The Authors Reply

We thank Davey Smith et al. for their comments (1). Regarding the relationship between the inflammatory marker C-reactive protein and coronary heart disease (CHD), Davey Smith et al. provide a cautionary example of how observational data—no matter how large the study—may still not amount to proof of causality (1). Although several meta-analyses have shown a robust association between the two factors (2–5), Mendelian randomization studies using genetic instruments suggest that C-reactive protein may not be a cause of heart disease (6, 7). In contrast, another inflammatory marker, interleukin 6, may contribute to the risk of heart disease (8).

Mendelian randomization can potentially inform the process of developing new therapies (and point to associations that are likely to be noncausal) prior to proceeding to expensive phase III trials. In social epidemiology, however, identifying a genetic instrument with which to explore causality is difficult for most exposures. For example, we are unaware of any genetic variants that can be used as instruments for evaluating job strain. We have previously examined this exposure using nongenetic instruments, such as rates of hospital-ward bed occupancy for a study of job strain among nurses (9), but in general, finding a convincing instrument is hard and relies on the chance availability of natural experiments.

For the purpose of establishing cause and effect, randomized controlled trials (RCTs) remain the gold standard. Demonstrating a robust association between exposure and outcome through meta-analysis of observational data may inform the design of RCTs. First, meta-analysis provides an evaluation of the expected effect size informing decisions about the size of trials to be implemented. This is likely to be an upper- rather than a lower-bound effect, as many observational associations have been refuted or found to be inflated when tested in RCTs (10, 11). For job strain, for example, the standardized effect size for CHD risk based on individual-participant meta-analysis of published and unpublished observational data was only one-seventh that for lifestyle factors such as smoking, physical inactivity, and obesity (12, 13). This suggests that very large RCTs are needed to confirm or refute a causal job strain-CHD association (Table 1). Second, meta-analytical information on expected effect size may facilitate the evaluation of more fundamental questions, such as whether the logistical challenges and financial requirements of large-scale RCTs are justified. Job strain has been examined in relation to employees' mental well-being (14). In the light of current evidence (Table 1), adding randomization and sensitive surrogate markers of cardiovascular risk to such interventions might be a more feasible next step than a large-scale RCT with CHD incidence as the primary outcome.

Table 1.

Current Evidence on the Association Between Job Strain and Coronary Heart Disease

Type of Evidence	Study Quality (Protection Against Bias)	Consistency of Results	Precision of Findings	Absolute Risk Difference (Exposed vs. Unexposed)
Intervention studies: No RCTs, some nonrandomized interventions on work characteristics and CHD risk factors, some natural experiments	Moderate/poor	No systematic reviews/meta-analyses with which to assess consistency	Low (underpowered)	Unknown
Observational studies: 2 literature-based meta-analyses, 1 individual-participant meta-analysis (total n > 250,000)	Moderate	High	High (relative risk = 1.3)	4 CHD events per 1,000 in 10 yearsa
				Population attributable risk 3%–4%

Abbreviations: CHD, coronary heart disease; RCT, randomized controlled trial.

^a In the total working population. Based on data from Kivimäki et al. (13).