Data
Emergency Room (E.R.) visit public-use data were obtained from the California Office of Statewide Health Planning and Development (OSHPD) for community-based hospitals, excluding federal hospitals. There was one row of data for each of the 97,639,808 million ER visits between January 1, 2005 and December 31, 2014. Patient-level data obtained included year treated, expected source of payment, primary diagnosis, up to 24 secondary diagnoses, and disposition of discharge [20]. Through 2011 the public-use data had included patient gender, race, ethnicity, age category, and five-digit ZIP Code of residence. From 2012 to 2014 public-use data included only the five-digit ZIP Code of residence. Since the OSHPD data did not have reliable patient ZIP code data, area-based socio-demographic data at the hospital level were obtained. This included those in poverty and college graduates among adults 25 years of age and greater at the Congressional District level [21] and population density and percent non-Hispanic white at the Medical Service Study Area [22].

OSHPD provides hospital-level data via annual files which include information such as bed size in addition to hospital-level summary patient demographic data and geographic identifiers for each hospital, such as the city, ZIP Code, county, Congressional district, and Medical Service Study Area in which the hospital is located [23]. There were 357 hospitals that provided data during the study period.

Subjects
All ER visits were examined, with a focus on those having an International Classification of Diseases, Ninth Edition (ICD-9) diagnosis of any diabetes mellitus (250.x).

Measures
Mental illness was measured using the Clinical Classification Software (CCS) schema [24] developed by the Agency for Healthcare Research and Quality. CCS categorizes thousands of individual ICD-9 codes into 270 mutually exclusive and clinically meaningful categories schema [24]. All 25 diagnosis fields within the ER visit data were searched for diagnoses meeting the definition of the 14 CCS categories for mental illness and substance abuse. Hospital-level measures included bed size, academic affiliation, rural/urban status, and ownership type. Area-based measures included percent of population in poverty, college graduates among adults 25 years of age and greater, population density and percent non-Hispanic white. Patient-level measures included age group, gender, race and ethnicity. The primary outcome measure was non-routine discharge. This primarily includes hospitalization and transfer to another facility; but also includes leaving against medical advice and mortality.

Analysis
Analysis was guided by Andersen’s Behavioral Model of Health Services Utilization [25]. The conceptual model considers the effects of individual characteristics (predisposing, enabling, and need) on access to healthcare and also considers contextual effects, that is, factors related to the environment and health care system. Mental illness and substance abuse were considered as a personal predisposing – vulnerable domain measure [26].

A significant issue arose due to partially masked demographic data from 2005 to 2011 and completely missing demographic data from 2012 to 2014. A Markov Chain Monte Carlo method [27] was used to impute missing values of age, gender, race, and ethnicity, as has been done in other analyses involving public-use inpatient data [28,29]. Multiple imputations were done using known patient gender, age, race, and ethnicity data, and patient payer source. Also used in imputation were hospital-level summary measures such as the percent of discharges which were female, white, age less than 18 years, age greater than 60 years, and percent Spanish-speaking. ICD-9 codes were used to identify gender and age-specific visits, such as female, e.g., pregnancy-related codes accounted for 5% of all visits, male reproductive-related codes accounted for 1% of all visits, and newborn/infant-related for 0.2% of visits). Of the 97.6 million records, there was substantial masked/missing data for demographics: gender (37.75% missing data), age group (36.62%), race (45.94%), and ethnicity (48.26%).  Because expectation-maximization works best with continuous measures, race was imputed as white/non-white rather than as multiple categories and age as greater/less than 45 years of age. All imputed demographic measures were used as a continuous variable ranging from 0 to 1.

The entire 97 million record dataset was used to examined trends. A figure was created to show the annual percentage of ER visits having a mental diagnosis, both for cases having a diabetes diagnosis and those not having a diabetes diagnosis. Remaining analysis was limited to cases having a diabetes diagnosis. In order to determine population-based rates of ER visits, annual population estimates were obtained for 2005–2009 [30] and 2010-2014 [31]. Rates of visits per 100,000 population were calculated for visits with any diabetes diagnosis as well as rates for visits in which diabetes was the primary diagnosis.

Differences in patient, hospital, and area-based summary characteristics were first examined comparing those which had diabetes as a primary diagnosis compared to diabetes as a secondary diagnosis. Differences were assessed using the chi-square test for categorical variables and t-test for continuous variables. Next, multivariable logistic regression was used to examine the joint effect of all variables on discharge status, non-routine vs. routine. Data processing and analysis were performed using SAS 9.4 software (SAS Institute, Gary, North Carolina). This research was considered exempt from Institutional Review Board review at Loma Linda University due to being analysis of public-use data.