Depression is a highly prevalent condition, as confirmed by many national (Reference Kessler, McGonagle and ZhaoKessler et al, 1994; Reference AngstAngst, 1997; Reference Lepine, Gastpar and MendlewiczLépine et al, 1997) and international studies (Reference Ustun and SartoriusUstun & Sartorius, 1995), with many possible outcomes. Predicting outcome at the time of diagnosis can have a strong clinical impact, since it can help to distinguish people in need of specific treatment from those likely to recover spontaneously. Nevertheless, studies of predictors of outcomes in depression show mixed results (Reference Bagby, Ryder and CristiBagby et al, 2002). Research challenges include the need to determine if any specific predictor is independent of other predictors, and to know whether predictors identified in the USA and Europe are also valid in other settings.

Some of these challenges could be clarified through a cross-cultural longitudinal study of depression. The Longitudinal Investigation of Depression Outcomes (LIDO) is a multicentre, cross-national observational study which followed patients with depressive disorders in primary care settings for 12 months in six countries (Reference Chisholm, Amir and FleckChisholm et al, 2001; Reference Herrman, Patrick and DiehrHerrman et al, 2002). The objective of our study was to use the LIDO data to identify baseline factors associated with 9-month clinical outcomes across six sites.

METHOD

The design, instruments and methodology of the LIDO study are described in detail elsewhere (Reference Chisholm, Amir and FleckChisholm et al, 2001; Reference Herrman, Patrick and DiehrHerrman et al, 2002). Briefly, patients presenting at primary care centres in six sites (Barcelona, Spain; Be'er Sheva, Israel; Melbourne, Australia; Porto Alegre, Brazil; Seattle, USA, and St Petersburg, Russia) were screened for symptoms of depression. Those meeting inclusion criteria – new and/or untreated episode and a score of over 16 on the Center for Epidemiological Studies – Depression scale (CES–D; Reference RadloffRadloff, 1977) – were interviewed and assessed with a standardised diagnostic instrument for major depression, the Composite International Diagnostic Interview (CIDI; Reference Robins, Wing and WittchenRobins et al, 1988).

RESULTS

A total of 968 primary care patients across all sites were diagnosed as having major depression at baseline. Sample sizes ranged from 142 in Seattle to 185 in Barcelona. At the 9-month follow-up assessment the proportion of those in complete remission ranged from 25% in Porto Alegre to 48% in Barcelona (Table 1); the withdrawal rate was 15.4%. The likelihood of receiving potentially effective antidepressant or mental health treatment at 3 months or 9 months did not differ across the six sites between the patients who were in complete remission and those who were not (Table 2).

The comparison of scores at baseline for the complete remission and non-complete remission groups using univariate analysis is shown in Table 3. For all variables with a significant difference, the complete remission group did better at baseline. Comparison of socio-demographic data for the two groups showed that people in complete remission were younger, had more years of schooling and were more often employed at baseline. There was no difference in gender and marital status between the two groups. Comorbidity data show that those in complete remission had fewer medical conditions, dysthymia and anxiety disorders. No difference was found for alcohol risk. This group also had less intensity of depression as measured by the CES–D and better quality of life and functioning as measured by the QLDS, WHOQoL and SF–12 at baseline, as well as fewer previous episodes of depression. There was no difference in number of work days lost.

Univariate analysis was performed for each country separately (data not shown). The results showed the same tendency as in the pooled sample, although statistical significance was not obtained for all predictors in all sites, probably because of sample size limitations. Two predictors not significant for the pooled sample were identified for individual centres: low alcohol risk was a baseline significant predictor of complete remission in Be'er Sheva and St Petersburg, and a lower percentage of females was associated with complete remission in St Petersburg.

Multivariate analyses (logistic regression) were used to evaluate the relative weight of each predictor, using centres as a covariate in all models in the pooled sample (n=968). In the first step, socio-demographic data were entered adjusted for centres (Table 4); education and employment were the significant predictors. In step 2 (depression severity), education and employment continued to be significant predictors and CES–D score and number of previous depressive episodes were also significant predictors after adjusting for socio-demographic data (Table 5). In step 3, comorbidity variables were entered adjusted for CES–D and socio-demographic data: education and CES–D score remained significant and medical comorbidity was a significant predictor, adjusted for socio-demographic data and severity of depression (Table 6). In step 4 (quality of life), QLDS and the environment domain of WHOQoL–Bref were the significant predictors, adjusted for socio-demographic data, severity of depression and comorbidity (Table 7). In this analysis CES–D score and education were no longer significant, probably because of the high degree of co-linearity between quality of life measures, severity of depression and education. Finally, in step 5, when key life events were entered in the model adjusted for the preceding families of variables (socio-demographic data, severity of depression and comorbidity), education, QLDS and key life events were the significant predictors (Table 8). All significant predictors were modest in effect, since the odds ratios were close to 1.

When logistic regression was run for individual sites using selected predictors (Table 9), complete remission was predicted by key life events in three sites (St Petersburg, Porto Alegre and Seattle), medical conditions in Seattle, QLDS in Be'er Sheva and WHOQoL–Bref environment domain in Melbourne. No variable appeared as a consistent predictor across all sites. However, sample sizes at individual centres were modest. A logistic regression model for all sites including site×predictor interactions found that variation in predictors across sites was not statistically significant (data not shown).

DISCUSSION

There are few studies focused on factors associated with depression outcomes in primary care, and most of them have been conducted in North America and Europe. The data provided by the LIDO study have important implications since this is one of the few studies with longitudinal data in a cross-national perspective.

Our first major finding is the low proportion of complete remission among participants at the 9-month follow-up at most of the study sites. Since major depression (and even sub-syndromal depression) is highly associated with present disability (Reference Wells, Steward and HaysWells et al, 1989) and future relapse or recurrence (Reference Keller, Lavori and RiceKeller et al, 1986), the low prevalence of complete remission at follow-up in this cross-cultural naturalistic study (from 25% in Porto Alegre to 48% in Barcelona) is an important finding. Mynor-Wallis et al (Reference Mynor-Wallis, Gath and Baker2000) in the UK found higher proportions (56–66%) of complete remission at 1-year follow-up in a non-naturalistic study in a primary care setting. Simon (2000) found 45% remission from a 6-month follow-up period in a naturalistic primary care study in the USA, which was closer to the highest proportion found in the present study. Differences between studies in overall remission rates may, of course, reflect differences in screening, selection procedures or levels of treatment. In any case, major depression is not a benign condition for the majority of primary care patients.

The second important finding of our study is that some baseline characteristics emerged as modest predictors of outcome, using pooled data from the six study sites. Education, key life events and QLDS were the final predictors after adjusting for centres, socio-demographic data, severity of depression, comorbidity, general quality of life and key life events. Scores on CES–D and WHOQoL–Bref environment domain could also be considered as potential predictors since possible co-linearity with QLDS and education, respectively, could explain why these variables left the model in the final step. These findings were consistent with previous studies in which none of the reviewed variables was a particularly powerful or consistent predictor of remission in depression (Reference Bagby, Ryder and CristiBagby et al, 2002). Our study has supported this conclusion using a cross-cultural sample, suggesting that this conclusion could be applied for primary care patients of heterogeneous countries.

In reality we do not know what the relationship is between predictors and outcome in depression: are they actually moderators (or conditions which are necessary for the outcome to occur) or mediators (conditions closer to a causal model)? Multivariate analysis does not answer this important point. Two possible attitudes toward multivariate analysis could be assumed: the first is to throw in all the possible predictors to see what happens; the second is to construct theoretical models despite the lack of evidence but at least with a rationale to guide the analysis (Reference Victora, Huttly and FuchsVictora et al, 1997). We used the second option, with a hierarchical model entering variables in ‘families’ in a series of steps. For example, we found – interestingly – that quality of life rather than intensity of depression remained in the final step after adjusting for all variables. Some authors have suggested that quality of life and depression are in fact different facets of the same construct, or ‘tautological measures’ (Reference Katschnig and AngermeyerKatschnig & Angermeyer, 1997). As far as we know this is the first time that quality of life has emerged as a possible predictor of depression outcome. Although this might be seen as a statistical artefact (due to co-linearity), it could be regarded as a stimulus for the development of more complex models to elucidate the relationship between depression and quality of life.

There are some limitations to our study. Sites were selected on the basis of their research capacity and importance in terms of the primary care system of the city they were part of. Moreover, the screening involved a convenience sample of clinic attenders at each site. The sampling was relatively systematic; however, this was not a representative study of primary care attenders at the sites or in the six countries, nor was it a representative population study. Consequently, it is not possible to state that the sample is representative of the primary care population of each site. Rather, the LIDO study was designed to screen for and recruit patients with depression in primary care in six clinical and cultural settings as different from one another as possible, reflecting the heterogeneity of the countries involved in this project.

At least three possible hypotheses could be put forward to explain why we found only modest predictors of complete remission across a pooled sample from different sites. First, the design or measures used in the LIDO study might not have been adequate to identify more robust predictors. Although we cannot exclude this possibility, we can point to several important strengths: a systematically identified cohort, use of instruments well tested in cross-national research and a low rate of attrition. One area of potential difficulty with this conclusion is found in the heterogeneity of the study sites involved. Chisholm et al (Reference Chisholm, Amir and Fleck2001), comparing the assessment of local health systems and resource utilisation across the six LIDO study sites, concluded that:

Although this is an important limitation of the study, it also reflects the primary care setting from which the study population was drawn. The heterogeneity of the sites could explain why only modest predictors could be identified. The strategy we adopted to control this important confounding effect was to use logistic regression adjusted for centres. A second hypothesis to explain the lack of consistent predictors for complete remission across sites is that depression is really not predictable. It is impossible either to confirm or refute this hypothesis with the empirical data found in the LIDO study, since it involves the interpretation of conceptual issues. The concept of major depression has been criticised by many authors and researchers. For example, Van Praag considers that mood disorders as defined by the DSM system are not properly validated and are utterly heterogeneous and overlapping; this author states, ‘The assumption that such presumably pseudo-entities will be carried by single well-definable pathophysiology is implausible’ (Reference Van PraagVan Praag, 1993). As a result, one should consider that it would also be implausible to find consistent predictors across sites if we are dealing with a heterogeneous entity. Finally, a third possible explanation is that the course of depression is somewhat predictable, but that the predictors themselves vary or are other than those under investigation.

APPENDIX

Clinical Implications and Limitations