Statistical inference in behavior analysis: Some things significance testing does and does not do.

Branch (1999) wrote a plain-language essay. He asked one question: do p-values help or hurt behavior analysts?

He read psychology journals and spotted a pattern. Authors used big statistical tests to claim their graphs were “real.”

He argued this habit hides the actual data. The paper has no new numbers—just a warning.

The author found that p-values can fool us. A tiny p does not mean the behavior changed in a useful way.

He says the fix is simple: show the effect again in a new subject, a new room, a new day. If it repeats, it is real.

Lemons et al. (2015) counted every article in JEAB for 55 years. They saw inferential stats keep rising, just like Branch (1999) feared.

Busch et al. (2010) went the other way. They taught college kids to love t-tests with three short lessons. Branch (1999) would call this teaching the wrong skill.

The two papers seem to clash, but they talk past each other. Lemons et al. (2015) describe what we do; Branch (1999) tells us what we should do. Busch et al. (2010) show we can teach stats fast—yet that speed does not make the stats useful for single-case work.

You run sessions, not t-tests. Next time a graph looks good, skip the p-value. Run one more reversal or replicate with a new client. That second curve is your proof. Share that picture in your report and you follow N’s advice: let the behavior speak, not the statistics.