Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

anova_stats returns wrong partial eta squared values for repeated measures anova #111

Open
frankpapenmeier opened this issue Jan 21, 2022 · 1 comment

Comments

@frankpapenmeier
Copy link

frankpapenmeier commented Jan 21, 2022

I noticed that the anova_stats function returns wrong partial eta squared values for repeated measures ANOVAs.

Please see the following reproducible example for details regarding the issue and its source:

library(sjstats)

# generate some data

set.seed(325)

dat <- expand.grid(
  factor1 = c("A","B"),
  factor2 = c("X","Y"),
  participant = 1:10
)
dat$participant <- factor(dat$participant)

dat$value <- rnorm(length(dat$participant))

# add some main effects
dat$value[dat$factor1 == "A"] <- dat$value[dat$factor1 == "A"] + rnorm(dat$value[dat$factor1 == "A"], 5)
dat$value[dat$factor2 == "X"] <- dat$value[dat$factor2 == "X"] + rnorm(dat$value[dat$factor2 == "X"], 2)


# 2x2 repeated measures ANOVA
model <- aov(value ~ factor1*factor2 + Error(participant/(factor1*factor2)), dat)
summary(model)

# Error: participant
# Df Sum Sq Mean Sq F value Pr(>F)
# Residuals  9  17.41   1.935               
# 
# Error: participant:factor1
# Df Sum Sq Mean Sq F value   Pr(>F)    
# factor1    1  317.9   317.9   225.6 1.12e-07 ***
#   Residuals  9   12.7     1.4                     
# ---
#   Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
# 
# Error: participant:factor2
# Df Sum Sq Mean Sq F value  Pr(>F)   
# factor2    1  28.70   28.70   12.16 0.00686 **
#   Residuals  9  21.24    2.36                   
# ---
#   Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
# 
# Error: participant:factor1:factor2
# Df Sum Sq Mean Sq F value Pr(>F)  
# factor1:factor2  1  4.814   4.814   4.581  0.061 .
# Residuals        9  9.457   1.051                 
# ---
#   Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

# Thus, partial eta squared values should be:
#
# main effect factor 1: 317.9 / (317.9 + 12.7) = 0.961585
# main effect factor 2: 28.70 / (28.70 + 21.24) = 0.5746896
# interaction effect factor 1:factor2: 4.814 / (4.814 + 9.457) = 0.3373274
#

# This is the results of anova_stats from the sjstats package
anova_stats(model)

# main effect factor 1: 0.971
# main effect factor 2: 0.752
# interaction effect factor 1:factor2: 0.337

# --> partial eta squared of last effect is correct, but of the other effects is wrong

# the source of this issue lies in the following line 160 of the anova_stats.R script:
#
# "ss.resid <- aov.sum[["sumsq"]][nrow(aov.sum)]"

# --> this line assumes that the last row in the anova structure contains the residuals

# --> thus it calculates the partial eta squared effect size for all effects with the residuals of the last effect in the model

# So here is a reproduction of the false partial eta squared values as calculated by the sjstats package for the above example:

# false ss.resid (9.457) for all effects:
# main effect factor 1: 317.9 / (317.9 + 9.457) = 0.9711111 -> false result reported by anova_stats
# main effect factor 2: 28.70 / (28.70 + 9.457) = 0.7521556 -> false result reported by anova_stats
# interaction effect factor 1:factor2: 4.814 / (4.814 + 9.457) = 0.3373274 -> this one is correct


@strengejacke
Copy link
Owner

The development of effectsizes from anova models is better supported in the effectsize package, which is also used by the parameters package. Example:

library(parameters)

# generate some data

set.seed(325)

dat <- expand.grid(
  factor1 = c("A","B"),
  factor2 = c("X","Y"),
  participant = 1:10
)
dat$participant <- factor(dat$participant)

dat$value <- rnorm(length(dat$participant))

# add some main effects
dat$value[dat$factor1 == "A"] <- dat$value[dat$factor1 == "A"] + rnorm(dat$value[dat$factor1 == "A"], 5)
dat$value[dat$factor2 == "X"] <- dat$value[dat$factor2 == "X"] + rnorm(dat$value[dat$factor2 == "X"], 2)


# 2x2 repeated measures ANOVA
model <- aov(value ~ factor1*factor2 + Error(participant/(factor1*factor2)), dat)


model_parameters(model, eta_squared = "partial")
#> # participant
#> 
#> Parameter | Sum_Squares | df | Mean_Square
#> ------------------------------------------
#> Residuals |       17.41 |  9 |        1.93
#> 
#> # participant:factor1
#> 
#> Parameter | Sum_Squares | df | Mean_Square |      F |      p | Eta2 (partial)
#> -----------------------------------------------------------------------------
#> factor1   |      317.89 |  1 |      317.89 | 225.57 | < .001 |           0.96
#> Residuals |       12.68 |  9 |        1.41 |        |        |               
#> 
#> # participant:factor2
#> 
#> Parameter | Sum_Squares | df | Mean_Square |     F |     p | Eta2 (partial)
#> ---------------------------------------------------------------------------
#> factor2   |       28.70 |  1 |       28.70 | 12.16 | 0.007 |           0.57
#> Residuals |       21.24 |  9 |        2.36 |       |       |               
#> 
#> # participant:factor1:factor2
#> 
#> Parameter       | Sum_Squares | df | Mean_Square |    F |     p | Eta2 (partial)
#> --------------------------------------------------------------------------------
#> factor1:factor2 |        4.81 |  1 |        4.81 | 4.58 | 0.061 |           0.34
#> Residuals       |        9.46 |  9 |        1.05 |      |       |               
#> 
#> Anova Table (Type 1 tests)

Created on 2022-01-21 by the reprex package (v2.0.1)

I suggest using either effectsize or parameters. The anova-functions in sjstats are going to be deprecated.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants