Pass/fail decisions and standards: the impact of differential examiner stringency on OSCE outcomes

Variation in examiner stringency is a recognised problem in many standardised summative assessments of performance such as the OSCE. The stated strength of the OSCE is that such error might largely balance out over the exam as a whole. This study uses linear mixed models to estimate the impact of different factors (examiner, station, candidate and exam) on station-level total domain score and, separately, on a single global grade. The exam data is from 442 separate administrations of an 18 station OSCE for international medical graduates who want to work in the National Health Service in the UK. We find that variation due to examiner is approximately twice as large for domain scores as it is for grades (16% vs. 8%), with smaller residual variance in the former (67% vs. 76%). Combined estimates of exam-level (relative) reliability across all data are 0.75 and 0.69 for domains scores and grades respectively. The correlation between two separate estimates of stringency for individual examiners (one for grades and one for domain scores) is relatively high (r=0.76) implying that examiners are generally quite consistent in their stringency between these two assessments of performance. Cluster analysis indicates that examiners fall into two broad groups characterised as hawks or doves on both measures. At the exam level, correcting for examiner stringency produces systematically lower cut-scores under borderline regression standard setting than using the raw marks. In turn, such a correction would produce higher pass rates—although meaningful direct comparisons are challenging to make. As in other studies, this work shows that OSCEs and other standardised performance assessments are subject to substantial variation in examiner stringency, and require sufficient domain sampling to ensure quality of pass/fail decision-making is at least adequate. More, perhaps qualitative, work is needed to understand better how examiners might score similarly (or differently) between the awarding of station-level domain scores and global grades. The issue of the potential systematic bias of borderline regression evidenced for the first time here, with sources of error producing cut-scores higher than they should be, also needs more investigation.