Paradigm Shift in Causal Thinking in Epidemiology

Abstract

Professor Shapiro expressed concern with the general direction of epidemiology. Whereas the previous position was that we are unable to interpret limited data, this emphasis has shifted. In particular, the randomized controlled trial (RCT) now is unassailable in the US.

Shapiro identified the current issues as being (1) a loss of scepticism in the field, (2) the illusion of accuracy and validity fostered by large databases and their seduction into believing statistical significance, and (3) an over-interpretation of RCTs.

Examples of significant errors in epidemiology include spurious associations found for reserpine and breast cancer, calcium channel blocker and cancers in a follow-up study, and fertility drugs and ovarian cancers shown in meta-analysis. The lesson is that one must always be sceptical. The duty of epidemiologists engaged in research is to assume the null hypothesis. They should accept this until the null can be rejected with confidence.

Related to the loss of scepticism are the following. Small relative risks are often uninterpretable. If there are fragile data based on small numbers then the data need to be almost free of error or biases. If not, they can change the estimate in such a way as to make the result uninterpretable. Black box statistics might produce nontransparent results. Particularly the meta-analysis of data is often misleading.

When causal inference is made in epidemiology, there is a hierarchy of study types as to the level of proof they supply. From highest to lowest, these are RCT, cohorts, case-control, cross-sectional, ecological and descriptive studies. Case-control and cohort studies are simply alternate methods of sampling the population with neither method in principle superior to the other. RCTs are still considered to provide the highest level of proof, but in epidemiology, they cease to be RCTs. RCTs covering a time of 3-7 years become epidemiological follow-up studies, similar to regular cohort approaches.

An example of how a study designed to be a randomized trial turns into an observational study is the WHI study, which explored oestrogen plus progestin (EP) vs a placebo. The clinical prediction is that women will experience some specific symptoms during menopause. This is confirmed by the fact that the proportion that unblinded due to the occurrence of vaginal bleeding in the EP group was 44.4% and 6.2% in placebo. The relative risk (RR) of this is 6.5, the risk difference is 37.6%. In addition, there is the potential for detection bias, as women are going to examine their breast more carefully because they are on HRT, if they are unblinded. The cumulative proportion of discontinuation in the HRT group was 42%, in placebo 38%. There is a large pool of women with otherwise silent breast cancer. The annual incidence of breast cancer was 3.8/1,000 in WHI for EP, 3.0/1,000 for placebo, the hazard ratio (HR) was 1.26 (1.00–1.59; p > 0.05). The corresponding risk difference is 0.8/1,000 per year. Other outcomes show similarly small differences, with an excess of 0.07% for CHD, 0.08% for stroke, pulmonary embolism 0.08%, i.e. each with an approximate risk difference of 1/1,000. All these are enormously susceptible to bias, particularly pulmonary embolism.

The Million Women Study showed a risk estimate of 1.3 (1.22–1.38) for oestrogen and 2.0 for combined therapy; for tibolone it was 1.44. When seen on the population scale, this amounts to an incidence among EP users of 5.6/1,000 per year; the excess is 2.5/1000 per year.

It cannot be said from these data whether HRT does or does not increase the risk because epidemiology is too crude a science.