On the measurement of low self-control in Add Health and NLSY79

ABSTRACT

Limited attention has been devoted to the dimensionality of the low self-control scales commonly constructed in two nationally representative datasets routinely used to test self-control theory (SCT) – Add Health and NLSY79. We assess the measurement properties of the low self-control scales by comparing a series of exploratory and confirmatory models that are appropriate for the categorical nature of the observed items, including unidimensional, correlated factors, second-order factor, and bifactor models. Additionally, based on these results we explore the predictive validity of the respective scales on adolescents’ delinquent behavior. The results indicate that the low self-control scales in these data have acceptable levels of internal consistency but do not represent unidimensional latent factors. Rather, scales are best represented by a second-order factor structure. When measured this way, our Add Health scale is associated with delinquency in a cross-sectional context and our NLSY79 scale predicts delinquency longitudinally. This study reveals that low self-control is best conceptualized as a multidimensional construct within these data. The results of this study provide guidance to researchers measuring low self-control in either dataset (or other data sources) and inform the larger SCT measurement literature.

KEYWORDS:

• Low self-control
• measurement
• crime
• Add Health
• NLSY

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

1. Our intention is to point out some of the reasons why researchers use nationally representative datasets. Exploring the advantages of such data does not minimize the importance of studies that use smaller scale primary data collection efforts.

2. It is important to note that other operational low self-control scales have been constructed with the Add Health data but are used with much less frequency (e.g. Hoffmann et al., 2013; Jang & Franzen, 2013; Young, Barnes, Meldrum, & Weerman, 2011).

3. Turner and Piquero (2002) also constructed a six-item attitudinal measure of low self-control. The scale captured key dimensions of the construct (e.g. risk-seeking orientation and impulsivity) and evidenced decent internal consistency (α range = 0.59–0.64 across waves). The authors did not attempt to combine the attitudinal and behavioral indicators and assess measurement characteristics because the purpose of their analysis was to explore the stability of self-control using different operationalizations of the concept.

4. It is important to emphasize that the self-control items we examine do not represent the only options available to researchers using these data sources. In other words, there is not necessarily an agreed upon list of items that must be used in tests of SCT. Rather, an extensive review of the literature reveals that these sets of items are the most commonly employed in tests of the theory. As such, our goal is to provide an assessment of the most common self-control items available in these data. Our analysis does not preclude the use of other valid self-control items. Indeed, it is our hope that our analysis will provide useful tools for researchers who wish to do so.

5. We addressed only the behavioral low self-control items available in this wave of the NLSY79 for two reasons. First, and most importantly, a relatively small group of respondents provided answers to the attitudinal survey items which would have severely limited our sample size. In fact, when examining the attitudinal low self-control items and delinquency items the sample size was reduced to only about 300. Second, a majority of the literature using NLSY79 to test SCT relies on the behavioral questions.

6. The majority of studies using the Add Health or NLSY79 data to measure self-control have treated the manifest variables as continuous. However, a preferred approach is to utilize models that explicitly recognize their categorical measurement scheme (Barendse et al., 2015; Rhemtulla, Brosseau-Liard, & Savalei, 2012). This is particularly important given that much of the mass in most of the self-control items is centered on one end of the distribution. Nonetheless, the procedures used subsequently to estimate the measurement properties of the latent self-control variables do not require that the manifest variables be treated as categorical. It is simply good statistical practice to do so.

7. As noted earlier, the Grasmick et al. scale is one, albeit widely known, way to operationalize low self-control. Given that the items available in Add Health and NLSY79 are different than those in the Grasmick et al. scale, it is not surprising that the items tend to coalesce into groups not directly represented by their original dimensions. Importantly, however, the individual items and the groups of items overlap with many aspects of Gottfredson and Hirschi’s (1990) original articulation of self-control characteristics (i.e. our labels may differ but the items tap into the same characteristics).

8. Second-order factor models also provided a good fit to the delinquency items in both datasets, but, in the interests of a parsimonious presentation, we present here only the association between the low self-control latent constructs and the unidimensional delinquency constructs.

9. Using a standard factor analysis model, such as with the factor command in Stata or SPSS, one might also create predicted factor scales based on weighted approach, such as Bartlett’s method.