Signal-detection analyses of conditional discrimination and delayed matching-to-sample performance.

Alsop (2004) ran Monte Carlo simulations. He tested what happens when you apply standard signal-detection fixes to data with zero errors.

The study looked at conditional-discrimination and delayed-matching-to-sample tasks. These are common lab models for memory and concept learning.

The usual fixes for perfect scores put strong bias into the numbers. They can make you think a learner has good sensitivity when the data are actually empty.

In short, zero-error sheets can fool you into false conclusions.

Chou et al. (2010) extend the warning. They showed that live observers also drift when feedback or payoffs change. Brent flagged math bias; C et al. showed human bias.

Wirth et al. (2014) used the same Monte Carlo trick. They built error tables for interval sampling. Both papers shout the same message: routine data rules can distort.

Kangas et al. (2011) seem to clash. They praise 0 % commission errors in DTT, yet Brent says 0 % errors break the stats. The gap is level of analysis. D et al. tracked learner errors; Brent tracked the math after the learner is perfect. You need both views.

If you run conditional-discrimination probes or DMTS memory tasks, never trust a perfect score at face value. Report raw hits and false alarms. Add a tiny constant or use a bias-free estimator. Share the method so other BCBAs can judge the numbers. Clean data today keeps your conclusions safe tomorrow.