Anova - report sums of squares type + coding

[mattansb]

This matters a lot, and without this context there is no way to interpret the results.

[IndrajeetPatil]

Can you please modify and repost the output below to match your expectations? Will give a better idea about how we can achieve this.

library(report)

mod <- aov(formula = Sepal.Length ~ Species, data = iris)

report(mod)
#> The ANOVA (formula: Sepal.Length ~ Species) suggests that:
#> 
#>   - The main effect of Species is statistically significant and large (F(2, 147) = 119.26, p < .001; Eta2 = 0.62, 90% CI [0.54, 0.68])
#> 
#> Effect sizes were labelled following Field's (2013) recommendations.

^{Created on 2021-04-16 by the reprex package (v2.0.0)}

[mattansb]

It should be in the first line, before the actual stats.

I assigned myself, I'll make the needed changes (and try to also add support for afex / anova objects at the same time).

[strengejacke]

Fixed.

library(report)

mod <- aov(formula = Sepal.Length ~ Species, data = iris)

report(mod)
#> Don't use ANOVA. 

[bwiernik]

Related to @strengejacke's point, I wonder whether it would be good to also report the actual regression coefficients in addition to/instead of eta2?

[mattansb]

If anyone is using aov I'd say they're not interested in the coefs (and if they're using effects coding, they would have little information in them really...)

They can always report(lm(aovMod)) 🤷‍♂️

[mattansb]

Okay, I've lost track of the API here.... @IndrajeetPatil Can you help me out?

This first sentence should be changed:

report(mod)
#> The {type} ANOVA (formula: Sepal.Length ~ Species) suggests that:

Looking at R/report.aov.R, I see serval methods:

• aov - Type I tests (sequential)
• aovlist - Type I tests (sequential)
• anova - it depends... not all anova objects are the same. The printing method should tell you what type it is (if not, should default to Type I tests (sequential)). Can be:
  • Type II tests (marginal)
  • Type III tests (simultaneous)

This information seems to be stored in the "heading" attributes:

mod <- aov(Sepal.Length ~ Species * Sepal.Width, data = iris)
mod2 <- lmerTest::lmer(Sepal.Length ~ Sepal.Width + (1|Species), data = iris)


a0 <- anova(mod)
a1 <- car::Anova(mod, type = 2)
a2 <- car::Anova(mod, type = 3)
a3 <- anova(mod2)


attr(a0, "heading")
#> [1] "Analysis of Variance Table\n" "Response: Sepal.Length"
attr(a1, "heading")
#> [1] "Anova Table (Type II tests)\n" "Response: Sepal.Length"
attr(a2, "heading")
#> [1] "Anova Table (Type III tests)\n" "Response: Sepal.Length"
attr(a3, "heading")
#> [1] "Type III Analysis of Variance Table with Satterthwaite's method"

^{Created on 2021-04-18 by the reprex package (v1.0.0)}

[mattansb]

(people are gonna be doing type-1 tests with treatment-coding and think they're doing type-3 tests with effects-doing... daymn)

[bwiernik]

I think that most folks using R for ANOVA (why?) are used to using car::Anova() or es::aov_ex() instead of base R aov(). This is an oddness of S/R's aov() and anova() functions that has lamented by various folks for 50 years.

[DominiqueMakowski]

if anyone is using aov I'd say they're not interested in the coefs

I agree

I think that most folks using R for ANOVA (why?) are used to using car::Anova() or es::aov_ex() instead of base R aov()

I wouldn't be so sure, I've seen (too) many serious usage of aov() ^^ (maybe it's a psych thing 🤷)

[bwiernik]

if anyone is using aov I'd say they're not interested in the coefs

Well, perhaps not, but I'd argue they should be. 😜

[mattansb]

Well, perhaps not, but I'd argue they should be. 😜

Finally, a hill for me to die on!
_{(Seeing as how for proper SS in anova you should be using effects coding (or similar), coefs become virtually uninterpretable. The role of coef-inspection is replaced with simple-effects, expected means, and contrasts instead ^_^)}

[mattansb]

I wouldn't be so sure, I've seen (too) many serious usage of aov() ^^ (maybe it's a psych thing 🤷)

Yeah, this is very unfortunate 😔

[mattansb]

I think perhaps this should be supported elsewhere:

• parameters should have the "type" as a footer for the tests
• effectsize should have the "type" as a footer for the effect size (can pull from parameters?)
  Then report can pull from those.

@strengejacke what do you think?

[IndrajeetPatil]

Should the exact details of the type of test contained in title attribute of model_parameters output?

I think this is what we currently do:

mod <- stats::t.test(formula = wt ~ am, data = mtcars)

attributes(parameters::parameters(mod))$title
#> [1] "Welch Two Sample t-test"

^{Created on 2021-04-20 by the reprex package (v2.0.0)}

[mattansb]

Yes, just as this info appears in the title of anova objects:

anova(lmerTest::lmer(Reaction ~ Days + (Days | Subject), lme4::sleepstudy))
#> Type III Analysis of Variance Table with Satterthwaite's method
#>      Sum Sq Mean Sq NumDF DenDF F value    Pr(>F)    
#> Days  30031   30031     1    17  45.853 3.264e-06 ***
#> ---
#> Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

^{Created on 2021-04-20 by the reprex package (v1.0.0)}

[mattansb]

bump
This should probably be delt with before #342 - parameters and effectsize already return the type.