Study population
We included the first ICU admissions of patients aged 30 years and older registered in the Swedish Intensive Care Registry (SIR) [13] from 2009 to 2012 (eFig S1). The age restriction aimed to exclude younger age groups whose SES may be more dependent on their parents/family and whose own educational/income level does not accurately reflect their SES. The regional Human Ethics Committee approved the study (Approval no 2012/197).
Data sources
In 2012, SIR covered 92% of all ICU admissions in Sweden. SIR contains information on, e.g., the characteristics of patients admitted to an ICU, reasons for admission, and severity scores indicating baseline risk. Using unique patient identity numbers, SIR was linked to the other national healthcare registers [14]. Individual-level information on demographics, level of education, and income were obtained from the Statistics Sweden and the Longitudinal integrated database for health insurance and labour market studies (LISA) [15]. Hospital discharge diagnoses coded according to the Swedish clinical modification of the 10th revision of the International Statistical Classification of Diseases and Related Health Problems (ICD-10-SE) were retrieved from the Patient Register [16]. Reporting to this register is mandatory and has nationwide coverage regarding in-patient care since 1987. The date of death was retrieved from the Cause of Death Register [17].
Socioeconomic status
Educational level was categorized as low if restricted to compulsory school (9 years), medium if restricted to upper secondary school (additional 2–4 years), and high if it included education beyond that level (university). The educational level was defined from information reported in the calendar year preceding the date of admission to intensive care.
The patient’s annual disposable income was stratified into four arbitrary categories (< 10 000 €, 10 000–29 999 €, 30 000–49 999 €, and ≥50 000 €). It was defined from the income reported in the calendar year preceding the calendar year for admission to intensive care not to be affected by the acute condition causing ICU admission. In a sensitivity analysis, the income variable was defined from the income one year earlier than in the primary analysis.
Comorbidity and other baseline characteristics
We used a previously described and validated method to quantify baseline comorbidity [12]. Hospital discharge diagnoses from in-patient care during the five years preceding the index date for the ICU admission were divided into 36 predefined comorbidity categories. For each comorbidity category, we calculated two variables: (a) one quantifying the total length of hospital stay in days with a primary diagnosis within that category, and (b) the interval from the last hospital admission with a primary diagnosis within that category to the ICU admission date (1–6 months, 6–12 months, 1–3 years, > 3 years). The month preceding the index date was considered a grace period to reduce the inclusion of diagnoses directly related to the current hospitalization and ICU stay and not representing preexisting comorbidity. The described prevalence of comorbidity at baseline was based on both main and secondary hospital discharge diagnoses during the five years preceding the index date.
The simplified acute physiology score (SAPS) version 3 was extracted from SIR [18, 19]
Outcomes
The date of death was retrieved from the Cause of Death Register. Follow-up for all-cause mortality was available until December 31, 2016.
Statistical methods
Follow-up started on the day of ICU admission and ended on the date of death or December 31, 2016, whichever came first. The length of follow-up was calculated from a ‘reverse’ Kaplan-Meier analysis.
First, a Cox model was fitted for the association between all 72 comorbidity variables (two variables for each of the 36 comorbidity categories, quantifying the length of stay and the recency of previous hospital admissions within each category) and survival probability. This model was then applied to the study population, and the estimated linear predictor was used to provide a summary comorbidity score for each individual’s overall baseline comorbidity burden (eFig. S2).
In the regression models including the SAPS3 score, missing SAPS3 values were imputed (5 times) using multiple imputations [20]. The associations between education/income as exposures and survival probability were then estimated in Cox proportional hazard regression models, gradually adjusted for age, sex, SAPS3, and the summary comorbidity score. The continuous variables age, SAPS3 and the summary comorbidity score were modeled as restricted cubic splines with 5 knots. The results were presented as hazard ratios (HR) with 95% confidence intervals (CIs). A secondary analysis was restricted to complete cases.
Subgroup analyses were performed based on age, sex, country of birth, ICU type, and overall burden of comorbidity. In a complementary analysis a landmark was predefined to one year.
The proportional hazards assumption was evaluated by visual inspection of a log-log plot. Since age and some comorbidities are included as components in the SAPS3 score, potential collinearity was evaluated by calculating variance inflation factors (VIF) using linear regression. Data management was done in SAS version 9.4, and all statistical analyses were performed using R version 4.4.0.