- Enhanced general causation risk model quantifying scientific evidence supporting hypotheses of bodily injury
Enhanced General Causation Risk Model
The general causation (GC) risk model, a central component of Praedicat’s overall liability catastrophe model, quantifies the level of scientific evidence supporting a hypothesis that a Litagion agent causes bodily injury. This release represents the first significant revision to the GC model since it was first deployed by Praedicat in 2014.
The principal revision concerns how the model combines evidence obtained from studies conducted in different organisms – human, animal, and in vitro – to arrive at a single GC score. The model now estimates literature strength separately by organism type and then combines these sub-scores using a dynamic weighting system that determines the relative contributions of the human, animal, and in vitro literatures. As before, human studies receive more weight than animal studies which receive more weight than in vitro studies. And the model continues to give preference to literatures that have supporting evidence in multiple organisms. But literatures are now less penalized for being concentrated within a particular study type. Also, the point at which literatures become “saturated” with studies of a particular type has been relaxed, allowing GC scores to continue increasing over a broader range. This revised approach to combining evidence from different study types better reflects how the scientific community evaluates literatures and also permits a simple decomposition of the overall GC score by organism, allowing users to better understand how a literature derives its overall support.
The GC model revision also gives less weight to “coherence” than previously. Coherence is the idea embodied in the Hill Criteria that a causal hypothesis has the most support when studies conducted in multiple organism types support the hypothesis. We continue to calculate a coherence factor by comparing the proportion of positive studies across organism types but it is now weighted less than it was before because the dynamic weighting of the literatures from each of the three organism types now accounts for some of this same effect.
The overall distribution of GC scores is virtually unchanged with this model release, but there is movement in many individual GC scores. The median change in GC score is 0.05. About 16% of currently scored hypotheses see their GC score move by 0.15 or more. Literatures heavily weighted toward a particular study type see their scores increase the most. An example is nanomaterial literatures which tend to be dominated by in vitro study. Smaller literatures and highly equivocal literatures see their scores decrease the most.
The GC forecast model is largely unchanged with this release. The forecast model now assumes a secular increase in scientific publications of 3% per year, reflecting long term growth rates in scientific publishing. And the forecast model has been recalibrated based upon extensive backtesting of the model. As with historical GC scores, the distribution of forecasted GC scores has not changed appreciably.