Primary Clinical Outcomes For A Single Emergency Department 2017 - 2021 (part 4)
SMLR modelling of 1.9 million admissions records for the emergency departments of an undisclosed NHS Trust: predicting risk of treatment & hospitalisation in the vaccine era
To understand what I’m doing with all this staged multivariate logistic regression modelling (SMLR) of admissions to the ED of a single UK NHS Trust readers will need to go back and absorb part 1, part 2 and part 3. What it all boils down to is attempting to provide a level playing field for when we compare patient cohorts across periods of time. Differences in age, sex and diagnoses made are as certain as green energy taxes, and we have to face the tricky issue of clinical protocols changing over time that may alter the rate at which admissions are treated and/or hospitalised. Call that rate something fancy like ‘propensity’, ‘risk’, ‘likelihood’ or ‘probability’ and you arrive in the land of the applied statistician.
We can observe real world rates for treatment/admission and quote them as percentages, proportions, ratios or odds ratios, and we can use statistical modelling in an attempt to predict these real world rates from a bunch of factors. Those predictions, when churned out of a logistic regression that attempts to model binary (Yes/No) data, fetch up as continuous variables we might call ‘risk scores’, ‘probabilities’ or ‘propensity scores’. If these scores are derived using a control sample or control period then we end up with unbiased estimates of the likelihood of somebody being treated or hospitalised, ‘unbiased’ meaning that the scores will be untainted by any changes in procedure or protocol that may have taken place in the study period.
We all know that the NHS underwent the biggest and most traumatic change in its history during the draconian pandemic year of 2020 as wards and services were closed down and dubious COVID protocols implemented, and all with very little connection to evidence-based medicine. We must therefore expect strange contortions within treatment and admission rates during this period and, TBH, the following year of 2021 also, when the NHS was still recovering from the shock of totalitarian government. This is why it is valuable to have a set of predicted scores to hand that suggest what should be happening with treatment/hospitalisation rates under ‘normal’ circumstances. That’s the logic anyways, and my models seem to be saying sensible things that match experience.
What I’m going to do today is use the method to look at the vaccine era of 2020/w50 – 2021/w37 to see if any distortions are apparent. By ‘distortion’ I mean observed treatment and hospitalisation rates drifting away from predicted normality given the changing characteristics of the patient cohorts. If this makes sense I suggest we make a cuppa and treat ourselves to a slice of fresh-baked Madeira cake before proceeding to the tables.
Keep reading with a 7-day free trial
Subscribe to John Dee's Almanac to keep reading this post and get 7 days of free access to the full post archives.