NEED HELP with understanding when to use what method

[mgroves]

I have been trying to familiarize myself with statistics but I'm getting confused on what methods to use when answering what questions.

For instance, if I want to determine which of a set of tested parameters (A, B, C, ...) are significantly different between two groups, would I use a t-test, Mann-Whitney U test, or something else? Also, I read that when there are multiple significance tests going on at once, I need to do an adjustment (Bonferroni, Benjamini-Hochberg, Holm) on the data. However, when I do these adjustments, my BH p-values are crazy high (smallest is around 0.45). I'm not sure what that means. Interpretation of a lot of statistical tests is a bit of a challenge for me as well. Also, is this type of data set considered multivariate if I'm only grouping according to type (group) and nothing else, like age or sex?

What would I need to do if I wanted to answer questions such as:
1. Does group B show a significant change from group A?
2. Determine the dependence of the tested parameters.
3. Determine if the data follows a normal distribution. If not, what kind of distribution best fits the data?

General statistical questions:
4. What is the difference between a similarity matrix and performing various t-tests?
5. Why is clustering so important? What kinds of questions could clustering help answer?
6. How do similarity matrices and distance matrices differ?
7. How are QQ-plots best used?
8. When should data be transformed before performing an analysis on it? Are there certain transformations that are more ideal for answering certain questions about the data?

Clearly, I'm thoroughly confused. I am currently reading Numerical Ecology with R by Daniel Borcard. I am also referencing about 7 other books to try to help me to understand the things that Borcard doesn't explain. I'm getting very lost. Please HELP!!

 

[JesperHP]

It is very hard to help when you do not ask only one or perhaps two questions but rather a whole battery of questions ranging from very general questions such as choice of method - without delimiting and defining the methods to choose from - to question about the p value of a particular test you have run.

It may be that you are very stressed and confused - and probably anyone taking statistics seriously - have been in your state of mind, but if you want to get som answers you would probably be better of trying to deal with one problem at a time. And you probably also stand more of a chance getting an answer in this forum if you follow the philosophy: One post one question.

This is in no way meant to be disrespectfull I'm sure all you're questions are important and equally so.

Best regards

 

[terzi]

Hi mgroves,

As Jesper noticed, you certainly have a lot of issues that can't be fully addressed at once. I'll try to elaborate on a few details I noticed that may be confusing you.

* In order to test differences between two groups I'd recommend some Non parametric procedure (such as Mann-Whitney's U test), I find them easier since tend to have fewer assumptions involved.
* Cluster analysis is a multivariate technique that groups either observations or variables that have a similar "behavior", expressed within the data. In order to perform that analysis you need to decide which observations/variables are similar and for that you need a way to measure its similarity, thus similarity measures are needed. For simplicity, let's just say that distance matrices are some form of similarity measures.
* Normality is usually tested graphically (QQ-plots can be used for this but there are others) or with some special tests devised for the purposes. Try googling the shapiro wilk test.

My best suggestion would be to seek for some professional assistance if you are having troubles with your analysis. There are many statisticians who will gladly help you. Doing statistics wrong may be quiet serious, specially because usually no one will tell you that you are wrong and awful conclusions may be obtained from incorrect results. If the study will have a significant impact, I'd recommend an expert's advise.

Best regards,

 

[mgroves]

no one will tell you that you are wrong[/B] and awful conclusions may be obtained from incorrect results. If the study will have a significant impact, I'd recommend an expert's advise.

I am studying (on the job) to become one of these professionals you mentioned, but with my limited statistical (formal) education, it is quite mind-blowing learning everything at once. You have really helped to answer many of my questions though. Thank you so much!!