My goal is building a predictive model to give probabilistic forecasts.

My response variable has lots of zeros but otherwise looks close to a gamma.

I fit the whole dataset using some classification techniques and have probabilities for zero vs. nonzero for each record.

Now I run a gamma GLM on just the records with positive response.

I think in this case I have a mixture of a degenerate distribution at 0 and a gamma. How can I derive the cdf of a prediction?


I assume from the way your question is phrased you're conditioning on the estimated parameters (if you're not - i.e. if you're incorporating the impact of the uncertainty in the esitmation of the parameters, it's somewhat more complicated).

As you say, it's just a finite mixture of a degenerate distribution at 0 and a gamma.

So if your degenerate has distribution $H_0(x)$ and your gamma has cdf $G(x)$, and they occur with probabilities $p$ and $1-p$ respectively, then

$$F(x) = pH_0(x)+(1-p)G(x)$$


$$F(x)= \begin{cases} 0 & \; x \leq 0 \\ p+(1-p)G(x) & \; x \gt 0 \end{cases}$$

where $G(x) = \frac{\gamma\left(\alpha, \frac{x}{\beta}\right)}{\Gamma(\alpha)}$ (for the scale-shape form of the gamma, where $\gamma$ is the lower incomplete gamma function).


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.