Tuesday 25th April 2017

Causal inference, instrumental variables and their applications

Paul Clarke - Professor of Social Statistics, Institute for Social and Economic Research, University of Essex [Presentation]

Estimating causal effects on binary outcomes using instrumental variables

Instrumental variables analysis is synonymous with the use of two-stage least squares (2SLS) to estimate causal effects. However, 2SLS is consistent only if the relationship between exposure and outcome can be represented by a linear model, which poses problems for binary outcomes where this model is nonlinear (e.g. logistic regression). In this talk, I set out the specific problems faced when analysing binary outcomes, especially when we wish to estimate odds ratios, and critique the available alternatives to 2SLS in terms of robustness and ease of implementation. I will also argue that statisticians should consider including the so-called linear probability model in their toolboxes whenever they analyse binary outcomes using instrumental variables.

Peter Thwaites - Lecturer in Statistics, School of Mathematics, University of Leeds [Presentation]

Causal discovery and inference with Chain Event Graphs

It has been observed that causal models are much more robust than non-causal models to changes in the environment; and this has recently stimulated considerable interest in the learning of causal models. Chain Event Graphs (CEGs) were created for the representation and analysis of asymmetric discrete statistical models. They encode the full model conditional independence structure in the topology of the graph, and several papers on CEGs have been concerned with model selection and causal analysis. This talk draws on these two areas to show how asymmetric causal models can be learnt, and their causal hypotheses tested. This is done by scoring each CEG in a candidate set of possible models, and identifying the Markov equivalence classes of those CEGs which score highly. If there are collections of variables whose orderings are consistent throughout the equivalence class, we hypothesise that these orderings are causal. Establishing whether a probability distribution on a causally manipulated CEG is identifiable from the probabilities of the idle system is generally straightforward, so hypotheses can be tested via a small number of experimental manipulations, comparing results with those expected under the different models identified as plausible.

Adam Martin - Research Fellow in Health Economics and Chris Bojke, Chair in Health Economics, Leeds Institute of Health Sciences, University of Leeds [Presentation]

From muddied waters to causal clarity? The use and limitations of Instrumental Variables in health economics

Randomisation is rightly regarded as the gold standard is assessing the causal effect of an intervention or risk factor on an outcome. However in many cases, randomisation is not feasible, desirable or practical. In such cases observational data may be the only source of empirical evidence, but comes with a distinct health warning that bias may occur if potential confounding factors are not adequately addressed in the statistical analysis. Instrumental Variables are a ‘signature’ technique of economics, rarely used elsewhere, which effectively propose exploiting fortunate coincidences in the data generating process to resolve omitted variable bias. The key element is whether it’s possible to identify at least one variable which is correlated with the causal variable of interest, but not directly related to the outcome. From the origins of John Snow’s 1855 seminal work in identifying the mode of communication of cholera we chart the use of IVs in economics and health economics specifically, outlying the theory and the key requirements for successful application and the problems of weak or implausible instruments. We show that although a neat and promising concept, IVs should be used with caution and only applied where circumstances permit identification of plausible instruments.

The meeting will be held 3pm until 5pm, teas and coffees from 2:30pm until 3pm

Venue: Room 8.34a, Worsley Building, University of Leeds. See interactive campus map, (PDF copy available here).

The meeting is free to attend, open to all and no registration is required.

The meeting will begin with a short Annual General Meeting for the Leeds Bradford local group, if you are an RSS member and interested in becoming a member of the local organising committee or would like to know more about the role please contact me before the meeting, Sarah Fleming