Research Interests
Sara Geneletti - Lecturer in Statistics
Causal Inference
Causal inference is the name given to the the area of
statistical methodology aimed at identifying and estimating
effects of interventions. The reason we care about the effects of
interventions is that generally in social science and
epidemiology, we are actively interested in intervening. For
example, we want to know if a particular social programme works
because if it does, it might result in new policies -- new
interventions. We want to know whether a drug works because we
want to prescribe it -- i.e. intervene.
The big problem in
causal inference is that most of our data do not involve
interventions but rather passive observation (except clinical
trials) and we have to go somehow extract information about
interventions from them. Essentially it boils down to overcoming
confounding.
There are number of approaches to causal
inference, all involve enhancing (explicitly or implicitly) the
language of statistics with ways of formalising interventions. The
most common approaches use counterfactuals or potential responses.
I prefer to use an approach which uses statistical decision theory
to formalise interventions as decisions as counterfactuals don't
really exist. My thesis was about exploring aspects of causal
inference in the statistical decision model.
The regression discontinuity design in
Public Health and Epidemiology
The RD design is a quasi-experiment which
takes advantage of an imposed threshold resulting in a treatment
(e.g. a drug prescription guideline) to identify the causal
effects of the treatment. Together with collaborators from UCL
(Gianluca
Baio and members of the UCL
Research Department of Primary Care and Population Health) we
are estimating the effects of a number of drugs, notably Statins,
in clinical practice at the population level. We have recently
been awarded an MRC grant to do this.
Connections between causal inference estimators
I am also currently trying to formalise
the relationships between different causal inference estimators
across the disciplines that use them (statistics, econometircs,
epidemiology).
Bayesian statistics
The Bayesian approaches appeal to me not only because
they are often easier to implement and understand than straight frequentist approaches but
especially because I espouse the philosphy of subjective probability. Bruno de Finetti said that "Probability
does not exist" and I tend
to agree.
Modelling bias
It
is important to try and come up with models that will adjust for
bias in observational studies. I am especially interested in
structural biases such
as confounding, mediator and selection bias. I am currently
working on models to adjust
for selection bias in case-control studies and have used the
statistical decision model for causal inference to tackle mediator
bias.
Evidence synthesis
Evidence synthesis is the rigorous statistical combination of
data from different sources to adjust for bias and improve
inference. Recently it has become increasingly feasible to merge
information from large routinely collected data sets (e.g. Census)
with data from small observational studies to adjust for biases in
the latter. A single data source, e.g. a survey or a study, cannot
fully answer questions of interest to policy makers as the
participating individuals are not representative of the target
population. However, combining small biased with large
representative data sets can sometimes provide better answers. I
am interested in developing evidence synthesis methods based
principally on re-weighting techniques.