Outcome of test in tests_newborn_outcomes depends on RNG state

[matt-graham]

While reviewing #267 @asif picked up a use of the global NumPy random number generator interface numpy.random rather than the local seeded RandomState in newborn_outcomes

TLOmodel/src/tlo/methods/newborn_outcomes.py

Lines 471 to 472 in 82a0c3b

	birth_weight = np.random.normal(loc=params['mean_birth_weights'][mean_birth_weight_list_location],
	scale=standard_deviation)

After changing this to

self.rng.normal(loc=params['mean_birth_weights'][mean_birth_weight_list_location],                                 
                         scale=standard_deviation)

the test_newborn_hsi_applies_risk_of_complications_and_delivers_treatment_to_facility_births test in test_newborn_outcomes fails at this line

TLOmodel/tests/test_newborn_outcomes.py

Line 433 in 82a0c3b

assert sim.population.props.at[child_id, 'nb_early_onset_neonatal_sepsis']

From what I can tell from digging through the code in newborn_outcomes , this property is set in

TLOmodel/src/tlo/methods/newborn_outcomes.py

Lines 521 to 527 in 82a0c3b

	# The linear model calculates the individuals probability of early_onset_neonatal_sepsis
	if self.eval(params['nb_newborn_equations']['early_onset_neonatal_sepsis'], child_id):
	self.newborn_complication_tracker['early_onset_sepsis'] += 1
	df.at[child_id, 'nb_early_onset_neonatal_sepsis'] = True
	logger.debug(key='message', data=f'Neonate {child_id} has developed early onset sepsis following delivery')

based on the outcome of a call to NewbornOutcomes.eval which uses a linear model to compute the probability for a Bernoulli outcome and then uses the pseudo-random number generator (RNG) to compute the outcome. This means that the value set for the nb_early_onset_neonatal_sepsis is dependent on the state of the RNG which would explain why this test fails when updating to use the module level RNG.

It seems like this might potentially also be related to this previous PR #260

@tamuri @tbhallett @joehcollins - tagging you all as you are all listed as authors on the commit which added the test

[tamuri]

Please take a look at this @joehcollins

[joehcollins]

Hi @matt-graham @tamuri - thanks for catching this. Thought i'd forced nb_early_onset_neonatal_sepsis == True in this test by setting the intercept value of the model to 1 but some lurking treatment effects reduced the probability to below one (but close enough that even a couple of runs with a different seed didnt pick it up)- if that makes any sense.

I've made changes in a separate branch, including the change to self.rng.normal, shall I submit as a PR @tamuri?

[tamuri]

Thanks, Joe. If you let Matt know the branch, he'll merge your changes into this PR.

[joehcollins]

Perfect just running the test file once more

[joehcollins]

@matt-graham - everything seems to be working ok now. The branch is jcollins_rng_fix_may2021. Thanks!

[matt-graham]

Thanks @joehcollins for the quick response and fix. I've merged your branch into my branch for #267. Once that gets merged we can close this issue too.