Using Linear Regression for Classification

Question

I am using Vowpal Wabiit to explore the power of different loss functions (e.g. squared, hinge, logistic, quantile) for classification. I've trained different models using each of these loss functions. Among many of LibSVM dataset, I have seen minor difference between these loss functions. My question is if linear regression can provide similar accuracy as logistic regression why we need to use the costly logistic regression? Is it only because of probability interpretation?