Second-stage non-response in the Swiss health survey: determinants and bias in outcomes

The study is a methodologically-focused secondary analysis of a population-based, cross-sectional health survey carried out in 2007 in Switzerland.

The Swiss Health Survey carried out by the Swiss Federal Statistical Office (SFSO) is a nationwide survey on health status, health service utilization, and health-related behavior in Switzerland. The SHS was first conducted in 1992 and is repeated every five years. In 2007, a multistage probability sample was drawn of the entire Swiss population after stratification by geographic region. The initial sample (n=28,319) included subjects aged 15 years or older living in private households. The sampling frame excluded people living in institutions, i.e. hospitals, homes for the elderly, prisons, monasteries, barracks, etc. Furthermore, only subjects fluent in German, French or Italian were included. The response rate of those in the initial sample was 66.2% (n=18’760); 13.2% (n=3^′730) could not be contacted after numerous contact attempts, and 20.6% (n=5’825) refused to participate in the SHS. Of the latter, 445 subjects refused to participate due to language problems (1.6% of the initial sample).

Three different types of interviews were conducted. Generally, computer-aided telephone interviews (CATI) were conducted for subjects aged 15–74 years (n=17,856). Subjects aged 75 years or older were given the choice between a CATI or a face to face interview. For those opting for face to face interviews, computer-aided personal interviews (CAPI) were conducted (n=75). Finally, CATI were conducted with proxy persons (PROXY, n=829), that is close friends or relatives of the target subject were interviewed if the target subject could not be interviewed because of old-age, illness, handicap, language problems or due to absence from home for more than 4 weeks. At the end of the forty-minute interview, CATI and CAPI participants (n=17,931) were invited to complete and return an additional written survey (MSHS). PROXY participants were not eligible to participate in the MSHS. Of those eligible to participate, 97.7% (n=17’511) consented to return the MSHS and the questionnaires were sent out to those willing to participate in the next 2–3 days following the SHS interview. The mail-back questionnaire which included questions on mental stress, mental illness, chronic diseases, vaccination status, social support, health literacy, health insurance coverage etc., was returned by 80.3% (n=14,393) of the 17,931 CATI & CAPI interviewees. Subjects who did not return the questionnaire were not re-contacted.

In order to prevent the loss of information and to prevent potential bias due to selective refusal to provide answers, item non-response was addressed by the statistical technique of multiple imputations (MI) and multivariate normal regression was used as the imputation method to estimate missing values [35‐38]. In contrast to other imputation procedures, MI is more likely to yield valid statistical inference than other methods of imputation because its simulation-based procedures treat imputed values as estimates with a random component rather than treating the imputed values as known with certainty. For SHS variables, the percentage of imputed values ranged from 0.04% to 5.64%. For MSHS variables, between 1.74% and 9.73% of the values were imputed (Table 1). To obtain estimates based on the pooled set of imputation results from each completed dataset (m=20), subsequent statistical analyses were carried out using Stata’s “mi estimate” prefix.

Multivariate logistic regression was used to assess the factors associated with MSHS participation. Crude participation rates and the adjusted odds ratios of participation were reported. The latter were adjusted for age (including age²), gender, the interaction between age and gender, socio-economic status (including educational level, occupational status, employment, and income), nationality, religious denomination, urbanity, geographic region, self-reported health and the number of contact attempts needed until the CATI or CAPI could be realized. Age was measured in years (whole-numbers), contact attempts were also represented in continuous whole-numbers. The remaining covariates comprised two or more categories: gender (female, male), educational level (primary, secondary, and tertiary education), occupational status (high level non-manual, medium level non-manual, low level non-manual, self-employed craftsmen, skilled manual, unskilled manual), employment (full time, part time, not in workforce), income (1^st, 2^nd, 3^rd, and 4^th quartile of household income), nationality (Swiss, non-Swiss), religious denomination (Protestant, Free Protestant, Roman Catholic, Christian Catholic, Christian Orthodox, other Christian, Jewish, Muslim, other, none). Urbanity comprised two categories: rural area and urban area. Urban areas included isolated cities (≥10,000 inhabitants) and urban agglomerations (≥20,000 inhabitants); all other areas were assigned to the category rural area. Geographic region comprised the seven greater regions of Switzerland (Midlands, Northwest, Zurich, East, Central, Tessin, and Lake Geneva Region). Self-reported health comprised two categories (very good or good health versus moderate, bad or very bad health).

To assess potential bias in MSHS outcomes, 8 variables from 3 domains (chronic diseases, prevention, psycho-social resources) were selected and the average outcome values in non-participants were predicted based on several different linear regression models which had been previously fitted on MSHS participants. From the 14 chronic diseases covered by the MSHS, 5 chronic diseases were selected based on the frequency of item non-response, i.e. outcomes with fewer missing values were preferred. The 5 disease related outcomes comprised the lifetime prevalence of arthrosis, osteoporosis, high blood pressure, myocardial infarction, and depression. Only 2 outcomes related to prevention were covered by the MSHS: contraception and influenza vaccination. The 12 month prevalence of influenza vaccination was selected because of its relevance in the contemporary Swiss public health debate, i.e. Switzerland faces a new law on mandatory vaccination. Finally, mastery and sense of coherence, the two psycho-social resources covered by the MSHS were selected because of their relevance when coping with stressful life events such as diseases [39]. While the mastery instrument was based on a brief version of the Pearlin coping questionnaire [40], the SOC instrument was based on a brief version of Antonovsky’s sense of coherence questionnaire [41]. All eight prediction models included socio-economic and sociodemographic covariates (age, gender, educational level, occupational status, employment, income, nationality, religious denomination, urbanity and geographic region) as well as self-reported health (all indicators measured as described above) and the Mental Health Inventory score (MHI-5). The MHI-5 is a 5-item subscale from the SF-36 [42, 43] which is used to capture four major dimensions of mental health: anxiety, depression, loss of behavioural/emotional control, and psychological well-being.

Outcome estimates related to the lifetime prevalence of arthrosis, osteoporosis, high blood pressure, myocardial infarction, and depression also included the body mass index (BMI) as a covariate. BMI calculation was based on self-reported height and weight (BMI=kg/m²). Finally, the existence of an acute major depression according to DSM-IV criteria (measured using CIDI-SF [44]) was added as a binary covariate to the model predicting the lifetime prevalence of depression. Crude average outcomes and age-standardized, sex-specific average outcomes were compared between participants and non-participants of the MSHS. Statistical significance was established at p ≤ 0.05.

Overall, more than 80% of the subjects who completed the SHS CATI or SHS CAPI also completed and returned the MSHS (Table 2). Participation rates were relatively high in all groups and never dropped below 60%. However, substantial differences between socio-economic as well as between sociodemographic groups were found. Generally, people from lower socio-economic strata exhibited lower participation rates. Similarly, non-Swiss nationals and members of some religious denominations showed lower participation rates. Participation rates varied also with age and their patterns were different between women and men (Figure 1). When adjusting for all other variables, these initial results were confirmed. Participation rates of both, men and women, increased gradually with age, eventually reached their maximum and decreased with age thereafter. The participation rates of women started to decrease at approximately 45 years of age, the participation rates of men began to decrease at approximately 60 years of age. At younger ages, the participation rates of women were generally higher than those of men whereas participation rates of men were equal to those of women at approximately 56 years of age and started to be higher than those of women at older ages. Overall, women were more likely to participate in the MSHS than men (OR=1.63), age² (OR=0.99), age (OR=1.06), and the interaction between age and gender (OR=0.99) were all significantly associated (p<0.001) with MSHS participation. People with a secondary (OR=1.48) or tertiary education (OR=1.76) as compared with people with primary education were more likely to participate in the MSHS as were white-collar workers (high level non-manual workers OR=1.29, medium and low level non-manual workers OR=1.26 and OR=1.25 respectively) as compared with unskilled manual workers. People who work full-time were less likely to participate than those without paid work (OR=0.76), and Swiss nationals were more likely to participate than were non-Swiss nationals (OR=1.60). Compared with people who had no religious affiliation, members of the Christian Catholic Church (OR=0.66), the Orthodox Church (OR=0.49), Muslim religious communities (OR=0.36) or the residual category other religious community (OR=0.52) were less likely to participate in the MSHS. Members of the remaining religious denominations, most notably members of the large Swiss State Churches (Protestant and Roman Catholic Church), did not differ from people without religious affiliation. People with better self-rated health (OR=1.35) were also more likely to complete and return the MSHS as compared with people who rated their health as moderate, bad or very bad.

Finally, no statistically significant differences in the likelihood of participation were observed between rural and urban areas, different geographic regions of Switzerland and household income quartiles.

In order to assess potential bias in MSHS outcome variables, outcomes of non-participants were estimated based on different linear regression models which had been previously fitted on MSHS participants. Table 3 shows the model fit as well as the strongest predictors (P<0.001) for each outcome variable. Generally, the percentage of variance explained by the models was moderate and ranged from 18.6% to 26.7%.

Crude predicted values for the MSHS non-participants (Table 4) were higher than the upper boundary of the 95% confidence interval (95%-CI) of the MSHS participants in four of the eight outcomes (influenza vaccination, arthrosis, high blood pressure, and depression) and they were lower than the lower boundary of 95%-CI of the MSHS participants in two outcomes (mastery and SOC). The predicted lifetime prevalence of osteoporosis and myocardial infarction of non-participants was within the boundaries of the 95%-CI of MSHS participants. Age-standardized and sex-specific figures (Table 5) give a more differentiated picture. For male non-participants, the predicted 12 month prevalence of influenza vaccination, the lifetime prevalence of arthrosis, osteoporosis, and high blood pressure were found to be lower than the lower boundary of the 95%-CI of male MSHS participants. For female non-participants, the corresponding predicted outcomes were higher than the upper boundary of the 95%-CI of female MSHS participants. The predicted lifetime prevalence of depression of female and male non-participants was higher than the upper boundary of the corresponding 95%-CI of MSHS participants. Finally, average mastery and SOC scores of male and female non-participants were lower than the lower boundary of the 95%-CI of corresponding male and female participants.