Abstract
Purpose: :
Sample size estimation is an essential aspect of clinical trial design. The techniques used to generate these estimates often involve complex statistical calculations. Using the classic dry eye assessments of corneal staining and ocular discomfort, we have developed tabulations of sample size and power for endpoints with mean difference (MD) ranges from 0.1 to 1 unit and standard deviation (SD) ranges from 0.5 to 1.5 units. This tool can be used by anyone regardless of statistical background during protocol development or clinical trial planning.
Methods: :
Our method assumes that 2 treatment groups are being compared with a 2-sided level of significance of α = 0.05. We then compare three methods of calculating appropriate sample size: 1) a 2-sample t-test using Proc Power (SAS 9.2®); 2) a 2-sample t-test with Bonforroni correction accounting for 2 or 3 pairwise comparisons with α = 0.025 and α = 0.0167, respectively, using Proc Power (SAS 9.2®): and 3) a Dunnett’s adjustment when there are 2 or 3 active treatments compared to the same control group. The third method employed a modified Randy Tobias macro with maximum sample size of 1000.
Results: :
Using our approach we are able to generate a full set of tables of sample size and power calculation where power ranges from 60% to 99%, MD from 0.1 - 1 unit and SD from 0.5 - 1.5 units. The sample size and power can also be calculated using the MD to SD (MD/SD) ratio. Regardless of the value of MD and SD, for the same MD/SD ratios with the same sample size, the power will be the same, and vice versa. If the target sample size or power does not show in the table, the sample size or power can be approximated using linear interpolation. For example, based on a 2-sample t-test, if the target sample size for the study is 20 per group and the MD/SD ratio is 1 (e.g., MD = 0.5, SD = 0.5 or MD = 2, SD = 2), the estimated power will be 85%.
Conclusions: :
The approach we describe provides a general method to identify power and sample size for mean differences of 0.1 - 1 unit and standard deviations of 0.5 - 1.5 units. This method should be easily applied to other models to determine appropriate sample size.
Keywords: clinical (human) or epidemiologic studies: biostatistics/epidemiology methodology