Describe the bug
When the acceptance rate is too high the auto-adjusting subject-specific epsilon can reach very high levels to try and reduce acceptance rate.
Example Output
[1] "Epsilon for subject 110 is very high:"
id: 104311
eps: 550.0075
The very large epsilon results in extreme values in particles drawn from the individual's last winning particles.
That is, new particles are either from the group level, individual last estimate or efficient proposal distribution mean. For the first and last the covariance used to generate new particles is either the group level covariance or the efficient proposal distribution covariance. For the middle (individual last estimate) the covariance used for draws is the group level covariance multiplied by the individual subject specific epsilon squared. So in the above case the variance for the individual particles draw is likely to be extremely high.
Describe the bug When the acceptance rate is too high the auto-adjusting subject-specific epsilon can reach very high levels to try and reduce acceptance rate.
Example Output
The very large epsilon results in extreme values in particles drawn from the individual's last winning particles.
That is, new particles are either from the group level, individual last estimate or efficient proposal distribution mean. For the first and last the covariance used to generate new particles is either the group level covariance or the efficient proposal distribution covariance. For the middle (individual last estimate) the covariance used for draws is the group level covariance multiplied by the individual subject specific epsilon squared. So in the above case the variance for the individual particles draw is likely to be extremely high.