The Empirical Status of Social Learning Theory: A Meta‐Analysis

Abstract

Social learning theory has remained one of the core criminological paradigms over the last four decades. Although a large body of scholarship has emerged testing various propositions specified by the theory, the empirical status of the theory in its entirety is still unknown. Accordingly, in the present study, we subject this body of empirical literature to a meta‐analysis to assess its empirical status. Results reveal considerable variation in the magnitude and stability of effect sizes for variables specified by social learning theory across different methodological specifications. In particular, relationships of crime/deviance to measures of differential association and definitions (or antisocial attitudes) are quite strong, yet those for differential reinforcement and modeling/imitation are modest at best. Furthermore, effect sizes for differential association, definitions, and differential reinforcement all differed significantly according to variations in model specification and research designs across studies. The implications for the continued vitality of social learning in criminology are discussed.

Keywords:

• social learning theory
• differential association
• meta‐analysis

Notes

1. 2003 was considered as a cut‐off date because it marked the publication of Akers and Jensen’s (2003b) edited volume in the Advances in Criminological Theory series, which was the last major theoretical and empirical statement of the theory.

2. We recognize that using only published work in a meta‐analysis brings with it a certain degree of controversy over the potential inferential errors that could be made concerning “publication bias” (see Rosenthal, 1979); in particular, that the effect sizes may be inflated because of the tendency of studies revealing non‐significant relationships to be more likely to be either rejected for publication or to remain unsubmitted to journals by authors. Nevertheless, the two works that scholars typically cite as evidence of this potential problem are quite problematic themselves. For example, Cooper, DeNeve, and Charlton’s (1997) survey of 33 psychologists found that null findings (which was not a mutually exclusive response category) were actually rarely invoked as an explanation among scholars for why they believed their work was not eventually published—lack of researcher interest and methodological problems were by far the most common reasons cited by their respondents. Second, Gerber and Malhotra’s (2008) assessment of publication bias in sociology journals was only able to assess 14% of the studies that were originally gathered. Indeed, the remaining 86% of those studies failed to meet the inclusion criteria necessary for the statistical method the authors chose to use. Thus, in an odd bit of methodological irony, it is difficult for Gerber and Malhotra (2008) to rule out in their own study the very form of publication bias that they were trying to expose. Taken together, we are confident that the large sample of studies we assess here does not suffer from any form of publication bias that would not also be present in any other review—narrative or quantitative—of the criminological literature. Further, at the very least, our meta‐analysis provides a baseline of results that future research can built upon using a wider sample of studies, including unpublished works. In the interim, we furnish the most systematic quantitative synthesis of findings on social learning theory.

3. These journals were selected because they have been consistently ranked as being among the “top tier” journals in the field of criminology and criminal justice in terms of journal prestige (see Sorensen, Snell, & Rodriguez, 2006).

4. The studies included in the sample for the meta‐analysis are those by: Adams (1974), Agnew (1991, 1993), Akers and Cochran (1985), Akers and Lee (1996, 1999), Akers et al. (1979), Akers, Greca, Cochran, and Sellers (1989), Alarid, Burton, and Cullen (2000), Ary, Tildesley, Hops, and Andrews (1993), Aseltine (1995), Bahr, Hawks, and Wang (1993), Bahr, Marcos, and Maughan (1995), Bellair, Roscigno, and Vélez (2003), Benda (1994, 1999), Benda and Corwyn (2002), Benda and DiBlasio (1991, 1994), Benda and Toombs (1997), Benda and Whiteside (1995), Benda, DiBlasio, and Kashner (1994), Boeringer, Shehan, and Akers (1991), Brezina (1998), Brezina and Piquero (2003), Brownfield and Thompson (1991, 2002), Bruinsma (1992), Burke et al. (1987), Burkett and Warren (1987), Burton, Cullen, Evans, and Dunaway (1994), Byram and Fly (1984), Costello and Vowell (1999), Covington (1988), Curran, White, and Hansell (1997), Damphousse and Crouch (1992), DiBlasio (1986, 1988), DiBlasio and Benda (1990), Dull (1983), Durkin, Wolfe, and Phillips (1996), Edwards (1992), Ellickson and Hays (1992), Ennett and Bauman (1991), Fagan, Weis, and Cheng. (1990), Farrell and Danish (1993), Flay et al. (1994), Goe, Napier, and Bachtel (1985), Gottfredson, McNeil, and Gottfredson (1991), Griffin and Griffin (1978a, 1978b), Grube, Morgan, and Seff (1989), Hawkins and Fraser (1985), Hayes (1997), Heimer (1997), Heimer and DeCoster (1999), Heimer and Matsueda (1994), Hochstetler, Copes, and DeLisi (2002), Hoffman (2002), Hunter, Baugh, Webber, Sklov, and Berenson (1982), Hwang and Akers (2003), Jackson, Tittle, and Burke (1986), Jaquith (1981), Jarjoura and May (2000), Jiang and D’Apolito (1999), Johnson (1988), Johnson, Marcos, and Bahr (1987), Kaplan, Martin, and Robbins (1984), Krohn, Lanza‐Kaduce, and Akers (1984), Krohn, Skinner, Massey, and Akers (1985), Krohn, Lizotte, Thornberry, Smith, and McDowall (1996), Kruttschnitt, Heath, and Ward (1986), Lanza‐Kaduce and Capece (2003), Lanza‐Kaduce and Klug (1986), Lewis, Sims, and Shannon (1989), Lo (1995, 2000), Makkai and Braithwaite (1991), Marcos, Bahr, and Johnson (1986), Massey and Krohn (1986), Matsueda (1982), Matsueda and Anderson (1998), Matsueda and Heimer (1987), McCarthy (1996), McCarthy and Hagan (1995), McGee (1992), Mears, Ploeger, and Warr (1998), Melby, Conger, Conger, and Lorenz (1993), Menard (1992), Michaels and Miethe (1989), Morash (1983), Napier, Goe, and Bachtel (1984), Neapolitan (1981), Needle, Su, Doherty, Lavee, and Brown (1988), Nusbaumer, Mauss, and Pearson (1982), Orcutt (1987), Paternoster and Triplett (1988), Reed and Rose (1998), Reed and Rountree (1997), Ried (1989), Reinarman and Fagan (1988), Rodriguez and Weisburd (1991), Sellers and Winfree (1990), Sellers, Cochran, and Winfree (2003), Simons, Miller, and Aigner (1980), Skinner and Fream (1997), Smith and Brame (1994), Smith and Paternoster (1987), Sommers, Fagan, and Baskin, (1994), Strickland (1982), Thornberry, Lizotte, Krohn, Farnworth, and Jang (1994), Wang and Jensen (2003), Warr (1993), Warr and Stafford (1991), White and LaGrange (1987), White, Johnson, and Horwitz (1986), Wiltfang and Cochran (1994), Winfree and Bernat (1998), Winfree and Griffiths (1983), Winfree, Theis, and Griffiths (1981), Winfree, Griffiths, and Sellers (1989), Winfree et al. (1993), Winfree, Mays, and Vigil‐Backstrom (1994), Winfree, Vigil‐Backstrom, and Mays (1994), Winslow, Franzini, and Hwang (1992), and Wong (1998).

5. Several either implicit or explicit tests of social learning theory could not be included in our sample because effect size estimates could not be calculated. For example, some studies reported maximum‐likelihood or metric estimates that could not be converted (Agnew & Huguley, 1989; Bailey & Hubbard, 1991; Jang, 1999; Lanza‐Kaduce, Akers, Krohn, & Radosevich, 1984; Skinner, Massey, Krohn, & Lauer, 1985; Sommers, Fagan, & Baskin, 1993; Tittle, Burke, & Jackson, 1986; Warr, 1998; White & Bates, 1995). Other studies either provided no inferential statistics, or not enough descriptive information was reported (e.g., univariate descriptive statistics only) so that an effect size estimate could be calculated (Curcione, 1992; Fagan, Piper, & Moore, 1986; Kandel & Davies, 1991; Laner, 1985; Rouse & Eve, 1991; Shukla, 1976; Wood, Gove, Wilson, & Cochran, 1997). Still others estimated interaction effects only (Benda & Corwyn, 1997; Conger, 1976; Dembo, Grandon, La Voie, Schmeidler, & Burgos, 1986; Jensen & Brownfield, 1983; Linden & Fillmore, 1981; Weast, 1972) or did not assess a crime or deviance outcome measure (Capece & Akers, 1995; Gainey, Peterson, Wells, Hawkins, & Catalano, 1995; Kandel & Davies, 1991; Linden, Currie, & Driedger, 1985; Orcutt, 1978).

6. Others in this tradition have used similar bivariate effect size estimates such as the Cohen’s d which is the difference between two group means divided by the pooled within‐group standard deviation (Cohen, 1977; see also Loeber & Stouthamer‐Loeber, 1986). Still other researchers have used the RIOC statistic (relative improvement over chance), which scales down certain descriptive statistics into a 2 × 2 table of whether or not a predictor variable is present and whether or not an individual engaged in delinquency (Loeber & Dishion, 1983). Both of these statistics, however, assume, at minimum, a quasi‐experimental research design, and are therefore not applicable to synthesizing correlational research based on statistical control.

7. Hedges’ and Olkin’s critique of using multivariate effect sizes in meta‐analysis was certainly valid when it was raised over two decades ago. Since then, however, advances in both quantitative methods and computer software (in particular, Hierarchical Linear Modeling methods and software) have resulted in new approaches for handling this potential problem in methodologically defensible ways, including the approach taken here.

8. The r coefficient was also chosen not only for its ease of interpretation, but also because formulae are available for converting other test statistics (e.g., chi‐square, t, F) into an r‐value (see Wolf, 1986).

9. The equation for the transformation of r‐values to z(r) values (see Blalock, 1972), which converts the sampling distribution of r to one that approaches normality, is:

10. In 1998, Akers offered a further expansion of his social learning theory, providing what he viewed as the social‐structural roots of the learning process. As of 2003, however, little research has been conducted on these elements and no results had entered the empirical literature as defined for this meta‐analysis.

11. Three separate coders were used to code the sample of studies. A random sample of 20 studies was selected for coding by each coder, and reliability analysis was conducted with regard to the coding scheme. The inter‐rater reliability coefficients across all dimensions of the analysis conducted here (e.g., effect size calculation, codes for moderating variables) exceeded .92.

12. The Level 1 effect size variances are computed using the formula for estimating the standard error of the Fisher r‐to‐z transformation (see, e.g., Hox, 1995; Lipsey, & Wilson, 2001):

The full model is estimated with the equation:

13. The confidence intervals for the overall effect size estimates were calculated according to the random effects method, recently advocated by Schmidt, Oh, and Hayes (2009).

14. While numerous methods for significance testing exist for meta‐analysis (see, e.g., Hedges & Olkin, 1985), the significance tests in the present case are based on the most conservative method using a “t” distribution for significant values of r (Blalock, 1972).

15. As a final note on the issue of publication bias in meta‐analysis, scholars have noted that such bias is potentially a greater threat in research involving clinical trials (see Dickersin, Min, & Meinert, 1992), yet even a large‐scale study of 745 manuscripts submitted to the Journal of the American Medical Association found “no significant difference in publication rates between those with positive vs negative results” (Olson et al., 2002, p. 2825). Thus, the issue of publication bias for meta‐analysis—even in a context where the risk of its presence is should be at its peak—is far from a “given.” Nevertheless, we encourage scholars in the future to revisit our work while including unpublished studies to determine whether the inferences we have made here would be changed in any meaningful way.