Power analysis for longitudinal multilevel models: powerlmm 0.2.0 is now out on CRAN
- Analytical power calculations now support using Satterthwaite’s degrees of freedom approximation.
Simulate.plcpwill now automatically create lme4 formulas if none is supplied, see
- You can now choose what alpha level to use.
- Treat cluster sizes as a random variable,
uneqal_clustersnow accepts a function indicating the distribution of cluster sizes, via the new argument
rnormcould be used to draw cluster sizes.
- Expected power for designs with parameters that are random variables,
can be calculated by averaging over multiple realizations, using the
- Support for parallel computations on Microsoft Windows, and in GUIs/interactive
parallel::makeCluster(PSOCK). Forking is still used for non-interactive Unix environments.
- Calculations of the variance of the treatment effect is now much faster for designs with unequal clusters and/or missing data, when cluster sizes are large. The calculations now use the much faster implementation used by lme4.
- Cleaner print-methods for
- Multiple power calculations can no be performed in parallel, via the
simulate.plcp_multinow have more options for saving intermediate results.
print.plcp_multi_powernow has better support for subsetting via either , head(), or subset().
icc_pre_subjectis now defined as
(u_0^2 + v_0^2) / (u_0^2 + v_0^2 + error^2), instead of
(u_0^2) / (u_0^2 + v_0^2 + error^2). This would be the subject-level ICC, if there’s no random slopes, i.e. correlation between time points for the same subject.
study_parameters(): 0 and NA now means different things. If 0 is passed, the parameters is kept in the model, if you want to remove it specify it as NA instead.
study_parameters(): is now less flexible, but more robust. Previously a large combination if raw and relative parameters could be combined, and the individual parameters was solved for. To make the function less bug prone and easier to maintain, it is now only possible to specify the cluster-level variance components as relative values, if the other parameters as passed as raw inputs.
- Output from
simulate_data()now includes a column
y_cthat contains the full outcome vector, without missing values added. This makes it easy to compare the complete and incomplete data set, e.g. via
batch_progressenables showing progress when doing multiple simulations.
- Fix bug in
summary.plcp_simwhere the wrong % convergence was calculated.
- Simulation function now accepts lme4 formulas containing ”||“.
- The cluster-level intercept variance is now also set to zero in the control group, when a partially nested design is requested.
- Fix incorrect error message from
icc_cluster_pre = NULLand all inputs are standardized.
- Fix bug that would cause all slopes to be zero when
var_ratioargument was passed a vector of values including a 0, e.g.
var_ratio = c(0, 0.1, 0.2).
- Fix bug for multi-sim objects that caused the wrong class the be used for,
res[]$paras, and thus the single simulation would not print correctly.
- Results from multi-sim objects can now be summarized for all random effects in the model.
- More support for summarizing random effects from partially nested formulas,
cluster_slopeis now correctly extracted from
(0 + treatment + treatment:time || cluster).
- When Satterthwaite’s method fails the between clusters/subjects DFs are used to calculate p-values.
Power.plcp_multiis now exported.
get_power.plcp_multinow shows a progress bar.
- Fix a bug that caused dropout to be wrong when one condition had 0 dropout, and
deterministic_dropout = FALSE.
Published March 21, 2018 (View on GitHub)