Quantifying Impact Series: Evaluation Using Observational Data

The Centre for Evaluation, supported by the Centre for Statistical Methodology, hosted a series of lunchtime seminars introducing methodological approaches beyond stratification and regression to address confounding in quasi-experimental or non-randomised designs. Each session described some examples as well as discussing the advantages and disadvantages over more commonly used methods.


Seminars in this series:

  • Seminar 1: ‘Using difference-in-differences in health systems research’. The difference-in-differences approach is a quasi-experimental method widely used to evaluate the impact of health policies and interventions. This seminar will introduce the method and discuss a number of applications in the field of health systems research with a view to highlighting both the pros and cons of the method. Speaker: Tim Powell-Jackson. You can watch the seminar online and download the slides.

 

  • Seminar 2: ‘Statistical issues in the application of the regression discontinuity design for causal inference from clinical administrative databases’. Electronic Health Records (EHRs) are increasingly popular in health care and public health research because they represent a cheap and often very rich source of information about clinical practice. Of course, observational data obtained through EHRs have also potential severe limitations, such as the presence of confounding and missing data. Thus, it is important that suitable methods or designs are used in order to obtain relevant causal estimates of the effects of interventions or policies. The Regression discontinuity Design (RDD) is a quasi-experimental design, originated in the 1960s in the field of econometrics and it has received some attention in biostatistics, in the recent year. The basic idea is that, when interventions are applied according to some external guideline (associated with a continuous assignment variable) so that individuals “just above” a given threshold experience the intervention and those “just below” do not. Close enough to the threshold, individuals can be reasonably considered as exchangeable and thus the analysis of the RDD mimics that of experimental settings. This talk addressed some of the issues associated with the application of the RDD using EHRs, particularly under a Bayesian framework. Speaker: Gianluca Baio. You can watch the seminar online and download the slides.

 

  • Seminar 3: Matching. Matching methods aim to balance pre-intervention characteristics between comparison groups, to minimise the bias in the estimation of treatment effectiveness that is due to observed confounders. The major advantages of matching are: conceptual simplicity, that it can be customised according to the causal question of interest, and that it can be conducted without seeing the outcome data in question. This seminar outlined the key concepts and requisite assumptions in undertaking matching. The seminar covered propensity score matching together with more flexible, data adaptive matching methods specifically designed to achieve covariate balance and discussed sensitivity analyses to the assumption of ‘no unobserved confounding’. The latest developments in matching methods including new approaches for combing matching with instrumental variable estimation were introduced. The seminar drew on a raft of examples from clinical and economic evaluation, health services and health systems research. Speaker:  Richard Grieve. You can download the slides.

 

  • Seminar 4: Interrupted time series. Interrupted time series is a powerful quasi-experimental design that is often applied in public health evaluation. This seminar will introduce the methodology, illustrate the modelling framework, discuss assumptions and advantages and limitation, and present extensions and current methodological developments. Speaker: Antonio Gasparrini. You can watch the seminar online and download the slides.

 

 

  • Seminar 6: Instrumental variables. The Instrumental Variables method (IV) was discovered in 1928 and has been widely used in economics to estimate causal effects in the absence of experimental data. Less known in other disciplines, the method can be applied to a wide variety of contexts and approaches. The seminar will introduce the theory underpinning the method, its limitations and the correct interpretations of its results. Examples from the evaluation of development programmes will be used to illustrate its potential and wide applicability. Speaker: Edoardo Masset. You can watch the seminar online and download the slides.
Back