compare_mi() applies the statistical test described by Parkin and Bray
(2009) to identify registry mortality-to-incidence (M:I) ratios that differ
from comparison registries. Each registry is tested in turn, using all other
registries in the same stratum as its comparison population.
Arguments
- x
An object with class
qualityorqualitiescontaining the registry identifierregicode, incident case countsfbs, and death countssws.- alpha
A single numeric significance level strictly between 0 and 1. Defaults to
0.05, corresponding to a chi-squared critical value of approximately 3.84 with one degree of freedom.- reference
An optional external reference table. It must contain regicode, fbs, and sws; registry and cases are accepted as aliases for regicode and fbs. At least two reference rows are needed in each matched stratum to estimate dispersion.
Value
A tibble containing the stratum identifiers, regicode, fbs,
sws, the observed mi, the number of comparison registries
reference_n, the estimated common ratio theta, the dispersion estimate
phi, the test statistic z2, p_value, critical_value, the logical
flag unusual, and an evaluation status.
Details
The function accepts either the merged quality data frame returned by
create_quality(..., collapse = TRUE) or the qualities list returned by
create_quality(..., collapse = FALSE). Strata are detected automatically
from identifier columns such as year, sex, and cancer.
By default, each registry is compared with the other registries in its stratum. Supply reference to use an external reference table instead. External rows are matched to each target stratum using the stratum columns shared by x and reference. For comparison registries indexed by \(i = 1, \ldots, n\), let \(d_i\) be the number of incident cases, \(m_i\) the number of deaths, and \(n_i = m_i + d_i\). The common M:I ratio and dispersion are estimated as $$\hat{\theta} = \frac{\sum_i m_i}{\sum_i d_i}$$ and $$\hat{\phi} = \frac{1}{n - 1} \sum_i \frac{(m_i - \hat{\theta}d_i)^2}{n_i\hat{\theta}}.$$ For the registry under test, indexed by \(j\), the statistic is $$Z^2 = \frac{(m_j - \hat{\theta}d_j)^2} {\hat{\phi}n_j\hat{\theta}}$$, which is compared with a chi-squared distribution with one degree of freedom.
A status other than "ok" indicates that the row could not be evaluated.
Possible values are "zero_target_incidence",
"insufficient_reference_registries", "zero_reference_incidence",
"zero_reference_mi", and "zero_reference_dispersion". An unusual
result indicates a statistical difference from the comparison registries;
it does not by itself establish incomplete registration.
References
Parkin, D. M., & Bray, F. (2009). Evaluation of data quality in the cancer registry: Principles and methods Part II. Completeness. European Journal of Cancer, 45(5), 756-764. doi:10.1016/j.ejca.2008.11.033
Examples
quality_data <- structure(
data.frame(
regicode = c("A", "B", "C", "D"),
year = "2024",
sex = "1",
cancer = "101",
fbs = c(100, 100, 150, 200),
sws = c(80, 40, 75, 80)
),
class = c("quality", "data.frame")
)
compare_mi(quality_data)
#> # A tibble: 4 × 15
#> regicode year sex cancer fbs sws mi reference_n theta phi z2
#> <chr> <chr> <chr> <chr> <dbl> <dbl> <dbl> <int> <dbl> <dbl> <dbl>
#> 1 A 2024 1 101 100 80 0.8 3 0.433 0.788 21.9
#> 2 B 2024 1 101 100 40 0.4 3 0.522 6.19 0.330
#> 3 C 2024 1 101 150 75 0.5 3 0.5 7.14 0
#> 4 D 2024 1 101 200 80 0.4 3 0.557 4.82 1.31
#> # ℹ 4 more variables: p_value <dbl>, critical_value <dbl>, unusual <lgl>,
#> # status <chr>
