Evidence does not support link

EDITOR—The summary points of Hemingway and Marmot's review of psychosocial risk factors for coronary heart disease may be misleading.1 The first of these states that “prospective cohort studies show a possible aetiological role for type A/hostility.” However, of the four prospective studies of hostility, only two show any significant association between hostility and coronary heart disease (one for women only). Six of the nine aetiological studies of type A behaviour also show no association with coronary heart disease. The other three studies give no information on completeness of follow up or whether outcomes such as angina were assessed in a blinded manner. One of these studies had minimal adjustment for confounding. No study showed any prognostic role for type A behaviour or hostility. Taken together, these studies do not represent robust evidence that these psychological variables have an important role in the development or prognosis of coronary heart disease.

Funnel plot of studies of hostility and coronary heart disease before (top) and after (bottom) adjustment for degree of confounding

• Download figure
• Open in new tab
• Download powerpoint

The authors also rightly suggest that the role of publication bias should be considered. We recently investigated this by re-examining a meta-analysis, also cited by Hemingway and Marmot, which pooled several prospective and retrospective studies of hostility and coronary heart disease and concluded that hostility was an independent risk factor.2 A funnel plot of these studies suggests publication and related biases (figure (top)), indicating underpublication of negative results However, the apparent publication bias is more likely to be a result of poor methodological quality—in particular, inadequate adjustment for confounding in the primary studies. When we took this into account (using regression to adjust the effect sizes according to the number of variables used as adjusters in the primary study) the relation between hostility and coronary heart disease disappeared and the funnel plot became symmetrical (figure (bottom)). This supports the view that publication bias among these studies should be considered only when other sources of bias, such as the internal validity of the included studies, are taken into account, as suggested by Egger et al.3

Claims about the toxicity of type A behaviour and hostility have led to the development of behavioural “treatments” for these putative risk factors. A recent editorial also went as far as to suggest that, as hostility appeared to be a “toxic component,” we should all be nice to each other.4 Perhaps in the light of Hemingway and Marmot's review, and the above comments, this advice should be reconsidered.

Authors' reply

EDITOR—We welcome Petticrew et al's application of methods more commonly used in the context of clinical trials to our review of observational follow up studies. We agree with their conclusion that these studies “do not provide robust evidence” of a relation between coronary heart disease and hostility or type A behaviour. This is consistent with the conclusion we reached, that the evidence for hostility and type A behaviour was less consistent than that for depression and anxiety and social supports.

However, their attempt to take the systematic review further raises some problems. Firstly, Petticrew et al rightly raise the issue of completeness of follow up and degree of blinding of outcomes such as angina. But they are wrong to say that the two studies incorporating these soft outcomes were unblinded. Furthermore, such methodological issues should be examined systematically across all studies, not just the positive ones, since the direction of any bias cannot be assumed.

Secondly, Petticrew et al reanalysed a hostility meta-analysis and showed an asymmetry in the plot of study size against effect size. Such asymmetrical funnel plots may be due to several potential biases Petticrew et al propose that inadequate adjustment for confounding is likely to be most important because their adjustment for the number of confounders used in the primary studies made the plot more symmetrical, although the difference was not formally tested. As discussed in our review, psychosocial factors may have effects on disease which are mediated by health related behaviours. Thus if the causal chain involved hostile people being more likely to smoke,1 then adjusting for smoking as a confounder would be misguided Furthermore, the number of confounders which were allowed for may itself be confounded. It is possible, for example, that studies which differ in respect to adjustment for confounding also differ in terms of overall methodological quality.

Empirical evidence of health benefit should not be included among the grounds for the ethical imperative “be nice to each other.”