# Reliability Analysis In SimplyAgree: Flexible and Robust Agreement and Reliability Analyses

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


# Background

Another feature of this R package is the ability to estimate the reliability of a measurement. This R package allows for the calculation of Intraclass Correlation Coefficients (ICC), various standard errors (SEM, SEE, and SEP), and coefficient of variation. All of the underlying calculations (sans the coefficient of variation) is based on the paper by @weir2005^[The paper by Weir also appears to heavily rely on the work of @shrout1979 and @mcgraw1996]. This is a fairly popular paper within my own field (kinesiology), and hence was the inspiration for creating this function that provides all the calculative approaches included within that manuscript.

# Code Demonstration

For this package, the test-retest reliability statistics can be calculated with the reli_stats function. This function allow for data to be input in a long (multiple rows of data for each subject) or in wide (one row for each subject but a column for each item/measure).

For the long data form, the column containing the subject identifier (id), item number (item), and measurements (measure) are provided. In this function I refer to items similar to if we were measuring internal consistency for a questionnaire (which is just a special case of test-retest reliability). So, item could also be refer to time points, which is what is typically seen in human performance settings where test-retest reliability may be evaluated over the course of repeated visits to the same laboratory. If wide is set to TRUE then the columns containing the measurements are provided (e.g., c("value1","value2","value3")).

To demonstrate the function, I will create a data set in the wide format.

  # Example from Shrout and Fleiss (1979), pg. 423
dat = data.frame(judge1 =
c(9,6,8,7,10,6),
judge2 =
c(2,1,4,1,5,2),
judge3 =
c(5,3,6,2,6,4),
judge4 =
c(8,2,8,6,9,7))


Now, that we have a data set (dat), I can use it in the reli_stats function.

test1 = reli_stats(
data = dat,
wide = TRUE,
col.names = c("judge1", "judge2", "judge3", "judge4")
)


This function also has generic print and plot functions. The output from print provides the coefficient of variation, standard errors, and a table of various intraclass correlation coefficients. Notice the conclusions about the reliability of the measurement here would vary greatly based on the statistic being reported. What statistic you should report is beyond the current vignette, but is heavily detailed in @weir2005. However, within the table there are columns for model and measures which describe the model that is being used and the what these different ICCs are intended to measure, respectively.

print(test1)


Also included in the results is a plot of the measurements across the items (e.g., time points).

plot(test1)


# Calculative Approach

## Model

The linear mixed model used for the calculations is specified as the following:

$$Y_{i} \sim\ N \left(\alpha_{j[i],k[i]}, \sigma^2 \right)$$

$$\alpha_{j} \sim\ N \left(\mu_{\alpha_{j}}, \sigma^2_{\alpha_{j}} \right) \text{, for id j = 1,} \dots \text{,J}$$

$$\alpha_{k} \sim\ N \left(\mu_{\alpha_{k}}, \sigma^2_{\alpha_{k}} \right) \text{, for items k = 1,} \dots \text{,K}$$

## Components of Variance

Mean Squared Error (MSE)

$$MSE = \sigma^2$$

Variance Between Subjects

$$MSB = n_k \cdot \sigma^2_{\alpha j} + \sigma^2$$

Variance Between Items/Judges

$$MSJ = n_{j} \cdot \sigma^2_{\alpha_{k}} + \sigma^2$$

Variance Within Subjects/Participants

$$MSW = \sigma^2 + \sigma^2_{\alpha_{k}}$$

## Intraclass Correlation Coefficients

$$ICC_{(1,1)} = \frac{MSB - MSW}{MSB + (n_j-1) \cdot MSW}$$ $$ICC_{(2,1)} = \frac{MSB - MSE}{MSB+(n_j -1) \cdot MSE + n_j \cdot (MSJ - MSE) \cdot n^{-1}}$$ $$ICC({3,1)} = \frac{MSB - MSE}{MSB + (n_j -1) \cdot MSE}$$ $$ICC{(1,k)} = \frac{MSB - MSW}{MSB}$$ $$ICC_{(2,k)} = \frac{MSB - MSW}{MSB + (MSJ - MSE) \cdot n_j^{-1}}$$ $$ICC_{(3,k)} = \frac{MSB - MSE}{MSB}$$

## ICC Confidence Intervals

### ICC(1,1)

$$F = \frac{MSB}{MSW}$$

$$df_{n} = n_j - 1$$

$$df_{d} = n_j \cdot (n_k - 1)$$

$$F_{L} = \frac{F}{F_{(1 - \alpha, \space df_{n}, \space df_{d})}}$$

$$F_{U} = F \cdot F_{(1 - \alpha, \space df_{d}, \space df_{n})}$$

$$Lower \space CI = \frac{(F_L - 1)}{(F_L + (n_j - 1))}$$

$$Upper \space CI = \frac{(F_U - 1)}{(F_U + n_j - 1)}$$

### ICC(2,1)

$$F = \frac{MSJ}{MSE}$$

$$vn = (n_k - 1) \cdot (n_j - 1) \cdot [(nj \cdot ICC_{(2,1)} \cdot F + n_j \cdot (1 + (n_k - 1) \cdot ICC_{(2,1)}) - n_k \cdot ICC_{(2,1)})]^2$$

$$vd = (n_j - 1) \cdot n_k^2 \cdot ICC_{(2,1)}^2 \cdot F^2 + (n_j \cdot (1 + (n_k - 1) \cdot ICC_{(2,1)}) - n_k \cdot ICC_{(2,1)})^2$$ $$v = \frac{vn}{vd}$$

$$F_{L} = F_{(1 - \alpha, \space n_j-1, \space v)}$$

$$F_{U} = F_{(1 - \alpha, \space v, \space n_j - 1)}$$

$$Lower \space CI = \frac{n_j \cdot (MSB - F_U \cdot MSE)}{(F_U \cdot (n_k \cdot MSJ + (n_k \cdot n_j - n_k - n_j) \cdot MSE) + n_j \cdot MSB)}$$

$$Upper \space CI = \frac{n_j \cdot (MSB \cdot F_L - MSE)}{(n_k \cdot MSJ + (n_k \cdot n_j - n_k - n_j) \cdot MSE + n_j \cdot F_L \cdot MSB)}$$

### ICC(3,1)

$$F = \frac{MSJ}{MSE}$$ $$df_{n} = n_j - 1$$ $$df_{d} = (n_j-1) \cdot (n_k - 1)$$

$$F_{L} = \frac{F}{F_{(1 - \alpha, \space df_{n}, \space df_{d})}}$$ $$F_{U} = F \cdot F_{(1 - \alpha, \space df_{n}, \space df_{d})}$$ $$F3L <- F31/qf(1 - alpha, df21n, df21d)$$ $$F3U <- F31 * qf(1 - alpha, df21d, df21n)$$ $$Lower \space CI = (F3L - 1)/(F3L + n_k - 1)$$

$$Upper \space CI = (F3U - 1)/(F3U + n_k - 1)$$

### ICC(1,k)

$$F = \frac{MSB - MSW}{MSB}$$ $$df_{n} = n_j - 1$$

$$df_{d} = n_j \cdot (n_k - 1)$$

$$F_{L} = \frac{F}{F_{(1 - \alpha, \space df_{n}, \space df_{d})}}$$ $$F_{U} = F \cdot F_{(1 - \alpha, \space df_{d}, \space df_{n})}$$

$$Lower \space CI = 1-\frac{1}{F_L}$$ $$Upper \space CI = 1-\frac{1}{F_U}$$

### ICC(2,k)

$$F = \frac{MSB - MSW}{MSB}$$

$$vn = (n_k - 1) \cdot (n_j - 1) \cdot [(nj \cdot ICC_{(2,k)} \cdot F + n_j \cdot (1 + (n_k - 1) \cdot ICC_{(2,k)}) - n_k \cdot ICC_{(2,k)})]^2$$

$$vd = (n_j - 1) \cdot n_k^2 \cdot ICC_{(2,k)}^2 \cdot F^2 + (n_j \cdot (1 + (n_k - 1) \cdot ICC_{(2,k)}) - n_k \cdot ICC_{(2,k)})^2$$

$$v = \frac{vn}{vd}$$

$$F_{L} = F_{(1 - \alpha, \space n_j-1, \space v)}$$ $$F_{U} = F_{(1 - \alpha, \space v, \space n_j - 1)}$$

$$L3 = \frac{n_j \cdot (MSB - F_U \cdot MSE)}{(F_U \cdot (n_k \cdot MSJ + (n_k \cdot n_j - n_k - n_j) \cdot MSE) + n_j \cdot MSB)}$$

$$U3 = \frac{n_j \cdot (MSB \cdot F_L - MSE)}{(n_k \cdot MSJ + (n_k \cdot n_j - n_k - n_j) \cdot MSE + n_j \cdot F_L \cdot MSB)}$$

$$Lower \space CI = \frac{L3 \cdot n_k}{(1 + L3 \cdot (n_k - 1))}$$

$$Upper \space CI = \frac{U3 \cdot n_k}{(1 + U3 \cdot (n_k - 1))}$$

### ICC(3,k)

$$F = \frac{MSB}{MSE}$$

$$df_n = n_j - 1$$

$$df_d = (n_j - 1) \cdot (n_k - 1)$$

$$F_L = \frac{F}{F_{(1 - \alpha, \space df_n, \space df_d})}$$

$$F_U = F \cdot F_{(1 - \alpha, \space df_d, \space df_n)}$$

$$Lower \space CI = 1-\frac{1}{F_L}$$

$$Upper \space CI = 1-\frac{1}{F_U}$$

## Standard Error Calculations

The standard error of the measurement (SEM), standard error of the estimate (SEE), and standard error of prediction (SEP) are all estimated with the following calculations.

$$SEM = \sqrt{MSE}$$

$$SEE = \sqrt{\frac{\sigma^2+\sigma^2_{\alpha j}+\sigma^2_{\alpha k}}{(n_j-1)}} \cdot \sqrt{ICC_{(3,1)} \cdot (1-ICC_{(3,1)})}$$

$$SEP = \sqrt{\frac{\sigma^2+\sigma^2_{\alpha j}+\sigma^2_{\alpha k}}{(n_j-1)}} \cdot \sqrt{1-ICC_{(3,1)}^2}$$

## Coefficient of Variation

The CV is calculated 3 potential ways within reli_stats. I highly recommend reporting the default version of CV.

1. From the MSE (default)

$$CV = \frac{MSE }{ \bar y}$$

1. From the SEM (most conservative)^[Also called "%SEM" in some texts]

$$CV = \frac{SEM}{ \bar y}$$

1. From the model residuals (most liberal)

$$CV = \frac{\sqrt{\frac{\Sigma^{N}{i=1}(y_i - \hat y_i)^2}{N{obs}}}}{ \bar y}$$

## Other Confidence Intervals

If the other_ci argument is set to TRUE then parametric bootstrapped confidence intervals (using) will be calculated for the CV, SEM, SEP, and SEE. The number of resamples can be set with the replicates argument (default is 1999; increase for greater accuracy or lower for greater speed). The reported confidence intervals are estimated using the percentile method type = 'perc', the normal type = 'norm', or basic methods type = 'basic'.

To ensure reproducibility, please use set.seed() when these confidence intervals are calculated.

# Inter-Rater Reliability

In some cases, the reliability of a categorical or ordinal scale may be worth investigating. For example, physicians may want to develop a diagnosis tool and ensure that the diagnosis is reliable (i.e., categorical designation) or severity of the disease (i.e., a Likert-type scale). Coefficients can be calculated to assess the degree of inter-rater reliability. In its simplest form, the percent agreement between all the raters can be calculated. All other coefficients of agreement are esssentially trying to "correct" for random guessing of the rater. The function to make these calculations in the SimplyAgree is agree_coef, and it produces 4 estimates: percent agreement, Gwet's AC, Fleiss' Kappa, and Krippendorff's Alpha. However, other packages provide much a much greater breadth of calculative approaches [@irrCAC].

In the agree_coef function, the user can specify weighted = TRUE. If this argument is set to true than the ratings have quadratic weights applied to them. Essentially, this penalizes values farther away from each other more than those close to each other. For example, a pair of values equal to 3 and 4 would be penalized less than a pair of values to 1 and 4. For more details, on these "agreement coefficients" I refer all users to Gwet's textbook on inter-rater agreement [@gwet].

As a demonstration, we can create a matrix of ratings.

  ratermat1 = ("Rater1 Rater2 Rater3 Rater4
1       1      1     NA      1
2       2      2      3      2
3       3      3      3      3
4       3      3      3      3
5       2      2      2      2
6       1      2      3      4
7       4      4      4      4
8       1      1      2      1
9       2      2      2      2
10     NA      5      5      5
11     NA     NA      1      1
12     NA     NA      3     NA")

row.names=1))


We can then perform the analysis without the weights.

agree_coef(data = ratermat2,
wide = TRUE,
weighted = FALSE,
col.names = c("Rater1", "Rater2", "Rater3", "Rater4"))


Or, perform it with weighting.

agree_coef(data = ratermat2,
wide = TRUE,
weighted = TRUE,
col.names = c("Rater1", "Rater2", "Rater3", "Rater4"))


# References

## Try the SimplyAgree package in your browser

Any scripts or data that you put into this service are public.

SimplyAgree documentation built on Aug. 24, 2022, 9:05 a.m.