Multivariable Logistic Regression
 

Anyone know how to run a multivariable logistic regression in JMP?
Or some other way to analyze my data?

I have three independent variables an one dependent. All are categorical.

Doubt anyone knows this junk, but I've seen a few scientists about the boards, figure someone might have a clue.

I'll give out more info if anyone is interested in helping.

I haven't the slightest idea of JMP, nor do I know the nature of the data, but I am pretty good at making programs that model data. Since you say regression I assume you are given a list of points, each point with four elements: Pi=xi,yi,zi,ri.

So ri ~ f(xi,yi,zi), what is f(x,y,z)?

Given f(x,y,z) isn't too complex I can make a simple algorithm that gives you the constants of your equation.

Not familiar with JMP, but do you have any access to MiniTab software?

Argg...I took that class in the '80's and certainly any software used is long obsolete. Thanks a bunch, gg. My post traumatic statistics disorder has kicked in again. :)

Man, I bet WillRavel knows something about this stuff.

I fucking hate statistics. I learned about this in my Regression class in 2006, but that all faded away about 2 hours after I took the final exam, unfortunately.

Sapiens is the hands-down statistical king of TFP. I would PM him if you need help.

Thanks, Augi, and others for your responses here! Unfortunately I haven't even taken a statistics course. They threw me into this field class with this JMP program, and told me to analyze my data with it. The graduate student who was helping me reached the limit of her knowledge when she realized that I did not have continous data.

Sorry, chassiswelder, no minitab access. Just JMP.

I'm supposed to be able to find a simple equation in y= mx+b format to plug into the raster calculator of ArcGIS (mapping), to make a predictive model of the population of my sedge.

I've spoken with a couple of faculty members about the issue, and they've given me a couple of other work-arounds. If/then statements to plug in, if I can figure out to plug them into the GIS program. Unfortunately now I haven't gotten an e-mail response from the program's GIS expert to tell me if that's possible. I'm thinking that I'll resort to making my layers transparent, overlaying them, and making a new polygon shapefile layer highlighting every occurance of the specific combination of abiotic factors that I'm dealing with.

Why would you assume that? His training is in psychology, not statistics.

From what my psychology professors tell me, psychology majors learn a lot of statistics, actually. They learn even more if they want to become researchers.

Now I am confused since you said you have three independent variables, not one like a simple y=mx+b graph implies, to me anyhow. Are you going to make 3 linear regressions of each combination of 1 independent and the 1 dependent? I won't go into detail on a solution that uses matrices to produce the best fitting lines if that is not what is happening.

Right, that's the problem. it's not as simple as y=mx+b. But that's what the mapping program supposedly needs. The mapping program could also handle something along the lines of y= mx + nx+ ox...+b.

If the equation is r=ax+by+cz+d the algorithm uses simple matrices:
(X^T * X) * (X^T * R) = [ a ; b ; c ; d ]
With:
X = [ x₁ , y₁ , z₁ , 1 ; x₂ , y₂ , z₂ , 1 ; ... ; x_n , y_n , z_n , 1 ]
and
R = [ r₁ ; r₂ ; ... r_n ]


Some notes:
For this to actually work, n must be equal to or greater than 4. If you have only 4 points then you are finding an exact equation, if n is greater than 4 then it is a best fit ... hyperplane.

If you have a graphing calculator it can perform these matrix equations. If you know some basic programming you can make the computer perform these equations. You can even make Excel perform these computations. It would be easiest to have a graphing calculator at the ready... those other two methods require a lot of work, luckily the amount of work is not determined by the number of points in your data.

Thank you, Augi!

I have a TI-83 graphing calculator. I know nothing about matrices. I really appreciate your help.

Here's a bit about my data:

I have 8 soil types. 83 locations total where the soil was collected. Two of the soil types only occur once. The rest have anywhere from 5 to 20 occurances.

I have 9 "average percipitation ranges". 83 locations. "average percipitation ranges" occur anywhere from 4 to 30 times each. (ie- 4 locations have 2000cm of precipitation/year, 30 locations have 4300cm/year).

I have 2 "average temperature ranges". 83 locations. One "average temperature range" occurs in 5 locations, another 79 locations.

I have presence/absence data for my plant for these 83 locations. Absent in 23 locations, present in 60 locations.

I'd like to make a predictive model of where you're likely to find my plant on the entire island, based on what I can gather from temperature, soil, and precipitation data.

What would X, R, and n be?

X is the matrix that includes your input data. R is the matrix of the results. n is just element index---in this case n goes from 1 to 83.

I am afraid that this method will not work, now seeing what your data are.

There are two non-linear elements in the data: soil type and plant presence/absence. The plant presence/absence can be generated as 1 for presence and 0 for absence, however that will give poor results. If the plant presence/absence data were replaced with a plant growth scale, like the height of the plant, then this would be more manageable. The soil type itself is no way a linear scale... my method would only work so far as to give you 8 linear models, one for each soil type. If you had numerical properties of the soil type then that could be modeled.

Then, the model itself is not a linear one. Both temperature and precipitation have an optimum range. Plant presence falls the further away you get from that region---not linear behavior. A second order polynomial would be better... This is not a big deal with matrices, but the data are not polynomial of any order---they are bell curves. I do not know how to begin to model this. Sorry.

Thank you for your efforts!