Study design and sample

A cross-sectional study of pediatric ED visits from five EDs in the upper Midwest was performed. Included were all visits by children aged 0–17 years to one of the five EDs between June 2011 and May 2012. Analysis was limited to three racial groups (African American (AA), AI, and White). Approval from all Institutional Review Boards providing data was obtained, along with a waiver of informed consent.

Statistical analysis

LWCET was the primary outcome measure, defined as leaving the ED without complete evaluation or treatment and recorded as a binary yes/no variable. Our primary exposure variable of interest was race, limited to AA, AI, and White since few other races were common at all of our study sites. Age was categorized as less than 1 year, 1–4 years, 5–10 years, and 11–17 years old. Insurance type was defined as private and medical assistance (MA)/other. A two-level triage variable was created where the three highest and two lowest triage levels were combined. High triage included emergent, critical and acute visits, while low triage included urgent and non-urgent visits. Sex was categorized as male and female. Timing of visit was developed based on day and time of visit to the ED. Categories included: weekday business hours, evening hours during a weekday, night hours during a weekday, and weekends. ED setting was categorized into rural or urban, where urban was an ED located in a city with a population of 100,000 people or more. Distance from the ED was based on the patient’s zip code, and categorized as 5 miles or less and more than 5 miles to the ED.

An important source of variation to consider in examining LWCET is how busy the ED is at the time the patient presents. For each site, we used partial least squares approach for structural equation models to construct a formative latent variable which measured the ED activity level. This variable takes into account the number of patients in each triage level at the emergency department as well as the time of day the patient arrived. The analysis was completed using the semPLS package in R with the centroid weighting scheme using R, version 3.2.3 (www.r-project.org). From this analysis, we created a standardized score for the activity level for each visit such that each site had a mean activity level near zero with positive number indicating increased activity and negative number indicating decreased activity. We checked basic validity of this construct by assessing its correlation with time to examination and found a Spearman’s rank correlation of 0.29 indicating a modest level of association.

Descriptive analysis including counts and percentages are reported. Univariate analyses were performed using logistic regression with only that variable and a random study site effect. Logistic regression with a random effect for study site was used to determine differences in LWCET by race and ED location controlling for other possible confounding factors including age, insurance type, triage level, sex, timing of visit, distance from the ED and ED activity level. Unadjusted and adjusted odds ratios are reported with 95% confidence intervals. Multiple imputations analysis was completed to examine the impact of missing data, which ranged from 0.04–1.08%. All logistic regression and imputations analyses were performed using SAS version 9.4, SAS Institute, Cary, NC.

Demographics

The proportion of less severe triage scores was higher among children who LWCET compared to those who did not LWCET (70.70% vs. 57.50%), while percentages of more severe triage scores were lower among children who LWCET compared to those who did not LWCET (42.50% vs. 29.30%; Table 1). The proportion of return visits to the ED at ≤24 and ≤ 48 h was higher among LWCET visits (1.56% vs. 5.21% and 2.99% vs. 8.14% respectively, P < 0.001).

Univariate analysis

AI children were more likely to LWCET compared to White children (Table 1). Children who LWCET were more likely to be triaged with less severe level (70.70%, P < 0.001), travel 5 miles or less to the ED (54.36%, P = 0.0014), be males (55.38%, P < 0.001), and visit during weekday evening hours (30.48%, P < 0.001) as compared to children who did not LWCET.

Multivariate analysis

Triage level was significantly associated with LWCET, where children with more severe triage scores had lower odds of LWCET compared to those with less severe scores (OR = 0.63, CI = 0.55–0.73). Visits that occurred during weekday night hours had odds 1.43 times higher of LWCET compared to visits during weekday business hours (CI = 1.18–1.71). Children who lived further than 5 miles from the ED had lower odds of LWCET (OR = 0.86, CI = 0.76–0.98). ED activity level was also found to be associated with LWCET, where a busier ED had higher odds of resulting in LWCET (OR = 1.32, CI = 1.24–1.33).

Imputations analysis

Results from imputations analysis revealed slight differences in associations with LWCET due to missing data values. For example, the odds of LWCET for children who lived further than five miles from the ED was found to be significantly lower compared to children who lived within five miles (OR = 0.86, CI = 0.77–0.98). In addition, the odds of LWCET for visits during weekend hours was significantly lower compared to visits during weekday business hours (OR = 0.80, CI = 0.68–0.94).

Strengths

This study had several strengths. The data encompasses five EDs with a larger proportion of AI children that allows us to look at patterns in LWCET in this specific minority group not currently explained within the literature. While limited by medical record data, we were still able to account for several confounders such as distance from ED, ED setting, timing of visit, and the level of traffic within the ED. The development of our ED activity level allows us to look more into the environment of the ED, while other studies have only focused on the day and time of visit without accounting for how busy the ED is at the time of visit.

Limitations

There are potential limitations to this study. First, these data may not be representative of the AI children outside of the Upper Midwest since many regional differences may affect the relationship between race and LWCET; therefore, no assumptions should be made about generalizability outside this region. Furthermore, while we did attempt to account for differences by site by including it as a random effect in our model, the diversity of each site may affect the results and the EDs included may not be a random sample of the population of EDs that serve AI children. Finally, medical record data did not allow us to determine at what point in the visit children LWCET or for what reason, therefore we are only able to suggest further exploration into this observed relationship to better inform possible intervention efforts.