Data Analysis Plan

This study tested the extent to which heterosexual and sexual minority adults differed in their willingness to use digital health tools for public health screening and tracking, a latent variable, and whether sexual minority adults’ willingness to use these COVID-19–related digital health tools was associated with age, gender, and race or ethnicity. Measurement invariance (ie, whether the measure means and assesses the same thing across groups) was tested across heterosexual and sexual minority adults.

Descriptive statistics and Cronbach α were computed using Stata (version 16; StataCorp) [52], and coefficient ω and all analyses of associations and latent variables were conducted using Mplus (version 8; Muthén & Muthén) [53]. Weighted least squares estimation with Delta parameterization was used to estimate model parameters [53]. This estimation method uses a diagonal weight matrix with SEs and mean- and variance-adjusted chi-square test statistics that rely on a full weight matrix (ie, ESTIMATOR=WLSMV in Mplus) [53]. It is particularly appropriate for ordinal and nominal data [53]. Model fit was assessed with several fit indices based on any 2 of the following 3 criteria: a root-mean-square standard error of approximation (RMSEA) value of 0.06, a comparative fit index (CFI) value of at least 0.95, and a standardized root-mean-square residual (SRMR) criterion of 0.08 or less [54,55].

The author tested the extent to which the 5-item measure was invariant across the 2 sexual orientation categories. The 3 levels of measurement invariance—configural, metric, and scalar—were tested to determine if the 5-item measure was invariant across sexual-orientation categories. For ordinal variables and weighted least squares estimation methods with Delta parameterization, configural invariance (ie, pattern invariance), the least strict form of invariance, shows that each group has the same indicators loading onto the same factors in the same direction (ie, positive versus negative). To model configural invariance: (1) factor loadings are free to vary across groups, (2) thresholds are free to vary across groups, (3) scale factors are fixed to 1 in all groups, (4) factor means are fixed to 0 for all groups, and (5) factor variances are free to vary across groups [53]. Metric invariance (ie, weak invariance) indicates the invariance of factor loadings across groups, wherein (1) factor loadings are constrained to be equal across groups, (2) the first threshold of each item is constrained to be equal across groups, (3) the second threshold of the item that sets the metric of the factor is constrained to be equal across groups, (4) scale factors are fixed to 1 in 1 group and free to vary in the other groups, (5) factor means a Muthén & Muthénre fixed to 0 in 1 group and free to vary in the other groups, and (6) factor variances remain free to vary across groups [53]. Scalar invariance (ie, strong invariance) indicates equivalence of item intercepts or thresholds, in the case of categorical or ordinal variables, across groups and is the minimum needed to proceed with using a measure to test for differences in latent factor means between groups [56,57]. Scalar invariance is the same as metric invariance, except that thresholds are constrained to be equal across groups [53].

To compare invariance models, the author used a difference in CFI (ΔCFI) equal to or greater than 0.01 to indicate noninvariance [56]. Thus, a lack of worsened model fit with increased constraints indicates measurement invariance. Although scaled chi-square difference tests scaled for the weighted least squares estimator were conducted, this test may detect small discrepancies in ways that are not practically or theoretically meaningful in sample sizes greater than 200 [56-58].

Upon determining measurement invariance, the latent factor mean difference between heterosexual adults and sexual minority adults in the underlying factor of willingness to use digital health tools for public health screening and tracking was tested. Specifically, the latent variable for willingness to use digital health tools was standardized such that its mean was fixed to 0 and SD set to 1. The factor mean remained 0 for the reference group, heterosexual individuals, but the factor mean was freely estimated for the comparison group, sexual minority individuals. Thus, the resulting mean for sexual minority individuals reflected the difference in the mean from the reference group on a standardized metric, or in SD units. To identify correlates of willingness to use digital health tools for public health screening and tracking among sexual minority populations, specifically, willingness to use digital health tools was regressed on sexual minority adults’ age group, gender, and race or ethnicity, respectively.

Within an intersectionality-informed analytic framework, as described by Jackson et al [59], we can use additive measures of interaction to test for joint, referent, or excess intersectional disparities. Using the present analyses as a guiding example, the outcome variable would be recoded such that higher scores reflect a more adverse or disparity-oriented outcome (ie, less willingness to use COVID-19 screening and tracking tools). The predictor, gender (women coded 1), and the moderator, sexual orientation identity (sexual minority identity coded 1), would be coded such that the reference category (coded 0) is the nonminoritized group in this instance (ie, men and heterosexual adults) and the active category (coded 1) is the minoritized group (ie, women and sexual minority adults). Thus, the code of 1 reflects an adverse social position. For gender and sexual orientation, the original equation in the primary analyses before recoding and not including other covariates would be:

Given that the outcome should reflect a negative outcome to identify a disparity, the analyses would be repeated with the outcome variable recoded to reflect an unwillingness to use digital health tools rather than a willingness to use these tools. For gender and sexual orientation, the equation after recoding and not including other covariates would then be:

The joint disparity compares outcomes from the cell or group at the intersection of 2 minoritized identities, in this case, sexual minority women, to the group at the intersection of the 2 corresponding nonminoritized identities, in this instance, heterosexual men. In our example, b₁ + b₂ + b₃ equals the joint disparity in unwillingness to use COVID-19 screening and tracking tools comparing sexual minority women to heterosexual men. Referent disparities are those that affect only 1 minoritized population or identity, in this case, women compared with men among heterosexual adults or heterosexual adults compared with sexual minority adults among men. It describes the disparity based on gender as a proxy for sexism or sexual minority identity as a proxy for heterosexism or homonegativity, but not both. Specifically, b₁ equals the referent gender disparity in unwillingness to use COVID-19 digital screening among heterosexual adults, and b₂ equals the referent sexual minority disparity among men. Finally, the excess intersectional disparity focuses on the intersection of minoritized identities and describes the extent to which the joint disparity exceeds the 2 individual referent disparities. Suppose it is greater than 0, or statistically significant. In that case, the strength of the association indicates the disparity at the intersection of minoritized gender and sexual orientation, that is, women who are also sexual minority adults, and b₃ equals this excess intersectional disparity. A more detailed explanation can be found in Jackson et al [59] and VanderWeele and Tchetgen Tchetgen [60].

Disparities are indicated if the regression coefficients are positive, reflecting direct associations (ie, disadvantages for the minoritized groups) as opposed to inverse associations (ie, advantages for the more minoritized group). An advantage on an outcome for a relatively disadvantaged group that otherwise disproportionately and systematically experiences worse health outcomes and greater health risks would not meet established definitions of a disparity [61,62].

Given the complex nature of these survey data, analyses were adjusted using a sampling weight based on the inverse of the probability of selection in the sample. These analyses also accounted for stratification using pseudostrata based on census tracts. The data producer, NORC, used pseudostrata to preserve confidentiality. Per NORC, they did not include cluster variables because there were negligible cluster effects, and excluding these variables better preserved confidentiality (personal communication; Jennifer Benz, May 14, 2021). Descriptive statistics for the present sample accounted for weighting and stratification to reflect the complex survey design and national representativeness of the sample along key raking variables (ie, age, gender, and race or ethnicity). Latent factor mean differences (ΔM) and regression coefficients (b) are presented with their 95% CIs. Missing data, which were up to 3.7% missing across analyses, were handled using listwise deletion.