rmcorr Estimates with NaN"
In rmcorr: Repeated Measures Correlation

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)

knitr::opts_chunk$set(echo = TRUE, tidy = FALSE)
options(width = 80)
library(knitr)
library(rmarkdown)
library(rmcorr) 
library(ggplot2)
library(ggExtra)

NaN estimates

This synthetic dataset produces NaN estimates with rmcorr. Thanks to Shreya Ghosh for this example.

Running Examples Requires ggExtra [@ggExtra]

install.packages("ggExtra")
require(ggExtra)

Load data, visualize, and model

load(file = "../man/data/ghosh_synth.rda")

#Look at data
ghosh_synth #Note lots of repeated zeros in A3 and A4

set.seed(40) #Make jittering reproducible 
p <- ggplot(ghosh_synth, aes(x = A4, y = A3)) +
            geom_point(alpha = 0.2) +
            geom_jitter(width = 2, height = 2) 
p1 <- ggMarginal(p, type="histogram")
p1

rmc.ghosh <- rmcorr(Subject, A3, A4, ghosh_synth)

rmc.ghosh

#The default rmcorr plot doesn't jitter values, this masks identical values because they are drawn on top of each other
plot(rmc.ghosh)

The NaN estimates appear to be due to insufficient varability in the dataset. A possible way to address this issue is adding a small amount of random noise.

Add random noise

set.seed(67) 
small.noise1 <- rnorm(dim(ghosh_synth)[[1]], 0, 0.2)
small.noise2 <- rnorm(dim(ghosh_synth)[[1]], 0, 0.2)

ghosh_synth$A3.noise <- ghosh_synth$A3 + small.noise1
ghosh_synth$A4.noise <- ghosh_synth$A4 + small.noise2

rmc.ghosh.noise <- rmcorr(Subject, A3.noise, A4.noise, ghosh_synth)

rmc.ghosh.noise

p2 <- ggplot(ghosh_synth, aes(x = A3.noise, y = A4.noise, 
       group = factor(Subject), color = factor(Subject))) +
       ggplot2::geom_point(ggplot2::aes(colour = factor(Subject), 
                                        alpha = 0.10)) +
       ggplot2::geom_line(aes(y = rmc.ghosh.noise$model$fitted.values),
                         linetype = 1) +
     theme(legend.position="none")

p3 <- ggMarginal(p2, type="histogram")
p3

Caveats

The results with rmcorr should be interpreted with some caution because the data are non-normal with zero-inflation. Still, these results provides at least a starting point: A common linear association around 0. A much more complicated alternative is fitting a multilevel model with an appropriate distribution for zero-inflated data (e.g., negative binomial distribution or zero-inflated Poisson).