Title | Power Analysis For One-Way Anova, Independent Samples, SAS |
---|---|
Course | Multivariate Statistical Analysis |
Institution | East Carolina University |
Pages | 5 |
File Size | 180 KB |
File Type | |
Total Downloads | 49 |
Total Views | 132 |
Power Analysis For One-Way Anova, Independent Samples, SAS...
Power Analysis For One-Way ANOVA, Independent Samples, SAS Background Although Proc Power has a “onewayanova” option, I prefer to approach this analysis as a multiple regression. If you use the “onewayanova” option, you must specify the effect size in terms of group means and standard deviations. Psychologists are accustomed to reporting the effect size in terms of Cohen’s f. f2 . Cohen considered an f of .10 1+ f 2 to be a small effect, .25 a medium effect, and .40 a large effect. These benchmarks correspond to η2 values of .0099, .0588, and .1379. The one-way ANOVA is no different from a multiple regression analysis where the predictors are k-1 dummy variables, where k is the number of groups in the analysis. The R2 from such a multiple regression analysis is identical to η2 in ANOVA. Solving for Sample Size Suppose that we wish to determine how many subjects we need to have 80% power for an ANOVA with three groups. Assume that we are using the usual .05 criterion of statistical significance. Here is the SAS code and output: Cohen’s f can be converted into η2 : η 2 =
proc power; multreg model = fixed nfullpred = 2 nredpred= 0 RsqFull = 0.0099, 0.0588, .1379 RsqRed = 0 ntotal = . power = 0.8; run;
Notice that I specified “fixed.” Power depends, in part, on whether the predictor variables are fixed or random. I nearly all cases psychologists assume that the predictors in ANOVA are fixed. Since we have three groups, we have 3-1 = 2 predictors (the treatment df from the ANOVA), . We are testing our model versus a model (the null hypothesis) that has no parameters other than the intercept (the overall mean), so nredpred= 0. As values for η2 (RsqFull), I have entered those corresponding to Cohen’s benchmarks for small, medium, and large. The value of RsqRed is that of the null hypothesis, 0. Since I want to solve for sample size, I specify “.” for ntotal. The POWER Procedure Type III F Test in Multiple Regression Fixed Scenario Elements Method Model Number of Predictors in Full Model Number of Predictors in Reduced Model R-square of Reduced Model
Exact Fixed X 2 0 0
Nominal Power Alpha
0.8 0.05
Computed N Total
Index
R-square Full
Actual Power
N Total
1 2 3
0.0099 0.0588 0.1379
0.800 0.802 0.805
967 158 64
Let’s compare our solution with that provided by G*Power. F tests - ANOVA: Fixed effects, omnibus, one-way Analysis: A priori: Compute required sample size Input:
Output:
Effect size f
= 0.1
α err prob
= 0.05
Power (1-β err prob)
= 0.8
Number of groups
= 3
Noncentrality parameter λ
= 9.6900000
Critical F
= 3.0050418
Numerator df
= 2
Denominator df
= 966
Total sample size
= 969
Actual power
= 0.8011010
SAS told us we need 967 subjects. That is 967/3 = 322.3 per group. Rounding up to 323 per group and multiplying by 3 give us the 969 reported by G*Power. Input:
Output:
Effect size f
= 0.25
α err prob
= 0.05
Power (1-β err prob)
= 0.8
Number of groups
= 3
Noncentrality parameter λ
= 9.9375000
Critical F
= 3.0540042
Numerator df
= 2
Denominator df
= 156
Total sample size
= 159
Actual power
= 0.8048873
SAS told us 158 subjects, = 52.67 per group; 3(53) = 159, as reported by G*Power. Input:
Effect size f
= 0.4
α err prob
= 0.05
Power (1-β err prob)
= 0.8
Number of groups
= 3
Output:
Noncentrality parameter λ
= 10.5600000
Critical F
= 3.1428085
Numerator df
= 2
Denominator df
= 63
Total sample size
= 66
Actual power
= 0.8180744
SAS told us 64 subjects, = 21.3 per group; 3(22) = 66, as reported by G*Power. Suppose we intend to have five groups. Changing one line of SAS code, we obtain:
nfullpredictors = 4,
Computed N Total
Input:
Output:
Input:
Output:
Input:
Output:
Index
R-square Full
Actual Power
N Total
1 2 3
0.0099 0.0588 0.1379
0.800 0.800 0.803
1199 196 80
Effect size f
= 0.1
α err prob
= 0.05
Power (1-β err prob)
= 0.8
Number of groups
= 5
Noncentrality parameter λ
= 12.0000000
Critical F
= 2.3793764
Numerator df
= 4
Denominator df
= 1195
Total sample size
= 1200
Actual power
= 0.8006464
Effect size f
= 0.25
α err prob
= 0.05
Power (1-β err prob)
= 0.8
Number of groups
= 5
Noncentrality parameter λ
= 12.5000000
Critical F
= 2.4179625
Numerator df
= 4
Denominator df
= 195
Total sample size
= 200
Actual power
= 0.8097710
Effect size f
= 0.4
α err prob
= 0.05
Power (1-β err prob)
= 0.8
Number of groups
= 5
Noncentrality parameter λ
= 12.8000000
Critical F
= 2.4936960
Numerator df
= 4
Denominator df
= 75
Total sample size
= 80
Actual power
= 0.8030845
As you can see, the SAS analysis is equivalent to that provided by G*Power. Solving for Power Suppose that we have four groups with 100 subjects in each group. What is power? proc power; multreg model = fixed nfullpredictors = 3 nredpred= 0 RsqFull = 0.0099, 0.0588, .1379 RsqRed = 0 ntotal = 400 power = .; run; Computed Power
Index
R-square Full
Power
1 2 3
0.0099 0.0588 0.1379
0.355 0.993 >.999
F tests - ANOVA: Fixed effects, omnibus, one-way Analysis: Post hoc: Compute achieved power Input:
Output:
Input:
Output:
Input:
Effect size f
= 0.1
α err prob
= 0.05
Total sample size
= 400
Number of groups
= 4
Noncentrality parameter λ
= 4.0000000
Critical F
= 2.6274408
Numerator df
= 3
Denominator df
= 396
Power (1-β err prob)
= 0.3552592
Effect size f
= 0.25
α err prob
= 0.05
Total sample size
= 400
Number of groups
= 4
Noncentrality parameter λ
= 25.0000000
Critical F
= 2.6274408
Numerator df
= 3
Denominator df
= 396
Power (1-β err prob)
= 0.9928175
Effect size f
= 0.4
α err prob
= 0.05
Output:
Total sample size
= 400
Number of groups
= 4
Noncentrality parameter λ
= 64.0000000
Critical F
= 2.6274408
Numerator df
= 3
Denominator df
= 396
Power (1-β err prob)
= 1.0000000
Return to Wuensch’s SAS Lessons Page Karl L. Wuensch, Dept. of Psychology, East Carolina Univ., Greenville, NC 27858 USA February, 2010...