In this paper, I explore the use of logistic regression for automatic digit recognition, which is a common method for improving efficiency in many areas. I train and test the models on two pairs of digits: 0 and 1, 3 and 5. The models are estimated using stochastic gradient descent. I also vary the sample size of the training and test data and observe the accuracies of the models. I find that the prediction accuracies stay consistent across all sample sizes for both pairs of digits. This could be attributed to the comprehensiveness of the data set which includes a great variety of handwritten digits for both training and testing.