We improve the design and analysis of medical research through applied statistics research. We focus on four key areas related to the medical research that we conduct.
1. PROGNOSTIC MODELS: VALIDATION & DEVELOPMENT
Prognosis is the forecast or estimate of the risk of something happening in the future. In medicine, we are interested in predicting a person's future health state using their current characteristics, such as their weight, history, blood markers, or X-ray results. We may be interested in whether they will develop a particular disease, whether they will have recovered from a current disease, or whether their pain levels will have changed, for example.
Prognostic models are mathematical models that relate a person's characteristics now to the risk of a particular future outcome. Doctors use them to get objective estimates of the probability that something will happen, to use alongside other clinical information. The ideal prognostic model uses many pieces of information from a patient to give an accurate risk estimate, but is also easy for a doctor to use during their day-to-day work. It also accurately predicts outcomes using different data from that used to develop it, which is tested in external validation.
We are interested in improving how prognostic models are developed and validated. We investigate and have published guidance on:
- Selecting the right sample size for developing and validating clinical prediction models
- Effect of missing data on model development and validation
- Impact of measurement error on model performance
- Effect of categorising continuous predictors
- How to use large databases for model validation
2. RANDOMISED CONTROLLED TRIALS: DESIGN & ANALYSIS
Randomised controlled trials are considered the 'gold standard' in primary medical research. However, the results that are generated from these studies can only be as good as the design they are based on. We conduct applied statistics research to improve the analysis methods used in trials. We also test whether researchers are using the most correct designs and methods for their trials, often using systematic reviews.
For example, we are currently running an epidemiological study of randomised trials. We are comparing the methodological quality of published randomised controlled trials conducted in high-income countries and low/middle-income countries.
3. OBSERVATIONAL STUDIES: PROPENSITY SCORE ANALYSIS
We investigate using propensity score matching analysis to balance confounders and estimate treatment effects in observational studies.
Sometimes ethical or logistical constraints mean that we cannot use a randomised controlled trial to test a drug or treatment. In observational studies, the researcher doesn't control who gets the intervention. Instead, the researcher identifies a group of people with the disease or characteristic they are interested in, who are selected for the intervention of interest for any reason. The researcher also collects a group of people with the same characteristic or disease who do not get the intervention, to act as the control.
As the researcher did not randomise which people got the intervention and which were the control, there may be some biases that influenced who ended up in each group. These biases will affect how good the intervention appears to be. We can get rid of bias by matching the people in the two groups on every characteristic that may have affected which group they are in. Matching is done using propensity scores.
We are interested in how many characteristics should be matched and how to deal with any bias left after matching, for example.
4. SYSTEMATIC REVIEWS: DESIGN & ANALYSIS
Systematic reviews are considered the strongest form of evidence for making clinical decisions. We conduct applied statistics research to improve how systematic reviews are done.
For example, we are currently studying the association between the methods used in systematic reviews and the estimates of prognostic importance that they report.
The DELTA2 project aimed to develop guidance for specifying the target difference (or effect size) in the sample size calculation for a randomised clinical trial (RCT). This guidance is aimed at researchers and funders of RCTs. It is hoped that the guidance will help improve the design of future RCTs.
Study with us
CSM hosts full-time DPhil students researching statistical methodology. One of our previous students, Bethan Copsey, shared her experiences of NDORMS and the University of Oxford:
"There are lots of opportunities for training and development at NDORMS with a wide variety of courses to build your skills and seminars to hear what other people are working on and provide inspiration."