Sample

We analyzed cross-sectional data from the population-based community survey of the Swiss Spinal Cord Injury Cohort Study (SwiSCI) [24]. The survey was conducted between late 2011 and early 2013 and included Swiss residents with a traumatic or non-traumatic SCI aged over 16 years. Persons with congenital conditions leading to SCI, with new SCI in the context of palliative care, with neurodegenerative disorders, and with Guillain-Barré syndrome were excluded from the study. The SwiSCI population was recruited through the national association for persons with SCI (Swiss Paraplegic Association), three specialized SCI-rehabilitation centers, and a SCI-specific home care institution [24]. The survey featured written or online questionnaires, and in special cases, telephone interviews. This study has been approved by the Medical Ethical Committee of the Canton Lucerne, Switzerland. In addition, the study protocol has been approved by the Steering Committee of the SwiSCI study. All participants have signed a written consent form.

Measures

Socioeconomic circumstances: Level of education, net equivalence household income and perceived financial hardship were chosen as three indicators of individual-level socioeconomic circumstances. Education was classified according to the International Standard Classification of Education as total years of formal education, combining school and vocational training [25]. For bivariate analyses, education was regrouped in four categories according to the Swiss National Cohort [26]. In multivariable models, education was introduced continuously. Income was measured by net equivalent household income in Swiss Francs, including information on disposable income, household size and number of adults and children according to the Organisation for Economic Co-operation and Development (OECD) criteria [27]. Income was reclassified into distribution-based quartiles for bivariate analysis and entered into multivariable models as continuous variable. Income was divided by 1000 in order to receive legible effect sizes in multivariable models. Financial hardship was assessed by the question ‘Did you experience financial difficulties that restricted your everyday life (participation) over the past four weeks?’. Answer categories were ‘not applicable’, ‘had no impact’, ‘has complicated my life somewhat’, and ‘has complicated my life massively’. By combining the first two categories, a three-categorical variable was computed.

Health conditions & body functions: To assess the prevalence of health problems in addition to SCI, a sum score of four dichotomous items on comorbidities (diabetes, heart disease, cancer, depression) based on the Self-Administered Comorbidity Questionnaire [28] was computed (range 0–4, higher scores indicating more comorbidities). Due to a low prevalence of persons indicating more than two comorbidities, the variable was classified into ‘having no comorbidities’, ‘having one comorbidity’, and ‘having two or more comorbidities’ for multivariable analysis. Information on the prevalence and severity of secondary conditions during the past three months was collected using the Spinal Cord Injury Secondary Conditions Scale (SCI-SCS) [29]. The SCI-SCS measures 16 SCI-relevant conditions by a 4-point Likert scale (2 items on pain and 1 item on diabetes were not assessed by SCI-SCS in this study as respective information was gathered in other questions). The SCI-SCS therefore contains 13 items and ranges from 0–39, higher scores indicating higher prevalence and severity. Referring to ICF Domains [21], 4 items are classified as health conditions (contractures, urinary tract infections, injury caused by loss of sensation, pressure sores), 6 items as body functions (sexual dysfunction, spasticity, bladder dysfunction, bowel dysfunction, autonomic dysreflexia, postural hypertension), and 3 items can be assigned to both categories (circulatory problems, respiratory problems, heterotopic ossification). Secondary conditions were classified into distribution-based quartiles (0–5; 6–10; 11–14; ≥15) for multivariable analysis (see section on Statistical analysis). Pain intensity (body function) was assessed using a ten-point Likert-scale [30]. This information has been categorized into ‘no pain’, ‘mild (1–3)’, ‘moderate (4–6)’ and ‘severe (>6)’ pain [31] for multivariable analysis. To measure mental health (body function), we used the 5-item Mental Health Index of the 36-item Short Form Health Survey (SF-36) and computed a sum score ranging from 0-100 with higher scores indicating better mental health [32]. For multivariable analysis, distribution-based quartiles were computed (0–60; 61–76; 77–84; ≥84).

Activity & participation was measured using the 11-item subscale on participation restrictions of the Utrecht Scale for Evaluation of Rehabilitation-Participation (USER-participation) [33]–[35] which employs a 4-point Likert type rating scale. A sum score of the ratings of items that were applicable to a person was computed and converted into a score ranging from 0–100. The higher the score, the lower the participation restrictions [33]. The score has been categorized into distribution-based quartiles for multivariable analysis (0–54.9; 55–72.9; 73–87.9; ≥88).

Quality of life was assessed by a single item rated on a 5-point Likert-scale ranging from 0 (‘very poor’) to 4 (‘very good’) [36]. Due to a low prevalence of ‘very poor’ quality of life, the categories ‘very poor’ and ‘poor’ were combined for multivariable analysis.

Control variables: Age, gender, lesion characteristics (years since injury, para- vs. tetraplegia, complete vs. incomplete lesion), aetiology (traumatic vs. non-traumatic), and social support were included as potential confounders in multivariable analyses. Severity of disability had been linked to diverse health indicators [37] and might be associated with downward social mobility, leaving those with higher lesion level and complete lesions at higher risk. Aetiology is an important potential confounder in associations with social inequalities of health as persons with traumatic SCI are privileged in the Swiss insurance system. Given its relevance for health inequalities [38], the degree of social support of SCI persons was introduced as confounder in multivariable models. Social support was assessed by asking about having a partner (‘yes’/’no’) and living arrangement (‘living alone’/’living with others’). Poor social support was coded if the two answers were negative.

Statistical analyses

Following descriptive analysis of the study population, we explored unadjusted bivariate associations between socioeconomic circumstances and health indicators. Given the fact that dependent variables were not normally distributed, we used a nonparametric ANOVA to evaluate differences between groups (Kruskal-Wallis) and a test for trend across ordered groups according to Cuzick [39].

Next, a set of regression analyses was applied to evaluate associations. As effect sizes of ordinal regressions are easier to interpret as compared to models for continuous dependent variables, ordinal logistic regression models were applied for all outcomes. Ordinal logistic regressions are extensions of logistic regressions [40], but allow to examine the dependence of a polytomous ordinal outcome on several independent variables without losing information through outcome dichotomization. As a prerequisite to apply ordinal logistic regressions, the parallel lines assumption indicating that betas are the same for each transition from an ordinal scale point must be confirmed [41]. This assumption, tested by Stata's gologit2 command and its autofit option, was confirmed in all cases. For education and income, two subsequent series of ordinal logistic models were computed. The first models were adjusted for age, gender, lesion characteristics, aetiology, social support, income and education. To assess the impact of financial hardship on the associations between education, income and health indicators, financial hardship was introduced separately in a second model. For financial hardship, only one model (adjusted for all control variables) was computed. Control variables were mean centered in all analyses.

In the main data analyses, we adjusted for both unit and item nonresponse [42]. Unit nonresponse, which refers to the complete absence of a response by an eligible person of the initial survey population, was 50.7%. Unit nonresponse was accounted for by using inverse probability weights (IPWs) as sampling weights in regression analyses. IPWs were derived from propensity scores in multivariable-adjusted logistic regression analyses. The propensity to participate in the survey was higher for members of the Swiss Paraplegic Association as compared to non-members; convexly shaped with increasing age (highest at age 52); concavely shaped with time since injury (least at 16 years post SCI); but not related to other variables including age, gender, lesion level (paraplegia vs. tetraplegia), and language (including German, French, and Italian). The average IPW used in weighted analyses was 2.03 (range: 1.02–6.65) [42].

Item nonresponse, which refers to the failure of survey respondents to answer a specific question, was addressed using multiple imputation (MI). We used MI by chained equations (MICE)[43] to impute different types of variables, including categorical, ordinal and linear variables. We imputed predictors (socioeconomic circumstances) and control variables, but not outcomes (health indicators). For each model, imputations were carried out for 10 datasets (see Appendix S1 for more detail on MI modeling). In the respective table, odds ratios, 95% confidence intervals, and p values from the conditional (equal fraction-missing-information, FMI) test are presented [44]. The equal FMI test is a likelihood ratio test suitable for multiply imputed datasets.

To assess the risk of bias due to list-wise exclusion of cases with missing data in health indicators, a sensitivity analysis was performed comparing four scenarios:

1. imputed data for socioeconomic circumstances/controls, but only full cases in health indicators (as displayed in the main results),
2. only full cases in all variables,
3. imputed data for socioeconomic circumstances/controls and replacing missings in health indicators by ‘best case’ (i.e. absence of complication, highest quality of life),
4. imputed data for socioeconomic circumstances/controls and replacing missings in health indicators by ‘worst case’ (i.e. presence of comorbidity, lowest quality of life).

Results of the sensitivity analysis are displayed in Table S1 in the Supporting Information. Our findings from scenario 1 were robust and mainly confirmed by the sensitivity analysis. Analyses were conducted using STATA Version 13 for Windows (College Station, TX, USA).

Strengths and limitations

This study has several limitations. First, findings are based on cross-sectional data, which does not allow to draw conclusions on causal associations. Second, our sample may not be fully representative of persons with an SCI living in Switzerland. Although several SCI-clinics and the national association for persons with SCI have been involved in recruitment [24], the exact number of persons with an SCI living in Switzerland is unknown. Moreover, the robustness of reported findings needs further exploration and the clinical significance of observed differences according to socioeconomic circumstances needs to be elaborated.

Despite these limitations, our study has several strengths. The study population is based on a large community sample rather than on some selected group, and data meet high quality standards. We analyse a comprehensive set of health indicators thus drawing a detailed picture on health problems, functioning and quality of life in persons with SCI. Moreover, we conducted additional sensitivity analyses to correct for item and unit nonresponse. Given the small number of aberrations (see Table S1), we conclude that our findings are robust against bias due to missing values and nonresponse.