Allow for NA values in the dependent variable throughout

[dwwolfson]

mcp Version 0.3.2
dplyr Version 1.0.8

When trying to plot a large fitted mcpfit object, I'm running into this error:
Error in dplyr::mutate():
! Problem while computing fitted = rlang::exec(...).
Caused by error in if (any(x < lower) || any(x > upper)) ...:
! missing value where TRUE/FALSE needed`

I also get the same error if calling fitted(model)

[lindeloev]

Thanks for reporting. Can you post code that reproduces this error? And can you post the stack trace? (use traceback() if it isn't returned automatically)

[dwwolfson]

The full traceback is pretty enormous. Here is the backtrace, hopefully that is helpful. I'll try to reproduce the error either by simulating data or posting data to a gist.

---

Backtrace:
x

1. +-base::plot(f1)
2. +-mcp:::plot.mcpfit(f1)
3. | -mcp local_pp_eval("fitted")
4. | +-... %>% dplyr::rename(!!yvar := !!type)
5. | -mcp:::pp_eval(...)
6. | -... %>% ...
7. +-dplyr::rename(., :=(!!yvar, !!type))
8. +-dplyr::mutate(...)
9. +-dplyr:::mutate.data.frame(...)
10. | -dplyr:::mutate_cols(.data, dplyr_quosures(...), caller_env = caller_env())
11. | +-base::withCallingHandlers(...)
12. | -mask$eval_all_mutate(quo)
13. +-rlang::exec(...)
14. +-mcp <fn>(...)
15. | -mcp:::assert_numeric(per)
16. -base::.handleSimpleError(...)
17. -dplyr h(simpleError(msg, call))

18. \-rlang::abort(...)
    

[dwwolfson]

Here is the dataframe:
`> str(df)
track_xyt [200 x 8] (S3: track_xyt/track_xy/tbl_df/tbl/data.frame)
$ x_ : num [1:200] 23.7 26.7 13.3 28.5 96.8 ...
$ y_ : num [1:200] -9.06 -11.33 2.95 26.63 121.78 ...
$ t_ : Date[1:200], format: "2000-01-01" "2000-01-01" "2000-01-01" ...
$ nsd : num [1:200] 0 14.1 254.4 1296.6 22454.4 ...
$ speed : num [1:200] 0.001087 0.000437 0.002327 0.002607 0.002199 ...
$ time : num [1:200] 0.18 0.22 0.74 0.88 1.4 1.41 1.48 2 2.15 2.16 ...
$ per : num [1:200] NA -0.001071 0.000105 0.002324 0.002607 ...
$ turning: num [1:200] NA 0.0001808 -0.0004243 -0.0001208 0.0000503 ...

• attr(*, "crs_")= chr NA`

And here is the mcp code:
m1<-list( per~1+ar(1), ~1+sigma(1)+ar(1), ~1+sigma(1)+ar(1), ~1+sigma(1)+ar(1), ~1+sigma(1)+ar(1) ) options(mc.cores=3) f1<-mcp(m1, data = df[,c("time", "per")], par_x = "time", adapt=20000, chains=4, cores=4, inits=list(cp_1=13, cp_2=19, cp_3=31, cp_4=47), prior=list(cp_1="dunif(5,15)", cp_2="dunif(15,25)", cp_3="dunif(25,35)", cp_4="dunif(45,55)") )

[dwwolfson]

Not able to reproduce with the example data. Here is a gist for the csv:
https://gist.github.com/dwwolfson/3291e124879029b52fe1c7412b01b56a

[lindeloev]

Thanks! This is definitely a bug. The post-processing functions should allow for NA values in the dependent variable and they don't. As a temporary workaround, use a dataset without NA in per. E.g., removing these rows or imputing values.

[dwwolfson]

Sounds good, thanks for your help!

[lindeloev]

Reopening to remember to fix this bug in the next release.

[lindeloev]

BTW, just looking at your data, perhaps remove the ar() terms in segment 2+. This will fit a single ar() parameter that apply to all segments. No ar() is implicitly ar(N, 0) where N is the order of the previous segment and "0" means "no change in intercept of the autocorrelation coefficient".