I have a question regarding the calculation of the variance of the sum of predicted health care costs using a two-part regression model. Details are below, but it boils down to how to calculate the covariance for a random variable which is itself the product of two random variables.

The two-part regression model consists of a logistic regression component for estimating the probability $p$ of cost > 0 and an OLS component predicting the cost $y$ given $y > 0$. The predicted cost $C_i$ for a given member $i$ is then the product $C_i=p_iy_i$. We can assume that $p_i$ and $y_i$ are independent.

I'm calculating the variance of the sum of costs $S=\Sigma_{i=1}^mC_i$ for members $i=1$ to $m$ belonging to a particular provider, assuming member level predicted costs are not independent given they're sharing the same two-part regression model.

So running through the math:

$Var(S)=\Sigma_{i=1}^m Var(C_i)+\Sigma\Sigma_{i\neq j}^mCov(C_i,C_j)$

where $Var(C_i)$ is:

$Var(C_i) = E(C_i^2) - [E(C_i)]^2$



and $Cov(C_i,C_j)$: is


My first inclination was to use a multivariate Taylor expansion to approximate the term $E(C_iC_j)$.

However I run into trouble with a first-order expansion:

$Z=g(C_i,C_j)\approx g(\mu_{C_i},\mu_{C_j}) + (C_i-\mu_{C_i})\frac{\delta g(\mu_{C_i},\mu_{C_j})}{\delta C_i} + (C_j-\mu_{C_j})\frac{\delta g(\mu_{C_i},\mu_{C_j})}{\delta C_j}$

since $E(Z)\approx g(\mu_{C_i},\mu_{C_j})=E(C_i)E(C_j)$ which means $Cov(C_i,C_j)\approx 0$, which can't be right.

Is there another method for approximating the covariance of $C_i$ and $C_j$? Or did I make a mistake in my calculations? Thanks!


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.