Data
This cross-sectional study is a secondary analysis of administrative and survey data from adult participants in the 2008 study, “Disparities and Barriers to Utilization among Minnesota Health Care Program Enrollees”; overall survey results were reported previously [18]. The original 2008 study, an update and partial replication of a similar 2003 effort [19], was funded by the Minnesota Department of Human Services and sought to identify problems with access, barriers, and self-reported utilization among publicly insured individuals in Minnesota. Administratively derived measures, for July 1, 2007 - June 30, 2008, pertained to selected health conditions and utilization, whereas survey data, collected July-December 2008, covered self-reported attitudes, utilization, access, and barriers to care.
Study sample
The study sample included non-institutionalized enrollees of Minnesota Health Care Programs (Minnesota’s Medicaid/CHIP and low-income public health insurance programs). MHCP includes three main programs: Medical Assistance (MA, Minnesota’s Medicaid program), General Assistance Medical Care (GAMC), a state-funded program covering low-income individuals (mainly adult men) not eligible for MA, and MinnesotaCare, a state and federally subsidized program for children and adults without insurance who are not eligible for MA or GAMC. The original study used a stratified design to oversample minorities and obtain comparable proportions of racial/ethnic groups. The survey was mailed, with telephone follow-ups for initial non-response offered in various languages as needed. Survey response rate was 44% in the study overall. The original study received approval from the Institutional Review Boards (IRBs) of Minnesota Department of Human Services and the University of Minnesota; the present project also has review clearance (defined “not human subjects research,” as a secondary analysis of de-identified data) from the Mayo Clinic and University of Minnesota IRBs.
Dependent variables
Our two dependent variables were “ED as usual provider” and number of ED visits. ED as usual provider was a dichotomous measure; the survey item asked where the individual “Which of the following places best describes where you usually go for your health care?,” with emergency department listed alongside seven other choices including primary care providers’ offices, urgent care, and an open-ended option for respondents to identify anything else. For this study, we coded the variable as 1 = ED as usual source of care and 0 = not. The second variable, ED visits, came from an administrative data count of ED visits for each enrollee over the past year. We coded ED visits as the following: 0 = no visits; 1 = 1 visit; 2 = 2+ visits. In sensitivity analyses, we modeled ED visits in various ways: as a binary (none versus any visits), a differently aggregated ordinal (0, 1–2, and 3+), or an uncategorized count. Most of these analyses produced substantively similar results, and so we used the 0-1-2+ coding based on a review of the variable’s natural distribution among this sample, the need to distinguish between single-visit users and other users, and high zero inflation which complicated count models.
Independent variables
Our key independent variables concerned financial and non-financial barriers to accessing health care from survey data (the question and items are shown, along with the rest of the survey, as an appendix to the original report [18]). Barrier-related items were developed primarily based on focus group data and involvement of community members in the original studies [18, 19] and a local survey as part of another project [20]. Respondents were asked whether each in a series of items was a problem in getting the health care they needed. Financial barriers represents a set of seven self-reported items pertaining to financial cost and coverage concerns, most closely representing Penchansky and Thomas’ “affordability” aspects of access [21]. These financial barriers included worries that: insurance won’t cover care, that the respondent will have to pay more than expected, that he/she will have to pay more than he/she can afford, that medications will cost too much, not being sure about being dropped from the public health care program, not knowing what the health plan covers, and not knowing where to go with questions about coverage. The survey asked whether each barrier was ‘a big problem,’ ‘a small problem,’ or ‘not a problem.’ Big and small problem were combined to indicate each financial concern. We used these measures to construct a dummy variable indicating any financial concern (1 = yes; 0 = no) and as a summed count of financial concerns (range: 0–7).
Non-financial barriers included seven practical, non-financial barriers to access (“non-financial barriers”), best understood as representative of accessibility and accommodation under the Penchansky and Thomas model [21]. Rather than coverage or payment, these self-reported barriers concerned practical hardships which complicated access to care, including: transportation difficulties, problems making appointments, not knowing where to go for care, work/family responsibilities, offices/clinics not being open at suitable times, obtaining childcare, and not being able to utilize one’s preferred provider. Items were coded similarly to financial concerns; again, we constructed a dummy variable for any non-financial barrier (1 = yes; 0 = no) and a summed count of barriers (range: 0–7).
Covariates
Because differences in health status may explain some variations in ED use among vulnerable populations [22, 23] we tested available binary indicators for International Classification of Diseases-9th revision (ICD-9-CM) diagnostic categories derived from administrative data. We included two indicators with significant effects: one for the presence of a mental health disorder (ICD-9-CM codes 290–329), and another for an injury or poisoning (codes 800–999), as a first-listed diagnosis. We did not include a third significant category (codes 780–799) because it pertained to general or ill-defined symptoms and conditions or abnormal test results, and had no clear interpretation, whereas the association of mental health problems and with more frequent ED use is supported by research literature [24]–[27] and injury/poisoning (e.g., broken bones, trauma, accidental poisonings) is an expected class of reasons for ED use. Also, because ED use could simply be an expression of a tendency for higher health care utilization overall [27], we included count of primary care office visits (capped at 25 to limit outlier effects) and an indicator for whether the enrollee had an inpatient stay from administrative data. Other covariates included age; education (1–8, with 4 indicating a high school degree and 7 representing a four-year college degree); being married or “living with a partner in a marriage-like relationship” (versus single/divorced/widowed); being employed (versus not); birth in the United States; and indicators for race/ethnicity (Black, Hispanic/Latino, Native American/Alaska Native, and Asian; “White, non-Hispanic” was the reference category). Race/ethnicity was based on administrative data initially; we used self-reported race from surveys to correct or fill in these values. Previous work has shown strong concordance between self-reported and administrative race/ethnicity in this population [28].
Analyses
Analyses were completed using StataSE 12 [29]. For all results shown, we used weights to correct for unequal selection probabilities and to ensure that the sample characteristics matched those of underlying population (i.e., non-institutionalized MHCP enrollees as of March 31, 2008).
Following descriptive statistics, we employed multivariate logistic regression to assess the associations of key independent variables with enrollee-reported ED as usual source of care. Next, we employed ordinal logistic regression to assess the associations of the same independent variables with 0, 1, or 2+ ED visits; we also added ED as usual source of care as a covariate to examine and control for people’s self-reported care tendencies (significance of other covariates did not change based on the inclusion or exclusion of this variable, however). Following both models, to improve interpretability for combinations of predictors, we used post-estimation predicted probabilities [30], based on regression-adjusted estimates.
We performed supplemental sensitivity analyses to assess the robustness of findings under alternative modeling strategies. First, regarding missing data, most variables had no or less than 2% missing; the only ones with higher missing were ED as usual provider (5.86%), and counts of financial (6%) and non-financial (5.95%) barriers. However, to address the effects of any larger amounts of missing and test robustness of findings, we first completed analyses using simple list-wise deletion, then completed sensitivity analyses employing multiple imputation using chained equations (MICE) for all missing data. Substantive results differed very little; in particular, findings for key independent variables (financial and non-financial barriers) remained the same. Therefore, we present the unimputed (yet still weighted) results here for the effective/non-missing sample of 1,737 individuals, but said results are robust to multiple imputation. Second, because ED as usual source of care was a relatively rare outcome and sample size was moderately small, we completed sensitivity analyses to adjust for possible bias. Specifically, this entailed the Firth method [31] to penalized maximum likelihood (“firthlogit” command in Stata, similar to King and Zeng’s proposed approach [32]); significance levels did not change, and odds ratios were within .01 of uncorrected estimates, indicating relatively limited bias in our estimates. Finally, due to sample size concerns and limited information on enrollment, we did not limit or exclude participants based on months of MHCP enrollment; in other words, administratively derived measures may or may not be misrepresentative due to lapses in coverage. As such, we performed sensitivity analyses with the basic indicators of enrollment by month that we did have. Most enrollees had 11 or 12 months of enrollment (73.68% of the effective sample of 1,737); this increased to 79.3% with 10 or more months (or 77.6% if limited to at least 10 consecutive months), and 93.2% with at least six months of enrollment (90.9% if limited to at least six consecutive months). When we limited analysis of ED visits to those with six or more (n = 1,579) and 10 or more consecutive months (n = 1,347) of enrollment, non-financial barriers remained significant in ordered logit models, as did all administratively derived diagnosis and utilization measures.