NEWS

NEWS and changes for the pvaluefunctions package

Bug fixes

x_scale = "logarithm" is now honored. Previously only the automatic logarithmic scale for trans = "exp" worked.
Pearson’s correlation coefficient: the exact distribution no longer fails with n larger than about 172 (overflow of the gamma function), and the counternull is now computed accurately for extreme tail probabilities. Several Pearson estimates in one call (n and estimate of length > 1) no longer fail.
The lengths of stderr, df, tstat and estimate are now checked correctly. Previously a mismatch of the third argument went unnoticed.
trans can now be a function as documented, not only the name of a function. Names are no longer forced to lower case when looking up the function (a lower case name is still tried as a fallback), functions defined inside the calling function are found, and the transformation function does not need an argument named x.
An invalid type now gives an informative error message.
One-sided confidence limits for proportions (type = "prop") now use the matching two-sided level (2 * conf_level - 1) like all other types. Previously, the two-sided interval of level conf_level was returned.
The labels for the confidence levels no longer trigger a deprecation warning with ggplot2 4.0.0 or later (label.size).
Null values that are outside of the plotting area or whose counternull is missing no longer cause errors when plotting.

Input validation

Inputs are now validated up front and give informative errors, e.g. for missing values, non-positive standard errors, degrees of freedom or sample sizes, non-logical flags, an invalid n_values, plot_p_limit, cut_logyaxis, nrow or ncol, and duplicated est_names.
Correlation coefficients of exactly -1 or 1 are rejected. The minimum sample size is 4 for Pearson’s correlation (n = 3 never worked).
t-statistics must be non-zero and have the same sign as the estimates.
A message is printed if confidence levels are ignored because they are too low for the alternative.

Internal changes

Code is split into several files, duplicated code was removed and imports are now explicit. ggplot2 (>= 3.5.0) and scales (>= 1.3.0) are now required (transform instead of trans, new_transform()).
README.Rmd and tooling files are now correctly excluded from the package build.
Added a testthat test suite (closed-form checks against stats, input validation, plot properties and regression tests against the results of version 1.6.3).

Deprecated arguments in ggplot2 were changed, such as “size” in “geom_line”. No more messages should appear now.
Small changes in the references in the vignette.

The calculations for Pearson’s correlation coefficient now use the exact distribution described in a preprint by Gunnar Taraldsen (2020). This makes the use of the gsl package necessary because the calculations involve the Gaussian hypergeometric function (2F1). Warning: This can make the calculations drastically longer if a high value of n_values is used.
The function conf_dist now also calculates the proportion of the AUCC that lies above any specified null values.
Small changes in the help page for conf_dist.

Fixed a bug concerning the calculation of Newcombe’s Wilson score interval with continuity correction for the difference of two proportions.
Removed the “cairo” device from pngs in the vignette.

The returned data frame is now sorted for convenience; this is purely cosmetic.
Dependence on R increased to R version 3.5.0
Added an option same_color to specify whether curves should be distinguished by colors or not if they are plotted together in the same graph. Can be useful if there are many curves plotted together.
Added an option plot_legend to specify whether a legend should be drawn if multiple curves are plotted together and distinguished by color (i.e. same_color = FALSE and together = TRUE).
Added an option col to specify the color of the curves if they are not to be distinguished by color (i.e. same_color = FALSE).
Areas under the confidence curves (AUCC) according to Berrar (2017) are now calculated and returned. They offer a way to compare multiple estimates with respect to their precision. The AUCC is calculated on the untransformed scale using numerical integration (trapezoidal integration) implemented in the pracma package which is now imported.
Removed data link to external data (UCLA) in an example (odds ratio). Certificates for this site apparently expired.

Vignette, README and DESCRIPTION updated to reflect that the newest version of ggplot2 (3.2.1) fixes the former bug with sec_axis.
Added an option plot to conf_dist that controls whether a plot is created or not. If users want to create their own plots, they can set this option to FALSE and use the returned data (res_frame) which is the basis for the plots to create them.

Added option inverted which allows users to plot p-value functions, s-value functions and confidence distributions with the y-axis inverted.
Added a new example showing a p-value function for an odds ratio with an inverted y-axis (cf. Bender et al. 2005).
The option xlim is now strictly enforced: Any null values that are outside of the specified x-axis-limits are not plotted and a corresponding message is printed out as information.
Added new option x_scale to manually force the scaling of the x-axis.
Added two more examples in the vignette replicating Figure 1 and Figure 2 from Bender et al. (2005).
Various smaller bug fixes and improvements: Fixed some checks, fixed plotting vertical lines for null values.

NEWS file was converted to an .md file.
Users can now provide a title for the plot (option title).
Users can now provide titles for the primary and secondary y-axis (options ylab and ylab_sec).
Users can now specify the number of rows nrow and columns ncol to be used in facet_wrap (ggplot2) when multiple estimates are plotted separately (option together = FALSE).
Changed some of the examples: Changed option log_yaxis = TRUE to log_yaxis = FALSE.
Code: Changed logical checks x == TRUE and x == FALSE to isTRUE(x) and to isFALSE(x).
Code: Checked and improved/fixed some of the initial consistency/input checks that are performed at the beginning of the function.

Added option plot_counternull to plot the counternull value(s) on the graphics if applicable and possible.
Effect sizes on the log-scale (e.g. Odds ratio, Hazard ratio, Incidence rate ratio) are now plotted on a logarithmic x-axis so that the p-value function is symmetric around the point estimate.
The default is now to omit a logarithmic part of the y-axis.
Warnings from ggplot2 are now suppressed during the function call.
Not interesting for practitioners: Source code for creating the plots was cleaned up and more comments were added.

Improved vignette with reduced figure size and several orthographic errors fixed. Added a new example (difference between proportions)
New estimate type implemented: Difference between two independent proportions (type = "propdiff") based on Wilson’s interval (see Newcombe (1998) for details)