Population

We conducted a population-based cohort study of all adults ≥18 years of age, alive and residing in Denmark at least 1 day between 1 January 2005 and 31 December 2013. The cohort was constructed using data from the Danish Civil Registration System,22 which includes data on sex, date of birth, vital status and emigration since 1 January 1968. In the register, Danish residents are each assigned a unique personal identification number which links to person-level data.22

Primary independent variable

Our primary independent variable of interest was depression as identified by either psychiatric diagnosis or filling at least one antidepressant prescription. Depression was treated as a time-dependent variable (ie, an individual without a recorded depression diagnosis or antidepressant prescription redemption at baseline could be diagnosed with depression or redeem an antidepressant prescription during the follow-up period, moving from the ‘unexposed’ to the ‘exposed’ group). Information on psychiatric diagnoses was obtained from the Danish Psychiatric Central Register23 (see online supplementary appendix 1), which includes diagnostic information on all psychiatric hospitalisations from 1969 onwards and outpatient specialty mental health visits from 1995 onwards.23 Prescription fills for antidepressants (ie, selective serotonin re-uptake inhibitors, monoamine oxidase inhibitors, and other non-tricyclic (TCA) antidepressants, see online supplementary appendix 1) were identified using the Danish National Prescription Registry.24 This register includes data on all prescriptions dispensed at Danish pharmacies since 1995, including purchase date and classification of drugs according to the Anatomical Therapeutic Chemical Classification.25 We excluded TCA prescriptions from our depression definition because of their frequent use for insomnia and/or pain. We also excluded bupropion or trazodone prescriptions since neither was approved for treating depression in Denmark during the study period. Individuals with schizophrenia, schizoaffective disorders or bipolar disorder were censored at date of diagnosis (see online supplementary appendix 2) and excluded from analyses.

Outcomes of interest

Our primary outcome of interest was hospitalisation for 1 of 12 ACSCs as defined by the Agency for Healthcare Research and Quality (AHRQ) (see online supplementary appendix 3).2 Register-based diagnoses were based on the Danish version of the International Classification of Diseases, 8th Revision (ICD-8) prior to 31 December 1993.26 From 1 January 1994 onwards, the Danish version of the ICD-1027 was used. Since the AHRQ-defined ACSCs were originally derived using ICD-9 diagnoses, we included eight AHRQ-defined ACSCs (ie, angina without concomitant cardiovascular procedures, COPD exacerbation, CHF exacerbation, diabetes with short-term complications, diabetes with long-term complications, uncontrolled diabetes, hypertension (HTN) and appendicitis with perforation) that were translated into ICD-10 diagnosis codes and validated in a previous study.28 We also included four AHRQ-defined ACSCs (ie, bacterial pneumonia, diabetes-related lower extremity amputations, UTIs and adult asthma exacerbations) based on ICD-10 codes used in prior Danish register-based studies.29–32 We further divided ACSCs into five ‘chronic’ ACSCs (ie, angina, CHF exacerbation, HTN, diabetes related, COPD/adult asthma exacerbation) and three ‘acute’ ACSCs (ie, appendicitis with perforation, pneumonia and UTI). We used the Danish National Patient Register,33 which contains information on all medical hospitalisations since 1 January 1977 and outpatient visits since 1 January 1995,33 to obtain information on hospitalisations with principal discharge diagnoses for ACSCs occurring between 1 January 2005 and 31 December 2013. If a discharge was followed by an admission within 1 day, it was considered a transfer and counted as one admission only. We excluded hospitalisations with secondary obstetric diagnoses (ICD-10 codes: O00.0-O99.9).

Our secondary outcome of interest was rehospitalisation for an ACSC within 30 days of discharge from the initial ACSC-related hospitalisation. We counted rehospitalisations that were for the same ACSC, or for a different ACSC, using data from the Danish National Patient Register.

Socioeconomic factors, comorbid medical conditions and substance abuse disorders

Information on marital/partnered status and education was obtained from Statistics Denmark and the Danish Educational Registers, respectively (see online supplementary appendix 4).34 ,35 We categorised marital/partnered status as living with a partner (ie, married, registered partnership or cohabitation) or living alone (ie, living without a partner, including widows/widowers). We classified maximum educational level attained into the following three categories based on the United Nations Educational, Scientific and Cultural Organisation's International Standard Classification of Education: low (<10 years), middle (10–15 years) and high (>15 years).36

For the five chronic ACSCs, we defined ACSC-predisposing medical comorbidity specific for each ACSC in question (see online supplementary appendix 5). Information on ACSC-predisposing medical comorbidity and non-ACSC predisposing medical comorbidity was obtained from the Danish National Patient Register and based on Charlson comorbidity index (CCI) categories37 (see online supplementary appendix 6) (eg, myocardial infarction as ACSC-predisposing medical comorbidity for angina hospitalisation, etc), with two exceptions. Diabetes diagnoses were obtained from the Danish National Diabetes Register between 1 January 1990 and 31 December 2013 (see online supplementary appendix 7).38 Chronic pulmonary disease was identified as either a diagnosis based on the CCI category obtained from the Danish National Patient Register or ≥2 prescription redemptions within a 6-month period for medications treating obstructive airway diseases (see online supplementary appendix 8) as obtained from the Danish National Prescription Registry. Non-ACSC predisposing medical comorbidity included all remaining CCI diagnostic categories. We did not define ACSC-predisposing medical comorbidity for the three acute ACSCs.

Data on substance abuse (excluding tobacco abuse) was obtained from the Danish Psychiatric Central Register or the Danish National Patient Register (see online supplementary appendix 9).

Primary care utilisation

We obtained information on daytime face-to-face visits with primary care physicians (PCPs) or other primary care staff from the Danish National Health Service Register,39 which has been collecting primary care administrative data since 1 January 1990. To reduce the chances of including a primary care visit that directly resulted in an ACSC-related hospitalisation, we constructed a time-dependent variable counting the number of primary care visits from 10 to 375 days before any given day. We categorised primary care visits into three equally sized categories of low, medium or high utilisation based on observed frequencies (ie, 0–2, 3–9 or ≥10 visits).

Statistical analysis

We compared individuals with depression to those without depression using Poisson regression models in order to estimate incidence rate ratios (IRRs) of hospitalisations for ACSCs and subsequent rehospitalisation within 30 days for an ACSC. We estimated corresponding 95% CIs using cluster robust variance estimation to account for interperson correlation and dichotomy of rehospitalisation. In these analyses, our outcomes of interest were a count of the number of hospitalisations for ACSCs. Age and calendar period were adjusted for using 2-year and 1-year age and time bands, respectively. All variables (including depression status), except sex, were treated as time-dependent. Individuals contributed at-risk time from 1 January 2005 or from their 18th birthday, whichever came last, in different time bands based on the different covariate combinations they enter with during follow-up. Within each of these combinations, we counted the number of ACSC-related hospitalisations. These methods allowed us to count only ACSC-related hospitalisations that occurred after registration of a depression diagnosis and/or redemption of an antidepressant prescription. Censoring occurred at date of death, emigration, date of bipolar disorder or schizophrenia diagnosis, or on 31 December 2013, whichever came first.

For each ACSC-related hospitalisation outcome, we fitted five risk models, adjusting sequentially for demographics (ie, age, sex and calendar period), socioeconomic factors (ie, marital/partnered status and education), ACSC-predisposing medical comorbidity (with each comorbid condition entered individually), other comorbidities (ie, non-ACSC-predisposing medical comorbidity entered individually and substance abuse) and primary care utilisation. All model covariates were chosen a priori based on prior studies identifying their potential associations with both depression and healthcare utilisation outcomes.12 ,16 ,17 ,40 To address missing data on education, we conducted multiple imputation using five imputed data sets according to methods developed by Rubin.41

We performed two pre-specified subanalyses. First, we examined whether the association between depression and risk of ACSC-related hospitalisations was modified by age. To do so, we repeated our Poisson regressions stratified by three age categories: ≤40, 41–64 and ≥65 years. Second, we examined the associated risk of hospitalisations for chronic and acute ACSCs based on time since depression diagnosis in models adjusted for demographics.

In order to determine if an association between depression and risk of hospitalisations for ACSCs was impacted by our depression definition, we performed a prespecified sensitivity analysis in which we repeated our regressions using three different depression definitions: antidepressant prescription alone, outpatient psychiatric visit-based diagnosis alone or psychiatric hospitalisation for depression.

We fitted three models examining risk of rehospitalisation within 30 days for an ACSC. The first model was adjusted for demographics, the second included adjustment for socioeconomic factors and the third for medical and substance abuse comorbidities. Our outcome of interest in these models was time to rehospitalisation for an ACSC within 30 days of discharge from the initial ACSC-related hospitalisation. Individuals were at risk of the outcome on the day of discharge from their ACSC-related hospitalisation. All variables in these analyses excluding sex were treated as time-dependent.

We used two-sided significance tests for all analyses with statistical significance set at p<0.05. Analyses were performed using STATA V.13 (Stata Corporation, College Station, Texas, USA).