This should not happen!
mode.respy.txt
est.respy.log
est.respy.txt

We need a section in the documentation that illustrates how to actually estimate a model on simulated dataset using the fit() function that emphasizes the parts of the initialization file that are relevant here but were not already discussed for the simulation step. For example, optimizer options, scaling, derviatives , etc. Please feel free to reach out to @janosg for questions in the process

... just for completeness, it allows for easier creation of some tables in paper and manuscript.

How to think about the fixed covariance matrix in light of mean shifts. Maybe it would be better to fix the correlation structure?

The following init file only generates type 3 draws in a simulation as the intercepts, that deterime the ordering, are all identical equal to zero. full.respy.txt

Starting state space creation

... finished

Starting calculation of systematic rewards

... finished

Starting backward induction procedure

**... solving period 48 with ***** states

... solving period 47 with ***** states**

In rare instances, the value of the criterion function might be too large and thus printed as a string. This occurred in the past, when the gradient preconditioning had zero probability observations. We now generate random initialization files with smaller gradient step sizes ... We need to revisit this and handle this case more explicitly by adding a minimum/maximum value. If this happens at the gradient calculation at the start values, then the scaling results in nonsense. We need to add a maximum value in addition to the minimum value. So, that a -HUGE log likelihood can still be processed properly.

The simulated dataset contains a variable WAGE_INTERNAL that writes out the full wage and not just two two digit precision as WAGE. This is useful for unit testing some relationships in the data. However, I would prefer to write out the WAGE variable just to full precision in the simulation case. This breaks the release tests as the likelihood of identical samples is slightly altered.

merge coeffs_a and coeffs_b to allow for loops in codes

Add the use of scripts

The set of admissible actions needs to be extended, very restrictive at the moment.

Revisit package for compliance https://www.python.org/dev/peps/pep-0257/

We need authorship information in the respy.init following this convention.

In general I want to turn off warnings if if production compiled.

respy/tests/test_integration.py::TestClass::()::test_15
/home/peisenha/ownCloud/office/workspace/software/repositories/organizations/restudToolbox/package/respy/clsRespy.py:266: ResourceWarning: unclosed file <_io.BufferedWriter name='solution.respy.pkl'>
pkl.dump(self, open(file_name, 'wb'))
/home/peisenha/ownCloud/office/workspace/software/repositories/organizations/restudToolbox/package/respy/clsRespy.py:266: ResourceWarning: unclosed file <_io.BufferedWriter name='solution.respy.pkl'>
pkl.dump(self, open(file_name, 'wb'))
/home/peisenha/ownCloud/office/workspace/software/repositories/organizations/restudToolbox/package/respy/clsRespy.py:266: ResourceWarning: unclosed file <_io.BufferedWriter name='solution.respy.pkl'>
pkl.dump(self, open(file_name, 'wb'))
/home/peisenha/ownCloud/office/workspace/software/repositories/organizations/restudToolbox/package/respy/clsRespy.py:266: ResourceWarning: unclosed file <_io.BufferedWriter name='solution.respy.pkl'>
pkl.dump(self, open(file_name, 'wb'))

-- Docs: http://doc.pytest.org/en/latest/warnings.html

• I think the covariate any_exp is calculated wrong. Why the constraint that exp > 0?
• I need to extend the code to allow for initial conditon actiity lagged not equal to schooling in the last period.
• Is the wage equation always properly calculated, I am worried about period -1. The code seems not to be consistent in iterating over period starting at 0 or 1.
  • It seems that there is a low-impact bug regarding periods treatment in line 158 of simulate_fortran.f90. There I should pass in periods + 1, right?

I would like that the debug flag also has consequences for the types of warnings for example that we show, i.e. only errors if not debug otherwise warnings as well.

This needs to be removed once the manuscript is published.

Add count for intial schooling shares, this allows to easier modification of initalization files

There seems to be a problem with the type counts, there are too many shifts specified compared to the shares.

Can I allow to pass in shares that do not sum to one? This would make it easier to iteratively add initial schooling levels by coping the initial schooling shares in the simulated dataset.

The order of education initial conditions is maintained. So, when the original user specifies for 10 years first then for 11 and then for 9 it would be nice if these were ordered.

Please add a summary in data.respy.info after simulation. Now, I always need to go back to the the very large ambiguity log file.

The processing of an initialization file fails when there is a space between the value and the sign of a coefficient.

property.respy.txt

The f2py tests seem to fail for ill-conditioned problems only.
The interpolation test takes very long to run, also, the interpolation tests should be refactored to also include the flag_ambiguity=True if called from test_ambiguity explicitly.

I should construct all derived information such as lagged activity and experiences simply inside the program ... I can still output it for simulated datasets for debugging and testing purposes.

/home/peisenha/.envs/2.0.0.dev16/local/lib/python2.7/site-packages/respy/python/process/process_auxiliary.py:60: FutureWarning: 'Identifier' is both a column name and an index level.
Defaulting to column but this will raise an ambiguity error in a future version

fort_evaluate.f90, num_rows seems not to be used and can thus be removed

Would be nice to have a more formal process, pickled storage of failed seed which can be removed when investigated etc. Hypothesis Package for Property-based Testing, contribute hour runs as open source contribution. There seems to be a lot of overlap with my needs for the test battery with random parameterizations. Also, I should think about using more and more pzytest features, such as skipif properties and such in my own run/property test setup. At the moment, I remove the test_parallels if not IS_Parallel available.

There were some issues for further investigation, that were all solved.

Mention Amdahls Law in scalability exercises, https://en.wikipedia.org/wiki/Amdahl%27s_law, nice graph on slide 4 at http://insidehpc.com/2016/10/parallel-programming-efficiency/

We need to add a new CHANGELOG file following the convention here. Please add one just with the header information

This concept does only make sense in the case of a single type, otherwise the likelihood is never one ... This does not make any sense either anymore:

            #If there is no random variation in rewards, then the observed wages need to be
            # identical their systematic components. The discrepancy between the observed
            # wages and their systematic components might be small due to the reading in of
            # the dataset (FORTRAN only).

Related to #70

https://launchbylunch.com/posts/2014/Feb/16/sql-naming-conventions/

It would be nice if we could also test the parallel implementation on TRAVIS-CI. This requires installation of MPICH. Also, we want to add a build matrix.

Add ability to be able to simulate from new points in est.respy.info without updating the model.respy.ini

The naming of preconditioning/scaling is not properly done throughout the code. For example, scaling now refers to performance exercise and preconditioning to the preconditioning of the parameters of the criterion function. However, the function doing the recording is still named record_estimation_scaling.

This eases reliability testing.

The SLSQP implementation involves intentional zero divisions which hinders the use of ffpe-trap=invalid as an option for the debug compilation. Maybe we can disable this for one particular file or find a better FORTRAN implementation.

Traceback (most recent call last):
File "/home/eisenhauer/.envs/structAmbiguity/bin/respy-modify", line 190, in
scripts_modify(*args)
File "/home/eisenhauer/.envs/structAmbiguity/bin/respy-modify", line 84, in scripts_modify
respy_obj = RespyCls(init_file)
File "/home/eisenhauer/restudToolbox/package/respy/clsRespy.py", line 147, in init
self.lock()
File "/home/eisenhauer/restudToolbox/package/respy/clsRespy.py", line 199, in lock
self._check_integrity_attributes()
File "/home/eisenhauer/restudToolbox/package/respy/clsRespy.py", line 737, in _check_integrity_attributes
np.testing.assert_almost_equal(np.sum(edu_spec['share']), 1.0)
File "/home/eisenhauer/.envs/structAmbiguity/lib/python2.7/site-packages/numpy/testing/utils.py", line 539, in assert_almost_equal
raise AssertionError(_build_err_msg())
AssertionError:
Arrays are not almost equal to 7 decimals
ACTUAL: 1.0000099999999998
DESIRED: 1.0

The test is too strict it might fail when reading in valid shares.

If I have an init file with all parameters either fixed or with bounds, then this throws an error. It should not if at least one parameter is free, even with bounds, for estimation.

Add ability to reactivate a evaluation in est.respy.log for est.respy.info to work with from there.

I would like to have a stricter sharing of global variables within the program unit. The variables required for the evaluation of the criterion functions should only be accessible there and not everywhere else.

We allow to run estimations on less or equal the number of estimation agents specified in the init file to ease adding iteratively different initial conditions, etc. However, this case in not part of the random init file generation. This needs to be explicitly included.

This is wrong if individuals enroll in school the very last period. We are only using the information on the schooling when they enter the last period.... This is a little more tricky than calculating average occupation experience for example as individuals differ in their level of initial schooling.

_ indicates a protected folder .

Why use the TRANSFER function and not simple DBLE() or DINT(). If there is not reason, remove use of function.

The automatic vectorization of the states within period calculation is not used due to the IF statement that checks whether we are in the last period. This can be fixed by adding compiler pragmas.