The Influence of Social Determinants of Health on Emergency Departments Visits in a Medicaid Sample

Abstract

Objective.

To evaluate the relationship between social determinants of health (SDH) and emergency department (ED) visits in the Medicaid Cohort of the District of Columbia.

Methods.

We conducted a retrospective cohort analysis of 8,943 adult Medicaid beneficiaries who completed a SDH survey at time of study enrollment. We merged the SDH data with participants’ Medicaid claims data for up to 24 months prior to enrollment. Using latent class analysis, we grouped our participants into four distinct social risk classes based on similar responses to the SDH questions. We classified ED visits as primary care treatable (PCT) or ED care needed (EDCN) using the Minnesota algorithm. We calculated the adjusted log relative PCT and EDCN visit rates among the social risk classes using generalized linear mixed effects models.

Results.

The majority (71%) of the 49,111 ED visits made by the 8,943 participants were EDCN. The adjusted log relative rate of both PCT and EDCN visit rates increased with each higher (worse) social risk class compared to the lowest class. Participants in the highest social risk class (i.e. unemployed and many social risks) had a log relative PCT and EDCN rate of 39% (28% – 50%) and 29% (21% and 38%) respectively, adjusted for age, sex and illness severity.

Conclusions.

There is a strong relationship between social determinants of health and ED utilization in this Medicaid sample that is worth investigating in other Medicaid samples and patient populations.

INTRODUCTION

Background.

As a critical component of the US healthcare safety net, emergency departments (EDs) treat a disproportionate share of patients with complex social and behavioral needs. To date, the majority of studies that have examined the relationship between social determinants of health (SDH) and ED visits have used community-level measures of both.^1,2 A few studies have evaluated the relationship between community-level SDH factors and patient-level ED outcomes.^3–5 For example, Beck et al. found that children with asthma who lived in the highest quartile of housing code violation census tracts were at increased odds of an ED revisit or rehospitalization within 12 months of an initial acute visit compared to those living in the lowest quartile census tracts.³

Importance.

Community-level SDH information represents a measure of the health-related risk factors associated with the community in which a person lives. Community-level SDH information is well-suited for community-level interventions such as improving food access and quality in communities with food deserts or improving green space and walkability in neighborhoods that lack these amenities.^6,7 Ecologic bias⁸ can result from incorrectly assuming that community-level and individual-level risk are the same.^9,10 Few studies have examined the relationship between patient-level SDH information and ED visits and all of them have focused on a specific SDH rather than a wide array of social risk factors.

Goals of this Investigation.

The objective of this study was to evaluate the relationship between multiple SDH and the number of ED visits, both measured at the individual level. We hypothesized that patient-reported SDH would significantly influence the number of ED visits after adjusting for age, sex and illness severity.

METHODS

Study Design and Setting.

We performed a retrospective cohort analysis using data from the Medicaid Cohort of the District of Columbia (MCDC) to examine the relationship between patient-reported SDH and the number of ED visits of each participant within a 2-year period prior to study enrollment. The study enrolled adult Medicaid beneficiaries who presented for medical care at the emergency departments, a primary care clinic or an obstetrics and gynecology clinic affiliated with one of two medical facilities located in Washington, District of Columbia (DC).¹¹ One of the facilities is a non-profit, public medical center that has 210 staffed hospital beds and an annual ED volume of 55,000. It is located in an area of DC with high unemployment and poverty rates. In contrast, the other hospital is a for-profit, academic tertiary care hospital with 370 staffed beds with an annual ED volume of 75,000. It is located in an affluent section of DC. This study was approved by the Institutional Review Board.

Selection of Participants.

To be eligible for the study, patients had to be between the ages of 18 and 64, insured by the DC Medicaid program, have access to a telephone and present at one of the aforementioned acute or ambulatory care sites affiliated with the two facilities between September 2017 through December 2018. Patients were excluded if they were unable to understand consent, were non-English speaking, too sick (i.e. triage acuity level 1) or also insured by Medicare. Patients willing to participate signed a written consent form and agreed to complete an interview during their medical encounter, allowed a research assistant (RA) to abstract relevant study data from their medical record and permitted the study team to obtain a copy of their Medicaid claims.

During the 16-month enrollment period, we screened 17,719 patients, 12,346 were eligible. The most common reason for ineligibility (85%) was the patient had already been enrolled. Among the 3,403 eligible patients that were not enrolled, 55% refused, 24% reported feeling too uncomfortable and 21% completed their medical visit before the RA was able to approach them, resulting in a final sample of 8,943 participants. Due to enrollment occurring at two EDs with much higher patient volumes and longer hours of operation than either of the clinics, the majority of study enrollment (87%) occurred at the time of an ED visit.

Measurements.

At enrollment, the RA documented each participant’s date of birth, sex and DC Medicaid beneficiary number from the patient registration system. The RA also conducted a SDH interview with each participant. Participants received a $5 gift card after survey completion.

The SDH survey was developed based on the World Health Organization (WHO) SDH conceptual model and included measures of structural and intermediary determinants of health.¹² The WHO model posits that structural determinants of health such as education, income and occupation define a person’s socioeconomic position in society. In turn, SES position shapes intermediary social determinants including material circumstances, health behaviors and psychological factors. Structural and intermediary SDH impact the incidence of illness and injury as well as the ability of people to manage their health problems.

Our survey included 35 questions and took approximately 10 minutes to complete and consisted largely of questions or short scales previously validated by others.¹¹ The survey included the following structural determinants of health: sexual orientation, highest level of education achieved, employment status and employment duration. The intermediary determinants we asked participants about included measures of material circumstances (i.e. food insecurity, financial strain, etc), health behavior questions (i.e. smoking, alcohol use, etc), and psychological factors (i.e. marital status, loneliness, etc).¹¹

Because structural and intermediary determinants of health are correlated with one another, considering all SDH factors simultaneously in a multivariate regression model may mask relationships and lead to important associations being missed. To address this, we used latent class analysis (LCA) to identify unique subgroups of participants with similar response profiles to the SDH assessment variables. We dichotomized all of the SDH factors to make them easier to interpret the class solution. We wanted a relatively small number of latent classes for ease of interpretation so we ran models with 2 – 6 class solutions. LCA assigns individuals to classes based on their probability of being in classes given the pattern of scores they have on the indicator variables, in our case, the SDH factors.

The LCA identified four distinct social risk classes within our cohort based on fit, classification diagnostics and ease of interpretation.¹¹ Class 1 members reported the fewest social adversities and were most likely to be employed; we refer to class 1 as “employed and fewest social risks.” Class 2 participants also had a high employment rate like participants in class 1 but they were most likely to report financial strain (i.e. difficulty paying bills); we refer to them as “employed and high financial strain.” Class 3 consisted of participants who were mostly unemployed and reported limited access to internet access at home and to a car for medical appointments; we named class 3 “unemployed and limited internet and car access.” Finally, class 4 participants (9%) reported the most social adversities including the highest rates of unemployment, smoking, food insecurity and housing instability; we named this class “unemployed and many social risks”. Figure 1 depicts the prevalence of the 21 SDH risk factors for each of the four latent social risk classes.

Figure 1.

The Probability Of Each Social Adversity By The Four Social Risk Classes.

We obtained permission to the Medicaid claims data for a 42-month period (24 months prior and 18 months post study enrollment). In this study, we used the Medicaid claims associated with a 24-month period before and up to the day of study enrollment. The claims data include eligibility information (dates of Medicaid coverage, reason for eligibility, date of birth, date of death, etc) as well as detailed billing information such as first and last date of service, type of claim (i.e. inpatient, outpatient, pharmacy, professional, etc), ICD-10 diagnosis codes, CPT codes, place of service, and provider name and national provider identification (NPI) number. We used the Medicaid claims to determine the total number of months each participant was covered by Medicaid during the 24-month observation period, to identify the total number of ED visits that occurred and to measure illness severity.

There is no widely adopted method of defining an ED visit using Medicaid claims data.¹³ We used a definition similar to the one developed by Venkatesh and colleagues. We defined an ED visit according to the following criteria: (1) an emergency medicine (EM) professional claim with a CPT code between 99281–99285 (EM evaluation and management) or a CPT code between 99290–99292 (critical care evaluation and management) and place of service=23 (emergency room); or (2) a facility claim with a revenue code between 0450–0459 (emergency room) or 0981 (ER professional fee); or (3) the presence of both a professional and facility claim with either same first dates of service or first dates of service within one day of each other as one unique ED visit. If a participant had more than one facility claim on the same date, we counted it as two ED visits if the facility names were different (i.e. treated at two different EDs). We characterized ED visits as admitted versus discharged based on claim type (i.e. inpatient versus outpatient). We excluded the ED visit that occurred on enrollment date for participants who were enrolled in the study during an ED visit to one of the participating hospitals.

After we identified all unique ED visits that occurred within the 2-year period, we used the Minnesota Algorithm to determine whether the ED visit was primary care treatable (PCT) or ED care needed (EDCN).¹⁴ The Minnesota algorithm defines an ED visit as EDCN under the following conditions: (1) ED visit results in admission; (2) ED visit has a high severity E&M code (i.e. 99284, 99285 or critical care code); or (3) medium severity E&M code (i.e. 99282, 99283) plus an ED procedure performed that indicated the patient needed to be seen in the ED. A PCT visit is defined by the Minnesota algorithm as a visit with a medium severity E&M code without the ED indicator procedure or a visit with a low severity E&M code (i.e. 99281).

Finally, we used all of the claims data to characterize the illness severity for each participant according to the Chronic Disability Payment System (CDPS).^15,16 The CDPS is a combination of a diagnosis and pharmacy-based risk adjustment system developed specifically for the Medicaid population. The CDPS classifies medical diagnoses using the ICD-10 diagnosis codes and national drug codes (NDC) documented in the encounter and pharmacy claims data into 20 major diagnosis categories and 15 major pharmacy groups based on body systems (i.e. cardiovascular, renal, pulmonary, etc) or disorder types (i.e. diabetes) and then within each category stratifies by level of severity. The level of severity reflects the healthcare needs of the individual with a diagnosis/disorder within a given category. The CDPS yields indicator variables for the 20 major categories and 15 pharmacy groups and their corresponding weights are applied to estimate an overall CDPS score for each individual. A higher CDPS score connotes more severe illness.

Outcomes.

The main outcomes for this study are the total number of PCT and EDCN visits per quarter enrolled in Medicaid over the 24-month period prior to study enrollment. We summed each type of visit over each of eight quarters so that we could evaluate the possibility of seasonal trends in ED visits over time. Of the 8,943 participants in this analysis, 82% were enrolled in the DC Medicaid program during the entire 24-month period; 95% were enrolled at least 12 months.

Analysis.

We described the relationship between each structural and intermediary SDH factor, and the quartiles of the CDPS by the mean number of ED visits during the study period and the 95% confidence interval (CI) using quasi-Poisson regression models since both types of ED visits were positively skewed. We also examined the relationship between the social risk classes and the total number of ED visits using a Kruskal Wallis test.

After the univariable analysis, we separately modeled the number of PCT visits and the number of EDCN visits using generalized linear mixed effects models (GLIMMIX) with a Poisson distribution. The GLIMMIX models included a random intercept for each subject, fixed effects for social risk class, age and gender as well as offset terms that reflected the total number of days each subject was enrolled in Medicaid during that quarter.¹⁷ The random intercept accounts for the correlation in repeated measures (i.e. ED visits per quarter) among individuals. In the multivariate models, we used the four social risk classes to estimate the relationship between SDH and ED visits rather than the individual structural and intermediary determinants of health because of the moderate to strong correlation among many of them.¹¹ These basic GLIMMIX models quantify the influence of social risk factors adjusted for age and gender only.

We then modeled the total number of PCT and EDCN visits adjusted for age, gender and illness severity using the log CDPS (included as a natural spline with 3 degrees of freedom). A priori, we expected the CDPS to have a strong relationship to ED visits. We used log CDPS to take account of its right skewed distribution and we used a natural spline to allow the effect of CDPS to be non-linear while requiring the effect to be smooth. After this, we modeled the total number of PCT and EDCN visits adjusted for age, gender and illness severity as well as the presence of selected medical conditions such as diabetes, cardiovascular disease, and psychiatric illness. For the condition-specific models, we calculated an overall CDPS score without that specific condition and included a separate indicator variable for the medical condition to estimate its effect on the ED visit rate. The regression coefficients represent the log relative rate of ED visits for one category (i.e. a higher social risk class) compared to the reference category (lowest social risk class). We also tested for an interaction effect between social risk class and medical condition in each of the condition-specific models.

Finally, we conducted a posthoc power analysis using a one sample z-test to determine the power we had to detect differences in the adjusted log ED visit rates between the first social risk class (employed and fewest social risks) and each of the three other social risk classes. The power calculations revealed that we had adequate power to detect the differences we observed for social risk class 3 and 4 compared to the lowest social risk class. However, we did not have adequate power to detect the smaller differences that we observed for PCT visits (i.e. < 8%) and EDCN visits (i.e. < 5%) between social risk class 2 (i.e. employed and high financial strain) and the lowest social risk class.

All analyses were completed using either Statistical Analysis Software (SAS) Version 9.4 (SAS Institute, Cary, NC) or R Version 4.0.2 (http://www.r-project.org/).

RESULTS

Characteristics of Study Subjects.

The median age of study participants is 36 (IQR 18 – 51). The majority of participants are female (66%), black (90%) and 52% were working at the time of study enrollment. Table 1 shows the mean number of PCT and EDCN visits during the two-year period by demographic and SDH factors. For many of the characteristics, the mean number of EDCN and PCT visits increases as the SDH characteristic worsens (see Table 1). For example, on average, participants with lower education have a higher number of EDCN and PCT visits compared to participants with more education. On average, people with limited to no internet access at home have more ED visits (of both types) compared to those with access by phone and computer. Furthermore, the average number of EDCN and PCT visits rises with increasing illness severity (CDPS quartiles).

Table 1.

Mean Number (95% Confidence Interval) of Emergency Department Care Needed (EDCN) And Primary Care Treatable (PCT) Visits Within Past Two Years By Socio-Demographic and Social Determinant of Health Characteristics.

Characteristic	N 8,943 (%)	Mean EDCN¶ (95% CI)	Mean PCTβ (95% CI)
A. Demographics
Age at Time of Enrollment*
18 – 24	1,425 (16%)	3.24 (2.78, 3.79)	1.66 (1.48, 1.88)
25 – 44	4,262 (48%)	3.76 (3.46, 4.09)	1.59 (1.48, 1.71)
45 – 64	3,256 (36%)	4.38 (4.01, 4.79)	1.55 (1.42, 1.68)
B. Structural Determinants
Sex
Male	2,997 (34%)	4.11 (3.75, 4.52)	1.74 (1.61, 1.89)
Female	5,946 (66%)	3.80 (3.54, 4.07)	1.51 (1.42, 1.60)
Sexual orientation
Heterosexual/bisexual	8,135 (91%)	3.84 (3.64, 4.04)	1.58 (1.50, 1.66)
Gay/lesbian/other	580 (6%)	3.20 (2.58, 3.97)	1.53 (1.26, 1.86)
Education*
< High school degree	1,536 (17%)	5.13 (4.56, 5.78)	1.96 (1.77, 2.17)
High school degree/GED/	4,525 (51%)	3.70 (3.41, 4.02)	1.66 (1.55, 1.77)
Some college/associate’s degree	2,203 (25%)	3.58 (3.18, 4.03)	1.34 (1.21, 1.49)
Bachelor’s degree+	677 (8%)	3.53 (2.85, 4.38)	1.07 (0.87, 1.32)
Work Status*
Working full-time (1 or 2 jobs)	2,962 (33%)	2.66 (2.38, 2.99)	1.29 (1.18, 1.41)
Working part-time	1,570 (18%)	3.01 (2.60, 3.49)	1.44 (1.28, 1.62)
Not working due to disability/ill health	1,412 (16%)	6.76 (6.10, 7.50)	1.88 (1.69, 2.09)
Not working for other reasons	572 (6%)	3.99 (3.21, 4.95)	1.64 (1.36, 1.98)
Not working and looking for work	2,420 (27%)	4.31 (3.90, 4.75)	1.87 (1.72, 2.03)
Duration of Working/Not Working*
Working for ≥ 5 years	870 (10%)	2.90 (2.38, 3.52)	1.28 (1.08, 1.51)
Working for 1 – 4 years	1,684 (19%)	2.73 (2.36, 3.16)	1.31 (1.16, 1.47)
Working for < 1 year	1,951 (22%)	2.79 (2.44, 3.19)	1.39 (1.25, 1.55)
Not working for < 1 year	1,673 (19%)	3.95 (3.50, 4.46)	1.67 (1.51, 1.85)
Not working for 1 – 4 years	1,183 (13%)	4.74 (4.15, 5.40)	1.65 (1.46, 1.87)
Not working for ≥ 5 years	1,518 (17%)	6.56 (5.94, 7.24)	2.15 (1.96, 2.37)
C. Intermediary Determinants
Marital Status*
Married/living together	1,109 (12%)	3.65 (3.09, 4.30)	1.33 (1.14, 1.54)
Single/Never married	6,589 (74%)	3.81 (3.57, 4.07)	1.65 (1.56, 1.74)
Widowed/separated/divorced	1,240 (14%)	4.60 (4.00, 5.28)	1.49 (1.31, 1.71)
Living Arrangements*
Lives alone	1,390 (16%)	5.24 (4.65, 5.92)	2.04 (1.84, 2.27)
Lives with other adults	2,955 (33%)	3.91 (3.55, 4.30)	1.63 (1.50, 1.77)
Lives with children only	1,311 (15%)	3.36 (2.87, 3.92)	1.49 (1.31, 1.69)
Lives with other adults and children	3,265 (37%)	3.54 (3.21, 3.89)	1.38 (1.27, 1.51)
Internet Access At Home*
Access by smart phone and computer	5,333 (60%)	3.31 (3.07, 3.58)	1.36 (1.28, 1.45)
Access by smart phone or computer	2,837 (32%)	4.52 (4.13, 4.95)	1.82 (1.69, 1.97)
No internet access	714 (8%)	5.98 (5.12, 6.99)	2.36 (2.06, 2.70)
Regular Doctor*
Yes	6,871 (77%)	4.06 (3.81, 4.32)	1.56 (1.47, 1.65)
No	2,042 (23%)	3.38 (2.99, 3.84)	1.68 (1.52, 1.86)
Has Access to Car for Medical Appointments*
Yes	5,336 (60%)	3.65 (3.38, 3.94)	1.42 (1.33, 1.52)
No	3,556 (40%)	4.29 (3.94, 4.68)	1.84 (1.72, 1.98)
Alcohol Per Week
None	5,557 (62%)	4.32 (4.04, 4.61)	1.7 (1.60, 1.80)
1 – 3 drinks	2,115 (24%)	3.00 (2.64, 3.40)	1.38 (1.24, 1.54)
≥4	1,271 (14%)	3.61 (3.11, 4.20)	1.43 (1.25, 1.64)
Uses Illicit Drugs
No	8,583 (96%)	3.84 (3.62, 4.07)	1.56 (1.48, 1.64)
Yes	360 (4%)	5.46 (4.32, 6.92)	2.21 (1.80, 2.71)
Smoking Status*
Not at all	5,282 (59%)	3.68 (3.41, 3.97)	1.48 (1.39, 1.58)
Less than daily	969 (11%)	4.12 (3.49, 4.87)	1.64 (1.42, 1.90)
Daily	2,671 (30%)	4.29 (3.88, 4.73)	1.78 (1.63, 1.94)
Vigorous Physical Activity*
None	2,891 (32%)	4.44 (4.05, 4.86)	1.54 (1.42, 1.68)
1 – 2 days per week	1,323 (15%)	4.03 (3.51, 4.64)	1.68 (1.49, 1.89)
3 – 4 days per week	1,508 (17%)	3.80 (3.32, 4.35)	1.47 (1.31, 1.66)
5 – 6 days per week	1,327 (15%)	2.90 (2.46, 3.42)	1.41 (1.23, 1.61)
7 days per week	1,888 (21%)	3.79 (3.36, 4.27)	1.81 (1.64, 1.99)
Mental Health Index*
Lowest quartile (0 – 56) worst	2,296 (26%)	4.80 (4.36, 5.29)	1.72 (1.57, 1.89)
Middle lower (57 – 76)	2,287 (26%)	3.81 (3.41, 4.25)	1.64 (1.49, 1.81)
Upper middle (77 – 92)	2,817 (32%)	3.44 (3.10, 3.82)	1.49 (1.36, 1.63)
Highest quartile (93+) best	1,516 (17%)	3.57 (3.11, 4.11)	1.5 (1.32, 1.69)
Loneliness*
None of the time	4,942 (55%)	3.50 (3.23, 3.79)	1.48 (1.38, 1.58)
Little to some of the time	2,317 (26%)	4.15 (3.73, 4.62)	1.56 (1.41, 1.72)
Good bit/Most/All of the time	1,621 (18%)	4.69 (4.16, 5.30)	1.97 (1.78, 2.18)
Food Insecurity*
None	4,441 (50%)	3.67 (3.37, 3.98)	1.46 (1.35, 1.57)
Sometimes true to one question	2,749 (31%)	3.79 (3.42, 4.21)	1.59 (1.46, 1.74)
Always true to one question	830 (9%)	4.52 (3.80, 5.37)	2.04 (1.77, 2.35)
Always true to both questions	916 (10%)	4.85 (4.14, 5.69)	1.80 (1.56, 2.08)
Trouble Paying Utilities
No	6,528 (73%)	3.94 (3.69, 4.20)	1.63 (1.54, 1.72)
Yes	2,415 (27%)	3.82 (3.42, 4.25)	1.48 (1.35, 1.63)
Trouble Paying Rent or Mortgage
No	6,979 (78%)	3.90 (3.66, 4.15)	1.60 (1.51, 1.69)
Yes	1,964 (22%)	3.93 (3.49, 4.43)	1.55 (1.40, 1.72)
Trouble Paying Phone Bill
No	6,951 (78%)	3.91 (3.67, 4.17)	1.56 (1.48, 1.65)
Yes	1,992 (22%)	3.88 (3.45, 4.37)	1.68 (1.52, 1.85)
Housing Situation
House	2,494 (28%)	3.39 (3.03, 3.79)	1.42 (1.29, 1.57)
Apartment	5,709 (64%)	3.84 (3.58, 4.12)	1.52 (1.43, 1.62)
Other	740 (8%)	6.16 (5.28, 7.19)	2.67 (2.35, 3.04)
Housing Condition Problems
None	5,802 (65%)	3.86 (3.60, 4.14)	1.59 (1.50, 1.69)
One	1,771 (20%)	3.93 (3.46, 4.45)	1.49 (1.33, 1.67)
Two or more	1,370 (15%)	4.08 (3.55, 4.69)	1.68 (1.49, 1.90)
Worried about housing stability
No	6,947 (78%)	3.72 (3.49, 3.98)	1.53 (1.45, 1.62)
Yes	1,996 (22%)	4.53 (4.06, 5.06)	1.78 (1.61, 1.96)
Length of Time Living at Present Place*
< 1 year	2,206 (25%)	4.33 (3.89, 4.82)	1.84 (1.68, 2.02)
1 – 2 years	1,273 (14%)	3.75 (3.22, 4.36)	1.61 (1.41, 1.83)
2 – 4 years	1,933 (22%)	3.97 (3.52, 4.47)	1.52 (1.37, 1.69)
5+ years	3,502 (39%)	3.64 (3.32, 3.99)	1.45 (1.34, 1.58)
Relationship with Someone Who Threatened or Physically Hurt You*
No	6,084 (68%)	3.46 (3.25, 3.68)	1.52 (1.43, 1.62)
Yes	2,626 (29%)	4.53 (4.17, 4.92)	1.70 (1.55, 1.85)
Ever Been in Jail or Prison*
No	5,940 (66%)	3.48 (3.26, 3.71)	1.41 (1.33, 1.50)
Yes	2,774 (31%)	4.44 (4.08, 4.83)	1.89 (1.75, 2.05)
Receiving Food Stamps
No	4,653 (52%)	3.33 (3.07, 3.62)	1.45 (1.35, 1.56)
Yes	4,290 (48%)	4.53 (4.20, 4.87)	1.73 (1.62, 1.85)
Receiving SSI or SSDI
No	7,198 (80%)	3.15 (2.96, 3.35)	1.46 (1.38, 1.55)
Yes	1,745 (20%)	7.03 (6.46, 7.66)	2.11 (1.92, 2.32)
Receiving Housing Assistance/Emergency Shelter
No	6,296 (70%)	3.76 (3.52, 4.03)	1.55 (1.46, 1.65)
Yes	2,647 (30%)	4.24 (3.84, 4.68)	1.67 (1.53, 1.82)
Chronic Disability Payment Score (quartiles)
0 – 24% (least sick)	2,237 (25%)	1.24 (1.09, 1.43)	0.95 (0.85, 1.07)
25 – 49%	2,231 (25%)	2.23 (2.01, 2.47)	1.35 (1.23, 1.49)
50 – 74%	2,238 (25%)	3.51 (3.24, 3.81)	1.85 (1.71, 2.01)
75% and above (sickest)	2,237 (25%)	8.63 (8.19, 9.09)	2.19 (2.03, 2.36)

^¶ED Care Needed.

^βPrimary Care Treatable.

^*Missing data on at most less than 0.7% of the sample.

During the 2-year period, the study participants made 49,111 ED visits. Almost three-quarters of the ED visits were EDCN (71%). Figure 2 displays the distribution of PCT and EDCN visits by the four social risk classes. The distributions of both types of ED visits are skewed to higher numbers of visits. The average number of ED visits increases by more than 50% across the social risk classes (p < 0.001).

Figure 2.

Box Plots of Total Primary Care Treatable And ED Care Needed Visits Over A Two-Year Period By Social Risk Class.

Main Results.

When the total number of PCT or EDCN visits are modeled as a function of age, gender and social risk class, both types of ED visit rates increase significantly by social risk class (see table 2). For example, participants in social risk class 4 (i.e. unemployed and many social risks) have a 59% higher PCT log visit rate compared to participants in the lowest social risk class (i.e. employed and fewest social risks). When illness severity (i.e. CDPS) is included in the models, the ED log visit rates still significantly increase as a function of social risk class, although the effect is smaller compared to the models that do not adjust for illness severity. The adjusted log relative PCT and EDNC visit rates are significantly higher for social risk class 3 and social risk class 4 compared to social risk class 1 (i.e. employed and fewest social risks) for all of the models. In contrast, the adjusted log relative PCT and EDNC visit rates between social risk class 2 (i.e. employed and high financial strain) and 1 (i.e. employed and fewest social risks) are not significantly different for any of the models that include the CDPS.

Table 2.

Adjusted Log Relative Rate (95% CI) Of ED Visits*.

ED Visit	No CDPS	CDPS Only	CDPS, Diabetes	CDPS, Cardiovascular	CDPS, Renal	CDPS, Pulmonary	CDPS, Psychiatric	CDPS, Substance Abuse
A. Primary Care Treatable (PCT)
Class 2 (employed and high financial strain)	0.13 (0.05, 0.22)‡	0.07 (0, 0.16)	0.07 (0, 0.15)	0.07 (0, 0.15)	0.07 (0, 0.16)	0.07 (0, 0.16)	0.04 (0, 0.13)	0.06 (0, 0.14)
Class 3 (unemployed and limited internet and car access)	0.37 (0.30, 0.45)‡	0.24 (0.16, 0.32)‡	0.24 (0.16, 0.31)‡	0.23 (0.16, 0.31)‡	0.23 (0.15, 0.31)‡	0.23 (0.16, 0.31)‡	0.20 (0.12, 0.28)‡	0.18 (0.10, 0.26)‡
Class 4 (unemployed and many social risks)	0.59 (0.48, 0.70)‡	0.39 (0.28, 0.50)‡	0.38 (0.27, 0.49)‡	0.38 (0.27, 0.49)‡	0.38 (0.27, 0.49)‡	0.38 (0.27, 0.49)‡	0.31 (0.20, 0.42)‡	0.30 (0.19, 0.41)‡
Medical condition type	-	-	0.05 (0, 0.13)	0.05 (0, 0.13)	0.09 (0, 0.19)‡	0.20 (0.13, 0.26)‡	0.36 (0.29, 0.42)‡	0.44 (0.36, 0.51)‡
B. ED Care Needed (EDCN)
Class 2 (employed and high financial strain)	0.17 (0.09, 0.25)‡	0.04 (0, 0.11)	0.04 (0, 0.10)	0.04 (0, 0.10)	0.04 (0, 0.10)	0.04 (0, 0.11)	0.03 (0, 0.09)	0.03 (0, 0.10)
Class 3 (unemployed and limited internet and car access)	0.47 (0.40, 0.54)‡	0.12 (0.06, 0.19)‡	0.12 (0.06, 0.18)‡	0.13 (0.06, 0.19)‡	0.11 (0.05, 0.17)‡	0.12 (0.05,0.18)‡	0.12 (0.05, 0.18)‡	0.07 (0, 0.13)‡
Class 4 (unemployed and many social risks)	0.80 (0.70, 0.90)‡	0.29 (0.21, 0.38)‡	0.29 (0.20, 0.38)‡	0.30 (0.21, 0.39)‡	0.28 (0.20, 0.37)‡	0.28 (0.19, 0.37)‡	0.27 (0.18, 0.36)‡	0.22 (0.13, 0.30)‡
Medical condition type	-	-	0.16 (0.09, 0.22)‡	0.30 (0.24, 0.36)‡	0.29 (0.21, 0.36)‡	0.54 (0.49, 0.59)‡	0.37 (0.32, 0.43)‡	0.49 (0.43, 0.55)‡

All models are adjusted for age and gender and the CDPS except for the first set of models labeled “No CDPS”.

^‡The adjusted log relative rate is significant, the 95% CI does not include zero.

When we added the interaction term between social risk class and a specific disease to the condition-specific models, the interaction terms were not significant. Thus, table 2 only shows the adjusted log relative ED visit rates for the main effects of social risk class and disease condition.

DISCUSSION

This retrospective cohort study of approximately 9,000 adult Medicaid beneficiaries revealed a strong association between SDH and ED use. Participants in higher social risk groups visited the ED more frequently, on average, relative to those in the lowest social risk class. This pattern was reliable regardless of type of ED visit. The influence of SDH on the number of ED visits was consistent across the medical conditions we examined. After adjusting for illness severity, social risk class remained significantly associated with ED utilization.

The strong association we observed between SDH and ED use is consistent with prior research that has examined this relationship at the individual level. For example, Hunt et al found that the odds of visiting the ED frequently (i.e. at least four times in a year) increased as income decreased.¹⁸ McConville et al. found that patients who were involved in the criminal justice system at the time of the ED visit (i.e. admitted from jail/prison to ED or discharged to jail/prison from ED) were significantly more likely to be a frequent ED user.¹⁹ Homelessness is another SDH associated with frequent ED use.^20,21 Finally, Balakrishnan and colleagues found that patients with limited health literacy had more than twice as many preventable ED visits in a two-year period compared to patients with adequate health literacy.²² The findings from our study are unique in that rather than focus on a specific SDH, we show the influence of a constellation of structural and intermediary determinants of health on ED utilization.

Interestingly, the relationship between SDH and ED visits did not vary significantly across the physical and behavioral conditions we examined. This may be because we examined SDH in aggregate, using a large set of SDH factors, rather than evaluating the influence of a specific SDH on different disease conditions. For example, it could be that food insecurity has a much stronger influence on ED visit rates for patients with diabetes compared to patients with psychiatric disease. This notion is supported by recent research designed to address a specific adverse SDH and patient population (e.g. food insecurity and diabetes) to improve health outcomes.^23,24

We suspect that the impact of SDH on healthcare use, including emergency health services, is by shaping the circumstances in which people manage their health problems. Social risks, such as eviction or food insecurity, compete with health problems. Patients who are dealing with stressful circumstances and chronically have inadequate resources may access medical care in the ED because EDs are open 24/7, they do not require an appointment and they provide a wide range of services.^25–27 Given that EDs treat a disproportionate share of patients with social vulnerabilities, identifying ways to address them in the ED is an important area of future research.^28,29 There is precedence for ED providers developing successful ways to support patients’ ability to engage and adhere to treatment despite adverse social circumstances.^30,31

The results of this study must be considered in the context of the following limitations. First, the generalizability of the study results are limited by several factors. Our participants enrolled in the study at the time of a healthcare visit, the majority of enrollment occurred at the time of an ED visit. The RAs did not enroll patients in the ED during the night hours. We excluded children, dual eligible, and non-English speaking Medicaid beneficiaries. Lastly, the DC Medicaid program is one of the most generous in the country in terms of access and coverage of services. We suspect that if we had enrolled a more representative sample of Medicaid beneficiaries that included those who did not use healthcare services at the time of enrollment or used different health care services, the relative rates of ED use by social risk class would be even larger.

Second, while we measured a wide array of social factors, we did not measure all SDH that may be important to healthcare utilization (i.e. health literacy, social support, discrimination, etc). Third, there is no widely accepted and validated SDH assessment tool. The majority of our questions came from validated scales and surveys (e.g. education, food insecurity, housing instability, etc), but not all of them (e.g. alcohol use, exercise). Fourth, in this analysis, SDH were measured at a single point in time and ED use was measured over a two-year period. This could lead to biased estimates of the relationship if the SDH changes were consistently worse or better than the single point-in-time measurements that we used. Fifth, we used the algorithm identified by Venkatesh et al to identify ED visits and the Minnesota algorithm to classify them as PCT versus EDCN. While these methods may have resulted in some misclassification it is not likely have biased the results. Sixth, the LCA assigns individuals to a class based on probabilities so proper class assignment is not guaranteed. Finally, we did not have adequate power to detect the smaller differences (i.e. 3% – 7%) in the adjusted log relative ED visit rates between social risk classes one and two.

In conclusion, within a cohort of Medicaid beneficiaries, all of whom are low-income, we observed a SDH gradient in ED use. This study adds to a large body of compelling evidence demonstrating the powerful and insidious role that SDH plays in a wide range of health indicators, including ED utilization.