Graphical Investigation of Threshold Choice Effect on Odds Ratio Related to Prognostic Factors in Stroke Recovery
Pages 348-355
Marco Iosa, Giovanni Morone, Augusto Fusco and Stefano Paolucci
DOI: http://dx.doi.org/10.6000/1929-6029.2014.03.04.3
Published: 06 November 2014

Abstract: The aim of this study was to assess the effects of the arbitrary choices of threshold-values for dichotomizing not binary factors on the computation of odds ratio (OR) for the identification of prognostic factors, in particular of motor recovery after stroke.

Data of a sample of 1000 patients with subacute stroke have been analysed. We considered as dependent variable the effectiveness of neurorehabilitation (i.e. the achieved level of independency in activities of daily living, measured using the Barthel Index, expressed in percentage of the maximum achievable improvement), and as independent variables age, time between stroke acute event and beginning of neurorehabilitation, gender, type of stroke (ischemic vs. haemorragic) and side of hemiparesis. We performed univariate analyses for computing OR with respect to different choices of threshold for dichotomizing age and time from stroke. In this analysis median value of effectiveness was used for dichotomizing subjects in good and poor responders. Then these analyses were repeated also varying the threshold-value of effectiveness. Finally multivariate analyses based on forward binary logistic regression were performed varying at the same time the thresholds of age and time from stroke. With respect to threshold choice, OR-values of age resulted stable, but those of time from stroke resulted more variable. Variability increased when also the threshold chosen for dichotomizing the independent variable was changed. Multivariate analyses showed that these choices could even make not statistically significant the effect of a binary prognostic factor such as gender. In conclusion, OR-values resulted affected by threshold choices. It can increase the difficulties in marking predictions of outcomes after stroke. In this study we reported a possible graphical evaluation of the variability of OR-values with respect of threshold choice, that can be helpful whenever threshold is arbitrary chosen.