On the third Thursday of every month, we speak to a recent graduate about their thesis and their studies. This month’s guest is Dr Alexander Thompson who has a PhD from the University of Manchester. If you would like to suggest a candidate for an upcoming Thesis Thursday, get in touch.
Title
Incorporating multiple conditions within decision-analytic modelling for use in economic evaluation
Supervisors
Katherine Payne, Matt Sutton
Repository link
https://www.research.manchester.ac.uk/portal/en/theses/incorporating-multiple-conditions-within-decisionanalytic-modelling-for-use-in-economic-evaluation(c1eee553-c910-4b90-90e5-92a370f5c635).html
Why do multiple health conditions pose a challenge for decision modellers?
As multiple conditions in populations becomes the norm, rather than the exception, it poses a challenge for decision modellers (and decision-makers) who want to believe model results. As we have all heard said, so many times it has become a bit of a cliché, all models are wrong but some are useful. But I personally would struggle to place a lot of faith in a model for an intervention, particularly for a long-term condition, which blindly ignored and didn’t acknowledge the likely presence of other conditions co-existing alongside it. If enough other people think similarly, then I suppose such a model stops being useful and change is needed.
In such a situation, the challenge for the modeller is what to do. Do we make models more complex to incorporate interactions between conditions or drugs (as an example of just a couple of the mechanisms that multiple conditions throws up)? Alternatively, do we just keep things simple whilst doing a better job of explicitly acknowledging the underlying assumptions and try to do a better job of grading the data that have been used (e.g. for ‘applicability’). My PhD was a bit of a journey and I came to the view, in the end, that it was probably better to keep things relatively simple and declare the assumptions up front.
What data sources could you use to explore multi-morbidity in your research?
I am fortunate enough that I work in a Centre in Manchester that has lots of different skills (economic evaluation and econometric evaluation) and, naturally running alongside that, a group that has lots of different datasets. So I used data such as NHS PROMs, which recorded EQ-5D for patients before and after undergoing routine planned interventions within the NHS. A lot of people have those interventions every year so it provides a huge amount of data to play with. I also used the GP Patient Survey which was again sent out to a huge sample of the general population every year, collecting the EQ-5D alongside some important self-reported conditions. Unfortunately, the GPPS doesn’t collect the EQ-5D anymore, I believe, which is a shame. There was also the Health Survey for England which includes the EQ-5D alongside self-reported conditions and a tonne of other variables.
Despite conducting the health economics part of a decent number of trials during the time of my PhD, I never actually used any of that data for my PhD. Thinking about it now, there was probably a good reason for that. Trial data often does quite a poor job of quantifying the other conditions that patients may have. Also, the sample size might just not be big enough.
Is there a consistent relationship between the presence of multiple conditions and quality of life?
I think so, yes. The more health conditions that you might have, the worse your health-related quality of life tends to be. Whilst that is fairly obvious, I was particularly interested in whether certain conditions which occurred together have a disproportionate effect or interactive effect. So, would condition A and condition B, when experienced together, have a similar impact as the two conditions appearing apart? Again, to answer this sort of question you might start interacting conditions within an econometric model. But, with multiple conditions, where do you stop? All conditions interacting simultaneously and every combination between?
Even with large datasets, you can quickly run into the high-dimensionality problem associated with multiple conditions. To overcome that and to avoid the danger of dredging the data, I opted to explore some relationships that would have some theoretical underpinnings. I explored the impact of physical health and mental health and found that when experienced together there was a negative interactive effect. That sort of made sense to me.
What are your practical recommendations for analysts dealing with multiple conditions?
If you are a decision modeller faced with a decision problem, likely involving multiple long-term conditions simultaneously, I would probably start by working with the clinical team to identify the simplest approach that you can get away with to give justice to the particular problem. Be guided by their judgements on whether conditions or interventions interact (because there is probably little data out there on it).
If you are working with utility data then try (and try again) to get it in the relevant population directly. If it is published data, then look at the baseline characteristics tables, etc. Go above and beyond and see if you can get breakdowns of the population and the different conditions experienced. It should really match the population the decision problem is for. If none of that is available, then try using one of the prediction methods we tested using the GPPS (multiplicative or additive). They do a decent enough job for relatively low combinations of conditions but at higher numbers they tend to either under- (additive) or over-predict (multiplicative) utility. A caveat to that is don’t assume you can combine a mental health condition utility with a physical condition utility. The simple combination approach doesn’t really work there because of the likely interaction described earlier.
How would you like to see this line of research developed in the future?
I would say there is a lot of untapped value exploring health economics within clinical guidelines. My PhD straddled a project exploring NICE clinical guidelines and ‘multimorbidity’ (just another term for multiple conditions). Whilst NICE makes binding decisions about expensive new (often cancer) drugs in fairly small populations, which grabs the headlines, clinical guidelines make (non-binding) recommendations on large populations and yet there isn’t a lot of interest. If we really do care about opportunity cost and the best allocation of scarce resources then, whilst keeping an eye on new technology appraisals, we should also be watching and helping to do more research for clinical guidelines.
