Analysis approach for dominance and Big 5 personality factor intercorrelations

This project is a secondary analysis of data from the second wave of the Midlife in the United States study (MIDUS).

Sample

MIDUS is a longitudinal study currently consisting of three waves. From the National Archive of Computerized Data on Aging:

“In 1995-1996, the MacArthur Midlife Research Network carried out a national survey of 7,108 Americans aged 25 to 74 (MIDLIFE IN THE UNITED STATES (MIDUS), 1995-1996 [ICPSR 2760]). The purpose of the study was to investigate the role of behavioral, psychological, and social factors in understanding age-related differences in physical and mental health. The study was innovative for its broad scientific scope, its diverse samples (which included twins and the siblings of main sample respondents), and its creative use of in-depth assessments in key areas (e.g., daily stress and cognitive functioning). A description of the study and findings from it are available at http://www.midus.wisc.edu. With support from the National Institute on Aging, a longitudinal follow-up of the original MIDUS samples: core sample (N = 3,487), metropolitan over-samples (N = 757), twins (N = 925 complete pairs), and siblings (N = 950), was conducted in 2004-2006. Guiding hypotheses for it, at the most general level, were that behavioral and psychosocial factors are consequential for physical and mental health. MIDUS 2 respondents were aged 35 to 86. Data collection largely repeated baseline assessments (e.g., phone interview and extensive self-administered questionnaire), with additional questions in selected areas (e.g., cognitive functioning, optimism and coping, stressful life events, and caregiving).” (Ryff et al., 2021)

Variables

The Big 5 personality domains were assessed using the Midlife Development Inventory (MIDI; Lachman & Weaver, 1997), which contains multiple adjective indicators for each standard domain, as well as a scale labeled “agency” in the MIDUS dataset (Table 1). Agency gets its name from the bipolar agency axis of the interpersonal circumplex (Lorr & Strack, 1990); the “high” end of this axis is referred to as “assured-dominance” in circumplex language. This MIDI scale uses assured-dominance adjectives for the construct; we will refer to it as “dominance” for short. See the online codebooks for all source information.

Table 1. Standard array of personality items assessed by the MIDI.

While the basic assured-dominance measure captures general behavioral proclivities relevant to the construct, aspects like leadership and fearlessness that are part of many dominance/assertiveness assessments are absent, so we looked to other scales in MIDUS2 that might contain relevant content and could be added to the dominance construct. Thus, in order to probe the wider potential meaning of the dominance construct, we will also fit a second model set for dominance, including several additional items from other scales measured in MIDUS2.

Table 2. Supplemental dominance items drawn from other MIDUS2 instruments.

Variables will be reverse coded so that high values (e.g. 4) represent “a lot” of a trait, and low values (e.g. 1) represent the least amount of a trait. Variables will be cleaned of missing values and inappropriate values – all set to NA.

Analysis

All analyses will be caried out in R (version 4.2.2; R Core Team, 2022). We will use confirmatory factor analysis (CFA) to evaluate fit of specific constructs and the correlational relationships between the constructs, using the R package ‘lavaan’ (Rosseel, 2011). We will also first fit simple Spearman correlations among the constructs, using the pre-constructed personality variables in the MIDUS2 dataset.

In CFA we will fit each of the six personality constructs in isolation, first. The latent variables (LVs) will be defined by the variables of each domain as specified in Table 1. We will treat indicator variables as ordinal and use weighted least squares with mean and variance adjusted test statistic (‘WLSMV’). Once we have adequate model fit for each construct in isolation, we will combine all six LVs into one model, assess fit, and examine the correlations between the LVs.

After fitting these models, we will refit dominance, in isolation, with the additional variables described in Table 2. We will then refit the wider model with this revised dominance latent variable and present the correlations from that model.

For determining best fit CFA models, we will use a combination of χ², CFI, and SRMR. We will not interpret models with fit statistics beyond the following minimum cut-offs (Hooper et al., 2008; Hu & Bentler, 1999): SRMR < 0.09, CFI > 0.90. In the initial phase of model fitting by individual factors, if fit is inadequate we will iteratively remove each item and refit the model to see if adequate model fit is achieved by any of the simpler models, and if so, we will move forward with the best fitting new model.