Fig. 5.

Decision tree for measuring socioeconomic conditions.

^aDiemer et al. (2013) provided an excellent set of pragmatic considerations when measuring many of these variables. Galobardes et al. (2006a, 2006b), Krieger et al. (1997), and Shavers (2007) provided a description of theoretical strengths and limitations of income, wealth, education, and other socioeconomic conditions.

^bFor example, Wright (1997) and Wright & Perrone (1977).

^cSee Haug (1977) for very serious concerns about the validity of existing prestige measures.

^dSee Coleman (1988) for a theoretical discussion of social capital. Tulin et al. (2018) provided one example of measuring social capital.

^eNote that procedures for selecting indicators for formative models are largely undeveloped (West & Grimm, 2014). Diamantopoulos and Winklhofer (2001) provided a set of recommendations for indicator selection. Their recommendation to use multiple-indicators multiple-causes (MIMIC) models for path estimation should be ignored, however, because MIMIC models are irrelevant to formative models (Lee et al., 2013; Muthén, 1989). Theory on formative models has proceeded as far as identifying when to use them and how to estimate them, but not on how to decide which indicators to use for them. One approach to selecting indicators begins with recognizing that a formatively measured variable is essentially a variable optimized to predict a set of outcomes. Because the formatively measured variable begins as the shared variance of the outcomes, its indicators’ weights reflect only the unique variance they contribute to this shared variance. Hence, their weights, and thus the formative variable they contribute to, are optimized to predict the outcomes. From this recognition, one approach to picking indicators is to choose those that are relevant to socioeconomic status (SES) and that are uniquely related to the outcomes. Hence, income and education may be relevant for some outcomes, whereas occupation and wealth may be relevant for others. A major issue with this approach is that the chosen indicators need not be a complete representation of SES but be only the set of variables that most fully account for SES’s relation to an outcome. Thus, using only predictive indicators to represent SES in a formative model could err and omit variables important for a complete representation of SES. Thus, a better approach might be to start with a set of indicators judged to represent the breadth of SES. When entered into the model, the indicators of SES from this broader set that do not uniquely predict the outcomes will receive low weights and may need to be dropped to obtain satisfactory model fit. To my knowledge, no guidelines exist for managing this tension between model fit and content validity. (Note that this logic follows that developed by Diamantopoulos & Winklhofer, 2001, for selection and retention of indicators.)

^fNote that variables that are reflectively measured (e.g., identity, subjective SES) should be modeled as reflective indicators of SES. Bollen and Bauldry (2011) and Bainter and Bollen (2014) provided examples of how to fit formative models. van Bork et al. (in Asendorpf et al., 2016, Figure 1, bottom half, p. 308) demonstrated how to test whether formatively measured variables affect outcomes over and above their indicators. I provide an example of these two steps in the Supplemental Material available online using the lavaan package in R (Rosseel, 2012).