power4mome

• Updated x_from_power() and related functions to allow users specifying how tests will be collapsed ("none" is not allowed), by setting the argument rejection_rates_args. (0.1.1.39, 0.1.1.41)

"Quick" Functions

To simplify the workflows for some

common mediation models, a set of

"quick" functions have been added.

Users can do common tasks using just one

function.

• Added q_power_mediation() and friends for common mediation models. Users can do most steps in one call. (0.1.1.18, 0.1.1.19)
• Updated the quick functions (q_power_mediation() and friends) to have one more mode, "n". Probabilistic bisection is the default algorithm for this mode. Other methods have been updated for this mode. (0.1.1.56)

Rejection Rates

• Changed the default method for rejection rate confidence intervals to Wilson's (1927) method. For backward compatibility, use options(power4mome.ci_method = "norm") to set the default method to normal approximation. (0.1.1.2)
• Updated summarize_tests() and rejection_rates() to use the extrapolation method by Boos and Zhang (2000) if the number of resamples for bootstrapping or Monte Carlo is of the supported values. (0.1.1.5, 0.1.1.6, 0.1.1.7)
• Added merge_all_tests to rejection_rates() to support merging all tests into one. The argument collapse can then be used for collapsing several different tests, not just for one test with several results. (0.1.1.28)
• Added the p_adjust_method argument to some tests, as well as the rejection_rates method and summarize_tests(). Users can adjust *p*-values using p.adjust() when there are more than one test in a test function set to test_fun, or when merging several tests in summarize_tests(). This feature is used to estimate power when multiple-comparison adjustment is used, such as false discover rate (FDR) or Bonferroni correction. (0.1.1.31)

Model Specifications

• Added two levels of effects, sm for small-to-moderate, and ml for moderate-to-large. (0.1.1.11)
• Added loading_difference and reference for sim_data(). The population standardized factor loadings of indicators of a factor can have different loadings, and the first indicator (the default reference indicator) can be the indicator with the strongest, weakest, or medium loadings. (0.1.1.70)

Tests

• Added the test_method argument for tests of indirect effects and their variants to use asymmetric *p*-values to do the tests. This change is for using the efficient method by Boos and Zhang (2000) to estimate the rejection rates for resampling methods. (0.1.1.3)
• Updated test_k_indirect_effects() to support computing and testing the total indirect effect. (0.1.1.65)

Data Processors

When calling the function power4test(),

the generated data can now be further

processed by *data processors*: functions

to manipulate the generated data, such

as introducing missing values or computing

scale scores. Some built-in processors

are included, but user-supplied functions

can also be used.

• Added scale_scores() for replacing the indicator scores by the corresponding scale scores before fitting a model. This allows estimating power when the data generating model is a latent factor model with indicators but the model will be fitted to the scale scores instead of the indicators. (0.1.1.43)
• Added missing_values() for generating missing values. This allows sample size determination that takes into account probable missing data. This also allows comparing the levels of power between two different methods to handle missing data, such as full information maximum likelihood versus listwise deletion. (0.1.1.66)
• Added ordinal_variables() for converting continuous indicator variables to ordinal variables. This allows estimating power when the indicators are ordinal variables, such as binary items or 5-point Likert scales. This also allows comparing the levels of power between different numbers of response options. (0.1.1.67, 0.1.1.69)
• The argument sim_data_name of process_data is now default to "data". It is not a required argument. (0.1.1.43)

Multigroup Models

Support for multigroup models have been

substantially improved. Users can

generate data from a

multigroup models. Some tests can also

be used for estimating power in detecting

group differences, such as the group

difference in an indirect effect.

• Added n_ratio to power4test() and related functions to supporting controlling the sample sizes of multigroup models using one single value for n. This allows functions such as n_from_power() and n_region_from_power() to support multigroup models. (0.1.1.44)
• test_cond_indirect() and test_cond_indirect_effects() now support multigroup models, although test_cond_indirect_effects() only support either a path with moderators (wlevels) or a path between groups, but not both. (0.1.1.45)
• Updated test_cond_indirect_effects() to support computing and testing group differences in indirect effects for multigroup models. (0.1.1.46)
• Updated test_parameters() to support doing likelihood ratio tests to test constraining pairs of parameters in multigroup models, by using the argument compare_groups. (0.1.1.47)
• Added the argument exclude_var to test_parameters() to exclude variances and error variances, useful for multigroup models. (0.1.1.47)
• Updated power4test() to support updating a parameter in one group when the population model is a multigroup model. Use "y ~ x.g2" to denote "y ~ x" in Group 2. If the suffix is omitted, the parameter is assumed to be in Group 1. The function power4test_by_es() now also support multigroup models due to this change. (0.1.1.48)
• Added test_group_equal() for testing equality constraints between groups. (0.1.1.49)
• If the model is a multigroup model, descriptive statistics for the generated data are printed by group, using psych::describeBy(). (0.1.1.76)

Probabilistic Bisection

The probabilistic bisection algorithm

proposed by Waeber et al. (2013) has

been added. For some scenarios, it allows

for a more efficient search for a

sample size with sufficient power, especially

when a high precision is desired (by

using a large value of nrep).

• Added the algorithm "probabilistic_bisection" for x_from_power() and friends. (0.1.1.51)

Miscellaneous

Quality-of-Life Changes

• Change the default of test_long to TRUE for the print method of power4test and related objects. (0.1.1.41)
• For some functions, nrep will be included in the output if its values vary across rows. (0.1.1.57)
• The default values of n of q-functions now depends on the mode and algorithm. (0.1.1.57)
• Added some helpers, R_for_bz() and Rs_bz_supported(), for users to use the Boos-Zhang-2000 method without remembering the number of resamples supported. (0.1.1.61)
• Updated the quick functions (q_power_mediation() and friends). When mode is "n" or "region", it is optional to set n. If not set, it will be determined internally. (0.1.1.63)
• Added the rejection_rates() method to x_from_power, n_region_from_power, and q_power_mediation objects. (0.1.1.71)
• Moved delta_tol and last_k to variants, such that users can change them through control when calling x_from_power() and friends. (0.1.1.72)
• Changed the default algorithms to be used in x_from_power(). (0.1.1.74)
• Response proportions will be printed with data_long = TRUE if a variable has 10 or fewer unique values (e.g., ordinal variables), using psych::responseFrequency(). (0.1.1.77)

Bisection

• Improved the function for extending the initial interval before doing a bisection search. (0.1.1.1)
• Optimized the search by bisection, to make use of value already tried and store all values tried. (0.1.1.14)
• The bisection algorithm has been improved in handling unusual intervals. (0.1.1.21)
• Skipped the check for combining objects in the bisection algorithm because they must be identical in the model. (0.1.1.23)
• Updated the bisection algorithm to use power_curve() to assist finding the solution. If estimated solution inside an interval, use it instead of the mean. (0.1.1.37)
• More checks for the solution in the bisection algorithm. When extending an interval, the power curve will also be used. (0.1.1.38)

Power Curve

• Modified the power curve algorithm to support goal = "close_enough" and all three types of what ("point", "lb", and "ub"). (0.1.1.34)

Probabilistic Bisection

• Adjust the initial_nrep in probabilistic bisection such that the rejection rate will not be exactly equal to the target power. (0.1.1.52)
• For probabilistic bisection, the initial interval will no longer be adjusted. This algorithm should be used with a wide enough initial interval because the interval will not be adjusted during the search (for now). (0.1.1.56)
• Increased delta_tol fo PBA (2 for n and .002 for es). (0.1.1.57)
• Added pba_diagnosis() to generate plots related to the search history of probabilistic bisection algorithm. For advanced users and developers. (0.1.1.58, 0.1.1.62)
• Updated the help page of x_from_power() to describe the probabilistic bisection method. (0.1.1.60)

Tests

• Updated test functions that used manymome to store the number of bootstrap or Monte Carlo samples and the number of estimates less than zero. Used by the Boos-Zhang-2000 method. (0.1.1.4)
• Updated all test functions to include *p*-values in the output. (0.1.1.30)
• Test functions using manymome will now automatically use "pvalues" as the test method if R is a value supported by the Boos-Zhang-method. (0.1.1.53)

Others

• Improved x_from_power() and friends (e.g., n_from_power() and n_region_from_power()) to make use of previous trials. (0.1.1.13)
• x_from_power() now detects whether a test has more than one result (e.g., testing two parameters but omnibus is "none"). If yes, it will throw an error. (0.1.1.10)
• Updated extend_interval() to handle intervals with nearly equal function values. (0.1.1.12)
• Fixed duplicated values of x when extending the interval. (0.1.1.15)
• Functions that print a call will replace object with <hidden> if it is not a symbol. (0.1.1.16)
• Functions that print a call will replace the function with the original function name if it is not a symbol. (0.1.1.17)
• The arguments final_nrep and final_R of x_from_power() and its wrappers will use stored values if available. (0.1.1.20)
• Revised c.power4test_by_n() to allow for minor differences in error variances when they are determined by Monte Carlo simulation. (0.1.1.24)
• Properly support a model with only one latent variable. (0.1.1.25)
• Vertically displace the labels of sample sizes in plot.n_region_from_power() to prevent overlapping. (0.1.1.26)
• Fixed the printing of effects in a multigroup model with within-group moderation. (0.1.1.27)
• The function power4test() now properly reuse arguments such as parallel and ncores when adding a new test to a power4test object. (0.1.1.29)
• Added one model to power_curve(). (0.1.1.34)
• Added nls_options to power_curve() to configure the use of nls(), such as when it should not be attempted. (0.1.1.36)
• Updated sim_data(). Lines for the indicators will not be added to the model syntax if scale scores are used. (0.1.1.43)
• The attribute number_of_indicators will be added to the generated data before passing to process_data. Used by data processors such as scale_scores(). (0.1.1.43)
• The arguments in control are now passed directly to the entry point of the algorithms. Potential conflicts due to partial matching should be prevented inside these algorithms (0.1.1.50)
• Updated power4test() to make one cluster that will be used in all stages. (0.1.1.55)
• Load balancing is no longer used by default, to ensure the results are reproducible. To enable load balancing, set the option "power4mome.use_lb" to TRUE using options(). (0.1.1.56)
• Updated the internal functions summarize_one_test_vector () and summarize_one_test_data_frame() to handle failed replications properly. (0.1.1.59)
• The internal helper do_FUN() for parallel processing now export functions defined in the global environment to clusters, because they may be used in arguments such as process_data. (0.1.1.64)
• The internal helper do_FUN() will reproduce the search path by loading the packages in the workers. (0.1.1.69)
• Added and updated technical appendices for the three algorithms used in x_from_power() and friends. (0.1.1.73)
• Fixed the help for the argument mode of q_power_mediation() and friends. (0.1.1.75)
• Simplify some examples to meet the five-second requirement. (0.1.1.78)
• Replaced a test using lavaan by a test using lm. Also improved a test for stability. (0.2.1)

Bug Fixes

• Fixed a bug in extending intervals in the bisection algorithm, and also improved the way intervals are extended. (0.1.1.35)
• Fixed fix_many_lm_model() to handle model syntax with a regression model spanning more than one line and fitted by lm(). (0.1.1.54)