Math Geek.... Power Analysis and Sample size
 

Anyone know about sample size and power analysis. I am writing a research paper and need to get a power analysis of 0.80. I would like my sample size to be large enough to yield this data. Would a sample size of n = 50 (group 1 = 25, group 2 = 25) provide this? This study will be reviewed by the IRB committee here in LC for approval. Realistically a sample size of 50 is unlikely in the sample i am researching (SQ insulin vs. IV insulin in the management of uncomplicated diabetic ketoacidosis) because this would require a year or so of data collection. Or if anyone can find a website that does the calculation I would be grateful.

Group 1 = IV insulin
Group 2 = SQ insulin

T test would be the given formula.

Excuse my potential ignorance, but is power analysis the same as confidence level

I'm no expert on this aspect but I believe you are going to need a rough estimate of the pop mean and std. deviation to test your sample size. You could play around with those inputs to back into your n.

You should be able to find a program online that you can trial to help you out. Don't know if Minitab can be used for this specific analysis but i have downloaded a trial in the past.

Power analysis allows someone to determine the sample size required to detect an effect of a given size with a given degree of confidence.

sooo to answer your question... IDK lmao. this is something i have to learn about to conduct my study. right now i am just trying to set up the study population and how i am going to measure it. after the data is collected I would probably pay someone to do the statistics

For this papers sake, all i need now is a population size (hopefully under 50) that would yield a .80 power analysis.

As suggested above, one needs to know the standard deviation and mean before one can estimate the uncertainty, confidence, or statistical power for a given sample size. One also usually assumes a normal (Gaussian distribution), but this then leads to the requirement of testing normality after the experiment and if the data is not normally distributed, then one needs an alternate approach to estimating uncertainties and confidence levels.

I am sort of assuming by statistical power, you mean the number given the symbol pi in this discussion:

Statistical power - Wikipedia, the free encyclopedia

Statistical power - Wikipedia, the free encyclopedia
Statistical power - Wikipedia, the free encyclopedia

See the interpretation section.Every experimental design has this same sort of chicken and egg problem. You don't know the mean or standard deviation before the experiment is performed; therefore, you don't know a priori what the sample size needs to be.

The way to address this paradox is to estimate the mean and standard deviation that are likely from past studies that are as close as possible to the current study. Past experience tells me I need to weigh and measure 100 fish to reduce the uncertainty in the relative condition factor below 1%. This works most of the time, because the standard deviation tends to be about the same regardless of if I'm measuring rainbow trout in Colorado or black drum in Louisiana. However, some species can have a bimodal distribution with males and females having significantly different means or infected and healthy fish having significantly different means. This would require a different experimental design.

Experience has also taught me that I need about 10 shots to measure a bullet's ballistic coefficient to 1%, but that plastic tipped bullets tend to have smaller standard deviations than hollow point match bullets, so the same level of uncertainty requires more samples of hollow point match bullets. Most physical systems have smaller variation than biological systems, so it is common for biological systems to require a lot more samples for a given level of accuracy.

Statistical power comes in when comparing two samples, because it is related to the confidence that the two samples (treatment and non-treatment, or male and female, etc.) are significantly different. (Here, I mean significant in the statistical sense.) The closer the two means (of the different sample groups), the smaller the standard deviation needs to be to say whether or not the means or significantly different for a given sample size.

Man this thread makes me feel like I wasted money the first 2 years of college!

In principle, a student should have some notion of experimental design after their first college statistics course. In practice, only a subset of the simpler sorts of experiments are covered, and many students pass the course based more on their understanding of other material.

Many four year science and engineering degrees require a junior or senior level course that has a much more significant focus on experimental design. However, the real research world is much broader than what can be considered in any single semester course, and few (if any) undergraduate courses do much in terms of addressing uncertainty and confidence levels in cases of non-normal distributions. Other science and engineering degrees work in experimental design and uncertainty analysis into their laboratory coursework.

None of it is terribly hard, but it is specialized, depending on the type of experiment and goals of the study. It requires focus and attention to detail to determine which calculation is needed and when.

Damn, I feel like a dummy

dont worry i do to... the last time i took statistics was in 2002. Drug calculations are the only math I have been using since then lol.

I measure my samples in CC's ... Works out great for me!! LOL