https://www.google.com/url?sa=t&source=web&rct=j&url=https://www.ilo.org/public/english/bureau/stat/isco/docs/publication08.pdf&ved=2ahUKEwjFvM-CqoDxAhXFKewKHfCsBKEQFjAAegQIAxAC&usg=AOvVaw0Zj_XjnHEiOacFSC3KcgsK

Joshua,

Yes, it is a good idea to design your experimental model with sufficient power, all the way down to the sub-scale levels. However, the vexing issue is how to juggle the three main factors involved in all of the power (beta error) calculations:

1. Effect size you are expecting your model to detect at a given level of alpha error (95% confidence threshold = p value of 0.05 or less required to Recject the Null Hypothesis that the Independent Variable(s) (predictive factor) had no statistically significant effect at the level of confidence. Effect sizes are small, medium, large.

2. Sample Size - roughly the smaller the sample size, the less likely that any effect (small, medium, or large) will be detected.

3. Accept/Reject Threshold (for Null Hypothesis) 80% Confidence, p =0.20; 90% Confidence, p = 0.10; 95% Confidence, p= 0,05; 99% Confidence p= 0.01.

What you will find as you explore the interactive dynamics among these three major determinatives, is that they all are mostly dependent upon the mathematical model that is valid to use, based on how the assumptions of the different hypothesis testing models that are available, "fits" the descriptive statistics of your analytical data's distribution. Is it normal? Is it uniform? Is it Poisson? If your data is not a normal distribution or a uniform distribution (KS test - Kolmogorov-Smirnov) then the only valid models avaiable will be the non-parametric tests (such as Mann-Whitney U instead of t-test, if the data is normally distributed).

Please understand your Question is a great one (a clasic one) that is always better asked sooner rather than later, when designing a predicive hypothesis testing model, especially so in human subjects and health care (where the avaible convenience sample, small sample, multi-modal data) is almost NEVER normally distributed.

On the other hand, there is a common heuristic ("rule of thumb") that if you have 3,500 observations (rows of data) in true "three normal format" (one unique row, is a unique non-duplicate observation, each column is a unique non-overlapping variable, and the Dependent Variables(s) versus the Independent Varaible(s) are clearly labeled with "Human Readable" variable names on the first row of the N 3,500 observation data set, then, there if there is a statistically detectable anomoly present in the data set, it will be discoverable, in at least one or more of the availble parametric and non-parametric statistical hypothesis testing methods available, given that the assumptions of the models are observed as being valid for the data set. That is about the best that I can sumarize, and I have had this exact conversation routinely with students and faculty, on a regular on-going basis, these many years.

At this point, I have advised approximately 280 DNP, PhD, Masters, and BS students on mathematical modeling of their academic project data, over the past 10 years that I have functioned in the role of a biomedical informatics data scientist.

I hope this helps you in your quest for the ideal model to use in your manuscript or publication and reporting of your findings.

Richard E. Gilder, MS RN-C

Bioinformatics Scientist

TTUHSC Recurrent Faculty

School of Nursing

[email protected]

https://www.google.com/url?sa=t&source=web&rct=j&url=https://www.ilo.org/public/english/bureau/stat/isco/docs/publication08.pdf&ved=2ahUKEwiOidKCtY3xAhWDOuwKHXE8BpEQFjACegQIGRAC&usg=AOvVaw0Zj_XjnHEiOacFSC3KcgsK