Effects of treatment‐integrity failures on a response‐cost procedure

St. Mundy et al. (2016) asked a simple question: how sloppy can you get with response cost before it stops working? They used an alternating-treatments design with neurotypical adults. Each session the experimenter either followed the script perfectly or broke it on purpose.

Some sessions left out half the fines. Others left out 80 percent. A third set added fines when the person had not actually misbehaved. The team watched how these errors changed the power of the token system.

Missing every other fine still kept behavior low. Missing four out of five killed the effect. Adding false fines also weakened the system and made people avoid the task more.

In short, response cost needs at least 50 percent accuracy to stay useful. Below that, you might as well not use it.

Cymbal et al. (2022) looked at 20 years of OBM studies and found that three-quarters never even report integrity data. St. Peter’s work shows why that matters: if you do not check, you may be running a broken intervention and never know.

Wilder et al. (2023) argue that counting only percent-correct hides drift. The 2016 data back them up: percent-correct fell, but the curve of behavior loss was steeper than the math alone would predict. Using response-rate metrics would have flagged the drop sooner.

Cook et al. (2020) warn that measurement error itself can look like treatment failure. St. Peter controlled measurement carefully, proving the drop in effect was real, not a scoring mistake.

If you use response cost in classrooms, clinics, or staff incentive plans, track at least every other consequence. A quick tally on your phone keeps you above the 50 percent line. When integrity slips, pause and retrain before the whole system collapses.