Study setting and design

The Medical Research Council Elderly Hypertension Trial was a randomised, placebo-controlled, single-blind trial of the treatment of moderate hypertension (mean systolic blood pressure=160-209 mmHg, mean diastolic blood pressure <115 mmHg) in 4396 subjects aged 65-74 years, conducted from 1983 to 1989 in 226 general practices across the UK (MRC Working Party, 1992). Exclusion criteria were known serious cardiovascular or cerebrovascular disease, cancer or dementia. Subjects were tagged via their National Health Service numbers. The date of death, if this occurred, was collected until the end of follow-up in September 1994. Thus, 11 years of follow-up data were available.

Data on a range of variables were collected at baseline. Some variables were also measured at the 1-, 9-, 21— and 54-month follow-up assessments. Patients were randomised to three groups: beta-blocker (atenolol, 50 mg daily), thiazide diuretic (hydrochlorothiazide, 25 mg or 50 mg, plus amiloride, 2.5 or 5 mg daily) or placebo. The results of the main study have been published elsewhere (MRC Working Party, 1992). In summary, both treatments reduced blood pressure compared with placebo. Furthermore, the diuretic group, but not the beta-blocker group, had significantly reduced risks of stroke, coronary events and all cardiovascular events, compared with placebo.

Participants

The subjects for the current paper were those who took part in the psychiatric substudy, namely the first 2680 recruits for the main study. Of these 2680, 29 refused to participate. Of the 2651 who agreed, 2584 (97.5%) contributed adequate data on the main explanatory measures. These constitute the current sample.

Measures

Statistical analyses

All analyses were carried out using STATA version 6 (StataCorp, 1999). Univariate associations between the potential explanatory variables and mortality were assessed using classical Mantel—Haenszel methods for rate ratios. Hazard ratios were derived using Cox's regression (Reference Clayton and HillsClayton & Hills, 1996). Assessment of the proportional hazards assumptions was carried out using an Aalen cumulative incidence plot, and a formal test for heterogeneity of the effect of depression over thirds of the follow-up (each with an equal number of deaths). The effect of each continuous and ordered categorical variable on mortality was modelled first to assess whether a factor or linear relationship was the best fit for each variable. Any departure from linearity was formally tested using a likelihood ratio test. The survival probability was plotted for those with and without depression at entry, with all other variables held constant at baseline. Multivariate analyses were carried out using Cox regression to build up a model for the effect of depression on mortality, adding in potential confounding variables one at a time and assessing the contribution of depression within the model following the addition of each confounder using the likelihood ratio test. Potential interactions between depression and any other explanatory variables were first identified by looking for any suggestion of unequal rate ratios using classical Mantel—Haenszel methods and then, if interaction appeared likely, by a formal likelihood ratio test using Cox's regression.

Descriptive characteristics of the sample

Table 1 shows the main socio-demographic and clinical characteristics of the sample. Of the 2584 subjects, 42% were male and 58% female. The mean age at entry to the trial was 70.3 years. During the follow-up period, 744 subjects (29%) died. On the depression scale at baseline, 8.0% scored ≥6 (possible depression) and 4.9% scored ≥7 (depression). The percentages who died in these groups were 32% and 36%, respectively.

Table 1

Baseline characteristics of participants (n=2584)

Socio-demographic variables
Age at baseline, years (s.d.)	70.3 (2.8)
Gender, n(%)
Male	1088 (42.1%)
Female	1496 (57.9%)
Urban residence, n(%)	1703 (65.9%)
Depression
Depression score at entry1
Median (interquartile range)	1.3 (0.3-2.5)
≥ 6	197 (8.0%)
≥ 7	121 (4.9%)
Antidepressant use	95 (3.8%)
Cardiovascular risk factors
Systolic BP at entry, mmHg (s.d.)	184.0 (12.9)
Cholesterol, g/l (s.d.)	6.5 (1.2)
Body mass index (s.d.)	26.6 (4.0)
Current smoker, n(%)	533 (20.6%)
Ischaemia on ECG, n(%)	363 (14.1%)
Arrhythmia on ECG, n(%)	139 (5.4%)
Cognitive risk factors
PALT at entry (s.d.)	17.0 (1.7)
TMT at entry (s.d.)	60.5 (29.9)
NART (s.d.)	31.1 (10.8)
RPM (s.d.)	15.4 (4.1)

Proportional hazards assumptions

A formal test for heterogeneity of the effect of depression over thirds of the follow-up (each with an equal number of deaths) showed no statistical evidence of departure from the proportionality assumption (χ²=0.99, P=0.609). Thus it was reasonable to use Cox's proportional hazards to model the effect of depression on mortality.

Univariate associations between baseline variables and mortality

Table 2 describes the association between the main variables and mortality. Depression was only weakly associated when coded as a binary variable (P=0.07). Possible depression was not associated with mortality. Not surprisingly, male gender, older age, smoking, raised systolic blood pressure and ECG ischaemia were all associated with mortality. Lower cholesterol levels were also associated with increased mortality. More died in the beta-blocker group than in the diuretic group (31% ν. 26%), with the placebo group intermediate.

Table 2

Univariate associations with mortality

	Hazard ratio (95% CI)	Likelihood ratio test χ2(d.f.)	P
Demographics
Female gender	0.52 (0.45-0.60)	80.01 (1)	<0.001
Urban residence	1.17 (1.01-1.36)	4.10 (1)	0.043
Age at baseline1	1.21 (1.11-1.32)	17.98 (1)	<0.001
Cardiovascular risk factors
Smoker at entry	1.82 (1.56-2.14)	51.32 (1)	<0.001
Systolic BP at entry2	1.00 (1.00-1.01)	1.55 (1)	0.210
Mean BP over trial1	1.34 (1.23-1.47)	42.18 (1)	<0.001
Cholesterol at entry1	0.80 (0.73-0.87)	25.55 (1)	<0.001
ECG ischaemia at entry	1.37 (1.13-1.66)	9.89 (1)	0.002
ECG arrhythmia at entry	0.82 (0.61-1.10)	1.73 (1)	0.190
Body mass index1	0.93 (0.85-1.01)	2.62 (1)	0.110
Treatment group
Diuretic v. rest	0.86 (0.72-1.02)	3.2 (1)	0.078
Beta-blocker v. rest	1.11 (0.94-1.31)	1.59 (1)	0.200
Depression
Depression score
≥ 6 out of 12	1.14 (0.88-1.47)	0.90 (1)	0.340
≥ 7 out of 12	1.32 (0.97-1.79)	2.96 (1)	0.073
Antidepressant use at baseline	1.07 (0.72-1.58)	0.12 (1)	0.730
Cognitive risk factors
NART score1	0.94 (0.86-1.02)	2.02 (1)	0.160
Raven's Progressive Matrices score1	0.97 (0.88-1.06)	0.40 (1)	0.530
PALT at entry1	0.95 (0.87-1.03)	1.26 (1)	0.260
PALT coefficient
Mid v. most decline	0.79 (0.66-0.94)
Least v. most decline	0.91 (0.77-1.09)	7.35 (2)	0.025
TMT1at entry	1.05 (0.96-1.14)	0.99 (1)	0.320
TMT coefficient
Mid v. most improved	0.79 (0.78-1.12)
Least v. most improved	1.25 (1.06-1.49)	12.40 (2)	0.002

There was no association between mortality and baseline measures of intelligence and cognition. However, decline in cognitive performance over the course of the study was associated with mortality.

Association between depression and mortality controlled for confounders

Table 3 shows the change in the effect of depression on mortality when various potential confounders are added to the model. Depression (i.e. scoring ≥7 on the depression scale) becomes more powerfully associated once gender is controlled for. This reflects the higher prevalence of depression in females and the higher death rate in males. There was some confounding by cardiovascular risk factors and by urban residence. The full model, which controlled for gender, age, cardiovascular risk factors, cognitive decline and urban residence, showed that depression was still modestly associated with mortality. The effect of depression scored as a continuous variable was similar. Assuming an overall prevalence of major depression of 5% and a hazard ratio of 1.43, we calculate the population-attributable risk percentage of depression on mortality to be 2.1%.

Table 3

Cox's regression showing the association between depression at trial entry and mortality, adjusting for confounding variables

Depression (scoring ≥7 on depression scale)	Hazard ratio for mortality (95% CI)	Likelihood ratio test χ2	P value
Depression (univariate)	1.32 (0.97-1.79)	2.96	0.086
Depression adjusted for gender	1.55 (1.14-2.11)	6.96	0.008
Depression adjusted for variable 1	1.56 (1.15-2.13)	7.18	0.007
Depression adjusted for 1, 2	1.50 (1.10-2.05)	5.99	0.014
Depression adjusted for 1, 2, 3	1.51 (1.11-2.05)	6.05	0.014
Depression adjusted for 1, 2, 3, 4	1.48 (1.08-2.01)	5.47	0.019
Depression adjusted for 1, 2, 3, 4, 5	1.43 (1.03-1.98)	4.21	0.040

Possible depression (i.e. scoring ≥6 on depression scale) had a modest effect on mortality after controlling for gender (hazard ratio=1.33; 95% CI 1.12-1.72, P=0.041). The relationship was confounded by cardiovascular risk factors and by urban residence. After adjustment for all these factors, the hazard ratio was 1.22 (95% CI 0.92-1.63,P=0.18).

Interaction between depression and gender and depression and age

We detected an interaction term (χ²=3.05, 1 d.f.,P=0.081) between depression and gender. Males with depression had considerably increased risk of dying (hazard ratio=2.12, 95% CI 1.28-3.52) even after adjustment for cardiovascular and cognitive risk factors and for urban residence. In females the risk was not significantly increased (hazard ratio=1.14, 95% CI 0.75-1.73). There was also a significant interaction term between age and depression. Younger people with depression were at increased risk (Table 4), but depression in those aged over 70 years at baseline did not appear to be associated with mortality.

Table 4

The interaction between depression on mortality, by age group¹

Depression (scoring ≥7 on depression scale)	Hazard ratio for mortality (95% CI)	Wald test	P
Effect of depression in youngest third (aged 65-69 years at study entry)2	2.26 (1.34-3.80)	3.06	0.004
Effect of depression in middle third (aged 69-71 years at study entry)2	1.27 (0.69-2.36)	0.77	0.44
Effect of depression in oldest third (aged 72-75 years at study entry)2	0.95 (0.54-1.68)	-0.18	0.89

Antidepressant use and mortality

There were no significant associations between antidepressant use at trial entry and mortality. The hazard ratio for antidepressant use at entry, adjusted for gender, cardiovascular risk factors, age, urban residence and depression score, was 1.01 (95% CI 0.65-1.55, P=0.99). The hazard ratio for antidepressant use during the trial (linear increase in quarters), adjusted for gender, cardiovascular risk factors, age, urban residence and depression score, was 1.06 (95% CI 0.99-1.13, P=0.084).

Study limitations

The study restricted the sample to a relatively healthy group. Thus, it is unlikely that the association is due to the presence of physical illness among those rated as depressed at baseline. However, there were potential methodological flaws in the design of this study. First, we concentrated on a high-risk group of elderly subjects with hypertension who had agreed to participate in a randomised trial; therefore the results may be of limited generalisability and may not apply to younger and normotensive populations. Second, although we were able to control for a wider range of confounders than most previous studies, smoking was only coded as a binary variable. This overaggregation might lead to residual confounding. However, this may not be of major concern because controlling for cardiovascular risk factors only reduced the hazard ratio from 1.56 to 1.50. One would expect a much more dramatic decrease in the hazard ratio when controlling for this confounder if residual confounding was capable of explaining the association. Third, our measure of depression was a questionnaire. It is possible that this is less valid than clinical interviews, although the SelfCARE—D has been shown to have acceptable validity and reliability in this age group (Reference Bird, Macdonald and MannBird et al, 1987; Reference Upadhyaya and StanleyUpadhyaya & Stanley, 1997). If there was misclassification of cases it is most likely that this would be random misclassification, and thus lead to underestimation of effect size.

Major versus minor depression

Our finding is consistent with recent findings from the longitudinal Amsterdam study reporting a greater effect on mortality for major depression than for minor depression (Reference Penninx, Geerlings and DeegPenninx et al, 1999). Given the finding by Blanchard et al (Reference Blanchard, Waterreus and Mann1994), using a comprehensive structured clinical assessment for depression in the elderly (GMS AGECAT; Reference Copeland, Dewey and Griffiths-JonesCopeland et al, 1986), that the mean score for cases of depression in an elderly community sample was ≥ 7 on the SelfCARE—D, it is likely that the ≥7 cut-off provides greater specificity and/or indicates greater severity than the ≥6 cut-off. The effect of depression that we found appeared greater and was more statistically significant at the higher cut-off of the SelfCARE—D (i.e. at ≥7 out of 12 compared with ≥6). At the ≥6 cut-off 8% were depressed, and at the ≥7 cut-off 4.9% were depressed. If ≥6 represents possible/minor depression, this prevalence is some-what lower than reported from community studies (e.g. 12.8%; Reference Penninx, Geerlings and DeegPenninx et al, 1999). Given that our population is from a randomised trial, this probably reflects the healthy volunteer effect.

Male gender and mortality from depression

The increased risk associated with depression was present in males (but not females) and in the younger age group in this sample. The interaction terms that we detected were of only borderline statistical significance (0.05 <P <0.1) and may represent type I error. However, our findings are consistent with recent papers suggesting that the effect of depression on mortality is greater in men than in women (Reference Mallon, Broman and HettaMallon et al, 2000). Furthermore, Pennix et al (1999: p. 63) reported that although major depression predicted mortality in males and females, minor depression predicted mortality only in males. We are not aware of any research to date that has detected age differences in the effect of depression on mortality, so this finding requires replication.

Mediating factors

Given the increasing evidence for the association between depression and cardiac mortality (Reference Penninx, Beekman and HonigPenninx et al, 2001) and the association between male gender and cardiovascular disease, it could be that the effect of depression is partly explained by cardiovascular and/or cerebrovascular disease, which has not been dealt with in the analysis. Although suicide may be another explanation for the excess mortality in males with depression, this is unlikely because it is a rare cause of death in community samples (Reference Penninx, Geerlings and DeegPenninx et al, 1999). It would be valuable to carry out further analysis in our sample to look at cause-specific mortality.

The literature that has previously linked depression and mortality has proposed two main mechanisms. The first relates to life-style, namely that depression may be a more distal risk factor for a number of risk factors that increase morbidity and mortality from cardiovascular and other physical disease. For example, people with depression may smoke more, exercise less, adhere less to treatment and drink more alcohol (Reference AllgulanderAllgulander, 1994; Reference Blanchard, Waterreus and MannBlanchard et al, 1994). We have been able to study markers for cardiovascular disease, including ECG ischaemia and smoking. Although these sorts of variable have been dealt with as confounders in this analysis, they are, strictly speaking, intermediary factors on the same causal pathway. Our analyses suggest that the association is unlikely to be entirely explained by these risk factors, although we were unable to control effectively for exercise or alcohol intake.

The second mechanism involves a group of risk factors that relate to possible biological change occurring in depression. Depression is known to be associated with increased cortisol, increased platelet co-agulability and changes in heart rate variability, perhaps indicating a cardiovascular system more prone to arrhythmias (Reference Dalack and RooseDalack & Roose, 1990; Reference Mikuni, Kagaya and TakahashiMikuni et al, 1992). These findings are predominantly derived from small clinical samples of patients with severe depression. The next step will be to explore such associations in large epidemiologically based samples. We were able to adjust for ECG evidence of arrhythmia and hence this does not appear to have mediated the association between depression and mortality.

Future research

In terms of public health implications, depression is a common exposure. Most people with depression in the community do not receive any specific treatment for depressive symptoms (Reference Schoevers, Beekman and van TilburgSchoevers et al, 2000). Two early studies suggest that treatment of depression may lower mortality, especially in men (Reference Avery and WinokurAvery & Winokur, 1976; Reference Craig and LinCraig & Lin, 1981). The findings here, together with the previous literature, suggest the need to investigate the effect on mortality of vigorous treatment of confirmed depression.