Investigation of therapeutic causal pathways in anti-hypertensive trials using longitudinal data.

Lead Research Organisation: University of Manchester
Department Name: Medical and Human Sciences


The most important test for a new drug - whether it reduces the risk of future disease or early death - is addressed in clinical trials. The question of how it works will have been addressed at earlier stages in its development but often much can still be learnt from data collected on patients, especially data collected from patients monitored in trials. Data that has already been collected does not involve any extra cost and thus offers an economic way of addressing such questions.

Under recent guidelines (NICE) to doctors for treating hypertension, beta blockers are no longer a preferred option. One hypothesis about the causal pathways of such medications is that they ?work? by lowering blood pressure which in turn reduces risk of cardiovascular disease, but there might be other effects ? positive or negative - say on cholesterol or blood glucose, operating simultaneously. Quantification of the strength of different pathways is important because it can strengthen previous hypotheses about mechanisms, suggest new ones, feed into the development of new better drugs and might even reveal unintended biological effects.

Correct statistical analysis of data from trials is important to prevent biased answers about effects. While scientists agree on the correct method to address: ?Does the treatment work?, methods for quantifying the strength of different causal mechanisms are more complex and less developed. We wish to strengthen this methodology by combining the best ideas from two fields of statistical research; we would then apply the chosen methods to two large trials of beta blocker therapy for hypertension. The study would provide answers to questions about direct and indirect effects for this group of drugs. It would also provide new methods for analysis and these could be used by other studies asking similar questions.

Technical Summary

Whilst ?intention to treat? analysis remains the preferred approach for addressing effectiveness of drug therapies, the elucidation of mechanisms linking drug to outcome, or potential cause to effect, is also important. In hypertension research, the extent to which a greater reduction in cardiovascular disease risk in one arm of a trial is mediated through greater blood pressure reduction, and the significance of unintended adverse changes, such as an increase in diabetes, are of interest. The recent update in NICE guidelines, recommending a switch from a blockers as a first-line treatment of choice, gives renewed interest to data from MRC trials of these therapies. These trials are rich in information on intermediary variables including repeated measures of blood pressure and blood glucose. Paradoxically such data, in this and similar trials, tends to be underutilised, perhaps for fear of violating the ?intention to treat? principle. However while ITT protects against bias when asking whether a treatment ?works?, it can under-estimate causal effects.

Two recent developments in statistical methods are relevant to the quantification of direct and indirect (ie via the intermediary) causal effects of drugs on disease risk. ?Joint modelling? methods combine models for repeated measures and models for ?survival? data; they would allow hypotheses about effects on mortality mediated through blood pressure and by other pathways to be expressed in a single model, and use data efficiently. Recent theoretical advances by us13,14,20 has introduced new, more flexible methods in this field. ?Causal modelling? describes a general class of methods for unbiased estimates of causal parameters, ie parameters of models for underlying causal relationships, which are direct measures of biological effect. This research has shown that conventional statistical modelling (including joint modelling) can produce biased estimates of causal parameters when there is time?dependent confounding. Time-dependent confounding of dose-response and other relationships in hypertension trials is likely if treatment affects blood pressure and vice versa. However development of causal models for the problem of quantifying direct and indirect effects is limited to simplistic datasets. We propose to develop new methods of analysis for estimating causal pathways which combine the different advantages of these two schools. We will apply these methods to data from MRC trials to quantify the role of key intermediate states and events, in particular blood pressure and diabetes, between a blocker therapy and cardiovascular events. Our methods will be applicable to other studies of this type.


10 25 50