Suppose that we plan to meta-analyze all published findings to compute a summary effect size estimate for the group difference between typically developing individuals and individuals with autism on a measure of face recognition ability. In order to plan the study accordingly, we must choose plausible values for the following:
…for our meta-analysis of face recognition deficits in autism
To do this with metapower
, we use the core function mpower()
print(my_power)
#>
#> Power Analysis for Meta-analysis
#>
#> Effect Size Metric: d
#> Expected Effect Size: 0.2
#> Expected Study Size: 20
#> Expected Number of Studies: 10
#>
#> Estimated Power: Mean Effect Size
#>
#> Fixed-Effects Model 0.2917841
#> Random-Effects Model (i2 = 50%): 0.1982394
The first part of the output shows the expected input values, where the main results are shown in the bottom portion, mainly, Estimated Power
. Under this set of values, our power to detect a mean difference under a Fixed-Effects model is 29.18%. Furthermore, we can look at the power under a Random-Effects model.
Given that power analysis require a lot of assumptions, it is generally advisable to look at power across a range of input values. To visualize the power curve for these set of input parameters, use plot_mpower()
to generate a ggobject
that is modifiable and by default, shows 5x as many studies as the user inputs.
For fixed-effects model, power curves are shown for a range of effect sizes, whereas random-effects model shows power across a range of heterogeneity values, \(\tau^2\)
For users wanting more flexibility in visualization, the mpower
object contains a data frame $power_range
containing all data populating the ggobject
,
str(my_power$power_range)
#> 'data.frame': 147 obs. of 9 variables:
#> $ k_v : int 2 3 4 5 6 7 8 9 10 11 ...
#> $ es_v : num 0.2 0.2 0.2 0.2 0.2 0.2 0.2 0.2 0.2 0.2 ...
#> $ n_v : num 20 20 20 20 20 20 20 20 20 20 ...
#> $ i2 : num 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 ...
#> $ c_alpha : num 1.96 1.96 1.96 1.96 1.96 ...
#> $ test_type : chr "two-tailed" "two-tailed" "two-tailed" "two-tailed" ...
#> $ variance : num 0.201 0.201 0.201 0.201 0.201 0.201 0.201 0.201 0.201 0.201 ...
#> $ fixed_power : num 0.0967 0.1207 0.145 0.1695 0.194 ...
#> $ random_power: num 0.136 0.136 0.142 0.149 0.158 ...
For Fixed-Effects Model, the test of homogeneity examines whether the amount of variation among effect sizes is greater than that of sampling error alone. To compute this, simply wrap the same arguments used in mpower
with homogen_power
. Since it is hard to justify selecting the standard deviation between a set of studies, homogen_power()
automatically computes power across a range of different standard deviation values (i.e., SD = 1, SD = 2.)
(my_homogen_power <- homogen_power(effect_size = .25, study_size = 20, k = 30,i2 = .50, es_type = "d"))
#>
#> Power Analysis for Test of Homogeneity in Meta-analysis
#>
#> Effect Size Metric: d
#> Expected Effect Size: 0.25
#> Expected Study Size: 20
#> Expected Number of Studies: 30
#>
#> Estimated Power: Test of Homogeneity
#>
#> Fixed-Effects Model (SD = 1) 0.06687224
#> Fixed-Effects Model (SD = 2) 0.1348743
#> Fixed-Effects Model (SD = 3) 0.2966788
#> Fixed-Effects Model (SD = 4) 0.55878
#> Fixed-Effects Model (SD = 5) 0.8167712
#>
#> Random-Effects Model (i2 = 50%): 0.9872307
For Random-Effects models, homogen_power()
computes power given the user-specified heterogeneity value (i.e., \(I^2\))
In this example, we specified i2 = .50
, but below are common benchmarks:
We can visualize power across a range with plot_homogen_power
Although researchers are primarily interested in conducting meta-analysis to quantify the main effect of a specific phenomenon, It is very common to evaluate the moderation of this overall effect based on a number of study- and/or sample-related characteristics such as task paradigm or age group (e.g., children, adolescents, adults). To compute the statistical power for the detection of categorical moderators, we use the function subgroup_power()
with a few additional arguments, mainly:
n_groups
):effect_sizes
):…for our meta-analysis of face recognition deficits in autism
We may expect that face recognition tasks have larger effect sizes then face perception tasks; therefore, we specify 2 groups and their respective expected effect sizes:
n_groups = 2
effect_sizes = c(.2,.5)
my_subgroup_power <- subgroup_power(n_groups = 2,
effect_sizes = c(.2,.5),
study_size = 20,
k = 30,
i2 = .5,
es_type = "d")
print(my_subgroup_power)
#>
#> Power Analysis for Subgroup analysis:
#>
#> Effect Size Metric: d
#> Number of Subgroups: 2
#> Groups:
#> Expected Effect Sizes: 0.2 0.5
#> Expected Study Size: 20
#> Expected Number of Studies: 30
#>
#> Esimated Power to detect subgroup differences
#>
#> Fixed-Effects Model: 0.4456862
#> Random-Effects Model (i2 = 50%): 0.2517633
plot_subgroup_power(my_subgroup_power)
Given, this set of expected values, we have 44.57% to detect between-group differences under a Fixed-Effects model. As expected, moderator effects are much harder to detect and more studies are required, especially when heterogeneity is high.