Abstract
Purpose :
When psychophysical measures are obtained from stimuli with a latent variable such as item difficulty, conjoint methods such as Rasch analysis are required to estimate both person and item measures. Previously (Adeyemo, ARVO #4688, 2017) we applied this approach to data from individuals with ultra-low vision (ULV) performing real-world activities with 2-, 3-, or 4-alternative force choice (AFC) judgments. However, since Rasch analysis ignores correct performance by chance, item measures from mixed m-AFC instruments should be questioned. A different approach, based on SDT and yielding person and item measures in d' units, has recently been introduced. Here we compare the two approaches.
Methods :
: Performance data from 17 activities, at 3 difficulty levels, each performed twice by 25 individuals with native ULV and 4 Argus II and 6 Brainport users, were submitted to both analyses. Confidence intervals (CIs) and fit statistics were obtained in Winsteps© for Rasch analysis, whereas the SDT-based analysis does not currently yield CIs for person measures with unequal probability test items.
Results :
Both person (left panel) and item (right panel) measures show excellent (r2>0.995) correlation between the two methods. However, under SDT-based item measures for 2-, 3-, and 4-AFC activities are shifted relative to Rasch item measures, as would be expected: In Rasch analysis 2-AFC activities are fitted as relatively easier, and 4-AFC activities as harder, than 3-AFC activities due to different chance correct rates, whereas SDT-based items measures are estimated correctly. This explains the 3 parallel linear clusters. In the conversion from logit intervals (Rasch) to d' intervals (SDT), item CIs scale roughly proportionally with the item intervals.
Conclusions :
Our results demonstrate that the SDT-based analytic approach can successfully handle data from mixed m-AFC instruments, and otherwise produces estimates that closely correspond to those obtained from Rasch analysis. It is thus a powerful tool for conjoint analysis of psychophysical data.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.