Power Calculations for ANCOVA

Aaron R. Caldwell

16 May, 2022

library("Superpower")

Introduction

Recently, Shieh (2020) demonstrated that most software uses a slightly flawed approach to estimating power for ANCOVAs, and focused on one-way ANOVA designs to compare his method to the method first mentioned by Cohen (1988).

One-way ANCOVA

As Shieh (2020) eloquently points out in their simulations, the method of Cohen (1988) overestimates power and the problem is exacerbated by having a high number of covariates with high proportion of the variance (\(R^2\)) explained by the covariates. The problem is worst (30% error between estimated and actual power) in the simulation when there are 10 covariates included in the model when explained variance is approximately 81% (\(\rho = 0.9\)) [Shieh (2020) see Table 3). While this may not seem like much of issue if you don’t expect to encounter this scenario, it still demonstrates the Cohen (1988) method is inconsistent in producing appropriate estimates of power. I believe this is reason enough (when provided with a sustainable and implementable alternative) to abandon the old method of Cohen (1988) for a newer method that provides exact, rather than approximate, estimates of power for ANCOVA.

Thankfully, Shieh (2020) was diligent in his work and demonstrated, showing both the math and simulations, how a new exact method could be utilized. The direct method described in the paper is implemented in the power_oneway_ancova function.

We can copy the example from Maxwell and Delaney (2004) that Shieh also used. In this example there are 3 groups with means (mu) of 400, 450, 500 respectively. The error variance is 10000 (sd = 100). Rather than simulating dozens of examples, I will demonstrate one scenario below where there are 3 covariates, and the \(R^2\) is equal to 0.5 (treatment effect excluded). This is demonstrated in Shieh (2020), Table 2.

For power_oneway_ancova we can demonstrate both the approximate and exact methods using the type argument. We can leave the n argument out in order to solve for the sample size required to reach 80% power. Please notice that round_up is set to TRUE since we want have a whole number for sample sizes (rather than a fractional sample size).

Cohen (1988) Method

power_oneway_ancova(
  mu = c(400,450,500),
  n_cov = 3,
  sd = 100,
  r2 = .25,
  alpha_level = .05,
  #n = c(17,17,17),
  beta_level = .2,
  round_up = TRUE,
  type = "approx"
)
#> 
#>      Power Calculation for 1-way ANCOVA 
#> 
#>             dfs = 2, 42
#>               N = 48
#>               n = 16, 16, 16
#>           n_cov = 3
#>              mu = 400, 450, 500
#>              sd = 100
#>              r2 = 0.25
#>     alpha_level = 0.05
#>      beta_level = 0.1877374
#>           power = 81.22626
#>            type = approx

Shieh (2020) Method

Notice that this method requires 3 more subjects in order to achieve a minimum of 80% power.

power_oneway_ancova(
  mu = c(400,450,500),
  n_cov = 3,
  sd = 100,
  r2 = .25,
  alpha_level = .05,
  #n = c(17,17,17),
  beta_level = .2,
  round_up = TRUE,
  type = "exact"
)
#> 
#>      Power Calculation for 1-way ANCOVA 
#> 
#>             dfs = 2, 45
#>               N = 51
#>               n = 17, 17, 17
#>           n_cov = 3
#>              mu = 400, 450, 500
#>              sd = 100
#>              r2 = 0.25
#>     alpha_level = 0.05
#>      beta_level = 0.1878274
#>           power = 81.21726
#>            type = exact

Extending to factorial ANOVAs

Now, Shieh (2020) mentioned something very interesting at the end of section 3.

“Although the prescribed application of general linear hypothesis is discussed only from the perspective of a one-way ANCOVA design, the number of groups G may also represent the total number of combined factor levels of a multi-factor ANCOVA design. Hence, using a contrast matrix associated with a specific designated hypothesis, the same concept and process of assessing treatment effects can be readily extended to two-way and higher-order ANCOVA designs.”

Therefore, all that is needed to extend the one-way ANOVA code provided by Shieh (2020) is to provide the appropriate contrast matrix for the main effect or interaction ANOVA-level effect that is desired. Superpower accomplishes this with the ANCOVA_analytic function which internally uses the model.matrix function to form the appropriate contrast matrix.

ANCOVA_analytic

This function operates similar to the ANOVA_power and ANOVA_exact functions. However, the ANCOVA_analytic function doesn’t require the use of ANOVA_design first and relies upon the closed formulas from Shieh (2020) rather than a simulation to calculate statistical power. Please note that unlike the power_oneway_ancova function there is no option to apply the approximation from Cohen (1988) for factorial designs.

Example: 2x3 Design

We can extend the previous scenario with 3 groups to a factorial design with 2 groups across 3 conditions.

# Run function
res1 = ANCOVA_analytic(
  design = "2b*3b",
  mu = c(400, 450, 500,
         400, 500, 600),
  n_cov = 3,
  sd = 100,
  r2 = .25,
  alpha_level = .05,
  #n = 17,
  beta_level = .2,
  round_up = TRUE
)

# Print main results
res1
#> Power Analysis Results for ANCOVA
#>     Total N Covariates   r2 Alpha Level Beta Level Power
#> a       102          3 0.25        0.05     0.1897 81.03
#> b        30          3 0.25        0.05     0.1389 86.11
#> a:b     180          3 0.25        0.05     0.1995 80.05

The results can also be printed as power.htest objects by accessing the individual effects in the res1 object.

res1$aov_list$a
#> 
#>      Power Calculation for ANCOVA 
#> 
#>             dfs = 1, 93
#>               N = 102
#>               n = 17, 17, 17, 17, 17, 17
#>           n_cov = 3
#>              mu = 400, 450, 500, 400, 500, 600
#>              sd = 100
#>              r2 = 0.25
#>     alpha_level = 0.05
#>      beta_level = 0.1897221
#>           power = 81.02779
#>            type = Exact
res1$aov_list$b
#> 
#>      Power Calculation for ANCOVA 
#> 
#>             dfs = 2, 21
#>               N = 30
#>               n = 5, 5, 5, 5, 5, 5
#>           n_cov = 3
#>              mu = 400, 450, 500, 400, 500, 600
#>              sd = 100
#>              r2 = 0.25
#>     alpha_level = 0.05
#>      beta_level = 0.1389288
#>           power = 86.10712
#>            type = Exact
res1$aov_list$ab
#> 
#>      Power Calculation for ANCOVA 
#> 
#>             dfs = 2, 171
#>               N = 180
#>               n = 30, 30, 30, 30, 30, 30
#>           n_cov = 3
#>              mu = 400, 450, 500, 400, 500, 600
#>              sd = 100
#>              r2 = 0.25
#>     alpha_level = 0.05
#>      beta_level = 0.1994719
#>           power = 80.05281
#>            type = Exact

We can also check the design by using the plot method.

plot(res1)

However, you may want to compare the power of ANOVA to an ANCOVA. In that case you can use the ANOVA_design function and pass it onto the ANCOVA_analytic function. But, this forces you to set the sample size.

des1 = ANOVA_design(  design = "2b*3b",
  mu = c(400, 450, 500,
         400, 500, 600),
  n = 17,
  sd = 100)

res2 = ANCOVA_analytic(
  design_result = des1,
  n_cov = 3,
  r2 = .25,
  alpha_level = .05,
  round_up = TRUE
)

res2
#> Power Analysis Results for ANCOVA
#>     Total N Covariates   r2 Alpha Level Beta Level Power
#> a       102          3 0.25        0.05  0.1897221 81.03
#> b       102          3 0.25        0.05  0.0001954 99.98
#> a:b     102          3 0.25        0.05  0.4701157 52.99

Contrasts

User specified contrasts can also be used for a power analysis. These can be provided in the cmats argument of the ANCOVA_analytic function or supplied directly to the ANCOVA_contrast function as an independent test. The ANCOVA_analytic function requires that the contrasts matrices are provided as matrices (i.e., as.matrix). The contrasts can be accessed in the con_list part of the results and can be named (in this case “test”).

ANCOVA_analytic(design = "2b",
                mu = c(0,1),
                n = 15,
                cmats = list(test = matrix(c(-1,1),
                                              nrow = 1)),
                sd = 1,
                r2 = .2,
                n_cov = 1)$con_list$test
#> 
#>      Power Calculation for ANCOVA contrast 
#> 
#>             dfs = 1, 27
#>               N = 30
#>               n = 15, 15
#>           n_cov = 1
#>        contrast = -1, 1
#>              mu = 0, 1
#>              sd = 1
#>              r2 = 0.2
#>     alpha_level = 0.05
#>      beta_level = 0.1744293
#>           power = 82.55707
#>            type = Exact
# Same result
ANCOVA_contrast(cmat = c(-1,1),
                n = 15,
                mu = c(0,1),
                sd = 1,
                r2 = .2,
                n_cov = 1)
#> 
#>      Power Calculation for ANCOVA contrast 
#> 
#>             dfs = 1, 27
#>               N = 30
#>               n = 15, 15
#>           n_cov = 1
#>        contrast = -1, 1
#>              mu = 0, 1
#>              sd = 1
#>              r2 = 0.2
#>     alpha_level = 0.05
#>      beta_level = 0.1744293
#>           power = 82.55707
#>            type = Exact

References

Cohen, J. 1988. Statistical Power Analysis for the Behavioral Sciences. 2nd ed. https://doi.org/10.4324/9780203771587.
Maxwell, S.E., and H.D. Delaney. 2004. Designing Experiments and Analyzing Data: A Model Comparison Perspective. 2nd ed. Mahwah, NJ: Lawrence Erlbaum Associates.
Shieh, G. 2020. “Power analysis and sample size planning in ANCOVA designs.” Psychometrika 85 (1): 101–20. https://doi.org/10.1007/s11336-019-09692-3.