Social determinants of duration of last nursing home stay at the end of life in Switzerland: a retrospective cohort study

We extracted data from three different sources covering basically all individuals living in Switzerland:

All three data sources encompass the entire population and are fully anonymized. For reasons of data protection and confidentiality there is no common personal identification number that would allow to directly link the three data sources from hospitals (MedStat), nursing homes (SOMED), and the Swiss National Cohort (SNC) on an individual level. An anonymous common person identifier is only available for two of the three data sources (MedStat and SOMED). The linkage with the third source (SNC) had to rely on common identification variables such as place of residence, date of birth and date of death. As a consequence, among nursing home residents and hospitalized people the linkage to the SNC was only possible for those who died in an institution. The detailed linkage process is described elsewhere [5]. We restricted to individuals ≥65 years old at the admission (i.e., born before 1942) and who had died in 2007 and 2008 in a nursing home, resulting in a study population of N = 11,486 men and N = 24,253 women. Ethics approval for this study was given by the Ethics Committee of the Canton of Zurich.

Our outcome variable was the number of days of the last nursing home stay before death. We stratified our analysis by gender because there are fundamental differences between men and women regarding institutional stays before death [8, 23‐25].

The independent variables were grouped into individual, familial/housing and structural/regional attributes. As control variables we included age (at the time of admission) and nationality (Swiss or foreign) as well as cause of death: malignant neoplasms (ICD 10: C00-C99), coronary heart disease (I20-I25), stroke (I60-I69), chronic obstructive pulmonary disease (COPD, J40-J47), dementia (F01, F03, G30) and all other causes combined. The care level was assessed at time of admission and can be grouped into four categories: unknown or no care needed, low (max. 40 min per day), medium (between 41 and 80 min per day) and high (more than 81 min per day). Please note that Switzerland has no homogenous care level classification for the whole country and respective information was not always complete. However, we had no alternative health variable for nursing home residents. From MedStat we derived information about multimorbidity (2+ chronic conditions), assessed from inpatient diagnoses 2–6 years before death. We defined chronic conditions using ICPC-2, of which 129 rubrics were classified as chronic conditions [26]. Additionally, we included a dummy to control for hospitalizations in the last 365 days preceding death. We used specific time windows in order to test the different impact of health indicators which are more close or more distant from death. For better understanding of our health variables, Fig. 1 presents examples of two persons with the corresponding time windows. From the 2000 census we derived the educational level according to the International Standard Classification of Education (ISCED), version 1997: no or low secondary education completed (ISCED 0–2), post-secondary non-tertiary (“medium”, ISCED 3–4), and tertiary education (“high”, ISCED 5). Also from the 2000 census we extracted information on homeownership (owner-occupier household yes vs. no) and having had children (assessed on an individual level for men and women, i.e., not necessarily the same for all couples). Marital status was assessed at the time of home admission (never married, married, widowed and divorced). Place of residence was categorized into the three main language areas of Switzerland, namely the German, French and Italian speaking parts. In order to account for geographical variation in nursing home bed availability, we included a variable with the density of nursing home beds per 100 inhabitants aged 65 years or older in 2010 on the level of 106 quite homogenous regions.

For descriptive analysis, we calculated means, frequencies and proportions of the respective variables. To point out the right-skewed distribution, we report mean, median, SD and 25th and 75th percentile of our outcome. The distribution of the outcome variable can be interpreted as a count variable. Therefore, a count regression model such as Poisson or negative binomial model is appropriate. Because of a strong overdispersion in the outcome variable, we preferred the negative binomial model over the Poisson model. Separate negative binomial regression models for both men and women, yielding estimated incident rate ratios (IRR) [27] were used to assess the impact of several independent variables on the number of days in a nursing home. For better understanding of the interpretation of IRR, we give a reading example using the estimated IRR of 1.17 for a men with “low” educational level (cf. Table 2): Compared to a man with “medium” educational level (reference category), the expected length of the last nursing home stay for a man with “low” educational level is 1.17 times higher. Negative binomial regression was used instead of Poisson regression to account for overdispersion in the data. In order to obtain global p-values for each variable in the model, we performed likelihood ratio tests. The level of significance was set to α = 0.05 (two-sided tests).

For sensitivity analysis, we also tested if some potentially relevant interaction terms (age and multimorbidity, age and cause of death, age and hospitalizations in the last 365 days preceding death) improved the model which was not the case. However, neither of those interaction terms improved the model with respect to the Bayesian information criterion (BIC). For this reason, they were omitted in the final model.