A note on the measurement of conditional discrimination.

Blough (1980) wrote a short warning note. He said accuracy scores can trick us when we test conditional discrimination.

The paper has no new data. It is a think-piece that tells researchers to look deeper than percent correct.

The author showed that 80 % correct can hide two very different patterns. Kids might get every trial right except one type, or they might guess across all types.

Accuracy alone cannot tell which stimulus truly controls the learner’s choice. You need extra checks, like error-type plots or probe trials.

Iversen (2021) extends the warning. He says skip group averages and graph each learner’s accuracy alone. This makes the 1980 advice doable in daily practice.

Perez et al. (2020) give a live example. They blocked the correct picture and saw errors jump, even though overall accuracy stayed near 70 %. The 1980 note predicts exactly this trap.

POLIDORNEVIN et al. (1963) looks like a clash but isn’t. Longer time-outs raised accuracy, so the authors trusted the score. Blough (1980) replies: check what the learner actually looked at before you celebrate.

Next time you run a matching-to-sample program, do not stop at “85 % today.” Break down errors by sample-comparison pair. Add a few probe trials with new pictures. If errors cluster on one type, stimulus control is narrow, not solid. Fix it before you move on.