"Exactly," Lena said. "And next time, if you can’t randomize, use a — give half the classes the software in Phase 1, the other half in Phase 2. Compare each against itself over time."
Result: The +7 points was statistically significant (p < .01) and practically meaningful. Lena presented to Hartley: "The software works, but only by 7 points, not the 15-point jump you saw in the raw comparison. The raw difference was inflated by Ms. Chen’s prior excellence." quasi-experimentation a guide to design and analysis pdf
Hartley nodded. "So we keep the software, but we train Mr. Abel on it too." "Exactly," Lena said
Hartley laughed. "You quasi-people have a workaround for everything." Lena presented to Hartley: "The software works, but
But to be rigorous, she added a and used Huber-White robust standard errors (because monthly scores from the same class aren’t independent — a key point from quasi-experimental guides).
Lena smiled. "That’s the guide to design and analysis. No randomization? No problem. Just more thinking." Quasi-experimentation isn’t “second-best.” It’s a toolkit for causal inference when experiments are impossible. Master the threats (history, selection, maturation, regression), choose a design (ITS, DID, nonequivalent groups), and analyze with care — robust standard errors and pre-trend checks are your friends.