Very OT: any statisticians in the house?

[CantDog]

I’m preparing a grant with an aim that requires me to do an ordinal logistic regression. And when I say that I mean I will have a biostatistician run it once I get funding. Only problem is I still have to submit a short research plan describing what I will do...and I have no idea what I’m talking about.

Any statisticians in the house that I could run what I’ve written by to see if it makes sense?

[Sirshredalot]

I’m preparing a grant with an aim that requires me to do an ordinal logistic regression. And when I say that I mean I will have a biostatistician run it once I get funding. Only problem is I still have to submit a short research plan describing what I will do...and I have no idea what I’m talking about.

Any statisticians in the house that I could run what I’ve written by to see if it makes sense?

My background is in econometrics. PM me. Maybe I can help.

[primate]

Find a few papers that use ordinal logistic regression and mix all the jargon in the relevant methods description sections into an impressive sounding word salad. If you don't understand what you want to get out of the statistics, you may have bigger problems than a limited statistical vocabulary. If you have a good idea of what you want to get from the statistics, you should be fine adapting verbiage from the literature and wikipedia. Not a statistician, but I'm happy to briefly look at your research plan if you PM me.

[iceman]

Find a few papers that use ordinal logistic regression and mix all the jargon in the relevant methods description sections into an impressive sounding word salad.

The ol' "baffle 'em with bullshit" approach.

[CantDog]

The ol' "baffle 'em with bullshit" approach.

That’s my typical approach to most things—but this is one instance where they will know more about the shit than I.

[Mazderati]

https://www.ctsi.ucla.edu/education/...roach_2015.pdf

- Study Design
- Sample Size and Power Analysis
- Statistical Methods

UCLA and BU SPH websites are great resources for all things biostatistics. If you know ordinal logistic regression is required and know why, you may have the bulk of information needed for the study design and methods sections. You will probably need a statistician to calculate sample size and power. Statisticians specialize like other professions so leveraging someone with grant writing experience will be most helpful. That shouldn't be a problem given every major hospital in the city has a research department with strong academic ties likely willing to help.

[dan_pdx]

I’m preparing a grant with an aim that requires me to do an ordinal logistic regression. And when I say that I mean I will have a biostatistician run it once I get funding. Only problem is I still have to submit a short research plan describing what I will do...and I have no idea what I’m talking about.

Any statisticians in the house that I could run what I’ve written by to see if it makes sense?

Why don't you find the statistician you'll have run it if you get funded and ask them to help you? Nice to have your incentives aligned with theirs.

[Long duc dong]

That’s my typical approach to most things—but this is one instance where they will know more about the shit than I.

Interestingly enough I have NEVER done an ordinal logistic regression. Not even in grad school. I can't count how many times I've done logistic regression or multinomial regression, but never an ordinal one. I'm pretty busy right now, but I might have some time tomorrow to look into this. PM me. Like I said, I don't regularly use this approach, but I should be able to look it up and give you some ideas about what to say.

What type of data do you have? Is it longitudinal or cross-sectional? If it is longitudinal you will want to use a generalized estimating equation or perhaps a generalized linear mixed effects model. It might be very important to specify this. Not taking into account correlation of response variables, which is what is found in data with repeated measurements (longitudinal), can really be an issue with the inferences you make. In a nutshell, do subjects have one measurement, or does each subject have a number of measurements? If it is the latter you will have to take that into account as there will be correlated response variables for each subject.

If there is significant individual variation you might want to include a random intercept and possibly a random slope as well as the fixed effects. In order to determine all of this you will actually have to start analyzing the data.

I'm pretty busy right now, and I want to get a dawn patrol session in tomorrow, but I might be able to help.

Are you analyzing data that has already been collected, or do you need to get the funding in order to perform the study? Sorry for all of the questions, but you want to show the funders that you have a thorough understanding of exactly how the data you collect will be analyzed, so that you can make proper inferences on the results of the study. This is true even if you are going to have a professional do the data analysis.

Like I said, if this was just a regular old logistic or multinomial model I could help you more easily. I can't count how many times I have done a data analysis using those techniques.

How many categories are there for the response variable? For a standard logit model you are just modelling the log of the odds, basically log(p/(1-p) = XB, where X is the design matrix and B is the vector of beta coefficients. With multinomial regression you are basically have multiple categories, and instead of modelling the log odds of being in, or not in a category, as with the logit model, you are modelling the log of the probability of being in each different category over the probability of being in a base category, which is sometimes called the pivot. With multinomial regression there is no order to any of the categories, while in the ordinal case the categories are ordered. An example of the former might be categorizing people by which region of the country they live in, southeast, southwest, northeast, northwest......While an ordinal situation might be one where people are categorized by, say, level of education. I believe, though I have to look this up, that ordinal logit regression models the log of the probability of being in a given category, or the categories beneath it, divided by the probability of being in the categories above said category. I'm am far from certain about this though.