So let's say I wanted to do a research with this topic:
"Factors affecting QOL of beta-thalassemic patients"
Study Design: Cross Sectional
Number of Independent Variables: 21
Dependent Variables: mean QOL score
Statistical Analysis: Descriptive statistics, bivariate analysis, followed by multiple linear regression
Power: 80%
Significance level: 0.05
Aim of study: figuring out which of the 21 independent variables included significantly affect the QOL of beta-thalassemic patients
First of all, is it correct for me to use G*power F test to do a power analysis with A priori to determine my minimum sample size or should I be using something else?
If yes, then which test do I use? Do I use the "Linear multiple regression: fixed model, R-squared deviation from zero", or "Linear multiple regression: fixed model, R-squared increase" or should I use something else entirely? Thank you for helping out!
I think if you plain to apply simple model, i.e. detect a significant relationship between your independent variables and the dependent variable in one step, you may choose linear multiple regression: fixed model, R-squared deviation from zero.
The second choice is useful when you want to evaluate whether adding more predictors significantly improves the explanatory power of your model.
The key issue is what's your purpose. Is it to determine whether your entire bag of predictors work, or do you want to restrict to the subset which do predict over and above other predictors which also work? I can't imagine that you want to credit a set of variables which includes both "winners" and "losers". If I'm correct on that, you'd want to use ""Linear multiple regression: fixed model, R-squared increase". When doing the actual analysis, you'd presumably follow a stepwise procedure.
@Paul Max Kohn
So the idea is to first do a bivariate analysis of all 21 variables. Then the ones that are significant are included in the linear regression analysis.
The ones who have colinearity are excluded, then backward step to exclude variables with the least significance will be removed until we get the model in which all variables are significant.
Does that mean I need to use the R-squared increase instead?
Hello Aldwin Edbert,
What I'm suggesting is that you do a stepwise multiple regression. This procedure successively deletes variables which do not add significantly to prediction. (There are three alternative procedures forward inclusion which progressively adds significant predictors, backwards elimination which progressively deletes nonsignificant predictors, and stepwise (proper) which starts forward but retests deleted alternatives with the addition of each new predictor.) It is because all stepwise alternatives are based on R-squared increase that I think it's the appropriate choice for estimating statistical power.
- Badges
- Science topic
- Similar topics
- Psychology