| getDesignAgreement {lrstat} | R Documentation |
Power and sample size for Cohen's kappa
Description
Obtains the power given sample size or obtains the sample size given power for Cohen's kappa.
Usage
getDesignAgreement(
beta = NA_real_,
n = NA_real_,
ncats = NA_integer_,
kappaH0 = NA_real_,
kappa = NA_real_,
p1 = NA_real_,
p2 = NA_real_,
rounding = TRUE,
alpha = 0.025
)
Arguments
beta |
The type II error. |
n |
The total sample size. |
ncats |
The number of categories. |
kappaH0 |
The kappa coefficient under the null hypothesis. |
kappa |
The kappa coefficient under the alternative hypothesis. |
p1 |
The marginal probabilities for the first rater. |
p2 |
The marginal probabilities for the second rater. Defaults to be equal to the marginal probabilities for the first rater if not provided. |
rounding |
Whether to round up sample size. Defaults to 1 for sample size rounding. |
alpha |
The one-sided significance level. Defaults to 0.025. |
Details
The kappa coefficient is defined as
\kappa = \frac{\pi_o - \pi_e}{1 - \pi_e},
where
\pi_o = \sum_i \pi_{ii} is the observed agreement, and
\pi_e = \sum_i \pi_{i.} \pi_{.i} is the expected agreement
by chance.
By Fleiss et al. (1969), the variance of \hat{\kappa} is given by
Var(\hat{\kappa}) = \frac{v_1}{n},
where
v_1 = \frac{Q_1 + Q_2 - Q3 - Q4}{(1-\pi_e)^4},
Q_1 = \pi_o(1-\pi_e)^2,
Q_2 = (1-\pi_o)^2 \sum_i \sum_j \pi_{ij}(\pi_{i.} + \pi_{.j})^2,
Q_3 = 2(1-\pi_o)(1-\pi_e) \sum_i \pi_{ii}(\pi_{i.} + \pi_{.i}),
Q_4 = (\pi_o \pi_e - 2\pi_e + \pi_o)^2.
Given \kappa and marginals
\{(\pi_{i.}, \pi_{.i}): i=1,\ldots,k\},
we obtain \pi_o. The only unknowns are the double summation
in Q_2 and the single summation in Q_3.
We find the optimal configuration of cell probabilities that yield the
maximum variance of \hat{\kappa} by treating the problem as a
linear programming problem with constraints to match the given
marginal probabilities and the observed agreement and ensure that
the cell probabilities are nonnegative. This is an extension of
Flack et al. (1988) by allowing unequal marginal probabilities
of the two raters.
We perform the optimization under both the null and alternative
hypotheses to obtain \max Var(\hat{\kappa} | \kappa = \kappa_0)
and \max Var(\hat{\kappa} | \kappa = \kappa_1) for a
single subject, and then calculate the sample size or power
according to the following equation:
\sqrt{n} |\kappa - \kappa_0| = z_{1-\alpha}
\sqrt{\max Var(\hat{\kappa} | \kappa = \kappa_0)} +
z_{1-\beta} \sqrt{\max Var(\hat{\kappa} | \kappa = \kappa_1)}.
Value
An S3 class designAgreement object with the
following components:
-
power: The power to reject the null hypothesis. -
alpha: The one-sided significance level. -
n: The total sample size. -
ncats: The number of categories. -
kappaH0: The kappa coefficient under the null hypothesis. -
kappa: The kappa coefficient under the alternative hypothesis. -
p1: The marginal probabilities for the first rater. -
p2: The marginal probabilities for the second rater. -
piH0: The cell probabilities that maximize the variance of estimated kappa under H0. -
pi: The cell probabilities that maximize the variance of estimated kappa under H1. -
rounding: Whether to round up sample size.
Author(s)
Kaifeng Lu, kaifenglu@gmail.com
References
V. F. Flack, A. A. Afifi, and P. A. Lachenbruch. Sample size determinations for the two rater kappa statistic. Psychometrika 1988; 53:321-325.
Examples
(design1 <- getDesignAgreement(
beta = 0.2, n = NA, ncats = 4, kappaH0 = 0.4, kappa = 0.6,
p1 = c(0.1, 0.2, 0.3, 0.4), p2 = c(0.15, 0.2, 0.24, 0.41),
rounding = TRUE, alpha = 0.05))