Checking in.

[rivka]

I'm back at work for the second day. Yesterday I had an unusually light clinic schedule, so despite some continuing shortness of breath I decided to go in. Then I wound up having to arrange for a patient to be hospitalized, so it wasn't the lightest day after all. I need to call and find out whether he was actually admitted - I walked him over to the ER, and left the kind of note for the ER physician that should always result in admission, but the patient is kind of confused and I suppose that a lazy doc trying to avoid an admission at all costs (it's a lot of extra work) could worm out of him a promise not to kill himself before he sees his outpatient psychiatrist again, and call that justification for sending him home. So we'll see.

I just got off the phone from an hour-long conversation with my dissertation advisor. I sent him a huge number of analyses almost a month ago, and he kept putting off looking at them. I finally cornered him into agreeing to a phone appointment today to discuss them. The good news: he agrees that there really is something there in my results. The bad news: he thinks I need special analytic tools that neither of us really understands. Anyone out there an expert in logit and probit models? How about nonparametric models?

Comments

[melebeth]

I had a semester on them in my biostatistics class, but I don't remember all that much. However, if you have specific questions I can pass them onto our biostatistics guru who I am currently, coincidentally, sharing an office with :)

[kalmn.livejournal.com]

are logit and probit and nonparametric models special bio-stats things, or are they regular stats things? (one semester of stats in college, during which my ass was saved by a total moment of zen enlightenment during the final. now i don't remember most of it.)

[johnpalmer.livejournal.com]

I'm not an expert in any of those things... but if you get me specific names ("the Bieran Pretzels test", or whatever, I might be able to learn, and might be able to translate math to English for you, if you feel you need to understand the theory before applying the test.

I can give you the advice that you can find (at least) two different explanations. One will be the Math Ubergeek explanation, which will give information about PDF and UDF and doctoral theses supporting it, and one will be "so, you take the numbers, see... and then you do this to them, see... then you compare them to this chart, and if they're higher/lower, you know the results mean suchandsuch at the .05 level."

You might want to keep the first kind of reference to cite, but the second is what will have some meaningful information.

(Apologies if you know all of this... I'm remembering my own experiences, and how frustrating it was to discover what a distribution *really was*, from the math theory supporting it.)

If you had to choose between two people to explain things to you,

1) a statistics professor is probably going to be better than most math professors, and
2) a person with (or working on) a master's degree is probably going to be better than most doctoral candidates.

The statistics prof is more likely to know the second reference type; the person working on the master's degree is more likely to be looking for applied statistics, rather than theoretical (again: more likely to come up with the second reference, rather than the first.)

[rivka.livejournal.com]

My problem is that the data for my principal outcome measure is not even close to being normally distributed: I've got mostly scores of 0, with a scattering of higher-than-zero scores. Obviously this violates the assumptions of most statistical tests - it's not really fair to compare mean scores of my two groups using ANOVA, for example, which sucks because the ANOVA comes out significant. And I can't manage to set up a meaningful regression equation to test the contributions of possible predictors, because the range of the outcome variable is so restricted and the scores are so heavily clustered at zero.

Inspecting the data, it really does seem that a group difference exists. The question is which statistics we can legitimately use to bring it out. I fear that if I just cite the significant ANOVA result I'll be crucified by my committee, and rightfully so. But I don't know how else to organize the data analysis. If your biostatistics guru officemate has good suggestions, I am willing to provide any reasonable sort of reward. (For example, Godiva chocolates.)

[rivka.livejournal.com]

Thanks. My problem is more in test selection than in understanding the math behind whichever test I eventually use - honestly, I'm not expected to know that stuff as long as I can cite evidence that it's an appropriate test to apply to this kind of data. But I've got an oddly-shaped data set, one that doesn't meet the assumptions of most common statistical tests, and I'm not at all sure what to use instead. *sigh*

[rivka.livejournal.com]

They're regular stats things, for special kinds of data. I fear that I'm floundering around not even sure what kinds of stats would best apply to my situation.

[hobbitbabe.livejournal.com]

Some of the universities I've been affiliated with had a "Statistical Consulting Service". I know we have one here, but I haven't needed one since moving here. If you have such a thing, this sounds like the kind of problem for them.

Where I did my M.A.Sc., the consulting service open hours were run by grad students, but if your problem was too much for them / interesting enough you could see the experimental-statistics professor for a brief consult for free (an hour or two), or a longer consultation for money.

[rivka.livejournal.com]

It turns out that the University of Iowa (which is where I'm actually enrolled in the doctoral program, although I now work at the University of Maryland-Baltimore) does have a Statistical Consultation Center, and that graduate students can have up to ten free consultation hours.

The bad news: They're not technically open right now because it's the break before the summer session, and they don't technically re-open until June 7. And I need to defend my dissertation by the end of July or re-take comps. I can't re-take comps. I just can't.

The good news: The person who answered the phone was willing to listen to my problem and give me some advice on test selection. She thinks I should dichotomize my outcome variable (which you don't do to continuous data without a good reason, but she thinks this is a good reason) and use logistic regression, so that's what I'm going to try. Assuming that melebeth's biostats guru doesn't come back with a different answer, of course.

[kalmn.livejournal.com]

i forwarded you an email from my dad about this. (my dad is a stats geek.) i hope it's helpful.

[therealjae.livejournal.com]

Get thee to a stats consultant. I'm serious; they're worth the effort I'm sure the U of Maryland has a centre, though you might have to go to College Park.

-J