| mantelhaen.test {stats} | R Documentation |
Performs a Cochran-Mantel-Haenszel chi-squared test of the null that two nominal variables are conditionally independent in each stratum, assuming that there is no three-way interaction.
mantelhaen.test(x, y = NULL, z = NULL,
alternative = c("two.sided", "less", "greater"),
correct = TRUE, exact = FALSE, conf.level = 0.95)
x |
either a 3-dimensional contingency table in array form where each dimension is at least 2 and the last dimension corresponds to the strata, or a factor object with at least 2 levels. |
y |
a factor object with at least 2 levels; ignored if |
z |
a factor object with at least 2 levels identifying to which
stratum the corresponding elements in |
alternative |
indicates the alternative hypothesis and must be
one of |
correct |
a logical indicating whether to apply continuity correction when computing the test statistic. Only used in the 2 by 2 by K case. |
exact |
a logical indicating whether the Mantel-Haenszel test or the exact conditional test (given the strata margins) should be computed. Only used in the 2 by 2 by K case. |
conf.level |
confidence level for the returned confidence interval. Only used in the 2 by 2 by K case. |
If x is an array, each dimension must be at least 2, and
the entries should be nonnegative integers. NA's are not
allowed. Otherwise, x, y and z must have the
same length. Triples containing NA's are removed. All
variables must take at least two different values.
A list with class "htest" containing the following components:
statistic |
Only present if no exact test is performed. In the classical case of a 2 by 2 by K table (i.e., of dichotomous underlying variables), the Mantel-Haenszel chi-squared statistic; otherwise, the generalized Cochran-Mantel-Haenszel statistic. |
parameter |
the degrees of freedom of the approximate chi-squared distribution of the test statistic (1 in the classical case). Only present if no exact test is performed. |
p.value |
the p-value of the test. |
conf.int |
a confidence interval for the common odds ratio. Only present in the 2 by 2 by K case. |
estimate |
an estimate of the common odds ratio. If an exact test is performed, the conditional Maximum Likelihood Estimate is given; otherwise, the Mantel-Haenszel estimate. Only present in the 2 by 2 by K case. |
null.value |
the common odds ratio under the null of
independence, |
alternative |
a character string describing the alternative hypothesis. Only present in the 2 by 2 by K case. |
method |
a character string indicating the method employed, and whether or not continuity correction was used. |
data.name |
a character string giving the names of the data. |
The asymptotic distribution is only valid if there is no three-way interaction. In the classical 2 by 2 by K case, this is equivalent to the conditional odds ratios in each stratum being identical. Currently, no inference on homogeneity of the odds ratios is performed.
See also the example below.
Alan Agresti (1990). Categorical data analysis. New York: Wiley. Pages 230–235.
Alan Agresti (2002). Categorical data analysis (second edition). New York: Wiley.
## Agresti (1990), pages 231--237, Penicillin and Rabbits
## Investigation of the effectiveness of immediately injected or 1.5
## hours delayed penicillin in protecting rabbits against a lethal
## injection with beta-hemolytic streptococci.
Rabbits <-
array(c(0, 0, 6, 5,
3, 0, 3, 6,
6, 2, 0, 4,
5, 6, 1, 0,
2, 5, 0, 0),
dim = c(2, 2, 5),
dimnames = list(
Delay = c("None", "1.5h"),
Response = c("Cured", "Died"),
Penicillin.Level = c("1/8", "1/4", "1/2", "1", "4")))
Rabbits
## Classical Mantel-Haenszel test
mantelhaen.test(Rabbits)
## => p = 0.047, some evidence for higher cure rate of immediate
## injection
## Exact conditional test
mantelhaen.test(Rabbits, exact = TRUE)
## => p - 0.040
## Exact conditional test for one-sided alternative of a higher
## cure rate for immediate injection
mantelhaen.test(Rabbits, exact = TRUE, alternative = "greater")
## => p = 0.020
## UC Berkeley Student Admissions
mantelhaen.test(UCBAdmissions)
## No evidence for association between admission and gender
## when adjusted for department. However,
apply(UCBAdmissions, 3, function(x) (x[1,1]*x[2,2])/(x[1,2]*x[2,1]))
## This suggests that the assumption of homogeneous (conditional)
## odds ratios may be violated. The traditional approach would be
## using the Woolf test for interaction:
woolf <- function(x) {
x <- x + 1 / 2
k <- dim(x)[3]
or <- apply(x, 3, function(x) (x[1,1]*x[2,2])/(x[1,2]*x[2,1]))
w <- apply(x, 3, function(x) 1 / sum(1 / x))
1 - pchisq(sum(w * (log(or) - weighted.mean(log(or), w)) ^ 2), k - 1)
}
woolf(UCBAdmissions)
## => p = 0.003, indicating that there is significant heterogeneity.
## (And hence the Mantel-Haenszel test cannot be used.)
## Agresti (2002), p. 287f and p. 297.
## Job Satisfaction example.
Satisfaction <-
as.table(array(c(1, 2, 0, 0, 3, 3, 1, 2,
11, 17, 8, 4, 2, 3, 5, 2,
1, 0, 0, 0, 1, 3, 0, 1,
2, 5, 7, 9, 1, 1, 3, 6),
dim = c(4, 4, 2),
dimnames =
list(Income =
c("<5000", "5000-15000",
"15000-25000", ">25000"),
"Job Satisfaction" =
c("V_D", "L_S", "M_S", "V_S"),
Gender = c("Female", "Male"))))
## (Satisfaction categories abbreviated for convenience.)
ftable(. ~ Gender + Income, Satisfaction)
## Table 7.8 in Agresti (2002), p. 288.
mantelhaen.test(Satisfaction)
## See Table 7.12 in Agresti (2002), p. 297.