I am looking at how pavement thickness variation parameters(absolute thickness, variance, slopes, critical points) relate to distress occurrence.

After some research, I have settled on the linear logistic model for regression testing. In order to get distress rates (pHat), I bin my binary data data using MATLAB's binning algorithm and then use the following function to fit a logistic linear model:

function[pihat] = fitRegression(X,Y)
    %X = [n x 1] array of predictor variable
    %Y = [n x 1] logical array of distress occurrence

    [N,edges,bin] = histcounts(X); %use MATLAB automatic binning algorithm
    X = (edges(1:end-1) + edges(2:end))/2; % get mean bin values for regression fitting.
    bin = bin(bin~=0); %remove 0 data        
    for i = 1:length(N)
        pihat(i) = sum(Y(bin==i))./N(i); % calculate pihat for each bin

    Y = pihat;
    glm = GeneralizedLinearModel.fit(X,Y,'distr','binomial', 'link', 'logit');
    logitFit = glmval(glm.Coefficients.Estimate,linspace(floor(min(X)),ceil(max(X)),numel(X)),'logit');

    ps = plotSlice(glm);

Is my binning and use of a logit GLM valid?

  • $\begingroup$ Why do you bin the data ? If I understand your code (I don't know mathlab) you have an explanatory variable $X$ and a dependent variable $Y$ where $Y$ is binary, so you can do logistic regression without binning ? or is $X$ categorical ? if $X$ is categorical, then binning should make no difference compared to using the binary $Y$ as an dependent. $\endgroup$
    – user83346
    Sep 18, 2015 at 6:21
  • $\begingroup$ You are correct, I do bin X. I originally thought that logistic regression needed Y to be a pHat. I have come across a new method (also in MATLAB) that does not require binning. I will post when completed. $\endgroup$
    – rconway91
    Sep 18, 2015 at 20:28


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.