# What assumptions of linear regression are violated if we use linear regression with threshold for classifiction?

I have read through the topic and have seen two major factors:-

1. I have seen the intuition around this and sort of understand it

However, I keep seeing two other points which are mentioned here.

1. The error term in ordinary least squares regression is assumed to be normally distributed for the sake of inference. With a binary dependent variable, the error term won't be normally distributed. Instead, it follows the Bernoulli distribution. Can someone please explain why the error in such a case would follow a Bernoulli distribution?

2. Without going into the details, the assumption that the error term has constant variance also breaks. The error term becomes heteroscedastic. Can someone please explain this as well. Why would the error term become heteroscedastic?

• The first claim (2.) is wrong. The error term of threshold OLS can e.g. also be negative, so it can't be Bernoulli. – Michael M Oct 3 '17 at 6:35
• I thought so too. Can you please explain the second point? That sort of looks valid but cant exactly create an intuition around the same. – Arc Oct 3 '17 at 7:02