Stats help needed

[Badly Shaved Monkey]

My stats brain has turned out to be more rusty than I thought.

I have non-gaussian data.

Two groups: treatment X and placebo. Two time points: pre and post treatment.

What I need is to test whether X differs from placebo.

Please feel free to patronise. My mind has gone blank.

[Jeff Corey]

Try a t-test on the pre-post differences.
Were the groups randomly assigned?

[sphenisc]

My stats brain has turned out to be more rusty than I thought.

I have non-gaussian data.

Two groups: treatment X and placebo. Two time points: pre and post treatment.

What I need is to test whether X differs from placebo.

Please feel free to patronise. My mind has gone blank.

If your groups are paired then Mann-Whitney U.
If not then Wilcoxon Two sample.

Apply separately to pre and post data

[Note - from casual stats user, I could be wrong.] :)

[sphenisc]

Try a t-test on the pre-post differences.
Were the groups randomly assigned?

I think t-test requires that paired differences have a gaussian distribution. I'm not sure if BSM's data fits this?

[Jeff Corey]

Alan Boneau(1960) did a number of Monte Carlo runs and found that the t is rather robust with non-normal distributions.
(The effects of violation of assumptions underlying the t test. Psychological Bulletin, 57, 49-64)

[Badly Shaved Monkey]

Alan Boneau did a number of Monte Carlo runs and found that the t is rather robust with non-normal distributions.

The data are quite non-normal- truncated at lower values not much below the median and, I suspect, bimodal.

[Jeff Corey]

In that case, try the Mann-Whitney U and the t and see what you get.

[Badly Shaved Monkey]

Try a t-test on the pre-post differences.
Were the groups randomly assigned?

Yes, they were randomly assigned.

I think a U-test on pre-post difference is the simplest.

The members of the two groups are different, but pre and post-treatment is within the same subject, so I think the U-test encapsulates what is needed.

I don't know where to begin with a power analysis of non-parametric stats. The problem, as usual in vet med, is that there is very little prior data so a pilot trial collides with a definitive test. If we are doing U-tests, do 20 subjects in each group present a small likelihood of fluke results? Coefficient of variation is quite high- up to 20% for repeat tests in the same subject.

[Jeff Corey]

Using the t would make the power analysis simpler.

[Badly Shaved Monkey]

In that case, try the Mann-Whitney U and the t and see what you get.

Given my well-known interests, you might guess rightly that this is 'politically sensitive'. The protocol must be made public before the tests are run, so there is no room for embarking on a fishing expedition. The U-test makes more conservative assumptions so it is probably better to state that this will be the method of analysis.

[Badly Shaved Monkey]

Using the t would make the power analysis simpler.
Yes, that's the problem

We seem to be posting about one post out of step. I need to go and eat!!

Thanks for the help. I'll catch up with this later.

[TruthSeeker]

this might help:

http://www.childrens-mercy.org/stats/size/mann.asp

[T'ai Chi]

1) what software are you working with?

2) what does a histogram of the data look like?

3) and how big are the groups?

[Art Vandelay]

So the issue of whether they have same mean is definitely not settled, but are you assuming that they otherwise identical distributions? When you say "I have data", you are speaking in anticipation of the future situation?

How important is it to you to get as much power out of your significance level as possible? One strategy would be to simply pair up drug and placebo subjects, and look at how many times the drug is better. That will be a binomial distribution. Of course, there are several problems with this, such as that it doesn't have as much power as other methods, and that it doesn't measure the mean. But then, the Mann-Whitney U test doesn't either.

[Badly Shaved Monkey]

When you say "I have data", you are speaking in anticipation of the future situation?

Sorry, I was being elliptical. It is a future situation, so we are showing good discipline in defining the analytical methods ahead of time.

I probably shouldn't have said "I" have data, because I'm helping someone else with this, so the data "I" have is secondhand. Also I've been asked for advice quite late on in the study design so various of the preliminaries are now beyond my control. I've been a little surprised that there seems to be a problem with checking the structure and statistical distribution of already available data, but I have no control over that and have to work with what I've been given, which is a sparse amount of published data and my own clinical knowledge of the subject, to try to make reasonable judgements.

Why would I be asked for help with this despite my evidently marginal competence? Because in our little world it is hard to find cheap help that has even marginal competence and I'm doing it for free for a friend, which makes my small knowledge quiet attractive to the study designer.

[Badly Shaved Monkey]

this might help:

http://www.childrens-mercy.org/stats/size/mann.asp

Excellent. The good news is that, at worst, n for the U test is only 1.16 times n for an equivalent t-test. This means that not only can you calculate it, it's close enough to 1.0 that, in the real world, if you have a rough idea of an appropriate n for a t-test it's not miles away from being OK for the U test. It's unlikely that a study will fail to get run for needing to recruit 16% more subjects, but it might do if it needed 4 times as many subjects.

[Q-Source]

Stats help needed too.

Anyone here knows what are the basic diagnostic tests that I should apply to a simultaneous equations system?

It is a large dynamic system (6 lags) estimated by 3SLS. I need to check for heteroskedasticity but it seems impossible as it is too large to test for it. Anyone can suggest how to do it?.

Not even one textbook of statistics or econometrics says anything about it. Why don't they ever address practical things?. Aggg!

[Jorghnassen]

Are you talking about doing statistics on differential equations? Because if so, it's not that basic, though I do believe there's a book by Ramsay and Silverman (http://www.psych.mcgill.ca/faculty/ramsay/fda.html) that might help...

[Elaedith]

There is a newsgroup called sci.stat.consult which I have found helpful when I get stuck.

[Q-Source]

Are you talking about doing statistics on differential equations? Because if so, it's not that basic, though I do believe there's a book by Ramsay and Silverman (http://www.psych.mcgill.ca/faculty/ramsay/fda.html) that might help...

No, it's about econometrics. I need to test if my estimates are valid.


Elaedith

There is a newsgroup called sci.stat.consult which I have found helpful when I get stuck.

I might check that. Thanks.

[Jorghnassen]

No, it's about econometrics. I need to test if my estimates are valid.

Well, when you mentioned dynamic systems I figured it'd be functional data analysis. Anyway, maybe this book (http://www.econ.queensu.ca/ETM/) on econometric theory can help...