Soc Psychiatry Psychiatr Epidemiol (2003) 38 : 290–296
DOI 10.1007/s00127-003-0649-9
ORIGINAL PAPER
Scott B. Patten
Recall bias and major depression lifetime prevalence
Accepted: 4 February 2003
SPPE 649
■ Abstract Background Epidemiological studies have reported lifetime prevalence rates for major depression that have typically ranged between 10 % and 20 %, and have typically been fairly stable across age groups. This contradicts the intuitive expectation that lifetime prevalence should accumulate with age. Some authors have hypothesized that a cohort effect may account for this puzzling pattern, but another possible explanation is recall bias. In principle, differential mortality could also account for a failure of age-specific lifetime prevalence to increase with age. The objective of this study was to describe the relationship between recall failure (failure to recall previous depressive episodes) and bias in the evaluation of age-specific lifetime major depression prevalence. Method A framework recently described by Hill et al. was used in this analysis. Hill’s model relates the slope of an age-specific prevalence curve for an irreversible disease to incidence and mortality. The model is applicable in the current context since lifetime major depression is, by definition, an irreversible condition (albeit one typically characterized by recurring episodes). In the current investigation, an extra term for recall failure was added to the model. In order to incorporate mortality, a structured literature review was conducted and the impact of plausible levels of differential mortality was explored. Results Relatively low rates of recall failure (e. g. 2–4 % per year) can account for a flat or declining age-specific lifetime prevalence curve in most age groups, with higher rates of recall failure being necessary in the adolescent age range. Among the elderly, where mortality rates are higher, differential mortality may also contribute to the pattern. Conclusions Available Dr. Scott Patten (), MD, FRCP(C), PhD Dept. of Community Health Sciences The University of Calgary 3330 Hospital Drive NW Calgary, AB, T2N 4N1, Canada Tel.: +1-4 03/2 20-87 52 Fax: +1-4 03/2 70-73 07 E-Mail:
[email protected] Website: www.ucalgary.ca/~patten
data about the incidence of major depression, and the mortality associated with this condition, suggest that plausible rates of recall failure can explain a flat or declining lifetime major depression prevalence across age groups.A corollary of this result is that many existing estimates of major depression lifetime prevalence may be too low. A cohort effect can be inferred from cross-sectional lifetime prevalence data only when the diagnostic instruments employed make recall failure very unlikely. ■ Key words depressive disorder – statistics and numerical data – mental disorders – epidemiology – prevalence studies
Introduction Lifetime prevalence refers to the proportion of members of a population who have experienced a condition at any time in their life up to the time of sampling [1]. The lifetime prevalence of major depression is expected to increase with age since older persons have had more time to develop episodes of major depression. In distinction to this expectation, reported prevalence rates for lifetime major depression have tended to remain approximately stable across age groups in recent studies [2–4]. There are a variety of possible explanations for this. For example, it is possible that a cohort effect has led to increasing major depression prevalence among young people [5]. A cohort effect is not the only possible explanation for lifetime prevalence rates that do not increase with age. Differential mortality could explain such an effect. If depressed persons experience elevated mortality, lifetime prevalence in the population will diminish as a result. Another possible explanation for a decline in lifetime prevalence with age involves bias, specifically the recall bias that could result from failure to recall previous episodes of depression (recall failure). If some people fail to recall previous episodes in response to probes contained in structured interviews, the resulting preva-
291
lence estimate will be biased in a negative direction. Recall failure can be regarded as an event that involves forgetting previous episode(s): some proportion of people who report past episodes at the beginning of an interval may no longer report these episodes at the end of the interval because they no longer recall them. If the interval chosen is one year, this would represent an annual rate of recall failure. The objective of this study was to describe the annual rate of recall failure required to reduce the slope of an age-specific lifetime prevalence curve to zero, in view of rates of incidence and mortality that have been reported in the literature.
Subjects and methods Recently, Hill et al. [6] proposed a simple mathematical model for describing the relationship between incidence, mortality and prevalence for an irreversible disease. Although major depressive disorder is an episodic condition, the lifetime prevalence is by definition irreversible.A person experiencing an episode of major depression at any time in their life, with the exception of those who go on to develop a bipolar disorder or schizophrenia, will qualify for a diagnosis of major depressive disorder for the rest of their life. The model proposed by Hill et al. was based upon the slope of the age-specific prevalence curve, formulated as the first derivative of the equation for the prevalence ratio, leading [6] to the following equation:
terval between 1957 and 1972, an age standardized rate per year of 3.7 per 1000/year was reported in men and 7.7 per 1000/year was reported in women [8]. Another set of estimates were reported in a 10year follow-up of the Baltimore Epidemiological Catchment Area study: 3.62 per 1000/year in women and 1.98 per 1000/year in men [9]. The Lundby study also reported an age-specific incidence gradient, an issue that was addressed in this study during a secondary analysis (described below). In order to obtain information about the relative mortality (Mx/My) associated with major depression, a formal literature search was conducted. This search utilized two electronic databases: Medline [1966–2000] and Psych Info [1984–2000] and incorporated a blind methodological review by two graduate-trained epidemiologists. The results are summarized in Table 1. With the exception of two studies, each with relatively short follow-up periods [10, 39], the literature generally pointed towards an elevated mortality in those with depressive disorders. The relative risks and standardized mortality ratios reported ranged between 1.3 and 2.5. These values seemed consistent with the conclusion of a recent review, which placed the relative risk at 1.7 [12]. Since almost all of the mortality studies identified in the literature review reported estimates of relative mortality, and since the models employed in this analysis require mortality rates (see equations 1 and 2), age- and gender-specific mortality rates for the general population were obtained from Statistics Canada (http://www.statcan.ca). Polynomial regression models predicting overall mortality by age, using a natural log transformation were fit for men (mortalitym) and women (mortalityf) separately: Equation 3: mortalitym(a) = exp(–6.422–0.073a + 0.002a2 – 0.00000602a3)
Equation 1: p(a) = i(a) * {1 – p(a)} + {mx(a) – my(a)} * {1 – p(a)} * p(a)
Equation 4: mortalityf (a) = exp(–6.250–0.193a + 0.006a2 – 0.0000401a3).
where p(a) represents the slope of the age-specific prevalence curve, i(a) refers to age-specific incidence, p(a) refers to the age-specific prevalence and mx(a) and my(a) refer to age-specific mortality rates among persons without and with the disease in question, respectively. Intuitively, the formula indicates that prevalence will increase (positive slope) with age to the extent that incidence at any given age exceeds differential mortality among existing cases [note that in the usual case where my(a) exceeds mx(a), the {mx(a) – my(a)} term will be negative]. If the possibility of recall bias is acknowledged, failure to recall previous episodes can also be regarded as a determinant of the slope of the age-specific prevalence curve.As with differential mortality, recall will diminish measured lifetime prevalence (denoted in the following equation using the subscript ‘m’).
In these expressions, age (a) was represented in years, at the midpoint of the 5-year age intervals in which mortality rates are reported by Statistics Canada, regressed over the interval 17.5 years to 62.5 years. These regression equations provided a good description of the sexspecific mortality rates for men and women within this age range. The gender-specific mortality rates were assumed to represent a weighted average of mx and my with the weighting factor being measured sexspecific lifetime prevalence (LTPm for men and LTPf for women). For men:
Equation 2: pm(a) = i(a) * {1 – pm(a)} + {mx(a) – my(a)} * {1 – pm(a)} * pm(a) – fpm(a) where pm (a) refers to the slope of a measured age-specific prevalence curve and is the measured lifetime prevalence at a specific age. In Equation 2, f is the annual rate of recall failure: the proportion of those with measured lifetime prevalence at the start of a year who will then fail to recall their episodes when re-evaluated one year later. In this conceptualization, f does not depend on age nor on the amount of time since the occurrence of an episode. The possibility of forgetting starts with the experience of an episode, rather than at any specific age. This formula suggests that the slope of the measured age-specific prevalence curve for major depression will be negative if the number of apparent new cases (new episodes among those not reporting prior episodes) is exceeded by differential mortality among persons reporting lifetime episodes and the number of persons with apparent lifetime major depression who would begin failing to recall prior episodes during a year. The estimates of major depression incidence used in this analysis derived from several sources. One source was a longitudinal study conducted in Nova Scotia, the Stirling County Study. Here, annual incidence rates of 4.5 and 3.7 per 1000/year were reported for two distinct cohorts. Significant differences by age and sex were not identified [7]. Another source of incidence data was a prospective study conducted in Lundby, Sweden. Here, during a 15-year follow-up in-
Equation 5: mortalitym(a) = {pm(a) * my(a)} + {1 – pm(a)} * mx(a) where mx(a) and my(a) are age-specific values that apply to men. At any specified value for lifetime prevalence, and with a specified relative mortality (R), mx(a) and my(a) could be calculated for men by replacing my(a) with mx(a)*R and rearranging Equation 5 to solve for mx(a): Equation 6: mx(a) = mortalitym(a) / {pm(a) * R} + {1 – pm(a)}. In order to generate models relating incidence, measured prevalence and recall failure in circumstances where lifetime prevalence does not increase with age, Equation 2 was rearranged to solve for f when the slope, pm(a) was set to zero: Equation 7: f = [i * {1 – pm} + {mx(a) – my(a)} * {1 – pm} * pm] / pm. Since the situation under evaluation is that where measured prevalence does not change with age, the notation pm(a) is simply depicted pm in Equation 7.As such, rates of recall failure that would be required to create a prevalence curve that does not increase with age (under various assumptions about incidence and mortality) were identified, given various plausible assumptions (from the literature) about incidence and mortality. Subsequently, these are referred to as “zero slope” age-specific prevalence curves. These calculated rates of recall failure constitute threshold values. Higher rates of recall failure would result in a negative slope, and lower rates would result in a positive slope for any set of values for incidence and mortality. Incidence in Equation 7 is not depicted as being age dependent since initial simulations used an age-invariant value for this parame-
292 Table 1 Summary of the mortality data literature review Study
Study Design
Measure of Association
Association
Fredman et al. [10]
2-year follow-up of ECA-Piedmont Health Survey respondents aged 60 + 20-year follow-up of Maudsely Hospital patients 6-year follow-up of elderly (60 +) depressed Finns 24-year follow-up of “endogenous” depression 3-year follow-up of community sample aged 75 + 5-year follow-up in a cohort aged 60–88
Relative Risk (for MDE, or severe Dysthymia)
0.41
SMR* Relative Risk (longitudinal study control group) SMR Relative Risk SMR
1.9 1.3 1.44 1.6 2.0 () 1.7 () – 1.74
Lee et al. [27] Pulska et al. [28] Thomson [29] Engedal [11] Burvill and Hall [30] Hoch et al. [31] Buchholtz-Hansen et al. [32] Burvill et al. [33] Murphy [34] Pulska et al. [35] Pulska et al. [36] Brodaty [37] Allgulander [38] Koenig et al. [39] Murphy et al. [40] Weeke [41] Black et al. [42] Coryell [43]
2-year survival in a mixed community sample 3- to 10-year follow-up of clinical trial subjects with “endogenous” depression 1-year follow-up of elderly (60 +) depressed patients treated by psychiatrists 16-year follow-up of a community sample Community follow-up of elderly persons 6-year follow-up of elderly persons 25-year follow-up of hospital discharges 7- to 17-year follow-up of hospitalized patients in a Swedish case register 5-month (mean) follow-up of elderly medical patients with MDE 4-year follow-up of depressed elderly (65 +) and matched controls 6-year (average) follow-up of subjects in a Danish case register 5.2 year (average) follow-up of discharges from a University Psychiatric Hospital 42-year follow-up of hospital discharges with “primary unipolar depression"
9 % of those with MDE died versus 0 % of controls SMR No measure reported, observed deaths “almost 3 times” expected SMR Relative Risk
–
Relative Risk SMR SMR – only cause-specific SMRs were reported
1.7 1.3 (recovered) 1.5 (long-term) 2.0 1.4 –
Relative Risk
1.0
Relative Risk (abstracted)
2.44
SMR reported only for bipolar patients
–
SMR
1.44 () 2.49 () 1.39
SMR (abstracted)
* Standardized mortality ratio
ter. However, to evaluate the impact of declining incidence of major depression with age, as reported by Hagnell et al. [8], linear regression equations were incorporated into the simulations during a secondary analysis (see below), in which incidence depended on age.
Results Fig. 1 shows rates of recall failure expected to result in a zero slope to the age-specific prevalence curve for men and women at two levels of relative mortality: 1.3 and 2.5. These specific values were chosen to reflect the range of mortality values observed in the literature search (see Table 1). The values for (measured) lifetime prevalence (pm) used in the simulations presented in Fig. 1 were 12.7 % and 21.3 % for men and women, respectively, from the National Comorbidity Survey [13]. The Baltimore ECA incidence rates were used in the specific simulations presented in Fig. 1: 3.62 per 1000/year in women and 1.98 per 1000/year in men [9].Fig. 1 shows that, under these assumptions, low rates of recall failure (less than 1.5 % per year) are expected to result in a flat lifetime prevalence curve. For older age groups, even lower rates will produce the same result. As age in-
Fig. 1 Rates of recall failure producing zero slope, by relative mortality (R)
creases, differential mortality makes an increasing contribution, and lower rates of recall failure are required to create a zero slope. As noted above, a variety of incidence estimates have
293
been reported in the literature. Figures 2 and 3 depict the rates of recall failure expected to result in a zero slope for age-specific lifetime prevalence curves for men and women, respectively, under a variety of possible rates for incidence. The incidence rates depicted for men in Fig. 2 are 4.5 per 1000/year (Stirling County, first Cohort), 3.7 per 1000/year (from Stirling County, second Cohort and the Lundby male estimate) and 1.98 per 1000/year from the ECA study. The incidence rates for women are 4.5 and 3.7 per 1000/year (Stirling County), and 7.7 per 1000/year (Lundby). The 3.62 per 1000/year (ECA) estimate does not appear in the figure because this closely resembled the second Stirling County estimate. In both sets of simulations, a relative mortality of 1.7 was used. Figures 2 and 3 show that the rate of recall failure leading to a zero slope must be higher at higher incidence rates, but that generally the rates are quite low. Also, in all of the curves, the rates of recall failure required to produce a zero slope decline at older ages as a result of differential mortality.
■ Secondary analysis Hill’s model refers to age-specific incidence, but the simulations presented so far have each used single or gender-specific incidence rates. One study, the Lundby study [8], reported age-specific rates that appeared to diminish with time both for men and women. In order to evaluate the impact of this phenomenon, linear regression lines were fit to each set of data points, as presented in Fig. 4. The linear regression equations were then added to the simulation algorithms in order to reflect their impact on the rates of recall failure needed to produce a zero slope across the age range. The resulting simulation is presented in Fig. 5, showing that if incidence rates are higher in young people, then higher rates of recall failure must occur in order to prevent the agespecific lifetime prevalence curve from climbing. The simulation presented in Fig. 5 used the same values for relative mortality and pm(a) as the simulations presented in Figs. 2 and 3. The predicted values for annual rates of recall failure are slightly higher than those predicted for the Lundby simulations in Figs. 2 and 3. This
Fig. 4 Age-specific major depression incidence (data from Hagnell et al., linear regression) Fig. 2 Rate of recall failure producing zero slope at three incidence rates: men
Fig. 3 Rate of recall failure producing zero slope at three incidence rates: women
Fig. 5 Rates of recall failure producing zero slope, based on age-specific incidence
294
is because the Lundby incidence estimates were agestandardized and included some age groups (e. g. 0–9 years) where depressive episodes were not observed. As a result, the age-specific incidence rates across the age range depicted in Fig. 5 are all slightly higher than the reported age-standardized rates that were employed in generating the simulations presented in Figs. 2 and 3. In turn, this results in a slightly higher estimated annual rate of recall failure.
Discussion The lack of fit between available incidence estimates and lifetime prevalence estimates for major depression has received comment from previous authors. Hagnell et al. [8] described a method for calculating the lifetime prevalence that would be expected to result from a given incidence rate by a particular age, assuming survival to that age. Applying these calculations, it is interesting to note that an incidence rate of 5.0 per 1000/year would be expected to result in a lifetime prevalence of 24.7 % in a cohort of subjects observed over a 50-year period, exceeding the lifetime prevalence estimates for elderly subjects that have been reported by community surveys. As noted above, this discrepancy could be due to a variety of factors. Using comparable calculations, Eaton et al. [9] concluded that the incidence rate seen in the Baltimore ECA follow-up would have resulted in 50 % lifetime prevalence. The incidence rates reported from the ECA studies and the Stirling County study cannot account for the very high lifetime prevalence estimates reported for young people in recent studies, for example, the very high lifetime prevalence in 15- to 24-year-olds the NCS (15.7 %) is inconsistent with incidence rates in the range of 2–8 per 1000/year. Studies of major depression incidence in adolescents have reported markedly higher rates. One recent German study reported a 20-month incidence of major depression in an adolescent sample [14], 3.7 % in male and 7.5 % in female respondents. These authors estimated that the 12-month incidence in their sample would have been approximately 4.3 %. Another prospective study of high school aged adolescents reported annual incidence rates for major depression of 7.1 % for female and 4.4 % for male subjects [15]. Extremely high incidence rates reported among young subjects in these studies indicate either a very powerful cohort effect in those age groups (a 6 % incidence rate would predict a lifetime prevalence of > 40 % after 9 years) or they may indicate that a large proportion of these early-life episodes are later forgotten. Either interpretation is plausible, acknowledging that the incidence estimates deriving from the ECA and Stirling County studies were restricted to subjects aged 18 and over. When 18.4 % [15] is used as an estimate of measured lifetime major depression prevalence in adolescents, along with an incidence rate estimate of 5.72 % [15], stable lifetime prevalence is achieved only with very high
rates of recall failure, in the range of 25 % each year. Future research should investigate whether such high rates of recall failure actually occur in this age group. In epidemiology, bias may be defined as a “deviation of results from the truth” [1]. Bias is differentiated from sampling error because it is systematic, as opposed to stochastic in nature. An important source of bias in epidemiological research is that related to the recollection of past events or experience: recall bias. The idea that the lifetime prevalence parameter may be vulnerable to incomplete recall is not a new one [9, 16], and contemporary structured diagnostic interviews are undergoing refinements in order to facilitate enhanced recall of symptoms and experience [17]. A growing set of published population-based estimates of major depression incidence predict that the lifetime prevalence of major depression among elderly subjects should be somewhat higher than what has been reported in the literature, and recall bias is one of several possible explanations for this. The current analysis confirms that relatively modest rates of recall failure can account for an observed tendency of lifetime prevalence not to increase with age and, thereby, account for an underestimation of lifetime prevalence. The original version of the Diagnostic Interview Schedule (DIS) identified only 63 % of those with lifetime major depressive episodes who were identified by a physician interview in validation studies [18]. This large difference could result from recall failure (in the way it was conceptualized in this study) if a proportion of persons with previous episodes stop recalling these episodes in response to the DIS items with each passing year since the occurrence of the most recent episode. Brugha et al. similarly reported that diagnostic agreement between structured interviews and semi-structured clinical interviews is low [19]. Newman and Bland reported data from a follow-up of a cross-sectional survey in Edmonton, Canada [20]: 9.5 % of subjects (n = 295) identified as having DIS-defined lifetime major depression in a baseline interview ceased to be lifetime cases when they were again interviewed (on average) 2.8 years later. The rates of apparent misclassification reported in these studies are sufficiently high to make recall bias seem to be a plausible explanation for the failure of lifetime prevalence to increase with age. A cohort can be defined as the component of the population born during a particular period and identified by period of birth [1]. As society changes, successive birth cohorts are exposed to unique environmental factors, which may impact disease incidence. If some shared environmental factor(s) have led to an increased incidence in cohorts born in recent decades, an elevated prevalence in the corresponding (young) age groups could occur. In principle, these prevalences could exceed that of older people who are members of earlier birth cohorts. A classical example of such an effect was reported using British lung cancer mortality data by Doll in 1971 [21]. Identification of cohort effects in epidemiology has
295
generally required data collection from large numbers of people over long periods of time, typically using vital statistics or registry data. While a cohort effect has been suggested as an explanation for the pattern of age specific lifetime major depression prevalence observed cross-sectionally [22], the current analysis suggests that recall failure is an alternate explanation. Some combination of factors, of course, may ultimately provide the best explanation. One solution to the problem of recall bias in measurement of lifetime major depression prevalence is more accurate measurement. If this is to be achieved, it may be necessary to develop structured interviews that incorporate enhanced methods of facilitating recall. Such work is ongoing, e. g. by Wittchen and others, who have described the introduction of sophisticated cognitive interviewing techniques in structured interview protocols [17], and by Brugha and others who have explored the use of semi-structured interviews by survey interviewers [23]. The extent to which enhancement of existing measurement instruments can diminish recall bias in the assessment of lifetime prevalence remains to be seen. Another solution may be to use other measures of prevalence, such as using point prevalence rather than period prevalence as a measure of disease burden. Point prevalence can reflect the impact of therapeutic and preventive interventions on disease burden because it depends both on incidence, relapse rates and duration, whereas lifetime prevalence, in theory, will not be impacted by interventions that reduce relapse or the duration of episodes. Point prevalence is also less dependent upon recall, since the diagnosis can be assisted by observable signs and symptoms, and by reporting current as opposed to past symptoms. On the other hand, the concept of lifetime prevalence will remain useful since major depression is a recurrent condition. Persons with multiple previous episodes have a high rate of relapse and are considered candidates for long-term preventive antidepressant treatment [24, 25]. Recommendations for such “maintenance” treatment are found in contemporary clinical practice guidelines for management of major depressive disorder [26]. These long-term treatment needs relate more closely to lifetime than to current prevalence. Also, in studies of risk factors and comorbidities, it will generally be more practical to use lifetime prevalence. Since the observed prevalence proportions will be greater. While the results reported here do not identify any one perspective as being inherently superior to another, they do highlight the vulnerability of lifetime prevalence measures to recall bias. ■ Acknowledgement This project was supported by a grant from the National Health Research and Development Project (NHRDP). Grant Number – 6609-09-1999/2640029. Dr. Pattens supported by the Alberta Heritage Foundation for Medical Research.
References 1. Last JM (1995) A dictionary of epidemiology. 3rd ed. Oxford University Press, Oxford 2. Spaner D, Bland RC, Newman SC (1994) Major depressive disorder. Acta Psychiatr Scand Suppl 376:7–15 3. Carta MG, Carpiniello B, Kovess V, Porcedda R, Zedda A, Rudas N (1995) Lifetime prevalence of major depression and dysthymia: results of a community survey in Sardinia. European Neuropsychopharmacology Suppl:103–107 4. Chen CN, Wong J, Lee N, Chan-Ho MW, Lau JTF, Fung M (1993) The Shatin community mental health survey in Hong Kong. II. Major findings. Arch Gen Psychiatry 50:125–133 5. Kessler RC,Walters EE (2000) Epidemiology of DSM-III-R major depression and minor depression among adolescents and young adults in the National Comorbidity Survey. Depression and Anxiety 7:3–14 6. Hill GB, Forbes WF, Kozak J (1999) A simple method for estimating incidence from prevalence. Chron Dis Can 20:151–153 7. Murphy JM, Laird NM, Monson RR, Sobol AM, Leighton AH (2000) Incidence of depression in the Stirling County Study: historical and comparative perspectives. Psychol Med 30:505–514 8. Hagnell O, Lanke J, Rorsman B, Ojesjo L (1982) Are we entering an age of melancholy? Depressive illnesses in a prospective epidemiological study over 25 years: the Lundby Study, Sweden. Psychol Med 12:279–289 9. Eaton WW, Anthony JC, Gallo J, Cai G, Tien A, Romanoski A, Lyketsos C (1997) Natural history of diagnostic interview schedule/DSM-IV major depression. The Baltimore Epidemiological Catchment Area follow-up. Arch Gen Psychiatry 54:993–999 10. Fredman L, Schoenbach VJ, Kaplan BH, Blazer DG, Sherman JA, Kleinbaum DG,Yankaskas B (1989) The association between depressive symptoms and mortality among older participants in the Epidemiological Catchment Area-Piedmont Health Survey. Journals of Gerontology 44:149–156 11. Engedal K (1996) Mortality in the elderly: a 3-year follow-up of an elderly community sample. Int J Geriatr Psychiatry 11: 467–471 12. Wulsin LR, Vaillant GE, Wells VE (1999) A systematic review of the mortality of depression. Psychosom Med 61:6–17 13. Blazer DG, Kessler RC, McGonagle KA, Swartz MS (1994) The prevalence and distribution of major depression in a national community sample: the National Comorbidity Survey. Am J Psychiatry 151:979–986 14. Oldehinkel AJ, Wittchen H-U, Schuster P (1999) Prevalence, 20month incidence and outcome of unipolar depressive disorders in a community sample of adolescents. Psychol Med 29:655–668 15. Lewinsohn PM, Hops H, Roberts RE, Seeley JR, Andrews JA (1993) Adolescent psychopathology: I. Prevalence and incidence of depression and other DSM-III-R disorders in high school students. J Abn Psychol 102:133–144 16. Regier DA, Kaelber CT, Rae DS, Farmer ME, Knauper B, Kessler RC, Norquist GS (1998) Limitations of diagnostic criteria and assessment instruments for mental disorders.Arch Gen Psychiatry 55:109–115 17. Wittchen H-U, Ustun TB, Kessler RC (1999) Diagnosing mental disorders in the community. A difference that matters? Psychol Med 29:1021–1027 18. Helzer JE, Robins LN, McEvoy LT, Spitznagel EL, Stoltzman RK, Farmer A, Brockington IF (1985) A comparison of clinical and diagnostic interview schedule diagnoses. Physician examination of lay-interviewed cases in the general population.Arch Gen Psychiatry 42:657–666 19. Brugha TS, Bebbington PE, Jenkins R, Meltzer H, Taub NA, Janas M,Vernon J (1999) Cross validation of a general population survey diagnostic interview: a comparison of CIS-R with SCAN ICD-10 diagnostic categories. Psychol Med 29:1029–1042 20. Newman SC, Bland RC (1998) Incidence of mental disorders in Edmonton: estimates of rates and methodological issues. J Psychiatr Res 32:273–282
296 21. Doll R (1971) The age distribution of cancer: implications for models of carcinogenesis. J Roy Stat Soc 134:133–155 22. Kessler RC,Walters EE (1998) Epidemiology of DSM-III-R major depression and minor depression among adolescents and young adults in the National Comorbidity Survey. Depression and Anxiety 7:3–14 23. Brugha TS, Bebbington PE, Jenkins R (1999) A difference that matters: comparisons of structured and semi-structured psychiatric diagnostic interviews in the general populations. Psychol Med 29:1013–1020 24. Keller MB, Hanks DL (1994) The natural history and heterogeneity of depressive disorders: implications for rational antidepressant therapy. J Clin Psychiatry 55 (suppl A):25–31 25. Thase ME (1999) Long-term nature of depression. J Clin Psychiatry 60 (suppl 14):3–9 26. Work Group on Major Depressive Disorder (2000) Practice guidelines for treatment of patients with major depressive disorder (revision). Am J Psychiatry 157 (Suppl):1–45 27. Lee AS, Murray RM (1988) The long-term outcome of Maudsley depressives. Br J Psychiatry 153:741–751 28. Pulska T, Pahkala K, Laippala P, Kivelae SL (1997) Six-year survival of depressed elderly Finns: a community study. Int J Geriatr Psychiatry 12:942–950 29. Thomson W (1996) Type of depression and results of mortality. Personality and Individual Differences 21:613–615 30. Burvill PW, Hall WD (1994) Predictors of increased mortality in elderly depressed patients. Int J Geriatr Psychiatry 9:219–227 31. Hoch CC, Reynolds CF, Buysee DJ, Fasiczka AL (1993) Two-year survival in patients with mixed symptoms of depression and cognitive-impairment: comparison with major depression and primary degenerative dementia.Am J Geriatr Psychiatry 1:59–66
32. Buchholtz-Hansen PE, Wang AG, Kragh-Sorensen P (1993) Mortality in major affective disorder: relationship to subtype of depression. Acta Psychiatr Scand 87:329–335 33. Burvill PW, Hall WD, Stampfer HG, Emmerson JP (1991) The prognosis of depression in old age. Br J Psychiatry 158:64–71 34. Murphy JM (1990) Depression in the community: findings from the Stirling County Study. Can J Psychiatry 35:390–396 35. Pulska T, Pahkala K, Laippala P, Kivela SL (1999) Follow-up study of longstanding depression as a predictor of mortality in elderly people living in the community. Br Med J 318:432–433 36. Pulska T, Pahkala K, Laippala P, Kivela SL (1998) Major depression as a predictor of premature deaths in elderly people in Finland: a community study. Acta Psychiatr Scand 97:408–411 37. Brodaty H, MacCuspie-Moore CM, Tickle L, Luscombe G (1997) Depression, diagnostic sub-type and death: a 25-year follow-up study. J Affect Disord 46:233–242 38. Allgulander C (1994) Suicide and mortality patterns in anxiety neurosis and depressive neurosis. Arch Gen Psychiatry 51: 708–712 39. Koenig HG, Shelp F, Goli V, Cohen HJ, Blazer DG (1989) Survival and health care utilization in elderly medical inpatients with major depression. J Am Geriatr Soc 37:599–606 40. Murphy E, Smith R, Lindesay J, Slattery J (1988) Increased mortality rates in late-life depression. Br J Psychiatry 152:347–353 41. Weeke A,Vaeth M (1986) Excess mortality of bipolar and unipolar manic-depressive patients. J Affect Disord 11:227–234 42. Black DW, Warrack G, Winokur G (1985) The Iowa record-linkage study. III. Excess mortality among patients with ‘functional’ disorders. Arch Gen Psychiatry 42:82–88 43. Coryell W (1981) Diagnosis-specific mortality. Primary unipolar depression and Briquet’s syndrome (somatization disorder). Arch Gen Psychiatry 38:939–942