rep$pdHess : $ operator is invalid for atomic vector — Error when fitting ssm about animotum

[ltyrone]

SOLVED: Thanks for the reply. I am running the same versions as you. Turns out after some digging it was TMB requiring the package Matrix, and version Matrix_1.6-5 in particular. Got there in the end. Could be edges cases, but maybe add Matrix as a dependency when you next update? Thanks again

[jamesgrecian]

Hi @sharkcatch and @tmillac here is a link to help get you started with producing reproducible examples: https://reprex.tidyverse.org

[ianjonsen]

Hi, can you provide a reproducible example - code AND data? Failing that, try adding the following argument to fit_ssm() - this will turn off estimation of the observation error correlation parameter:

map = list(rho_o = factor(NA))

this often helps when fitting to GPS data, but I’m just guessing without a reproducible example.

Ian

[tmillac]

Apologies, will this work?

Data:

> glimpse(df)
Rows: 29
Columns: 4
$ id   <dbl> 2000, 2000, 2000, 2000, 2000, 2000, 2000, 2000, 2000, 2000, 2000, 2000, 2000, 2000, 200…
$ date <dttm> 2021-02-05 05:57:36, 2021-02-05 11:03:06, 2021-02-10 03:54:11, 2021-02-11 16:10:34, 20…
$ lon  <dbl> 170.9615, 170.9776, 170.8871, 171.1078, 171.2163, 171.2587, 171.2142, 171.0744, 171.109…
$ lat  <dbl> -38.16957, -38.20939, -38.18083, -38.09306, -38.09743, -38.04951, -38.01200, -38.03030,…

Code:

fit_df <- 
  fit_ssm(df, # Filtered to just look at one tag
          model = "mp", # Move persistence state space model
          time.step = 6, # Six hour time step
          spdf = TRUE,
          control = ssm_control(verbose = 0))

I tried your previous suggestion but get the following error:

> fit_df <- 
+   fit_ssm(df, # Filtered to just look at one tag
+           model = "mp", # Move persistence state space model
+           time.step = 6, # Six hour time step
+           spdf = TRUE,
+           map = list(rho_o = factor(NA),
+           control = ssm_control(verbose = 0)))
Guessing that all observations are GPS locations.
fitting mp SSM to 1 tracks...
Error in MakeADFun(data, parameters, map = map, random = rnd, hessian = TRUE,  : 
  Names in map must correspond to parameter names

[ianjonsen]

Unfortunately, I can’t access your data that way - glimpse just gives a condensed summary of your data. Easiest would be to use write.csv() to write your data to a .csv file and attach it to your comment using the Paperclip icon - provided your .csv file is not 100’s of Mb.

The error you received when running that code was due to a missing bracket, your fit_ssm() call should be:

fit_ssm(df, # Filtered to just look at one tag
 model = "mp", # Move persistence state space model
 time.step = 6, # Six hour time step
 spdf = TRUE,
 map = list(rho_o = factor(NA)),
 control = ssm_control(verbose = 0))

[tmillac]

github.csv

Hopefully this works! Whoops on the bracket, analysis works with that addition thanks.

[AdriGooden]

Hi Ian

I am getting the same error code running fit_ssm using an 'mp' model. Below is the code I have used but I have also tried the additional code line map = list(rho_o = factor(NA)), as you suggested above
I have attached some of my data.

fit <- fit_ssm(locs, vmax = 3, model = "mp", time.step = 24, control = ssm_control(verbose = 0))

Error in rep$pdHess : $ operator is invalid for atomic vectors

Thank you for your help.

[ltyrone]

Posting here on behalf and it appears @sharkcatch is the OP of the issue that I am raising here. A repex:

df <- structure(list(id = c(3L, 3L, 3L, 3L, 3L, 3L, 3L, 3L, 3L, 3L, 
3L, 3L, 3L, 3L, 3L, 3L, 3L, 3L, 3L, 3L, 3L, 3L, 3L, 3L, 3L, 3L, 
3L, 3L, 3L, 3L, 3L, 3L, 3L, 3L, 3L, 3L, 3L, 3L, 3L, 3L, 3L, 3L, 
3L, 3L, 3L, 3L, 3L, 3L, 3L, 3L, 3L, 3L, 3L, 3L, 3L, 3L, 3L, 3L, 
3L, 3L, 3L, 3L, 3L), date = structure(c(1458088599, 1458088599, 
1458088599, 1458118100, 1458143894, 1458197040, 1458413604, 1470885420, 
1470885420, 1470885420, 1461019843, 1461026270, 1461026270, 1461469583, 
1461469583, 1461469583, 1461469583, 1460269560, 1460269560, 1460269560, 
1460269560, 1460269560, 1460269560, 1460269560, 1460278320, 1460418060, 
1460418060, 1460418060, 1460467980, 1460467980, 1460467980, 1460671891, 
1460671891, 1460841188, 1460841188, 1460841188, 1460841188, 1460870005, 
1460949186, 1461016210, 1461016210, 1461016210, 1461016210, 1461016210, 
1461016210, 1461016210, 1461016210, 1461016210, 1461016210, 1461469583, 
1461469583, 1461469583, 1461469583, 1461469583, 1461469583, 1461469583, 
1461469583, 1461469583, 1462979940, 1462952040, 1462952040, 1462952040, 
1462952040), class = c("POSIXct", "POSIXt"), tzone = "UTC"), 
    lc = c("A", "A", "A", "B", "B", "B", "B", "1", "1", "1", 
    "B", "B", "B", "3", "3", "3", "3", "2", "2", "2", "2", "2", 
    "2", "2", "B", "A", "A", "A", "A", "A", "A", "B", "B", "0", 
    "0", "0", "0", "B", "B", "3", "3", "3", "3", "3", "3", "3", 
    "3", "3", "3", "3", "3", "3", "3", "3", "3", "3", "3", "3", 
    "B", "1", "1", "1", "1"), latitude = c(-31.50391, -31.50391, 
    -31.50391, -39.82062, -39.93783, -40.0844, -41.0345, -25.46735, 
    -25.46735, -25.46735, -40.23962, -40.20943, -40.20943, -40.43101, 
    -40.43101, -40.43101, -40.43101, -40.57273, -40.57273, -40.57273, 
    -40.57273, -40.57273, -40.57273, -40.57273, -40.58898, -40.60004, 
    -40.60004, -40.60004, -40.53807, -40.53807, -40.53807, -40.5596, 
    -40.5596, -40.30304, -40.30304, -40.30304, -40.30304, -40.30302, 
    -40.35905, -40.2418, -40.2418, -40.2418, -40.2418, -40.2418, 
    -40.2418, -40.2418, -40.2418, -40.2418, -40.2418, -40.43101, 
    -40.43101, -40.43101, -40.43101, -40.43101, -40.43101, -40.43101, 
    -40.43101, -40.43101, -40.58412, -40.5991, -40.5991, -40.5991, 
    -40.5991), longitude = c(106.90178, 106.90178, 106.90178, 
    146.76394, 146.81515, 146.89076, 148.13549, 153.41662, 153.41662, 
    153.41662, 147.57244, 147.56543, 147.56543, 147.94571, 147.94571, 
    147.94571, 147.94571, 148.41413, 148.41413, 148.41413, 148.41413, 
    148.41413, 148.41413, 148.41413, 148.4267, 148.32712, 148.32712, 
    148.32712, 148.3033, 148.3033, 148.3033, 148.37321, 148.37321, 
    147.74462, 147.74462, 147.74462, 147.74462, 147.72642, 147.66374, 
    147.57622, 147.57622, 147.57622, 147.57622, 147.57622, 147.57622, 
    147.57622, 147.57622, 147.57622, 147.57622, 147.94571, 147.94571, 
    147.94571, 147.94571, 147.94571, 147.94571, 147.94571, 147.94571, 
    147.94571, 148.49763, 148.50535, 148.50535, 148.50535, 148.50535
    )), row.names = c(NA, -63L), class = "data.frame")

Pre-model, NOTE that the cols were called longitude" and "latitude" prior to running this line, and they get converted to "lon" and "lat" after running this line:

df <- format_data(df, date = "date", coord = c("longitude", "latitude"))

The model:

fit <- fit_ssm(df,
               vmax = 3,
               model = "mp",
               time.step = 24,
               control = ssm_control(verbose = 0))

The error:

Error in rep$pdHess : $ operator is invalid for atomic vectors

As per your suggestion, I also tried:

fit <- fit_ssm(df,
               vmax = 3,
               model = "mp",
               time.step = 24,
               spdf = TRUE,
               map = list(rho_o = factor(NA)),
               control = ssm_control(verbose = 0))

but I get the same error.

If I convert the lon lat columns to EPSG:3857 the aforementioned error does not persist, but I get this:

Warning in pf_sda_filter(x, spdf, vmax, ang, distlim) :
  
trip::sda produced an error on id 3 using trip::speedfilter instead
Error in approx(x = select(d, date, x), xout = d.all$date, rule = 2) : 
  need at least two non-NA values to interpolate

I also tried with the projected data without running format_data() but still not working.

[ianjonsen]

Thank you for the reprex. Unfortunately, I can't reproduce your error so that makes me wonder what version of aniMotum and TMB you are using? The latest aniMotum version is 1.2-03.

Running aniMotum 1.2-03 & TMB 1.9.10 with R 4.3.2 on MacOS:

fit <- fit_ssm(df,
               vmax = 3, 
               model = "mp",
               time.step = 24, control = ssm_control(verbose = 0))

I get the following warning:

Warning messages:
1: In sqrt(as.numeric(object$diag.cov.random)) : NaNs produced
2: Hessian was not positive-definite so some standard errors could not be calculated.

This indicates that SE's could not be calculated for some model parameters and/or random variables. In this case, the model parameters all have SE's, though rho_p is close to -1 and its SE is very small which is suggestive of an issue estimating the bivariate process errors that describe variability in the movements:

summary(fit)

 Animal id Model Time n.obs n.filt n.fit n.pred n.rr converged  AICc
         3    mp   24    63     45    18    150    .      TRUE 522.3
print individual model parameter tables? [enter / esc to stop]

--------------
3 
--------------
 Parameter Estimate Std.Err
     rho_p  -0.9994   5e-04
   sigma_x  81.6253 23.4401
   sigma_y  20.7539  5.9069
     rho_o  -0.6713  1.9217
     tau_x   0.1126  0.4798
     tau_y   0.0661  0.1625
   sigma_g   0.1136  0.0785

But looking at the fitted location estimates, some SE's for the estimated x & y coordinates could not be calculated:

grab(fit, "f")

# A tibble: 18 × 11
   id    date                  lon   lat      x      y    x.se     y.se logit_g logit_g.se     g
   <chr> <dttm>              <dbl> <dbl>  <dbl>  <dbl>   <dbl>    <dbl>   <dbl>      <dbl> <dbl>
 1 3     2016-03-16 00:36:39  107. -31.5 11900. -3676.   2.69    1.46     2.20        1.71 0.900
 2 3     2016-03-16 08:48:20  147. -39.8 16338. -4813. NaN       1.90     2.20        1.44 0.900
 3 3     2016-03-17 06:44:00  147. -40.1 16352. -4850.   3.38    2.03    -0.230       1.02 0.443
 4 3     2016-03-19 18:53:24  148. -41.0 16490. -4989.   4.92    2.14     1.74        2.08 0.851
 5 3     2016-04-10 06:26:00  148. -40.6 16521. -4922.   0.173   0.0852   1.58        3.92 0.829
 6 3     2016-04-10 08:52:00  148. -40.6 16523. -4923. NaN     NaN        1.58        3.91 0.829
 7 3     2016-04-11 23:41:00  148. -40.6 16512. -4926.   2.69    1.45     1.41        3.80 0.804
 8 3     2016-04-12 13:33:00  148. -40.5 16509. -4917.   2.69    1.45     1.44        3.62 0.808
 9 3     2016-04-16 21:13:08  148. -40.3 16447. -4882.   1.54    0.985    1.28        3.40 0.783
10 3     2016-04-17 05:13:25  148. -40.3 16445. -4883.   1.82    1.76     1.28        3.37 0.783
11 3     2016-04-18 03:13:06  148. -40.4 16438. -4890.   2.66    1.94     1.30        2.97 0.785
12 3     2016-04-18 21:50:10  148. -40.2 16428. -4874.   0.113   0.0661   1.58        1.99 0.830
13 3     2016-04-18 22:50:43  148. -40.2 16427. -4872. NaN     NaN        1.59        1.99 0.830
14 3     2016-04-19 00:37:50  148. -40.2 16427. -4870. NaN     NaN        1.58        1.99 0.830
15 3     2016-04-24 03:46:23  148. -40.4 16469. -4901.   0.113   0.0661   1.57        4.19 0.828
16 3     2016-05-11 07:34:00  149. -40.6 16532. -4926.   0.419   0.169    3.41        8.00 0.968
17 3     2016-05-11 15:19:00  148. -40.6 16531. -4923.   4.98    2.14     3.42        8.00 0.968
18 3     2016-08-11 03:17:00  153. -25.5 17078. -2915.   0.419   0.169    6.06       19.0  0.998

This is an indication that there's an issue with the data, and a look at the model fit shows that the data are highly sporadic with huge data gaps relative to the 24-h prediction interval:

plot(fit, "p")

Highly sporadic data like this are far less than ideal for fitting any SSM and especially the mp model because you are asking the model to do far too much interpolation of locations and this in turn affects the move persistence estimates. It's difficult to provide more informative error and warning messages, but I can try to improve them so that you get a better idea of why the model is not fitting as expected. As I've tried to show in this reply, some exploration of the data and model fit(s) usually help identify the problem.