This might be a stupid question but hear me out if you have time.
Prediction rules (a.k.a multi-component prediction scores/indices) give individual-level predicted risk of mortality within a certain time frame (typically 3 years in COPD, some are 10 years). However, if you measured a certain prediction rule several times at different time points in a single patient you would get several predictions. If a patient receives two measurements and changes from 10% to 20% predicted risk of mortality within 3-years then maybe more treatment can be introduced. For certain prediction rules (depending on the components used) a 10% to 20% change may indicate progression of the disease.
Question: Is there any study out there that defines a clinically relevant change in predicted mortality?
I know that sounds odd but I need a threshold without looking a p-values (multiple testing problems).
Problems with this question: Certainly, to the patient any increase is relevant. Importantly, the expense and side effects of certain treatments may make somewhat large increases more tolerable.
I don't see that you could get such a cut off. Wouldn't it depend on both the patient and the disease? Best, D. Booth
The first question that I think should come to mind is: why does the predicted risk change between T and T+1? This might be due to factors that have nothing to do with a true shift in risk, such as measurement error variations and circadian rhythm. Obviously, a shift due to these factors isn't relevant to the patient, it's noise.
Even if it is a true shift in risk I doubt it is generally helpful to dichotomize into "relevant" vs "not relevant" predicted risk shift. The only setting where I can imagine a relevant vs not relevant shift categorization might be relevant is when the risk makes a shift passing a treatment threshold. To be useful (and regardless of whether it is a true risk shift or due to measurement error), such a clinically relevant treatment threshold should already exist and be used, which in my experience remains rare.
This idea of relevant shifts in risk categories relates to reclassification tables, see for instance here: Article Quantifying the added value of new biomarkers: how and how not