Strictness

Strictness criteria for model selection can be specified in the AMD tools. Models that do not fulfill the strictness criteria will be excluded from the model ranking and will therefore not be able to be selected as best models. The strictness argument in the AMD tools consists of a string of logically arranged criteria. Implemented strictness criteria are:

Strictness criterion	Type	Description
minimization_successful	Boolean	True if minimization was successful
rounding_errors	Boolean	True if minimization terminated due rounding errors
sigdigs	Numeric	Number of significant digits
maxevals_exceeded	Boolean	True if minimization terminated due maximum evaluations exceeded.
rse	Numeric	Relative standard errors of the parameters.
rse_theta	Numeric	Relative standard errors of the theta parameters.
rse_omega	Numeric	Relative standard errors of the omega parameters.
rse_sigma	Numeric	Relative standard errors of the sigma parameters.
condition_number	Numeric	Condition number of the covariance matrix
final_zero_gradient	Boolean	True if at least one parameter has a final zero gradient or if final gradient is nan
final_zero_gradient_theta	Boolean	True if at least one theta parameter has a final zero gradient or if final gradient is nan
final_zero_gradient_omega	Boolean	True if at least one omega parameter has a final zero gradient or if final gradient is nan
final_zero_gradient_sigma	Boolean	True if at least one sigma parameter has a final zero gradient or if gradient gradient is nan
estimate_near_boundary	Boolean	True if at least one parameter estimate is near its boundary (maximum distance to 0 = 0.001, maximum distance to non-zero bound = 2 significant digits
estimate_near_boundary_theta	Boolean	True if at least one theta parameter estimate is near its boundary (maximum distance to 0 = 0.001, maximum distance to non-zero bound = 2 significant digits
estimate_near_boundary_omega	Boolean	True if at least one omega parameter estimate is near its boundary (maximum distance to 0 = 0.001, maximum distance to non-zero bound = 2 significant digits
estimate_near_boundary_sigma	Boolean	True if at least one sigma parameter estimate is near its boundary (maximum distance to 0 = 0.001, maximum distance to non-zero bound = 2 significant digits

The strictness criteria can be arranged logically, e.g.:

"(A or B) and C < n"

where n is a number and A, B and C are strictness criteria.

Allowed logical operators are: and, or, not, <, <=, ==, >, >=, !=.

If the statement evaluates to False then the strictness criteria are not fulfilled and the model will be excluded from the results.

Examples

strictness = "minimization_successful or (rounding_errors and sigdigs >= 0.1)"

In this example the strictness criteria states that either the minimization must be successful or else the minimization was terminated due to rounding errors and the number of sigdigs is greater or equal to 0.1. This example is the default strictness criterion for AMD and all subtools.

strictness = "minimization_successful and rse < 0.4"

This means that the minimization must be successful and that all parameters must have an RSE smaller than 0.4.