grasmick, Tittle et al. (1993), in the second step, using factor analysis, most authors have shown that, in their sample, these sub-constructs are highly correlated, and have constructed a single self-control score. Yet some studies separately use single aspects of self-control as explanatory variables. 59 do so for risk seeking, 49 for im- pulsivity, 15 for temper, 9 for self-centeredness, 7 for physicality and 3 for simple tasks. Further studies have measured related constructs: a lack of premeditation (21), present orientation (14) and frustration tolerance (5). Happily if I add a dummy variable that is 1 if the original study has used only a single aspect of self-control as the explanatory variable, it never is significant. This makes it possible to also use these data points; otherwise my sample would reduce by 189 obser- vations.

I had already used the mean deviance in the respective sample to explain the differential effect of age (Table 6). Model 1 of Table 10 does not find a significant effect of this explanatory variable when not simultaneously controlling for other explanatory variables. By contrast, if one controls for the standard deviation of deviance, one predicts less sensitivity to self-control (models 2 and 3). Yet these two explanatory variables are most interesting in their interaction with other ex- planatory variables. The higher the level of deviance in a population, the more sensitively it re- acts to changes in self-control provided criminal acts are at stake (two-way interaction between delinquency and mean deviance). Yet sensitivity drops strongly if, on top, deviance varies great-

ly (three-way interaction with the standard deviation). There is the opposite picture for partici- pants of student age. Being in this age bracket dampens sensitivity of deviance to self-control the more the higher level of deviance. Yet the dampening effect is weakened the more the more de- viance varies.

model 1 model 2 model 3 model 4 mean_dv -. (.00855)

. (.02013)

. (.02117) sd_dv -.01666* (.00650)

-.02264** (.00810)

. (.02475) mean_dv*sd_dv -. (.04357)

-. (.05931) delinquency -. (.00301) delinquencymean_dv .14808** (.03491) delinquencysd_dv -. (.02541) delinquencymean_dvsd_dv -.22352* (.07556) student. (.00622) studentmean_dv -.09100* (.03397) studentsd_dv. (.02469) studentmean_dvsd_dv .24555 (.10985) adult .02407*** (.00582) adultmean_dv -. (.02859) adultsd_dv -.09008** (.03048) adultmean_dvsd_dv .13520+ (.07707) cons .00609** (.00196)

.00961*** (.00248)

.00923** (.00278)

. (.00243) N 694 597 597 597 p model .4991 .0121 .0203 <. R^2 .0091 .0969 .1127.

Table 10 Mean and Standard Deviation of Deviance OLS, robust standard errors, clustered at the level of publications standard errors in parenthesis, *** p < .001, ** p < .01, * p < .05, +^ p <.

I had already used the mean of self-control in the respective population to explain gender effects (Table 7). Controlling for this explanatory variable, or interactions with it, does not yield other interesting insights. Yet there is a significant positive main effect. The more pronounced self- control problems are in a population the more deviance of this population is sensitive to self- control (coef .02204*, cons -.00493, N = 644).

In the methodology section I have reported that many of the papers covered by the meta-study have estimated a non-linear model: a logit or probit model since the dependent variable was bi- nary (120 cases); a Tobit model since the dependent variable was censored (108 cases); a Pois- son or a negative binomial regression since the dependent variable consisted of counts (42 cas-