In investigating trends in surveillance (especially cancer). Joinpoint (https://surveillance.cancer.gov/joinpoint/) is often used. Joinpoint simply estimates a segmented regression with cutpoints (chosen using various methods). We might as well end up with 4 joinpoints (cutpoints) on a sample size of 20 (one data point per year cancer incidence). To me, it seems just wrong. I was wondering what are the thoughts of other people and what else they think is a good alternative to joinpoint (I was thinking something like a GAM might be better). Assuming we are also interested in calculating the annual percent change (https://surveillance.cancer.gov/help/joinpoint/setting-parameters/method-and-parameters-tab/apc-aapc-tau-confidence-intervals/average-annual-percent-change-aapc).

  • 1
    $\begingroup$ There's nothing wrong with this per se; these are just change point models. The assumptions are different to those of the GAM; the GAM assumes smoothness (typically measured by a wiggliness penalty on the second derivative of the spline) whilst this method assumes discontinuities. Implementation will matter; how do you choose/identify the number & location of changepoints for example. $\endgroup$ Oct 1, 2019 at 17:43


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.