Jonas Kristoffer Lindeløv
Jonas Kristoffer Lindeløv
There should be a seed-argument for all functions involving randomness cf this question (and comment-answer) on StackOverflow: https://stackoverflow.com/questions/73416610/why-is-set-seed-not-working-for-mcp-r-package Suggested behavior for `mcp()`: * `mcp(..., seed = 42)` sets a seed...
Each segment should take an arbitrary number of linear predictors. As with the `segmented` package, the only requirement is that one continuous predictor (say, `x`) is the dimension of the...
- [ ] `fit$simulate(fit, data, ...)` apparently doesn't respect the link function in the family? Try e.g. simulating from `mcp_example("demo")` with another link function. - [ ] Try `mcp_example("demo")` with...
mcp 2.0 will support `stan` in addition to JAGS. It is far out in the future but this issue collects working points. - [ ] Obviously, generate a `stan` model,...
The overview should include the yet-to-be-reviewed/tested packages mentioned at the top, and perhaps an additional example: https://lindeloev.github.io/mcp/articles/packages.html. Also update: - [ ] `strucchange::breakpoints` does take time series and returns confidence...
As raised in #115, #107, and elsewhere, users increasingly modify the `jags_code` generated by `mcp` but get stuck when introducing new variables. We should make it easier to do this....
Include optional random effects for all parameters. [Here is a great guide for JAGS and stan](https://www.ling.uni-potsdam.de/~vasishth/JAGSStanTutorial/SorensenVasishthMay12014.pdf) Or add a varying slope too: list(y~ 1 + x, (1|id) ~ 0 +...
Test accuracy of fits: - [ ] Add strong priors so that parameter estimates are (should be) altered considerably. - [ ] Validate parameter recovery in one-segment models against `rstanarm`,...
Family-specific functions are currently scattered over several locations in the code and dealt with in an if-else like fashion. This hard-coding of response families makes it harder to implement new...
As discussed in the last part of [this article](https://lindeloev.github.io/mcp/articles/predict.html) `mcp` can be "hacked" to model change points in regions with no data. This is useful when forecasting. A more user-friendly...