Products of correlational analyses of MIDUS2 personality data

Analytic code

A single analytic script written in R reproduces all results when given data from MIDUS2 project 1 (Ryff et al., 2021). The code imports, cleans, processes, and analyzes the data. The script is available on GitHub at https://github.com/dmaltschul/MIDUS_augmented_dominance

Spearman correlations

First, we wished to examine simple Spearman rank correlation coefficients, which are effective at analyzing ordinal data or data that are composed of ordinal indicators. The Spearman correlations among the Big 5 personality factors as well as dominance are shown in Table 1. The code script also gives a matrix of Ns and p-values for each coefficient.

Table 1. Bivariate Spearman correlation coefficients for personality constructs from the MIDI.

Note. MIDI = Midlife Development Inventory. All N > 3971. All coefficients are significant at p < 0.00001.

Confirmatory Factor Analyses (CFA)

Following our analysis plan, we fit each personality construct in its own model with one latent variable. Model fit for all CFA models is described in Table 2.

Table 2. Fit statistics for all models.

Note. SRMR = Standardized Root Mean square Residual, CFI = Comparative Fit Index.

The 1-factor models all had adequate fit, but the cumulative 6-factor model did not, which meant that the numeric results of that model were not interpretable. We went back and refit models of openness, since it had far and away the poorest fit of the 1-factor models. By removing each individual item and comparing fit between 7 such models, we found the best fitting reduced model (which excluded ‘Imaginative’) had good fit (Table 2). Nevertheless, fitting this reduced openness factor in the 6-factor model did not substantially impact fit; SRMR was above the acceptable threshold.

At this point, it was unclear how to proceed. The revised openness model did not have the poorest fit and all of the individual factor models fit quite well, so it was unclear how we might improve model fit in a principled way. We thus could not proceed with the CFA of all the constructs in our models, as they were.

Augmented dominance model

We planned to evaluate an expanded dominance construct, so we carried out an analysis of an augmented 1-factor dominance model as planned, although we were not confident that a broader model would succeed when it had previously failed for reasons of fit.

The new 1-factor model of dominance fit well (Table 2). Thus, we proceeded to fit this augmented dominance factor within the 6-factor CFA. This version of the 6-factor model fit adequately; fit statistics were just past our pre-defined thresholds. Correlations between personality factors are presented in Table 3.

Table 3. Correlations among Big 5 and augmented dominance, derived from CFA.

Exploratory analyses

At this point, the available results left us in an uncertain position. The Spearman correlations among pre-constructed variables could not be directly compared to the CFA correlations from the augmented dominance 6-factor model. The Spearmen correlations were intended for comparison to the unaugmented CFA model, and the 2 CFA models’ correlations were meant to be compared.

The next best option was to construct new variables in the dataset that were drawn from the structure of the augmented model. We could then calculate Spearman correlations for these variables, and compare those to the Spearman correlations in Table 1 and the CFA correlations in Table 3.

Thus, we created new dominance and openness variables. Openness as the sum of 6 items (excluding ‘Imaginative’) and dominance as the sum of 9 items, its original 5, 3 from the Multidimensional Personality Questionnaire, and 1 from the MIDUS psychological well-being questionnaire (see the method section).

Spearman correlations compared favorably with each other (Table 1 vs 4); correlations for dominance did not differ by any more than 0.05. The largest coefficient, openness, exceeded 0.5. In the case of conscientiousness, the correlation coefficient was almost exactly the same. According to these correlations, augmented dominance appeared to align better with openness than extraversion, and more with neuroticism than the unaugmented dominance construct.

Table 4. Spearman correlation coefficients for MIDI-based personality constructs, with augmented dominance and reduced openness.

Note. MIDI = Midlife Development Inventory. All N > 3831. All coefficients are significant at p < 0.0000001.

The CFA correlations were more revealing (Table 3 vs 4). The largest coefficient was extraversion at 0.35, well above openness; this disagrees with the Spearman correlations. In general, correlations were smaller, which could be expected (Westfall & Yarkoni, 2016), although the (albeit smallest) correlations with agreeableness and neuroticism compared favorably between the two methods.

As an additional exploratory analysis, we calculated Omega hierarchical, Cronbach’s alpha, and Guttman’s lambda (Revelle & Zinbarg, 2009) statistics for MIDI dominance and the augmented dominance domain as well (Table 5).

Table 5. Reliability statistics for dominance.

Note. α = Cronbach’s alpha, λ6 = Guttman’s lambda, ωh = omega hierarchical, ωt = omega total.