obtain probability prediction for binary logistic regression

I am trying to obtain probability predictions for a binary logistic regression model. However, I am getting output only as 0 or 1s.

The spec says:

•For a binomial model, the meaning of the output values in ypred depends on the value of the 'BinomialSize' name-value pair argument.

•If 'BinomialSize' is 1 (default), then each value in the output ypred is the probability of success.

•If 'BinomialSize' is not 1, then each value in the output ypred is the predicted number of successes in the trials.

https://uk.mathworks.com/help/stats/generalizedlinearmodel.predict.html

But this does not appear to be the case.

I need the input to be such that fit( X, y) with X (Y_init) being the features, and y (X) being the class vector. Here are my initial code:

mdl = fitglm(Y_init,X,'Distribution','binomial');

predictProb = predict(mdl, Y_init(500:544,:);

[the cyclist]

Can you upload your data, or a subset that gives the identical issue you are seeing?

If your model is effectively perfect, then the predicted probabilities could be all 0's and 1's.

[Tim Dong]

@the cyclist, I have attached the first 100 rows of the dataset.

This is the message I got when I run the model earlier.

Warning: Iteration limit reached.

> In glmfit (line 332)

In GeneralizedLinearModel/fitter (line 592)

In classreg.regr/FitObject/doFit (line 94)

In GeneralizedLinearModel.fit (line 905)

In fitglm (line 146)

Warning: The estimated coefficients perfectly separate failures from successes. This means the

theoretical best estimates are not finite. For the fitted linear combination XB of the predictors,

the sample proportions P of Y=N in the data satisfy:

XB<-4.1764e-09: P=0

XB>-4.1764e-09: P=1

> In glmfit>diagnoseSeparation (line 545)

In glmfit (line 338)

In GeneralizedLinearModel/fitter (line 592)

In classreg.regr/FitObject/doFit (line 94)

In GeneralizedLinearModel.fit (line 905)

In fitglm (line 146)

Warning: Regression design matrix is rank deficient to within machine precision.

> In classreg.regr/CompactTermsRegression/checkDesignRank (line 35)

In GeneralizedLinearModel.fit (line 924)

In fitglm (line 146)

>>

[the cyclist]

How many observations do you have in your full dataset? What you have uploaded here might not be that helpful, because you have more than 3,000 explanatory variables available to explain just 100 observations. So, it is no surprise that you have a "perfect" model, and that you are getting warnings.

In the following code, I use only 20 columns of Y_example, and you can see that the predicted probabilities are almost all very close to 0 or 1.

load X_example.mat
load Y_example.mat

columnSubset = 2:21; % Did not use first column of Y_example, because all values are equal
mdl =  fitglm(Y_example(:,columnSubset),X_example,'Distribution','binomial');
predictProb = predict(mdl, Y_example(:,columnSubset))

predictProb = 100×1

0.9954 0.8793 0.9978 0.9619 0.9548 0.9942 0.0544 0.9988 0.5806 0.9902

[Tim Dong]

@the cyclist, I have 544 rows in the entire dataset. Thanks for the suggestion, I can now see that these are indeed predicting probabilities. I am expecting to reduce the number of predictor columns to a lower number of ~10 though.

[Ive J]

On a different note, your independent variables are highly correlated, you should be extra cautious for multicollinearity.