You have taught us already.THANKS

Hi Abiodun

I could not understand what you mean ! Please respond honestly to my honest question...

First a question

are these the a logit coefficient or the probabilities? I think the logits - you then have to do any calculations on that scale and then finally transform to probabilities at the end.

The other thing is that the estimate of the intercept is the log-odds for when all the X's are zero which may be outside the range of the data (hence negative value on the logit scale - that is a probability below 0.5). One possibility is then to set the Xs to their average value and range x4 the variable of interest- over (say) deciles to calculate the logits and then turn into probabilities.

We have written software in MLwiN to automate this but once you have extracted the estimated coefficient you could easily do it in excel

Data Manual supplement for MLwiN Version 2.14

Dear Kelvyn

Thank you very much. I know my limitation to understand high sounding statistics. Here is my sample data-set for your reference. Any support on it will be highly appreciated.

Are any of these predictors categories or they all continuous variables? If they are categories give them short labels (eg near , some way, far , very far) and re-post- and which ones are of primary interest (the exposures) and which ones are controls?

Dear Kelvyn

They are all continuous variables. This data was collected from sample plots in forest where cut and uncut stems were recorded as binary. Except DBH, all the cut and uncut stems within a plot received same value, i.e. VilDist, RivDist, Slope, RdDist etc. My interest is to predict the probability of a tree being cut as predicted by these factors/variables. So, the data-set does not have any control variable/s., nor they are categorized into classes. I hope we can go ahead this data.

Thank you very much for your prompt response and initiative to my problem

Lila

Here are some results - are multiple trees in same plot - if so need an indicator of this as likely to be dependency

Hi Lila,

as, in your message, you named your variables X1, X2 etc. I am not sure of my following comment.

When you use a model with all the 6 predictors you obtain a result that could be different when using only one predictor at a time. When you plot cut vs. a given predictor you ar implicitly using this sole predictor. I ran a model in R with all the predictor and then the six models with only one predictor. Results are

all together ; one at a time

DBH ; -0.0446739399 ; -0.032742052

dUser ; 0.1264257260 ; 0.152767966

Slope ; -0.0264489008; -0.038388939

VillDist; -0.0018479015; -0.004168453

RdDist;0.0019185388 ;0.001479506

RivDist;0.0003425596 ; -0.001651943

You see that the sign has changed for the RivDist.

Was this variable the source of your trouble?

NB - The coefficients when all predictors are used are the same that KJ obtained.

I agree with KJ that a correlation within "plots" might be a concern.

Excellent question waiting for answer

Dear Kelvin

I am really overwhelmed by the results and plots you produced from the sample data. This is what I actually wanted out of my data. you have shown me my destination to reach, could you please provide me the steps that i can follow to carryout similar analysis.

I am really grateful to you kelvin!

Lila

I did it in MLwiN

there is a 30 day trial download here

http://www.bristol.ac.uk/cmm/software/mlwin/download/

I attach a long worked example of what can be done with a binary logit model - if there are multiple trees in plots then it is worth doing this type of multilevel modelling as there is likely to be unexplained depdenecy within a plot- mapping the unexplained residuals at the plot level could also give you some insight as to what plot level variables might be needed for a fuller explanation of what is going on.

Here is a free course on logit models - you need module 6 and possibly module 7

http://www.bristol.ac.uk/cmm/learning/online-course/course-topics.html