If someone could help me determine which statistical test I need, I would be appreciative. Here's a basic outline... \(f(w_1,x_1, y_1,z_1) = K_1\) \(f(w_2,x_2, y_2,z_2) = K_2\) ... The way function f() is set, for any set of values any variable may have the greatest effect on K. I want to test for a series of K's which variable predominates as the greatest impactor of K. These are the tests I am considering and their why's and why-nots: Pearson Correlation - This should quantify the assocation between 2 variables, so I could run multiple: x vs f w vs f ... I'm not sure if it's a Gaussian distribution or not...so I might choose a Spearman Correlation Since more than 2 variables are involved...should I be considering the Repeated-measures ANOVA or Friedman test as opposed to the afforementioned tests? ------- I'm also having difficulty understanding the difference between Matched and Unmatched, because my K is dependent on the variables is it considered matched?
You could try doing a Q-Q plot to test the data for normality, and then proceed from there. R-Project is good open source software that can do it for you, but it's written by geeks for geeks (it can also do geostatitsics when you get more familiarized with it), however R Commander is a good GUI written for R that will help you learn the commands at the same time.
Luckily SPSS does Q-Q plots and I won't have to learn these complicated programs, and I will probably be using them by doing a series of tests such as: f(x=constant,y=constant,z)=g(z) verses Actual K....it's a good test. If you know of a way to do a dynamic set that would be great too. Because f(g(y,z),h(x,z),i(y,z))=K where it would plot the theorietical value as being ever possible solution. However, I'm looking for something more along the lines of: How much of K does variable X explain for the series of K data, how much does Y explain, etc etc.
I had had a very similar thought, actually, but was going to look at a couple of things before I mentioned it (of course, I no longer remember what those things were. Okay. Question time. From what I gather, you have two data sets? Predicted and observed? Three variables in those two data sets? (or is it 4?) Is one of them time (in otherwords, is it a 3 dimensional time series)? If the data is three, rather than 4 variables, have you tried doing a 3-D residual error scatter plot? Or are you specifically trying to determine something? Finally, if the data is indeed 4 dimensional (as opposed to a 3-d time series) have you tried binning the data according to one (or each) of the variables, and producing a series of 3-d scatter plots? I'll have a poke around and see if I can find anything that might be more useful.
I'm not likely to be much help here, but I recall that a colleague (in social sciences) did something similar using Factor Analysis. I think its usefulness depends on how many variables are involved (better with more).
You could just run a general Spearman's correlation in SAS. Or try backwards regression. Back it up with normality testing and Pearson's.
