<!-- README.md is generated from README.Rmd. Please edit that file -->

presize <img src='man/figures/logo.png' align="right" width="200">

<– –>

Bland (2009) recommended to base study sizes on the width of the confidence interval rather the power of a statistical test. The goal of presize is to provide functions for such precision based sample size calculations. For a given sample size, the functions will return the precision (width of the confidence interval), and vice versa.

Installation

presize can be installed from CRAN in the usual manner:

install.packages("presize")

You can install the development version of presize with:

install.packages('presize', repos = c('https://dcr-unibe-ch.r-universe.dev', 'https://cloud.r-project.org'))

Overview

presize provides functions for

| Measure | Function | Methods available | |---------------------------------------|------------------|--------------------------------------------------------------------------------------------------------------------------| | Descriptive measures | | | | Mean | prec_mean | | | Proportion | prec_prop | Wilson, Agresti-Coull, exact, Wald (see Brown, Cai, and DasGupta 2001) | | Rate | prec_rate | Score, variance stabilizing, exact, Wald (see Barker 2002) | | Absolute differences | | | | Mean difference | prec_meandiff | | | Risk difference | prec_riskdiff | Newcombe (Newcombe 1998), Miettinen-Nurminen (Miettinen and Nurminen 1985), Agresti-Caffo (Agresti and Caffo 2000), Wald | | Relative differences | | | | Odds ratio | prec_or | Gart, Wolff, independence smoothed logit (see Fagerland, Lydersen, and Laake 2015) | | Risk ratio | prec_riskratio | Koopman (Koopman 1984), Katz (Katz et al. 1978) | | Rate ratio | prec_rateratio | Rothman (Rothman and Greenland 2018) | | Correlation measures | | | | Correlation coefficient | prec_cor | Pearson, Kendall, Spearman (see Bonnett and Wright 2000) | | Intraclass correlation | prec_icc | Bonnett (2002) | | Limit of agreement | prec_lim_agree | Bland and Altman (1986) | | Cohen’s kappa | prec_kappa | Rotondi and Donner (2012) | | Cronbach’s alpha | prec_cronb | Bonett and Wright (2015) | | Diagnostic measures | | | | Sensitivity¹ | prec_sens | As per prec_prop | | Specificity¹ | prec_spec | As per prec_prop | | Area under the curve | prec_auc | Hanley and McNeil (1982) | | Negative likelihood ratio² | prec_neg_lr | Simel, Samsa, and Matchar (1991) | | Positive likelihood ratio² | prec_pos_lr | Simel, Samsa, and Matchar (1991) | | Generic likelihood ratio | prec_lr | Simel, Samsa, and Matchar (1991) |

¹ Simple wrappers for prec_prop.

² Wrappers for prec_lr with values provided via sens and spec

Example

Suppose we want to estimate the proportion of hospital admissions with diabetes. Diabetes has a prevalence of approximately 10% (Emerging Risk Factors Collaboration et al. (2010)). We assume a slightly higher proportion of diabetics, 15%, as diabetes is a risk factor for a wide range of conditions. We want to estimate the prevalence of diabetes to within 5% (plus/minus 2.5%). With presize, this is simple. We use the prec_prop (precision of a proportion) function and pass our 15% and 5% as arguments p and conf.width:

library(presize) # load the package
prec_prop(p = 0.15, conf.width = 0.05)
#> Warning in prec_prop(p = 0.15, conf.width = 0.05): more than one method was
#> chosen, 'wilson' will be used
#> 
#>      sample size for a proportion with Wilson confidence interval. 
#> 
#>      p      padj        n conf.width conf.level       lwr       upr
#> 1 0.15 0.1517077 783.4897       0.05       0.95 0.1267077 0.1767077
#> 
#> NOTE: padj is the adjusted proportion, from which the ci is calculated.

In the n column, we see that we would need to ask 784 (rounding 783.5 up) patients to achieve the desired CI width. Disappointingly, we also know that we only have funds to collect the data from 600 patients. We wonder if 600 patients would yield sufficient precision - we could also accept a CI width of 6% (plus/minus 3%). In such a case, we can pass the arguments p and n.

prec_prop(p = 0.15, n = 600)
#> Warning in prec_prop(p = 0.15, n = 600): more than one method was chosen,
#> 'wilson' will be used
#> 
#>      precision for a proportion with Wilson confidence interval. 
#> 
#>      p      padj   n conf.width conf.level       lwr       upr
#> 1 0.15 0.1522266 600 0.05713404       0.95 0.1236596 0.1807936
#> 
#> NOTE: padj is the adjusted proportion, from which the ci is calculated.

Now we see that with 600 patients, the CI would have a width of 5.7%. We are happy with this and continue planning our study with those values. All of the functions listed in Table 1 can be used similarly.

We can also look at a range of scenarios simulatenously by passing a vector to one of the arguments, which could be used to create something analogous to a power curve:

prec_prop(p = 0.15, n = seq(600, 800, 50))
#> Warning in prec_prop(p = 0.15, n = seq(600, 800, 50)): more than one method was
#> chosen, 'wilson' will be used
#> 
#>      precision for a proportion with Wilson confidence interval. 
#> 
#>      p      padj   n conf.width conf.level       lwr       upr
#> 1 0.15 0.1522266 600 0.05713404       0.95 0.1236596 0.1807936
#> 2 0.15 0.1520563 650 0.05489329       0.95 0.1246097 0.1795030
#> 3 0.15 0.1519102 700 0.05289705       0.95 0.1254617 0.1783588
#> 4 0.15 0.1517835 750 0.05110386       0.95 0.1262316 0.1773355
#> 5 0.15 0.1516726 800 0.04948148       0.95 0.1269319 0.1764133
#> 
#> NOTE: padj is the adjusted proportion, from which the ci is calculated.

Shiny app

An online interactive version of the package is available here. The app can also be launched locally via launch_presize_app() in RStudio.

<img src="man/figures/app.png" alt="" width="1034" />

Getting help

The package website, including more details on the functions, can be found here.

If you have a question, feel free to make a thread on the discussion page.