Fitting probability distribution connection to regression

208 Views Asked by At

My question is whether there is some connection between fitting probability distribution on some data set and linear regression? Or this two tools are for different problems?

By fitting probability distribution I mean that I have some data $x_{1},...,x_{n}$ and I believe they came from for example normally distributed population with parameters $\mu$ and $\sigma^2$. And I will estimate these parameters with likelihood estimation technique or method of moments technique.

By regression, I mean that I have random variable $X$ that is modelled for example as $X \sim a+bY+cZ$ where $Y$ and $Z$ are some random variables and I am estimating parameters $a,b$ and $c$.

So my question is whether there is some link between these two "mathematical topics"? Whether for example, I can look on fitting probability distribution as some specific problem of linear regression?

Thank you for your answer.

1

There are 1 best solutions below

0
On BEST ANSWER

Yes, if you assume that your model is $$ Y_i = \beta_0 + \beta_1 x_i + \epsilon_i, $$ where $ \epsilon_i|X \sim \mathcal{N}(0, \sigma^2) $. Hence, fitting a regression model is the same as estimating the parameters of the (conditional) distribution of $y_i$. I.e., you assume - by imposing the distribution of the noise term $\epsilon_i$ - that $$ y_i |X \sim \mathcal{N}(\beta_0 +\beta_1x_i, \sigma^2), $$ thus estimating the coefficients $\beta_0$ and $\beta_1$ and the variance $\sigma^2$ is the same as estimating the conditional expectation and variance of $y|X$.

Moreover, in a process of model selection - selecting a model from a set of possible models is the same as fitting a distribution to $y|X$. Namely, from statistical POV fitting probability model and regression analysis are very closely related topics.