|
Abstract:
|
A forest plot, graph of regression coefficients, and other plots of effects and associated uncertainty provide intuitive visual means of assessing and comparing effect sizes or magnitudes of association with a regression outcome, assuming all share or are transformed to a common scale. A difficulty with plotting on a linear scale is that inclusion of more extreme values or confidence bounds may obscure the effects of more moderate values.
It is proposed that, to improve visualization, variable estimates and intervals on a linear scale may be treated as if they were on a logit scale and transformed accordingly to a probability scale. This transformation constrains the bounds from (-infinity,infinity) to (0,1).
The resulting graphs on a probability scale facilitate comparison of relative effect sizes even in the context of extreme effects or confidence bounds, since all are constrained within the (0,1) interval, and confidence bounds excluding .5 support interpretation of significance. A weakness of this transformed scale is a proportional distortion of the effect sizes, as with any nonlinear transformation.
|
