Standard errors as NaN in estimated HMM models

[xjtusjtu]

Hi Brett,

I estimated a HMM model but noticed some of the parameter estimates have standard errors (SEs) as NaN.

I wanted to ask:

1. How can I interpret NaN SEs for estimated parameters? What are the possible reasons?
2. Are there any methods to avoid this issue?

Some more details:

> mix2_1$CIbeta$delta
$est
                                            active
(Intercept)_mix1                       -13.9085573
I(cumsum_voteup + Wcumsum_voteup)_mix1  19.9623148
Wcumsum_voteup_mix1                     -6.0683358
(Intercept)_mix2                        -0.7539192
I(cumsum_voteup + Wcumsum_voteup)_mix2   8.6506310
Wcumsum_voteup_mix2                     -1.4783231

$se
                                           active
(Intercept)_mix1                       0.12183164
I(cumsum_voteup + Wcumsum_voteup)_mix1 0.70749799
Wcumsum_voteup_mix1                    1.48762388
(Intercept)_mix2                              NaN
I(cumsum_voteup + Wcumsum_voteup)_mix2        NaN
Wcumsum_voteup_mix2                    0.06691746

> mix2_1$CIbeta$answer
$est
     mu_1:(Intercept) mu_1:Wvoteup mu_1:cumsum_voteup mu_1:Wcumsum_voteup mu_1:mScore mu_2:(Intercept) mu_2:Wvoteup
[1,]        -6.230847     1.077521           2.839896          -0.3516859    3.655003         4.729057    0.6009367
     mu_2:cumsum_voteup mu_2:Wcumsum_voteup mu_2:mScore size_1:(Intercept) size_1:cumsum_voteup size_2:(Intercept)
[1,]        -0.08480876          -0.4721896 0.008675344          -1.841169            -5.979677         -0.2832883
     size_2:cumsum_voteup
[1,]           0.01745645

$se
     mu_1:(Intercept) mu_1:Wvoteup mu_1:cumsum_voteup mu_1:Wcumsum_voteup mu_1:mScore mu_2:(Intercept) mu_2:Wvoteup
[1,]         1.378275    0.2859571           1.065087           0.2131327         NaN       0.04573145  0.006273422
     mu_2:cumsum_voteup mu_2:Wcumsum_voteup mu_2:mScore size_1:(Intercept) size_1:cumsum_voteup size_2:(Intercept)
[1,]        0.009873569         0.006725937   0.2099942           1.714647             2.515318         0.01051128
     size_2:cumsum_voteup
[1,]           0.01006015

I'd appreciate any guidance on interpreting these results or tweaking the estimation to avoid NaN SEs. Please let me know if any other info would be helpful here.

Thanks in advance for your help!

[bmcclintock]

This can be a common problem with complex (and overparameterized) models. If you google "R NAN standard errors hessian" or something similar you can find some recommendations for diagnosing and/or fixing the issue. For example:

https://openmx.ssri.psu.edu/thread/1393

https://r-help.stat.math.ethz.narkive.com/sIRlkDOH/r-zeroinfl-problem-cannot-get-standard-errors-hessian-has-nan

[xjtusjtu]

Thanks for your help!