• if you can, pick a continuous outcome over a binary outcome
  • Why? For a binary outcome, you'll need a much larger sample size. Continuous outcomes also allow more precision.
• logistic regressions stink!

logistic regression = linear model for the log-odds of the outcome

if we don't take effect modification into account, we get an over-generalized estimate of the relationship between the outcome and the exposure for the entire co-hort

• Breslow-Day Test examines if evidence of a differential association between two variables across the level of a third variable
  • similar limitations to Cochran-Mantel-Haenszel test

• limitations
  • can only adjust for one variable at a time

looks at two binary categorical variables while adjusting for the value of a third categorical variable

• one-sample proportion test
  • do NOT use Yate's continuity, so specify:
    • prop.test(…, correct = FALSE)
• $\Chi^2$ goodness of fit test
  • to ensure sufficient sample size: $n \cdotp_{0} > 5$
  • don't use continunity corrections!
    • chisq.test(…, correct = FALSE)

NOTE: one-sample single proportion test gives a 95% CI – $\Chi^2$ does not!

Joint, Marginal and Conditional Probabilities

• QI
  • SQUIRE 2.0 for QI Reporting
  • Stepped-wedge trial
    • Link
• Linear regression
  • Q-Q plot
  • plot of residuals
  • Cook's Distance
  • these are different!
    • correlative
    • descriptive
    • predictive
    • associative
• confounding vs. effect modification

to assess a paired difference

• create histogram
• plot as box plot
• make Q-Q plot