Module 06 Summary

What this module established

scipy's ttest_ind defaults to equal_var=True; R's t.test defaults to Welch. On groups with SDs of 3.60 and 16.89 and sizes 60 and 15, the default gave p = 0.0019 on 73 df while Welch gave p = 0.1086 on 14.3 — and the default is the wrong one.

Carry forward

• Pairing subtracts the person out. With a between-person SD of 14.07 and a change SD of 2.80, the unpaired test found nothing and the paired test gave an interval 6.9 times narrower.
• chi2_contingency detects and explains nothing; the standardised residuals showed premium churning at +3.04 and basic at −2.72. On 2x2 tables it applies Yates' correction by default.
• f_oneway has no post-hoc at all. Follow it with pairwise_tukeyhsd, which adjusts for the comparisons it makes.
• Shapiro-Wilk detects genuine skew only 42% of the time at n = 8, and a real 5-point effect is missed 88% of the time at n = 10 per group.

Before moving on

Move on when you can state, for any null result you report, what effect the study could have detected.

Mark complete