-
Notifications
You must be signed in to change notification settings - Fork 13
/
README.Rmd
179 lines (144 loc) · 6.25 KB
/
README.Rmd
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
---
output:
md_document:
variant: markdown_github
---
<!-- README.md is generated from README.Rmd. Please edit that file -->
```{r, echo = FALSE}
knitr::opts_chunk$set(
collapse = TRUE,
comment = "#>",
fig.path = "man/figures/README-"
)
```
# powerlmm
[](https://travis-ci.org/rpsychologist/powerlmm)
[](https://cran.r-project.org/package=powerlmm)
Power Analysis for Longitudinal Multilevel/Linear Mixed-Effects Models.
## Overview
The purpose of `powerlmm` is to help design longitudinal treatment studies (parallel groups), with
or without higher-level clustering (e.g. longitudinally clustered by therapists,
groups, or physician), and missing data. The main features of the package are:
* Longitudinal two- and three-level (nested) linear mixed-effects models, and partially nested designs.
* Random slopes at the subject- and cluster-level.
* Missing data.
* Unbalanced designs (both unequal cluster sizes, and treatment groups).
* Design effect, and estimated type I error when the third-level is ignored.
* Fast analytical power calculations for all designs.
* Power for small samples sizes using Satterthwaite's degrees of freedom approximation.
* Explore bias, Type I errors, model misspecification, and LRT model selection
using convenient simulation methods.
## Installation
`powerlmm` can be installed from CRAN and GitHub.
```{r, eval = FALSE}
# CRAN, version 0.4.0
install.packages("powerlmm")
# GitHub, dev version
devtools::install_github("rpsychologist/powerlmm")
```
## Example usage
This is an example of setting up a three-level longitudinal model with random slopes at both the subject-
and cluster-level, with different missing data patterns per treatment arm. Relative standardized inputs are used,
but unstandardized raw parameters values can also be used.
```{r three-level-setup, fig.width=9}
library(powerlmm)
d <- per_treatment(control = dropout_weibull(0.3, 2),
treatment = dropout_weibull(0.2, 2))
p <- study_parameters(n1 = 11,
n2 = 10,
n3 = 5,
icc_pre_subject = 0.5,
icc_pre_cluster = 0,
icc_slope = 0.05,
var_ratio = 0.02,
dropout = d,
effect_size = cohend(-0.8,
standardizer = "pretest_SD"))
p
```
```{r, eval=FALSE}
plot(p)
```
```{r three-level-power, message= FALSE, warning=FALSE}
get_power(p, df = "satterthwaite")
```
### Unequal cluster sizes
Unequal cluster sizes is also supported, the cluster sizes can either be random (sampled), or the marginal distribution can be specified.
```{r three-level-power-unequal, message= FALSE, warning=FALSE}
p <- study_parameters(n1 = 11,
n2 = unequal_clusters(2, 3, 5, 20),
icc_pre_subject = 0.5,
icc_pre_cluster = 0,
icc_slope = 0.05,
var_ratio = 0.02,
effect_size = cohend(-0.8,
standardizer = "pretest_SD"))
get_power(p)
```
Cluster sizes follow a Poisson distribution
```{r three-level-power-unequal-random, message= FALSE, warning=FALSE}
p <- study_parameters(n1 = 11,
n2 = unequal_clusters(func = rpois(5, 5)), # sample from Poisson
icc_pre_subject = 0.5,
icc_pre_cluster = 0,
icc_slope = 0.05,
var_ratio = 0.02,
effect_size = cohend(-0.8,
standardizer = "pretest_SD"))
get_power(p, R = 100, progress = FALSE) # expected power by averaging over R realizations
```
### Convenience functions
Several convenience functions are also included, e.g. for creating power curves.
```{r three-level-power-curve, eval = FALSE}
x <- get_power_table(p,
n2 = 5:10,
n3 = c(4, 8, 12),
effect_size = cohend(c(0.5, 0.8), standardizer = "pretest_SD"))
```
```{r, eval = FALSE}
plot(x)
```
## Simulation
The package includes a flexible simulation method that makes it easy to investigate the performance of different models. As an example, let's compare the power difference
between the 2-level LMM with 11 repeated measures, to doing an ANCOVA at posttest. Using `sim_formula` different models can be fit to the same data set during the simulation.
```{r}
p <- study_parameters(n1 = 11,
n2 = 40,
icc_pre_subject = 0.5,
cor_subject = -0.4,
var_ratio = 0.02,
effect_size = cohend(-0.8,
standardizer = "pretest_SD"))
# 2-lvl LMM
f0 <- sim_formula("y ~ time + time:treatment + (1 + time | subject)")
# ANCOVA, formulas with no random effects are with using lm()
f1 <- sim_formula("y ~ treatment + pretest",
data_transform = transform_to_posttest,
test = "treatment")
f <- sim_formula_compare("LMM" = f0,
"ANCOVA" = f1)
res <- simulate(p,
nsim = 2000,
formula = f,
cores = parallel::detectCores(logical = FALSE))
```
We then summarize the results using `summary`. Let's look specifically at the treatment effects.
```{r}
summary(res, para = list("LMM" = "time:treatment",
"ANCOVA" = "treatment"))
```
We can also look at a specific model, here's the results for the 2-lvl LMM.
```{r}
summary(res, model = "LMM")
```
## Launch interactive web application
The package's basic functionality is also implemented in a Shiny web application,
aimed at users that are less familiar with R. Launch the application by typing
```{r, eval = FALSE}
library(powerlmm)
shiny_powerlmm()
```
