MEETING OF THE PRIMARY HEALTH CARE STUDY GROUP
Monday 22nd November 2010
Royal Statistical Society
Errol Street, London
2-5pm
All welcome. No pre-registration is necessary
Theme: Causal /Explanatory Statistics
Speakers:
Jon Williamson (School of European Culture & Languages, University of
Kent)
Interpreting causal claims in medicine. Causal claims in biomedical
contexts are ubiquitous albeit that they not always made explicit. This
talk, based on joint work with Federica Russo, addresses the question of
what causal claims mean. It is argued that in medical contexts causality
ought to be interpreted according to the epistemic theory. According to
this approach, causal claims tell us about which inferences (e.g.,
diagnoses and prognoses) are appropriate, rather than about the presence
of some physical relation analogous to distance or gravitational
attraction. It is shown that the epistemic theory has important
consequences for medical practice, in particular with regard to the
notion of evidence.
Richard Emsley (School of Community Based Medicine, University of
Manchester)
Causal mediation analysis in psychological treatment trials. Mediation
analysis is increasingly being used as a statistical tool for examining
mechanisms within psychological treatment trials, where potential
mediators can include compliance and fidelity to treatment, aspects of
the treatment process and use of concomitant medication. The literature
on statistical mediation analysis is dominated by methods for analysing
a single mediator and single endpoint. These methods make strong
assumptions regarding the absence of unmeasured confounding between the
mediator and the outcome, and causal interpretations are unlikely to be
valid if these assumptions do not hold.
Alternative approaches (referred to as causal mediation analysis) have
been proposed which address this issue, and provide valid causal
estimates of direct and indirect effects under different assumptions.
In this talk, we extend the application of causal mediation analysis to
repeated measures of mediators and outcomes. We use a finite mixture
model approach, fitting random coefficient models within latent classes
using maximum likelihood estimation; an approach known as growth mixture
modelling. We illustrate the methods with application to psychological
treatment trials in people with psychosis.
Lisa Hampson (Dept. Of Social Medicine, Bristol University)
In randomised controlled trials (RCTs) not all patients will receive
their allocated treatment, sometimes for prognosis-related reasons. We
consider RCTs where compliance is all-or-nothing and the novel
intervention is unavailable on the control arm. Whilst the
intention-to-treat analysis estimates effectiveness, we may also wish to
estimate the causal effect of intervention in those who receive it.
We discuss the problem of making such causal inferences in survival
trials and compare three methods for estimating the causal effect of
treatment in a proportional hazards model: the C-PROPHET estimator of
Loeys & Goetghebeur (2003), the method of Cuzick et al. (2007) and the
rank-preserving structural accelerated failure time model of Robins &
Tsiatis (1991). We describe how each method is extended to estimate the
effect of treatment in a proportional hazards model adjusting for
baseline prognostic factors in the presence of censoring; incorporating
baseline information is important in this context since if hazards are
proportional adjusting for covariates, they will not be marginally.
We assess the performance of each method via simulation and investigate
the impact of basing inferences on a mis-specified model omitting needed
covariates. We find that while each of the three estimators performs
similarly in terms of bias, there are differences in precision. We
conclude that presenting estimates of causal effects alongside
intention-to-treat estimates can be useful in illustrating the potential
benefits of a new intervention.
Followed by Discussion in groups
5pm Close
Further details from www.rss.org.uk/diary
Meeting organiser: Maurice Marchant ([log in to unmask])
You may leave the list at any time by sending the command
SIGNOFF allstat
to [log in to unmask], leaving the subject line blank.
|