Free

Applied Statistics to Health and Medicine (YSM prize winners) General Admis...

Event Information

Share this event

Date and Time

Location

Location

Room 116, M&S Building, University of Glasgow

University Place

Glasgow

G12 8SQ

United Kingdom

View Map

Event description

Description

Jointly hosted with the Royal Statistical Society's Young Statistician's Section. Drinks and nibbles will be provided after the presentations of the speakers.


Propensity score diagnostics: the challenges we face when providing evidence that a propensity based estimate is unbiased.

Emily Granger1, Dr. Jamie C Sergeant1,2, Dr. Mark Lunt1

1Arthritis Research UK Centre for Epidemiology, Division of Musculoskeletal and Dermatological Sciences, Manchester Academic Health Science Centre, The University of Manchester

2Centre for Biostatistics, Division of Population Health, Health Services Research and Primary Care, School of Health Sciences, The University of Manchester

Observational data are useful for studying comparative effectiveness or safety of medical treatments; however they are prone to bias due to confounders. Confounding bias arises when covariates which predict the outcome of interest have different distributions across treatment groups. Propensity scores (defined as the patient’s probability of receiving treatment conditional on their baseline characteristics) are becoming an increasingly popular method used to account for confounders. Conditioning on the propensity score balances the distribution of covariates across treatment groups and hence eliminates confounding bias. However, there is some debate as to whether propensity scores are reliable. One concern is that an inadequately estimated propensity score may not effectively balance covariate distributions, which could lead to biased results. There are a variety of balance diagnostics being used to assess the adequacy of propensity scores, but no consensus on the best method and the implementation in the applied medical literature is far from optimal.

During this talk I will discuss reasons why some of the commonly applied balance diagnostics are not appropriate and highlight conflictions in the literature which could mislead applied researchers using propensity scores. Additionally, I will present a simulation study designed to compare diagnostics in terms of their ability to identify different types of propensity score misspecification. Results indicate that commonly used diagnostics such as standardised differences can be unreliable. Finally, I will discuss plans regarding development of new propensity score diagnostics.


Network Time Series (Kathryn Leeming, Marina Knight, Guy Nason and Matthew Nunes)

"A network time series is formed of data collected over time at nodes of a graph, or network. These networks arise in a wide range of settings, such as environmental, financial, and medical.

In this talk, analysis of a disease incidence data set is presented using the Network Autoregressive (NAR) framework. As an extension to the univariate autoregressive time series model, NAR allows for dependence on neighbouring nodes according the underlying network structure.
The network associated with the time series could be un-weighted and un-directed, or have more complex features such as edge weights and directions, or even be time-varying. We demonstrate a method to recover an un-weighted graph from a multivariate time series using the NAR model.

Advantages of using network-based methods for such data will be discussed, and performance compared to methods which do not use network structure."

Share with friends

Date and Time

Location

Room 116, M&S Building, University of Glasgow

University Place

Glasgow

G12 8SQ

United Kingdom

View Map

Save This Event

Event Saved