[英]simr: how to specify expected effect size in lm() or aov() models?
I am trying to use simR
to assess the power of simple GLMs to detect a particular effect size, given a set of pilot data.在给定一组试验数据的情况下,我正在尝试使用simR
来评估简单 GLM 检测特定效果大小的能力。 For example:例如:
library(simr)
m1 = lm(y ~ x, data=simdata)
powerSim(m1)
I have no problem doing this when testing power to detect the "observed" effect size (ie whatever effect size is present in the pilot data), however I would like to specify an "expected" effect size.在测试检测“观察到的”效果大小(即试验数据中存在的任何效果大小)的能力时,我这样做没有问题,但是我想指定一个“预期的”效果大小。 This is easy to do when dealing with LMER
models, using the fixef
function, for example:这在处理LMER
模型时很容易做到,使用fixef
function,例如:
m2 = lmer(y ~ x + (1|g), data=simdata)
fixef(m2)['x'] = <expected effect size>
Unfortunately this function does not work with aov()
or lm()
models.不幸的是,这个 function 不适用于aov()
或lm()
模型。 For example, using...例如,使用...
fixef(m1)['x'] = <expected effect size>
Results in the following error:导致以下错误:
Error in UseMethod("fixef") :
no applicable method for 'fixef' applied to an object of class "c('aov', 'lm')"
Is there another method/package/workaround I can use to change effect sizes for aov()
or lm()
?我可以使用另一种方法/包/解决方法来更改aov()
或lm()
的效果大小吗? I imagine this might entail "hacking" the summary output in a way that alters the F value (for aov()
) or coefficient value (for lm()
), however I haven't had any luck getting this to work.我想这可能需要“破解”摘要 output 以改变 F 值(对于aov()
)或系数值(对于lm()
),但是我没有任何运气让它工作。
Any advice would be greatly appreciated!任何建议将不胜感激!
To clarify, by 'effect size' I mean the fixed effect coefficient generated by the model.澄清一下,“效应大小”是指由 model 生成的固定效应系数。 So in the following output:所以在下面的output中:
# Call:
# lm(formula = y ~ x, data = simdata)
# Coefficients:
# (Intercept) x
# 10.6734 -0.2398
The 'effect size' of x
is -0.2398. x
的“效果大小”为 -0.2398。 In the context of power analysis, changing the effect size should directly affect statistical power (because large effects require less power to detect, and vice-versa).在功效分析的背景下,改变效应大小应直接影响统计功效(因为大效应需要较少的检测功效,反之亦然)。 For example, when using LMER, changing the effect size with fixef()
directly affects statistical power:例如,在使用 LMER 时,使用fixef()
更改效果大小会直接影响统计功效:
m2 = lmer(y ~ x + (1|g), data=simdata)
summary(powerSim(m2, progress=F, nsim=100)
# successes trials mean lower upper
# 1 96 100 0.96 0.9007428 0.9889955
Specify smaller effect size and re-assess power:指定较小的效应大小并重新评估功效:
fixef(m2)['x'] = 0.05
summary(powerSim(m2, progress=F, nsim=100)
# successes trials mean lower upper
# 1 12 100 0.12 0.0635689 0.2002357
I have tried to modify the coefficient values for lm()
with the following approach:我尝试使用以下方法修改lm()
的系数值:
m1 = lm(y ~ x, data=simdata)
m1$coefficients['x'] = <expected effect size>
However this has no effect on power, eg when changing the coefficient from 0.9 to 0.09但是,这对功率没有影响,例如,将系数从 0.9 更改为 0.09 时
m1$coefficients['x'] = 0.9
summary(powerSim(m1, progress=F, nsim=100))
# successes trials mean lower upper
# 1 22 100 0.22 0.1433036 0.3139197
m1$coefficients['x'] = 0.09
summary(powerSim(m1, progress=F, nsim=100))
# successes trials mean lower upper
# 1 24 100 0.24 0.1602246 0.3357355
So I suppose a more accurate wording of my question would be: how do I change effect sizes for aov()
/ lm()
models in a way that reflects changes in statistical power?所以我想我的问题的更准确的措辞是:如何以反映统计能力变化的方式更改aov()
/ lm()
模型的效果大小?
You need to use:你需要使用:
coef(m1)['x'] = <expected effect size>
Instead of代替
fixef(m1)['x'] = <expected effect size>
The simplest solution to this is to avoid powerSim
altogether and instead use pwr.f2.test
from the package pwr.最简单的解决方案是完全避免使用powerSim
,而是使用pwr.f2.test
pwr 中的 pwr.f2.test。 This provides a precise measure of power (as opposed to simulated power), given particular model parameters and the expected effect size.考虑到特定的 model 参数和预期效果大小,这提供了功率的精确测量(与模拟功率相反)。
m1 = lm(y ~ x, data=simdata)
anova(m1)
# Analysis of Variance Table
#
# Response: y
# Df Sum Sq Mean Sq F value Pr(>F)
# x 1 14.231 14.2308 1.5943 0.2171
# Residuals 28 249.925 8.9259
Use the df values from anova(m1)
for the u
and v
arguments to pwr.f.test
将 anova anova(m1)
中的 df 值用于u
和v
arguments 到pwr.f.test
pwr.f2.test(u=1, v=28, f2=<expected effect size>)
Thanks to @StupidWolf for figuring this out!感谢@StupidWolf 解决了这个问题!
声明:本站的技术帖子网页,遵循CC BY-SA 4.0协议,如果您需要转载,请注明本站网址或者原文地址。任何问题请咨询:yoyou2525@163.com.