NEWS.md
New features
fairness_selection() to aid in the decision between fairness metrics. The function has an accompanying print() and plot() method.plot(x, type = "sequential") to the evaluation() function when used with prior, materiality and data. The plot shows Bayes factors as a function of the sample size.Bug fixes
evaluation() function using method = "hypergeometric" where the user could provide broken taints. These taints are now rounded to the nearest integer value.Minor changes
auditPrior() function, method = "factor" has been replaced with method = "power" to be consistent with existing terminology.selection() now always performs sampling with replacement when using items = "values" and method = "random" (Thanks to @lottemensink for the suggestion). This should not lead to substantial differences in typical audit populations.New features
planning() function. This can be done by providing a vector of integers to the expected argument that specifies the intended sampling plan. For instance, expected = c(1, 0) gives the sample size for a two-stage sampling plan in which the auditor is allowed to extend the initial sample after finding one misstatement and is allowed to approve the population if the initial sample contains no misstatements.method = "inflated.poisson" and method = "hurdle.beta". Since these methods rely on mcmc sampling, they can take any form of prior. They can also be fitted using maximum likelihood estimation with prior = FALSE.Minor changes
type = "estimates" is now the default option.Major changes
report() function as it did not align with the rest of the functionality in the package.Bug fixes
selection() where the numeric data in the sample was returned as character if the original data was entered as character only (i.e., selection(data.frame(x = c("hello", "bye", "hi")), size = 2)).N.units is now only taken from the prior if the user does not provide it manually to the planning() and evaluation() functions.New features
model_fairness() to compute fairness metrics for algorithm auditing on the basis of predictions of an algorithm.Minor changes
evaluation() function, the typical stringer bound has been changed from method = "stringer" to method = "stringer.binomial". Furthermore, additional methods stringer.poisson and stringer.hypergeometric are added that apply the stringer bound using the Poisson and hypergeometric distributions, respectively. The stringer method will remain for now but redirect to stringer.binomial.rstantools 2.3.1.New features
digit_test() function now returns element estimates containing observed proportions and their corresponding confidence intervals. The plot output also shows these intervals around the estimates.Bug fixes
New features
evaluation() function.digitTests package via the digit_test() and repeated_test() functions.Bug fixes
summary(selection) was displayed incorrectly.Minor changes
Minor changes
Warning: <img> attribute "align" not allowed for HTML5.R 4.2.0.Minor changes
x in evaluation() and expected in planning() is now automatically ceiled (while throwing a warning) for method = "hypergeometric".Major changes
jfaPredictive. The probabilities for the prior and posterior predictive distributions can be obtained by calling the predict() function.Minor changes
N.units in auditPrior(), planning(), and evaluation() is now automatically ceiled for likelihood/method = 'hypergeometric'.likelihood in the prior does not match with the likelihood/method inputs for planning() and evaluation(). The likelihood from the prior is leading in these cases and will overrule the other options.New features
randomize to the selection() function, which allows the user to randomly shuffle the items in the population before selection. Note that specifying randomize = TRUE overrules order.Bug fixes
planning stopped at max - 1).Major changes
order argument in selection() no longer accepts a logical, but instead takes the column name of the ranking variable in the data.Minor changes
New features
alternative with possible options less (default), two.sided, and greater to the evaluation() function that allows control over the type of hypothesis test to perform and the type of confidence / credible interval to calculate.predict.jfaPrior() and predict.jfaPosterior() that produce predictions for the data under the prior or posterior distribution.method = 'param' to function auditPrior() which takes as input the raw alpha and beta parameters of the prior distribution.method = 'strict' to function auditPrior() which constructs an (improper) prior distribution that yields the same results (with respect to sample sizes and upper limits) as classical procedures.method = 'sieve') to selection().jfa.auditPrior(), planning(), and evaluation() now contain information about the posterior predictive distribution when N.units is specified.Major changes
jfa 0.5.7 to jfa 0.6.0 there has been a major overhaul in the names of function arguments. This is done so that the calls integrate better with general R syntax and the package gets more user-friendly. I apologize for any inconvenience this may cause. The following names have been changed:
median -> impartial (in auditPrior())sampleK -> x (in auditPrior())sampleN -> n (in auditPrior())N -> N.units (in auditPrior())maxSize -> max (in planning())increase -> by (in planning())withReplacement-> replace (in selection())ordered -> order (in selection())ascending -> decreasing (in selection())intervalStartingPoint -> start (in selection())algorithm -> method (in selection())expectedErrors -> expected (in auditPrior() and planning())confidence -> conf.level (in auditPrior(), planning(), and evaluation())pHmin -> p.hmin (in auditPrior())minPrecision -> min.precision (in auditPrior(), planning(), and evaluation())population -> data (in selection())kSumstats -> x (in evaluation())nSumstats -> n (in evaluation())sample -> data (in evaluation())bookValues -> values (in selection() and evaluation())auditValues -> values.audit (in evaluation())counts -> times (in evaluation())popBookValues -> N.units (in evaluation())rohrbachDelta -> r.delta (in evaluation())momentPopType -> m.type (in evaluation())csA -> cs.a (in evaluation())csB -> cs.b (in evaluation())csMu -> cs.mu (in evaluation())records -> items (in selection())mus -> values (in selection())hypotheses -> hyp (in auditPrior())poisson is now the default likelihood / method for all functions since it is the most conservative.method = 'interval' is now the default selection method.method = 'default' or prior = TRUE are now set to the gamma(1, 1), beta(1,1), and beta-binomial(1, 1) priors.times (former counts) argument in evaluation() must now be indicated as a column name in the data instead of a vector.nPrior and kPrior have been removed from the planning() and evaluation() functions. All prior distributions must now be specified using prior = TRUE (noninformative priors) or using a call to auditPrior().auditBF() function since its value is available through evaluation(materiality = x, prior = auditPrior(method = 'impartial', materiality = x))
Minor changes
x and n to have the same value in evaluation().Minor changes
?jfa-package help file.Bug fixes
print.jfaEvaluation() call if there was no performance materiality specified and prior = TRUE.New features
print() functions now return a more concise description of the relevant output.summary() functions for all returned objects that take over the former (elaborate) output of the print() functions.auditBF() which computes Bayes factors from summary statistics of an audit sample.Bug fixes
evaluation() in which the likelihood stored in the prior was not properly passed to the function.Minor changes
New features
jfaPosterior as returned by evaluation()$posterior and planning()$expectedPosterior can now be used as input for the prior argument in the planning() and evaluation() functions.Bug fixes
method = 'bram' in the auditPrior() function where the prior parameters would go off to infinity when expectedError = 0.Major changes
planning() function does not require a value for the materiality anymore when planning with the hypergeometric likelihood.Minor changes
MUS package to the unit tests.New features
expectedErrors > 0 available for method = 'hypotheses' in the auditPrior() function.method = 'hypotheses' and method = 'impartial' in the auditPrior() function available for likelihood = 'hypergeometric'.bram as a method for the auditPrior() function. method = 'bram' computes a prior distribution with a given mode (expectedError) and upper bound (ub).Bug fixes
evaluation() function in which +1 was added to the beta parameter, resulting in slightly lower modes than the correct ones.N has the correct value of N = N - n (current) instead of N - n + k (before).Major changes
confidence = 0.95 in all applicable functions. confidence currently has no default value so that the user is required to give an input.likelihood = 'poisson' in the planning() function to likelihood = 'binomial' to be consistent across all functions.materiality and minPrecision are among the first ones to be shown.Minor changes
New features
Bug fixes
phyper() function instead of the qhyper() function, which resulted in lower bounds than usual.Minor changes
Bug fixes
selection() where if population is sorted or modified, bv still retained the old ordering and data. The resulting sample was overweighted towards small values and/or still contained negative values (Thanks to @alvanson).New features
report() that automatically generates an audit report.Major changes
sampling() function, which is now replaced entirely with the selection() function.evaluation() function when an estimator is used.New features
digits argument in the internal jfa:::print.jfaPrior(), jfa:::print.jfaPlanning(), jfa:::print.jfaSelection(), and jfa:::print.jfaEvaluation() functions to control rounding in printing.description, statistics, specifics and hypotheses to the output of the auditPrior() function.jfaPosterior with print() and plot() methods.expectedPosterior of class jfaPosterior to the output of the planning() function, includes description, statistics and hypotheses.posterior of class jfaPosterior to the output of the evaluation() function, includes description, statistics and hypotheses.Bug fixes
auditPrior() function for method = impartial when expectedErrors > 0.Major changes
sampling() function that it will be deprecated from 0.5.0 onward. You can use selection() instead, since sampling() causes namespace issues with other packages.Minor changes
jfaSampling to jfaSelection. This should not have any consequences.Bug fixes
planning() function that did not allow the user to plan for a monetary sample when their population size was too low.planning() function that did not allow the user to select a non-integer number of expected errors when there was a prior involved.Minor changes
planning() and evaluation() functions against benchmarks.New features
counts in the evaluation() function that quantifies how many times each observation should be evaluated due to being selected multiple times in the selection stage.New features
default, impartial, hypotheses, sample, and factor in the auditPrior() function. In addition to the already supported arm method, these methods allow the auditor to incorporate more sources of audit information into the prior distribution.minPrecision argument in the planning() function that allows auditors to calculate a sample size so that the difference between the posterior upper confidence bound and the most likely error is lower than the set minimum precision. Also implemented in the evaluation() function as a requirement to approve the population.mle from the evaluation() function, which quantifies the most likely error. Also return the value of the precision from this function.increase argument in the planning() function that allows the user to increase the sample size with a set amount each step of the iterations.Minor changes
New features
jfa package. The package provides four functions: auditPrior(), planning(), sampling(), and evaluation().