# Is the product of conditional probability distributions a valid distribution?

In probabilistic machine learning, the likelihood of the data is usually computed as the product of the individual likelihoods of seeing each data point given the parameters $$\theta$$. In logistic regression, the likelihood of the data given the parameters $$\theta$$ is equal to $$P(Y|X,\theta) = \prod_{i=1}^m[\frac{1}{1+e^{-x_i\theta}}]^{y_i}[1-\frac{1}{1+e^{-x_i\theta}}]^{1-y_i}$$ This is just the product of $$m$$ bernouli distributions. I have two questions.

1. What is the relation between a conditional density distribution and the likelihood that is used in baye's theorem ?
2. Is the product of distributions in this setting, bernouli distributions going to result in a valid conditional distribution ?

EDIT

Regarding the first question, suppose I am able to define a joint distribution for $$P(Y,\theta|X)$$. This is assuming $$\theta$$ random variable is follows some distribution. Then, given a fixed data $$Y$$, we can slice the joint distribution at $$Y=y$$ and obtain the marginal density $$P(\theta|Y=data,X)$$. Then $$argmax_{\theta}P(\theta|Y=data,X)$$ represents the $$\theta_{MLE}$$. In this case the conditional probability is the likelihood function for $$\theta$$. Integrating this yields a value of 1. However, if $$\theta$$ does not follow some distribution then integrating this likelihood is not equal to 1 ?

The entire (Bayesian and classical) analysis of a generalised linear model is conditional on the regressor vector $$X=(x_1,\ldots,x_n)$$.
The joint distribution of the $$Y_i$$'s in the logit model $$p(y|x,\theta) = \prod_{i=1}^m[\frac{1}{1+e^{-x_i\theta}}]^{y_i}[1-\frac{1}{1+e^{-x_i\theta}}]^{1-y_i}$$where $$y=(y_1,\ldots,y_n)\quad\text{and}\quad x=(x_1,\ldots,x_n)$$ is a valid joint pmf [on the components of $$Y$$] conditional on the vector $$X=(x_1,\ldots,x_n)$$, assuming the $$Y_i$$'s are independent given $$X$$ and that $$\mathbb P(Y_i=1|X=x,\theta)=\frac{1}{1+e^{-x_i\theta}}$$ As a joint distribution, it defines a likelihood function $$\ell(\theta|X,Y)=\prod_{i=1}^m[\frac{1}{1+e^{-x_i\theta}}]^{y_i}[1-\frac{1}{1+e^{-x_i\theta}}]^{1-y_i}$$ that can be used in a Bayesian analysis.
As a function of $$\theta$$, the likelihood is not a pdf, it does not integrate to one except in some specific cases (not including the logit model). The same applies when given a prior $$\pi(\theta)$$ one considers the product $$\ell(\theta)\pi(\theta)$$: it does not integrate to one. The joint distribution of $$\theta$$ and $$Y$$ is $$p(y|\theta)\pi(\theta)$$ which integrates to one in $$(y,\theta)$$ and the conditional distribution of $$\theta$$ given $$Y=y$$ (and $$X$$) is $$\dfrac{p(y|\theta)\pi(\theta)}{\int_\Theta p(y|\eta)\pi(\eta)\,\text{d}\eta}$$ which integrates to one in $$\theta$$. The marginal $$\int_\Theta p(y|\eta)\pi(\eta)\,\text{d}\eta$$ integrates to one in $$y$$ [except that it is a summation since $$Y$$ is discrete].
• The scenario I gave in the edit section where $P(\theta|Y=data,X)$ represent the likelihood. If i integrate over $d\theta$ i should get 1 because it is a marginal distribution ? – calveeen Aug 8 '20 at 16:39
• why is $p(y|x,\theta)$ a joint distribution ? I thought it is the conditioned on $x$ and $\theta$ ? – calveeen Aug 8 '20 at 16:41
• Because $\theta$ often is not a random variable, calveeen, there isn't any joint distribution to condition on. You need to distinguish conditioning of random variables from the presence of parameters, especially because they often employ similar (or identical) notation. – whuber Aug 8 '20 at 16:50
• @whuber. $\theta$ is not a random variable when used in the likelihood probability in baye’s theorem ? $p(data|\theta)$ ? – calveeen Aug 8 '20 at 17:07