Setting

Data from the Gerontological Regional Database (GERDA) survey were used. The survey targeted the 15 municipalities in the county of Västerbotten in Sweden and the 17 municipalities in the county of Österbotten in Finland. The survey was sent out to 65-, 70-, 75- and 80-year-olds in 2010. All individuals in their respective age groups were contacted in the rural municipalities, but only a randomly selected third of potential participants were contacted in the two urban municipalities in Sweden (Umeå, Skellefteå) and half in the urban municipality in Finland (Vaasa). The Regional Ethical Review Board of Umeå/Sweden approved the study (2010-220-32Ö).

Participants

The survey was sent to 5425 people in Sweden and 5271 in Finland, of whom 3779 (69.7 %) and 3059 (58.0 %) respectively responded, resulting in an overall response rate of 6838 out of a possible 10696 (63.9 %). Of those who responded, the 6098 (89.2 %) who answered the questions about having had a stroke and being depressed or the four relevant questions from a depression assessment scale were selected for the final sample. The final sample included 2441 of the 3925 (62.2 %) 65-year-olds, 1496 of the 2507 (59.7 %) 70-year-olds, 1246 of the 2310 (53.9 %) 75-year-olds and 915 of the 1954 (46.8 %) 80-year-olds, to whom the survey was sent.

Procedure

The postal mail survey was posted in October 2010 in Sweden and Finland. It included 75 items, covering demographic information and interests, and questions about health, attitudes and disabilities. One reminder letter was sent to participants who did not respond.

Stroke definition

In the present study, having a previous stroke was defined as answering “yes” to the question “Have you had a stroke?”. In a German demographic cohort of people aged 65 years and older [21], a single stroke question resulted in a sensitivity of 65.8 against registered diagnoses in medical records and neurological examinations in uncertain/equivocal cases. The single stroke question resulted in more false negatives compared with a concurrently tested 6-item stroke assessment scale, but also had notably fewer false positives [21].

Depression definition

The Geriatric Depression Scale (GDS) 4-item version [22] is a short assessment for depression comprising four yes/no questions. The four questions are drawn from the widely used GDS 15-item assessment scale [23]. Multiple 4-item versions of the GDS exist, depending on which of the 15 questions of the GDS-15 that are being used. The version used in this paper consisted of the questions “Are you basically satisfied with your life?”, “Do you feel that life is empty?”, “Are you afraid that something bad is going to happen?” and “Do you feel happy most of the time?”. In a primary care setting, with a cut-off at two or more points, the GDS-4 was found to have a sensitivity of 61 % and a specificity of 81 % measured against a comprehensive diagnostic interview [22].

The present study also included the yes/no question “Do you feel depressed?”. In a stroke disease cohort [24], the question “Do you often feel sad or depressed?” from the Yale-Brown Obsessive compulsive scale [25] resulted in a sensitivity of 86 % and a specificity of 78 % against a ≥6 cutoff on the Montgomery Åsberg Depression Rating Scale.

Depression was defined as answering “yes” to the yes/no question, or a GDS-4 score ≥2. This produced an overall figure of 739 people with depression; 290 people selected on the result of the GDS-4 cutoff alone, 294 people on the yes/no question alone, and 155 people who met both criteria. A combination measure was selected to increase the sensitivity for depression.

Other definitions

Malignancy, diabetes and a previous myocardial infarction were defined as answering “yes” to yes/no questions.

Dependence in personal activities of daily living (ADL) was defined as answering “no” to the question “Do you shower without human assistance?”. Dependence in instrumental ADL was defined as answering “no” to any of the following questions: “Do you clean your residence by yourself, without human assistance?”, “Do you buy your own groceries without human assistance?”, “Do you utilize public transportation without human assistance?” and “Do you cook your own food without human assistance?”. Vision impairment was defined as answering “no” to the question “Are you able to read the text in a newspaper?”. Hearing impairment was defined as answering “no” to the question “Are you able to hear what someone tells you in a normal tone of voice from a distance of about 1 meter?”. Weight loss was defined as answering “yes” to the question “Has your weight decreased in the last three months?”.

Living alone was defined as answering “I live alone” or “I am living with someone, but not in the same residence” to the question “Do you live together with someone?”. Having no one to talk to was defined as answering “no” to the question “Do you have someone you can talk to about anything, someone you can share your concerns as well as your joys with?”.

Having financial difficulties was defined as answering “It is pretty difficult” or “It is very difficult”, but not “With some difficulty” or “Without any difficulty”, to the question “Do you make ends meet?”. Having a pain problem was defined as answering “yes” to the question “Have you had any aches or pain in the last week?”. Having a current life crisis was defined as answering “yes” to the question “Have you experienced something that you would describe as a life crisis the last year (12 months)?” A comprehensive list of the questions and answers that the present study is based on is provided as Additional files 1 and 2.

Statistics

R version 3.0.2 was used for statistical analyses. The prevalence of stroke and depression and the corresponding prevalence ratio in the population were estimated using an R software library for complex surveys called “Survey”. These epidemiological results were weighted based on the proportion of participants compared with known population totals in demographic strata based on home municipality, sex and age and, adjusted for finite population correction. Differences in prevalence between age groups, men and women, and Sweden and Finland were tested using ChiSquare tests, and the prevalence ratios between stroke and depression were tested using log-binomial regression models. Log-binomial regression models were used instead of the more widespread logistic regression models in order to decrease the likelihood of distorted rate ratios.

Factors associated with depression were also tested using log-binomial regression models. These analyses were not adjusted for the composition of the underlying population. Out of 75 items in the survey, 18 that had been associated with depression in previous studies [17, 20] or were thought to be relevant based on clinical experience were entered into univariate models. Separate models were used for people with and without stroke.

The influence of stroke status on the univariate association between depression and the study items was explicitly tested using interaction analyses. Likelihood-ratio tests were used to compare interaction models of the form “item + stroke status + item * stroke status (interaction term)” with the corresponding models without interaction terms.

Factors significantly associated with depression in univariate models, and age and sex, were entered into two multivariate models: one for those with and one for those without stroke. Multivariate models were analyzed according to a previously published methodology [26], utilizing a restricted optimization routine to increase the probability of regression convergence. A value of p < 0.05 was considered statistically significant.

Strengths and limitations

Stroke prevalence was determined by self-reporting using a yes/no question. While some mischaracterized cases are to be expected when using self-reported data, our age-specific stroke prevalence estimates are similar to previous self-reported results [12, 13], and only slightly higher than self-reported cases confirmed by medical documentation [12]. One benefit of using self-reported data is that it made it possible to reach a large number of people representing both urban and rural community-dwellers.

Depression prevalence was determined by a combination of self-reporting with a yes/no question and the use of a depression assessment scale. While depression assessment with a yes/no question has been used successfully in stroke patients [24], depression is known to be underdiagnosed in elderly people [30], meaning that patients who are not diagnosed with depression and do not consider themselves depressed may be overlooked. In addition, our clinical experience suggests that elderly people who are depressed may not be open about it when asked a direct question. We decided to combine self-reporting with a short version of a depression assessment scale to increase the sensitivity for depression. Our estimated depression prevalence was similar to the pooled prevalence of depressive disorders reported in a recent meta-analysis [15].

The investigation of factors associated with depression was limited to items suitable to a postal survey. This precluded the use of any radiological assessment of stroke severity, blood tests or review of medical charts. There was no assessment of stroke severity at the time when the participant had his/her stroke, although long-term functional outcome was assessed with survey items for ADL dependency.

Associations with depression were calculated using log-binomial regression, making it possible to present prevalence ratios. These are easier to interpret than the widely used odds ratios as they are not distorted when analyzing prevalent outcomes. One possible negative aspect of this choice is that it is more difficult to compare with previously published cross-sectional results, as they mainly report associations as odds ratios. Using log-binomial regression is also difficult in most statistical software because of non-convergent optimization procedures in multivariate analyses, and the method used here, involving constrained optimization, is not widely utilized.

In the analytical part of the study, the rural areas were oversampled compared to the urban areas, which might have influenced the results. While survey weights were available, they were used only for the estimates of stroke and depression prevalence that make up the epidemiological part of the study. There is a scientific consensus that descriptive results should be weighted if the survey sampling was unbalanced, but there is no such consensus regarding analytical results [31]. The response rate in the present study was, however, fairly high, including a substantial number of people living both in urban and rural settings.