reporting interactions and main effects

[duncanbradley]

Several options:

1. Use summary() to report z-statistics or anova() to report t-statistics

   • pro: consistency in assessment of all fixed effects
   • con: no assessment of goodness-of-fit; tests may be less accurate (see Christensen, 2019), or at least different (see https://www.ssc.wisc.edu/sscc/pubs/MM/MM_TestEffects.html#testing-mixed-models-parameters)

2. Use drop1(model, scope=c("axis","denominator","axis:denominator"), test="Chisq")

   • pro: computes Chisq for all terms
   • con: Error: Only marginal terms can be dropped from the model (terms that can be dropped while respecting the hierarchy of terms in the model)

3. Use car::Anova(model, type="III", test.statistic="Chisq")

   • pro: computes Chisq for all terms
   • con: not convinced that these are accurate for my models, given bullet 2 above, and the fact they are identical to the F statistics from anova()

4. Create individual comparison models to test each fixed effect using anova() (used in Choropleth paper)

   • pro: goodness of fit test for all fixed effects
   • con: main effects cannot be assessed within the context of the interaction (see https://stats.stackexchange.com/questions/484148/r-lmer-model-degree-of-freedom-and-chi-square-values-are-zero)

5. Assess interaction term using a Chisq goodness-of-fit test, but assess main effects using z- or t-statistics

   • pro: uses Chisq where possible
   • con: variation in reporting of effects may be inconsistent

[duncanbradley]

Solve:

• conflict between params object and quarto params argument ✅
• ugly parameter names ✅
• avoid : in name (str_replace(x, ":", "_"))) ✅
• use group_map() group_split() and map() for simplicity ✅

[duncanbradley]

completed in 407dd1c