scoreci: Score confidence intervals for comparisons of independent...

Description Usage Arguments Value Author(s) References Examples

View source: R/scoreci.R

Description

Score-based confidence intervals for the rate (or risk) difference ("RD") or ratio ("RR") for independent binomial or Poisson rates, or for odds ratio ("OR", binomial only). Including options for bias correction (from Miettinen & Nurminen), skewness correction ("GNbc" method from Laud & Dane, developed from Gart & Nam, and generalised as "SCAS" in Laud 2017) and continuity correction (for strictly conservative coverage). Also includes score intervals for a single binomial proportion or Poisson rate. Based on the Wilson score interval, when corrected for skewness, coverage is almost identical to the mid-p method, or Clopper-Pearson when also continuity-corrected. Hypothesis tests for superiority or non-inferiority are provided using the same score, to ensure consistency between test and CI. This function is vectorised in x1, x2, n1, and n2. Vector inputs may also be combined into a single stratified analysis (e.g. meta-analysis), either using fixed effects, or the more general random effects "TDAS" method, which incorporates stratum variability using a t-distribution score (inspired by Hartung-Knapp-Sidik-Jonkman). For fixed-effects analysis of stratified datasets, with weighting = "MH" for RD or RR, or weighting = "IVS" for OR, omitting the skewness correction produces the CMH test, together with a coherent confidence interval for the required contrast.

Usage

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
scoreci(
  x1,
  n1,
  x2 = NULL,
  n2 = NULL,
  distrib = "bin",
  contrast = "RD",
  level = 0.95,
  skew = TRUE,
  simpleskew = FALSE,
  ORbias = TRUE,
  RRtang = NULL,
  bcf = TRUE,
  cc = FALSE,
  theta0 = NULL,
  precis = 6,
  plot = FALSE,
  plotmax = 100,
  hetplot = FALSE,
  xlim = NULL,
  ylim = NULL,
  stratified = FALSE,
  weighting = NULL,
  MNtol = 1e-08,
  wt = NULL,
  sda = NULL,
  fda = NULL,
  dropzeros = FALSE,
  tdas = NULL,
  random = FALSE,
  prediction = FALSE,
  warn = TRUE,
  ...
)

Arguments

x1, x2

Numeric vectors of numbers of events in group 1 & group 2 respectively.

n1, n2

Numeric vectors of sample sizes (for binomial rates) or exposure times (for Poisson rates) in each group.

distrib

Character string indicating distribution assumed for the input data: "bin" = binomial (default), "poi" = Poisson.

contrast

Character string indicating the contrast of interest: "RD" = rate difference (default), "RR" = rate ratio, "OR" = odds ratio. contrast="p" gives an interval for the single proportion or rate x1/n1.

level

Number specifying confidence level (between 0 and 1, default 0.95).

skew

Logical (default TRUE) indicating whether to apply skewness correction (for the SCAS method recommended in Laud 2017) or not (for the Miettinen-Nurminen method).

simpleskew

Logical (default FALSE) indicating whether to use the "simplified" skewness correction instead of the quadratic solution. See Laud 2021 for details. NOTE: this version of the score is only suitable for obtaining confidence limits, not p-values.

ORbias

Logical (default is TRUE for contrast="OR", otherwise NULL) indicating whether to apply additional bias correction for OR derived from Gart 1985. (Corrigendum to Laud 2017, published May 2018). Only applies if contrast is "OR".

RRtang

Logical indicating whether to use Tang's score for RR: Stheta = (p1hat - p2hat * theta) / p2d (see Tang 2020). Default TRUE for stratified = TRUE, with weighting = "IVS" or "INV". Forced to FALSE for stratified = TRUE, with fixed weighting. Experimental for distrib = "poi".

bcf

Logical (default TRUE) indicating whether to apply bias correction in the score denominator. Applicable to distrib = "bin" only. (NB: bcf = FALSE option is really only included for legacy validation against previous published methods (i.e. Gart & Nam, Mee, or standard Chi-squared test). Ignored for contrast = "p".

cc

Number or logical (default FALSE) specifying (amount of) continuity correction. Numeric value is taken as the gamma parameter in Laud 2017, Appendix S2 (default 0.5 if cc = TRUE). IMPORTANT NOTES: 1) This is a 'continuity correction' aimed at approximating strictly conservative coverage, NOT for dealing with zero cell counts. Such 'sparse data adjustments' are not needed in the score method, except to deal with double-zero cells for RD (& double-100 binomial RD & RR) with IVS/INV weights. 2) The continuity corrections provided here have not been fully tested for stratified methods.

theta0

Number to be used in a one-sided significance test (e.g. non-inferiority margin). 1-sided p-value will be <0.025 iff 2-sided 95% CI excludes theta0. If bcf = FALSE and skew = FALSE this gives a Farrington-Manning test. By default, a two-sided test against theta0 = 0 (for RD) or 1 (for RR/OR) is also output: if bcf = FALSE and skew = FALSE this is the same as Pearson's Chi-squared test.

precis

Number (default 6) specifying precision (i.e. number of decimal places) to be used in optimisation subroutine for the confidence interval.

plot

Logical (default FALSE) indicating whether to output plot of the score function

plotmax

Numeric value indicating maximum value to be displayed on x-axis of plots (useful for ratio contrasts which can be infinite).

hetplot

Logical (default FALSE) indicating whether to output plots for evaluating heterogeneity of stratified datasets.

xlim

pair of values indicating range of values to be plotted.

ylim

pair of values indicating range of values to be plotted.

stratified

Logical (default FALSE) indicating whether to combine vector inputs into a single stratified analysis. IMPORTANT NOTE: The mechanism for stratified calculations is enabled for contrast = "p", but the performance of the resulting intervals has not been fully evaluated.

weighting

String indicating which weighting method to use if stratified = "TRUE": "IVS" = Inverse Variance of Score (see Laud 2017 for details), "INV" = Inverse Variance (bcf omitted, default for contrast = "OR"), "MH" = Mantel-Haenszel (default for contrast = "RD" or "RR"), "MN" = Miettinen-Nurminen iterative weights. For CI consistent with a CMH test, select skew = FALSE and use MH weighting for RD/RR and IVS for OR.

MNtol

Numeric value indicating convergence tolerance to be used in iteration with weighting = "MN".

wt

Numeric vector containing (optional) user-specified weights.

sda

Sparse data adjustment to avoid zero variance when x1 + x2 = 0: Only applied when stratified = TRUE. Default 0.5 for RD with IVS/INV weights. Not required for RR/OR, default is to remove double-zero strata instead.

fda

Full data adjustment to avoid zero variance when x1 + x2 = n1 + n2: Only applied when stratified = TRUE. Default 0.5 for RD & RR with IVS/INV weights. Not required for OR, default is to remove affected strata.

dropzeros

Logical (default FALSE) indicating whether to drop uninformative strata for RR/OR, even when the choice of weights would allow them to be retained for a fixed effects analysis. Has no effect on estimates, just the heterogeneity test.

tdas

(deprecated: parameter renamed to random)

random

Logical (default FALSE) indicating whether to perform random effects meta-analysis for stratified data, using the t-distribution (TDAS) method for stratified data (defined in Laud 2017). NOTE: If random = TRUE, then skew = TRUE only affects the per-stratum estimates.

prediction

Logical (default FALSE) indicating whether to produce a prediction interval (work in progress).

warn

Logical (default TRUE) giving the option to suppress warnings.

...

Other arguments.

Value

A list containing the following components:

estimates

a matrix containing estimates of the rates in each group and of the requested contrast, with its confidence interval

pval

a matrix containing details of the corresponding 2-sided significance test against the null hypothesis that p_1 = p_2, and one-sided significance tests against the null hypothesis that theta >= or <= theta0

call

details of the function call

If stratified = TRUE, the following outputs are added:

Qtest

a vector of values describing and testing heterogeneity

weighting

a string indicating the selected weighting method

stratdata

a matrix containing stratum estimates and weights

Author(s)

Pete Laud, p.j.laud@sheffield.ac.uk

References

Laud PJ. Equal-tailed confidence intervals for comparison of rates. Pharmaceutical Statistics 2017; 16:334-348.

Laud PJ. Corrigendum: Equal-tailed confidence intervals for comparison of rates. Pharmaceutical Statistics 2018; 17:290-293.

Laud PJ, Dane A. Confidence intervals for the difference between independent binomial proportions: comparison using a graphical approach and moving averages. Pharmaceutical Statistics 2014; 13(5):294–308.

Miettinen OS, Nurminen M. Comparative analysis of two rates. Statistics in Medicine 1985; 4:213-226.

Farrington CP, Manning G. Test statistics and sample size formulae for comparative binomial trials with null hypothesis of non-zero risk difference or non-unity relative risk. Statistics in Medicine 1990; 9(12):1447-1454.

Gart JJ. Analysis of the common odds ratio: corrections for bias and skewness. Bulletin of the International Statistical Institute 1985, 45th session, book 1, 175-176.

Gart JJ, Nam Jm. Approximate interval estimation of the ratio of binomial parameters: a review and corrections for skewness. Biometrics 1988; 44(2):323-338.

Gart JJ, Nam Jm. Approximate interval estimation of the difference in binomial parameters: correction for skewness and extension to multiple tables. Biometrics 1990; 46(3):637-643.

Tang Y. Score confidence intervals and sample sizes for stratified comparisons of binomial proportions. Statistics in Medicine 2020; 39:3427–3457.

Examples

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
# Binomial RD, SCAS method:
scoreci(
  x1 = c(12, 19, 5), n1 = c(16, 29, 56),
  x2 = c(1, 22, 0), n2 = c(16, 30, 29)
)

# Binomial RD, MN method:
scoreci(
  x1 = c(12, 19, 5), n1 = c(16, 29, 56),
  x2 = c(1, 22, 0), n2 = c(16, 30, 29), skew = FALSE
)

# Poisson RR, SCAS method:
scoreci(x1 = 5, n1 = 56, x2 = 0, n2 = 29, distrib = "poi", contrast = "RR")

# Poisson RR, MN method:
scoreci(
  x1 = 5, n1 = 56, x2 = 0, n2 = 29, distrib = "poi",
  contrast = "RR", skew = FALSE
)

# Binomial rate, SCAS method:
scoreci(x1 = c(5, 0), n1 = c(56, 29), contrast = "p")

# Binomial rate, Wilson score method:
scoreci(x1 = c(5, 0), n1 = c(56, 29), contrast = "p", skew = FALSE)

# Poisson rate, SCAS method:
scoreci(x1 = c(5, 0), n1 = c(56, 29), distrib = "poi", contrast = "p")

# Stratified example, using data from Hartung & Knapp:
scoreci(
  x1 = c(15, 12, 29, 42, 14, 44, 14, 29, 10, 17, 38, 19, 21),
  x2 = c(9, 1, 18, 31, 6, 17, 7, 23, 3, 6, 12, 22, 19),
  n1 = c(16, 16, 34, 56, 22, 54, 17, 58, 14, 26, 44, 29, 38),
  n2 = c(16, 16, 34, 56, 22, 55, 15, 58, 15, 27, 45, 30, 38),
  stratified = TRUE
)

# TDAS example, using data from Hartung & Knapp:
scoreci(
  x1 = c(15, 12, 29, 42, 14, 44, 14, 29, 10, 17, 38, 19, 21),
  x2 = c(9, 1, 18, 31, 6, 17, 7, 23, 3, 6, 12, 22, 19),
  n1 = c(16, 16, 34, 56, 22, 54, 17, 58, 14, 26, 44, 29, 38),
  n2 = c(16, 16, 34, 56, 22, 55, 15, 58, 15, 27, 45, 30, 38),
  stratified = TRUE, random = TRUE
)

# Stratified example, with extremely rare instance of non-calculable skewness
# correction seen on plot of score function:
scoreci(
  x1 = c(1, 16), n1 = c(20, 40), x2 = c(0, 139), n2 = c(80, 160),
  contrast = "RD", skew = TRUE, simpleskew = FALSE,
  distrib = "bin", stratified = TRUE, plot = TRUE, weighting = "IVS"
)

ratesci documentation built on Dec. 11, 2021, 9:36 a.m.