The data were collected as a component to the larger ‘Aging in the Time of COVID-19’ study, a longitudinal, web based multi-wave study, conducted in 2020. Participants were recruited through online advertisements on social media platforms (Twitter, Facebook) posted by the Center for Innovation in Healthy and Resilient Aging, online list-serves, and university forums. Individuals were eligible to participate if they were aged 50 years or older and were English speaking. Given the English nature of the survey, and to minimize the variability in healthcare plans and policy responses from other countries, for this study, we focus on individuals who reside in the United States and completed the first wave of data collection. We included adults aged 50 and older to examine access issues specific to late midlife and older adulthood. Although the number of such survey respondents was 1443, we focus our study on a subset of these cases, as described later. Given rapid changes in COVID-19 policies occurring at the time, the present research was conceptualized and implemented within a 1.5 week span. This final survey did not undergo pilot or validity testing, but the assessments included were either previously validated or based on previously validated measures, with minor adjustments to increase relevance during COVID. The online survey included questions about a variety of experiences during the time of COVID-19. Data collection for the reported findings was completed using REDCap and ran from April 13–May 15th, 2020 [17]. The information was collected during the height of the stay-at-home orders in the United States. Following completion of the survey, participants had the option to be entered into a raffle to win one of five $25.00 gift cards. The study was approved by the Arizona State University IRB. The methods have been described more fully elsewhere [18, 19].
After agreeing to the informed consent, participants were asked basic demographic questions. Participants were also asked to indicate the presence or absence of chronic conditions. Using items from the 2017 Behavioral Risk Factor Surveillance Survey (BRFSS) Disease Scale [20]. Additionally, participants were asked about healthcare access related to several key access points (housing, transportation, support groups, legal services, dental and vision services, in-home health services, occupational and physical therapy, healthcare provider/doctor, medication, emergency room, counseling/therapy, and food assistance). Specifically, “We would like to ask you about services you may or may not need and may or may not have been able to obtain in the last two weeks due to COVID-19. Thinking about the services listed, please check the box that is appropriate.” Participants could select one of four options: needed and received, needed and did not receive because of COVID-19, needed and did not receive due to reasons other than COVID-19, and did not need. We focused on access to a healthcare provider and medication as these two access points are seen as important for the management of chronic conditions. These were considered the most representative of the access AHRQ access dimensions (e.g., coverage, services, timeliness and workforce), available on the survey. These measures are not exclusive nor exhaustive, but two important points to consider [21]. In addition, we eliminated 962 cases that responded not needing access to a healthcare provider and 678 who reported not needing access to medications because we were interested in those who needed such access. As such, the analytic sample size for the healthcare provider outcome was 481 and 765 for the medication’s outcome.
Demographic and health-related variables were collected for this study. The demographic variables included age (numeric, with ages ranging from 50 to 87), sex (1 = female, 0 = male), race (1 = White, 0 = Other), educational status (represented by two-dummy coded predictors, with those having not attained a Bachelors’ degree serving as the reference group and those with a Bachelor’s degree and those with a graduate degree comprising the other two groups), employment status [1 = employed (full or part-time or self-employed), 0 = not employed (unemployed, homemaker, student, retired, unable to work)], relationship status [1 = partnered (married or member of an unmarried couple), 0 = other], total annual household income (represented by two dummy-coded predictors with income < $50,000 serving as the reference group and those with income between $50,000 and $100,000 and then income > $100,000 comprising the other two groups), and caregiving responsibility (1 = yes, 0 = no). Note that with the exception of age, all of the demographic variables were categorical.
Several health-related predictors were collected and included the study. Participant self-rated health (numeric) was assessed through the single item health question, where participants were asked, “In general, how would you rate your health today?” Responses options were: 1 = very good, 2 = good, 3 = moderate, 4 = bad, and 5 = very bad. Scores were reverse coded for all analyses such that higher scores were indicative of better health (i.e., 1 = very bad and 5 = very good health) [22]. Other health-related predictors include the sum of the chronic health conditions indicated (numeric, with scores ranging from 0 to 10), UCLA loneliness (numeric, with scores ranging from 25 to 79 out of a possible 20 to 80, with higher scores equated to increased loneliness.), [23] PROMIS social isolation (numeric, with scores ranging from 34.8 to 74.2 out of a possible 34.8 to 74.2 with a population average of 50, and scores over 50 indicating higher rates of social isolation), and primary care provider status (1 = has a primary care provider; 0 = otherwise) [24].
Statistical analysis
We first examined basic descriptive statistics and frequencies to identify if unusual values were present and determine the extent of incomplete data. No unusual scores were present for a given variable, but participants had missing data for most [11] of the variables, including the outcomes, where 13 participants had missing data for provider access and 11 had missing data for medication access. Although the percentage of missing data did not exceed 4% for any given variable (with this missingness rate occurring for age and income), dropping cases with missing data (as listwise deletion does) would have resulted in removing up to 14% of the participants from the analyses, resulting in potential estimation bias and loss of power. As such, for the primary analysis, we used a modern missing data treatment, as described below, that effectively treats missing data for predictor and outcome variables. We also examined index plots (i.e., Mahalanobis’ distance by case number) to assess if multivariate outliers were present and examined values of the variance inflation factor to assess if multicollinearity was present. Although no troublesome multicollinearity was present (each variance inflation factor < 3), a multivariate outlier was detected. Inclusion or exclusion of this case did not alter any study conclusion. As such, this case was included in all analyses.
To examine associations between the demographic and health-related predictors (i.e., chronic health conditions, self-perceived health, loneliness, social isolation, established relationship with a primary care provider) and the access outcomes (i.e., access to a healthcare provider and access to medications), we first computed descriptive statistics to describe these variables by each of the outcome categories, as well as conducting bivariate statistical tests (i.e., independent samples t tests and chi-square tests of association). For the primary analysis, we estimated a logistic regression model for each outcome while simultaneously treating missing data on the outcomes and predictors. For this purpose, we obtained model parameters using Bayesian Markov Chain Monte Carlo (MCMC) estimation, which provides unbiased parameter estimates when data are missing at random [25, 26], with this missing data treatment also resulting in retaining in the analysis each participant who reported data for any study variable. Analogous to Firth estimation procedures [27], Bayesian estimation is also recommended for logistic regression when sparse data are present, because such estimation can remove bias in the estimates of regression coefficients and their standard errors [28, 29]. Note, though, that to avoid potential estimation problems that could arise due to a very small number of events, we combined the two “not received” categories due to the small number of participants who responded they did not receive access due to reasons other than COVID-19 (n = 28 and n = 23 for the access to a health provider and medication outcomes, respectively). Thus, for the logistic regression models, the two outcome categories were received or did not receive access. This analysis was implemented in Mplus software, Version 8.6 [30], using weakly informative prior distributions for the regression coefficients, as recommended for logistic regression [28, 29]. We monitored model convergence with the potential scale reduction factor, with a value less than 1.10 indicating convergence. When we estimated the models, the maximum potential scale reduction factor obtained was smaller than 1.01 for all parameter estimates for each of the iterations used to obtain parameter estimates, indicative of superior estimation.
Unlike traditional analyses, Bayesian estimation produces a distribution of values for each model parameter, and we requested 10,000 random draws to build these posterior distributions (after 10,000 burn-iterations). The median of these posterior distributions was used to represent final parameter estimates (i.e., logistic regression coefficients). Further, although we report two-sided p-values associated with the regression coefficients, the 2.5th and 97.5th values from the posterior distributions were used to form 95% Bayesian highest density credible intervals, which, when not containing a value of zero, is comparable to achieving statistical significance in traditional analyses with an alpha level equal to 0.05. Note that, analogous to bootstrapping, the use of such intervals does not rely on distributional assumptions or large-sample theory. To convey the practical importance, or meaningfulness, of the analysis results we computed and graphed model estimated probabilities of access for significant predictors, with the estimated probabilities and graphs obtained via SAS® software, version 9.4 M7 [31].