Prevention of Late-Life Depression in Primary Care: Do We Know Where to Begin?

The Amsterdam Study of the Elderly is a prospective cohort study on the incidence of depression in a large and representative sample of noninstitutionalized older persons (65+). Informed consent was asked and obtained from each participant before they began the study. The study was approved by the Medical Ethics Committee of the Vrije Universiteit Amsterdam. The population base for Amsterdam Study of the Elderly included all noninstitutionalized individuals in the 65–84-year age bracket who lived in the city of Amsterdam and were registered with a general practitioner at baseline. The study was actively supported by the general practitioners, whose lists were used as the sampling frame. In the Netherlands, general practitioners are the gatekeepers of the health care system, and almost every citizen is on the list of a general practitioner. In this role, general practitioners in the Netherlands generally provide social support and have a long-standing personal relationship with their patients. Thus, the source population consisted of almost all of the noninstitutionalized population. The sample was drawn from a list of 30 general practices spread throughout the city; practices were selected from all of those registered within the city of Amsterdam, 22 randomly and eight by convenience from a network of general practitioners participating in medical research. The profile of the general practice population over age 65 in terms of age and gender, corresponds to the noninstitutionalized population in Amsterdam (22, 23) . Within each practice, respondents were randomly selected from four age strata spanning 5 years each (65–69 and 80–84). In order to obtain equally sized age strata at follow-up, the older old were oversampled. Out of a total sample of 5,666, 4,051 (71.5%) responded and formed the baseline sample. Interviews with these subjects were conducted between May 1990 and November 1991. The profile of the general practice population over age 65 in terms of age and gender corresponds to the noninstitutionalized Amsterdam population.

At follow-up 3 years later (median=38 months), 2,244 (55.4%) were reinterviewed, 656 (16.2%) people were deceased, 662 (16.3%) refused to cooperate further, 282 (7.0%) were too ill or cognitively impaired to respond, and 207 (5.1%) were not available for interview because of a variety of other reasons. For this study, 523 subjects with depression (12.9%) or 261 with dementia (6.4%) at baseline were excluded. The study cohort consisted of all 1,940 respondents at follow-up who were neither depressed nor demented at baseline for whom complete data were available at follow-up ( Figure 2 ).

A 1-hour interview was developed to gather information on psychiatric symptoms, demographic and medical status, personal history of depression, and family history. The interview consisted of the Dutch translation of the Mini-Mental State Examination (24) ; all Geriatric Mental State examination items related to organic, affective, and anxiety syndromes (25, 26) ; the Activities of Daily Living scale (27) ; the Instrumental Activities of Daily Living scale (28) ; and part of the CAMDEX interview (29) . The interview was administered during home visits by lay interviewers who were trained to use video sessions and regularly supervised. The same Geriatric Mental State AGECAT package with an identical algorithm was used in the second wave of the study. When reinterviewing subjects, the raters were unaware of previous data and diagnoses.

Diagnoses of dementia, depression and generalized anxiety disorder were made according to the Geriatric Mental State AGECAT system. Diagnostic levels 3–5 correspond reliably to cases of depression requiring clinical attention in both the community (30) and in elderly hospital patients (31) . Geriatric Mental State AGECAT has proven reliability in epidemiological studies, as became evident through replication studies (26) . Geriatric Mental State AGECAT generates both a nonhierarchical syndrome level and a more narrowly defined diagnostic level. The diagnostic case level is calculated from the syndrome level with hierarchy from organic to depression to anxiety disorder. To be able to also assess overlapping comorbidity influences, syndrome levels were used in the analyses because otherwise the diagnostic hierarchy would bias the results. Depression “caseness” at baseline and follow-up was defined as a Geriatric Mental State AGECAT level 3 or higher (26) . Subjects not depressed at baseline who had become depressed at follow-up were considered to be incident cases. Subjects with depression case levels 1 or 2 at baseline were classified as having “subsyndromal depression.”

Sociodemographic variables included age, gender, and educational level, the latter of which was dichotomized into lower (primary school or less) and higher (more than primary school) education. Living situation was assessed with the associated questions in the Geriatric Mental State AGECAT. A personal history of depression was ascertained by a relevant CAMDEX question.

The question was considered affirmative if treatment had previously been requested. The presence of chronic diseases was assessed with the pertinent CAMDEX questions on myocardial infarction, stroke, cancer, lung disease, diabetes, Parkinson’s disease, arthritis, and epilepsy. Cognitive status was assessed by MMSE score. Cognitive impairment was defined as an MMSE score below 24. Sleeping disorder was determined by two questions from the Geriatric Mental State on “trouble falling asleep” (yes or “I would have problems if I did not use sleeping medication”) and “early wakening” (at least two times per week at least 2 hours earlier than usual). Disability was measured with a combined scale consisting of all activities of daily living and Instrumental Activities of Daily Living items (32, 33) . Subjects were considered disabled if the total score was two or more points below the maximum. This indicates that subjects were “in need of help” to perform at least two of the tasks mentioned or were “unable to perform’” at least one task.

The analyses were conducted in several steps. First, standard epidemiological techniques were used to calculate the absolute and relative risk of becoming a case for subjects with or without being exposed to a certain risk indicator.

In a second step, the importance of risk indicators for the onset of disorder at population level was quantified by taking into account the level of exposure to a risk factor in the population, also called the exposure rate, and the population-attributable fraction. The attributable fraction indicates by how many percentage points the incidence rate of depression in the population are lowered when the adverse effect of the risk factor (or combination of risk factors) can be completely blocked. Similarly, a measure of the efficiency of a preventive intervention can be calculated: the number needed to be treated (34) . The number needed to treat can be interpreted as a measure of efficiency, showing how many people must be protected from the adverse influence of a risk factor to avoid one new case of depression. Because the number needed to treat values presented here are based on the assumption that an intervention is able to completely block the adverse influence of a risk factor on the onset of depression, these should be adapted depending on the actual efficacy of each preventative intervention. If an intervention has 50% efficacy in containing the effect of a risk factor, the actual number needed to treat should thus be multiplied by 2. Using Stata (35) statistics such as relative risk, attributable fraction and the number needed to treat were obtained in a multivariate context with control for competing risks (10) .

In the third step, the relative risk, attributable fraction, the number needed to treat and E statistics were used to find the best values for efficiency (i.e., low needed to treat values) and the best values for health gains (i.e., highest attributable fraction values) for combinations of exposures. Targeting preventative interventions on groups that have been exposed to specific combinations of risk factors may yield higher health gains (relative risk, attributable) against less effort (number needed to treat).

To trace the effect of multiple exposures in all their combinations, we used a method known as classification and regression tree analysis (36, 37) and adapted it for our purposes. Classification and regression tree analysis can graphically be represented as tree-like figures ( Figure 3 ). First, a parent node is selected that optimizes the number needed to treat, the relative risk, and attributable fraction values in preferably the smallest exposure rate groups. This parent node branches off in two directions, creating child nodes in which people are not only exposed to the risk factor in the parent node but also to an additional risk factor in the child node. Therefore, the first generation of child nodes captures the effect of double exposures, the second generation of child nodes represents the effect of triple exposures, and so on. At each node, both branches (for subjects with and subjects without that specific risk indicator) are followed up to determine the number needed to treat, the relative risk, and the attributable fraction values of combined exposures. Thus, the sample is split into smaller subgroups at each step. To determine the optimal child nodes for our purpose, the following hierarchy of decision rules was used:

The following rules were employed for the termination of branches:

The baseline sample characteristics are shown in Table 1 and have been described in more detail elsewhere (13) . In multivariate logistic regression analysis, nonresponse at the follow-up assessment was predicted by higher age, male gender, lower education, living alone, chronic disease(s), disability, organic caseness, and depression (dichotomized as depression case level versus no depression case level). When subjects who died between measurements were excluded from these analyses, only lower education, living alone and baseline organic syndrome predicted nonresponse. At follow-up, 309 subjects (15.9% of the study sample) had become depressed. The following analyses investigate the associations of risk factors with incident depression.

Table 2 shows the predictor model and the associations with the onset of depression, with 95% confidence intervals. Statistically significant associations between risk indicators and depression incidence were found for the following risk factors: medical illness, disability, a personal history of depression, recent loss of spouse, sleep disturbance, generalized anxiety disorder, and subsyndromal depression. From these, the absolute risk of developing depression when exposed to a risk factor was calculated. The number needed to treat values for the risk factors in the model ranged from 242 (living alone, 95% CI not statistically significant) to 4.1 (95% CI=2.3–19.6) for subjects with generalized anxiety disorder. For each risk factor, the associated attributable fraction was also calculated, ranging from 24.6 (subsyndromal depression, 95% CI=18.8−30.0) to 1.1 (living alone, 95% CI not statistically significant).

Following the aforementioned algorithm, subsyndromal depression clearly was the preferred primary node ( Figure 3 ). Although the number need to treat of 5.8 was higher than that of generalized anxiety disorder (4.1) ( Table 2 ), the associated attributable fraction of depression was far greater (24.6%) and that of generalized anxiety disorder (2.0%) was below the threshold of 5% for the termination of branches. The prevention model was then continued for each branch, starting with subjects who had and those who did not have subsyndromal depressive symptoms. In subjects with subsyndromal depression, the next branching candidate was having a disability. This step reduced the number needed to treat to 3.9, but it also limited the attributable fraction to 9.7%. The model was then further subdivided with living alone and female sex as following risk indicators. Because the attributable fraction was below 5% at the next step (all→subsyndromal symptoms→disability→living alone→female sex→), the branch was terminated here. In subjects with subsyndromal depression who were not disabled, having a chronic illness (absolute risk=30.9%, number needed to treat=6.0) and living alone (absolute risk=32.9%, number needed to treat=5.6, attributable fraction=4.9%) were chosen as subsequent child nodes. In subjects who did not have depressive symptoms, spousal death was the strongest predictor, with an absolute risk of 31.4% of developing depression, a number needed to treat of 6.1, and an attributable fraction of 5.4%. Further branching was not performed because the attributable fraction became too small (<5%).

The way in which the number needed to treat, the attribution fraction, and exposure rate are interrelated is graphically illustrated in Figure 4 . It shows how adding more risk factors to the model results in a higher absolute risk, a lower number needed to treat, but also a lower attributable fraction.

Figure 5 shows the model without subsyndromal symptoms of depression. Now the focus is on selective prevention and easy case recognition by general practitioners. Following the algorithm described above, the first branching candidate on the basis of number needed to treat and attributable fraction scores is spousal death, with a number needed to treat of 5.3 and an attributable fraction of 8.2%. This was chosen as the primary (or parent) node. The selective-prevention model then yielded medical illness as the next branching candidate in subjects who lost their spouses. This step reduced the number needed to treat to 3.7 but also limited the attributable fraction to 5.8%. Because of this, further branching was not performed. In subjects who did not recently lose their partner, being disabled was the most important risk indicator, with a number needed to treat of 11.5 and an attributable fraction of 9.0%. Further risk indicators were the presence of a chronic medical illness and female gender. The latter step reduced the attributable fraction but only slightly affected the number needed to treat (from 8.6 to 8.3).

The Amsterdam Study of the Elderly is a large prospective cohort study with long follow-up intervals, and it incorporated a wide range of risk factors for late-life depression measured independently of depression onset. The findings are representative for urban populations of community dwelling older persons, and are highly relevant for primary care.

Nonetheless, a number of limitations deserve mentioning. First of all, studies with a limited number of measurements tend to overrepresent chronic forms of disorder. Between assessments, subjects may have had both onset and remission of shorter episodes of depression. At the same time, it may be concluded that even with a 3-year interval, the set of risk factors employed in this study can identify a large number of future cases of depression. Second, although the risk factors cover many domains relevant to depression, biological (genetic) risk indicators are less well represented in the data set. Other factors that would ideally have been available are medication use, substance abuse or dependence, and specific medical conditions, such as thyroid disorders that may be associated with depression onset. Third, it should be concluded that this study does not prove that preventative interventions are successful in community-living elderly. Available studies on indicated prevention were mostly performed in younger adults. However, because other forms of treatment of depression are also effective in the elderly, there would be no reason to doubt the efficacy of preventative interventions in elderly subjects. When considering the potential benefits of prevention, there is an urgent need for randomized, controlled trials to address this matter.