5
$\begingroup$

I'm running a two-part model on a health insurance claims dataset where I predict the probability of nonzero health care costs using a logistic regression (1st part), then predict the magnitude of the log health care costs using an OLS regression (2nd part).

Question: When calculating the expected health care costs = (prob. nonzero claim)x(expected claim amount)x(Duan Smear factor) what is the value for the expected claim amount for the zero $ claims?

For example in the following table:

    member  observed cost  prob. of nonzero cost  expected cost  predicted cost                         
    1       0              .3                     ?              (.3) x (?) x (ds factor)      
    2       0              .2                     ?              (.2) x (?) x (ds factor)      
    3       1,000          .6                     900            (.6) x (900) x (ds factor)  
    4       5,000          .7                     7,000          (.7) x (7,000) x (ds factor)   

what is entered into the expected cost column for members 1 and 2 if the OLS regression is only applied to the nonzero cost data? Do we score the zero cost claims using the OLS regression parameters to calculate expected cost?

$\endgroup$
3
  • $\begingroup$ Consider adjusting your title to reflect the general statistical issue you are having. $\endgroup$ Feb 12, 2013 at 20:52
  • 1
    $\begingroup$ I think I've just spotted the issue here - you want to know for the purposes of calculating the Duan Smear factor correct? $\endgroup$
    – Corvus
    Feb 12, 2013 at 21:18
  • $\begingroup$ Yes - I should have specified for purposes of calculating the DS factor as well as calculating MSE for predicted vs. actual costs. $\endgroup$
    – RobertF
    Feb 12, 2013 at 21:27

2 Answers 2

4
$\begingroup$

The OLS model is the model of expected cost given that there is a non-zero cost. Therefore by conditionality principal you simply don't use the data that has zero observed cost when fitting that part of the model.

Reading between the lines I'm guessing your real issue is how to calculate the Duan factors. The duan factors are usually calculated by using the residuals and so without getting the residuals how do you calculate the duan factors?

The key here is to look carefully at the model and think about what the Duan factors are there for.

You want to know the expected loss $E[L]$. This is given by the expected loss given non-zero loss times the probability of loss, $E[L] = E[L|L>0]P(L>0)$. The Duan factors come in because you are not fitting your model to $E[L|L>0]$, but rather to $E[\log(L)|L>0]$. The Duan Smearing attempts to correct for this bias using the residuals of the fit. It should now be clear that the correct thing to do here is to only use the residuals from the OLS fit when calculating the DS factor.

Effectively treat this as two separate bits of analysis. First discard all the zero losses, then with the data that remains fit your OLS model and calculate the DS factor. The put your dataset back together and fit the logistic model.

$\endgroup$
5
  • $\begingroup$ Thanks Corone, this is helpful. If we're ignoring the data w/ observed cost of zero, won't the predicted cost be biased? If we're modeling nonzero health costs via an OLS regression, then multiplying the expected cost by the prob. of nonzero cost, seems like on average we'll undershoot the actual cost since the prob. will always be < 1.00, right? $\endgroup$
    – RobertF
    Feb 12, 2013 at 21:33
  • $\begingroup$ It depends what you want - do you want the expected cost, or the expected cost given cost > 0? I am assuming these zero costs are "real" - not just "missing data" $\endgroup$
    – Corvus
    Feb 12, 2013 at 21:37
  • $\begingroup$ Right, the zero costs = no health care costs. I'm attempting to predict future expected cost, given past cost data & other covariates. So using 2010 claims data, I'm predicting 2011 health costs, the idea being I can then apply the model to predict 2013 costs. However I won't know a priori which 2013 claims will be zero & which will be non-zero, so the part where we throw away the zero claims is still a bit confusing. $\endgroup$
    – RobertF
    Feb 12, 2013 at 21:57
  • 1
    $\begingroup$ Ok, think of it like this. What is the average of 4,6,8,0,0,0? Well, $\frac{4+6+8+0+0+0}{6} = 3$. We could instead have thrown away the zeros and found the conditional expectation $\frac{4+6+8}{3} = 6$, then we multiply by the probability of being non-zero $6 \times 0.5 = 3$. Does that help? $\endgroup$
    – Corvus
    Feb 12, 2013 at 22:23
  • $\begingroup$ Wonderful, thank you for the example - this clears up my confusion. $\endgroup$
    – RobertF
    Feb 13, 2013 at 16:39
4
$\begingroup$

I have not seen the Duan smearing correction used in this way, but at first blush, it seems sensible. Here's the intuition.

The general re-transformation problem we have is that we want to get \begin{equation}E[y_i \vert x_i]=\exp (x_i'\beta) \cdot E[\exp (u_i)].\end{equation} For the first term, you can use the exponentiated prediction from the logged model. The second is more tricky. If we assume normality and independence, we can approximate the second term with $\exp (\frac{\hat \sigma^2}{2}),$ where we use the RMSE from the logged regression for the unobserved $\sigma$. Or we can use a weaker assumption of $iid$ on $u_i$, and use the sample average of the exponentiated residuals from the logged model for the second term.* That's the Duan "smearing" approach.

If you have access to Cameron and Trivedi's Microeconometrics Using Stata, take a look at chapter 16, in general, and table 16.2, in particular. There they give expressions for the conditional and unconditional means of expenditure when the outcome is in logs.

However, they are using a probit for the first stage, and making assumptions about joint normality and homoskedasticity of the errors in the two stages, so it's a slightly different model than the weaker $iid $ assumptions of the Duan approach with the logit. But the intuition carries over.

Combining the intuition from the first paragraph and their table, we have that \begin{equation}E[y_2 \mid x,y_2>0 ] = \exp (x_2' \beta) \cdot \exp( \frac{ \sigma^2_2}{2} )\end{equation} and \begin{equation}E[y_2 \mid x ] = \exp (x_2' \beta) \cdot \exp(\frac{ \sigma^2_2}{2} ) \cdot \Phi(x_1' \beta).\end{equation} The $\sigma_2$ is the variance in the logged expenditure equation, which you can approximate as above. The subscripts indicate the stage (since you can have different predictors). The $\Phi$ is just the probability of going to see the doc from the probit first stage, which is like your logit.

Thus, it seems that as long as you maintain the assumption that the two stages are independent, all you need to do is multiply by $\frac{1}{N}\sum_i^N \exp (\hat u_i)$ if you want the expected positive expenditure. You will also need to create out of sample predictions for the expenditure for those who did not get to see the doc.

Addendum:

Stata now has the a user-written tpmcommand that deals really nicely with this problem.


* If you're worried about heteroskedasticity, you might try smearing by group. For example, you might try an age-group specific average residual.

$\endgroup$
3
  • $\begingroup$ Excellent, thank you Dimitriy. I use SAS but the Stata manual looks like a good resource to have handy. $\endgroup$
    – RobertF
    Feb 12, 2013 at 21:48
  • $\begingroup$ If you're a SAS user, I would recommend their other Microeconometrics book. It's not SAS-specific, but it might be more general. On the other hand, the Stata book from above covers less ground, but is much more applied, with lots of examples that could be "translated". $\endgroup$
    – dimitriy
    Feb 12, 2013 at 21:57
  • $\begingroup$ In particular, many or most of the examples are health-care related. $\endgroup$
    – dimitriy
    Feb 12, 2013 at 22:02

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.