def branch_or_leaf(dag: LocationDAG, location_id: int, sex: int, model_version_id: int,
parent_location: int, parent_sex: int,
n_sim: int, n_pool: int, upstream: List[str], tasks: List[_CascadeOperation]):
"""
Recursive function that either creates a branch (by calling itself) or a leaf fit depending
on whether or not it is at a terminal node. Determines if it's at a terminal node using
the dag.successors() method from networkx. Appends tasks onto the tasks parameter.
"""
if not dag.is_leaf(location_id=location_id):
branch = branch_fit(
model_version_id=model_version_id,
location_id=location_id, sex_id=sex,
prior_parent=parent_location, prior_sex=parent_sex,
child_locations=dag.children(location_id), child_sexes=[sex],
n_sim=n_sim, n_pool=n_pool,
upstream_commands=upstream,
ode_fit_strategy=True
)
tasks += branch
for location in dag.children(location_id):
branch_or_leaf(dag=dag, location_id=location, sex=sex, model_version_id=model_version_id,
parent_location=location_id, parent_sex=sex,
n_sim=n_sim, n_pool=n_pool, upstream=[branch[-1].command], tasks=tasks)
else:
leaf = leaf_fit(
model_version_id=model_version_id,
location_id=location_id,
sex_id=sex,
prior_parent=parent_location,
prior_sex=parent_sex,
n_sim=n_sim, n_pool=n_pool,
upstream_commands=upstream,
ode_fit_strategy=True
)
tasks += leaf
class MeasurementInputs:
def __init__(self, model_version_id: int,
gbd_round_id: int, decomp_step_id: int,
conn_def: str,
country_covariate_id: List[int],
csmr_cause_id: int, crosswalk_version_id: int,
csmr_process_version_id: Optional[int] = None,
location_set_version_id: Optional[int] = None,
drill_location_start: Optional[int] = None,
drill_location_end: Optional[List[int]] = None):
"""
The class that constructs all of the measurement inputs. Pulls ASDR,
CSMR, crosswalk versions, and country covariates, and puts them into
one data frame that then formats itself for the dismod database.
Performs covariate value interpolation if age and year ranges
don't match up with GBD age and year ranges.
Parameters
----------
model_version_id
the model version ID
gbd_round_id
the GBD round ID
decomp_step_id
the decomp step ID
csmr_process_version_id
process version ID for CSMR
csmr_cause_id: (int) cause to pull CSMR from
crosswalk_version_id
crosswalk version to use
country_covariate_id
list of covariate IDs
conn_def
connection definition from .odbc file (e.g. 'epi') to connect to the IHME databases
location_set_version_id
can be None, if it's none, get the best location_set_version_id for estimation hierarchy of this GBD round
drill_location_start
which location ID to drill from as the parent
drill_location_end
which immediate children of the drill_location_start parent to include in the drill
Attributes
----------
self.decomp_step : str
the decomp step in string form
self.demographics : cascade_at.inputs.demographics.Demographics
a demographics object that specifies the age group, sex,
location, and year IDs to grab
self.integrand_map : Dict[int, int]
dictionary mapping from GBD measure IDs to DisMod IDs
self.asdr : cascade_at.inputs.asdr.ASDR
all-cause mortality input object
self.csmr : cascade_at.inputs.csmr.CSMR
cause-specific mortality input object from cause csmr_cause_id
self.data : cascade_at.inputs.data.CrosswalkVersion
crosswalk version data from IHME database
self.covariate_data : List[cascade_at.inputs.covariate_data.CovariateData]
list of covariate data objects that contains the raw covariate data mapped to IDs
self.location_dag : cascade_at.inputs.locations.LocationDAG
DAG of locations to be used
self.population: (cascade_at.inputs.population.Population)
population object that is used for covariate weighting
self.data_eta: (Dict[str, float]): dictionary of eta value to be
applied to each measure
self.density: (Dict[str, str]): dictionary of density to be
applied to each measure
self.nu: (Dict[str, float]): dictionary of nu value to be applied
to each measure
self.dismod_data: (pd.DataFrame) resulting dismod data formatted
to be used in the dismod database
Examples
--------
>>> from cascade_at.settings.base_case import BASE_CASE
>>> from cascade_at.settings.settings import load_settings
>>>
>>> settings = load_settings(BASE_CASE)
>>> covariate_id = [i.country_covariate_id for i in settings.country_covariate]
>>>
>>> i = MeasurementInputs(
>>> model_version_id=settings.model.model_version_id,
>>> gbd_round_id=settings.gbd_round_id,
>>> decomp_step_id=settings.model.decomp_step_id,
>>> csmr_process_version_id=None,
>>> csmr_cause_id = settings.model.add_csmr_cause,
>>> crosswalk_version_id=settings.model.crosswalk_version_id,
>>> country_covariate_id=covariate_id,
>>> conn_def='epi',
>>> location_set_version_id=settings.location_set_version_id
>>> )
>>> i.get_raw_inputs()
>>> i.configure_inputs_for_dismod(settings)
"""
LOG.info(f"Initializing input object for model version ID {model_version_id}.")
LOG.info(f"GBD Round ID {gbd_round_id}.")
LOG.info(f"Pulling from connection {conn_def}.")
self.model_version_id = model_version_id
self.gbd_round_id = gbd_round_id
self.decomp_step_id = decomp_step_id
self.csmr_process_version_id = csmr_process_version_id
self.csmr_cause_id = csmr_cause_id
self.crosswalk_version_id = crosswalk_version_id
self.country_covariate_id = country_covariate_id
self.conn_def = conn_def
self.drill_location_start = drill_location_start
self.drill_location_end = drill_location_end
self.decomp_step = ds.decomp_step_from_decomp_step_id(self.decomp_step_id)
if location_set_version_id is None:
self.location_set_version_id = get_location_set_version_id(gbd_round_id=self.gbd_round_id)
else:
self.location_set_version_id = location_set_version_id
self.demographics = Demographics(
gbd_round_id=self.gbd_round_id,
location_set_version_id=self.location_set_version_id)
self.location_dag = LocationDAG(
location_set_version_id=self.location_set_version_id,
gbd_round_id=self.gbd_round_id
)
# Need to subset the locations to only those needed for
# the drill. drill_locations_all is the set of locations
# to pull data for, including all descendents. drill_locations
# is the set of locations just parent-children in the drill.
drill_locations_all, drill_locations = locations_by_drill(
drill_location_start=self.drill_location_start,
drill_location_end=self.drill_location_end,
dag=self.location_dag
)
if drill_locations_all:
self.demographics.location_id = drill_locations_all
self.demographics.drill_locations = drill_locations
self.exclude_outliers = True
self.asdr = None
self.csmr = None
self.population = None
self.data = None
self.covariates = None
self.age_groups = None
self.data_eta = None
self.density = None
self.nu = None
self.measures_to_exclude = None
self.dismod_data = None
self.covariate_data = None
self.country_covariate_data = None
self.covariate_specs = None
self.omega = None
def get_raw_inputs(self):
"""
Get the raw inputs that need to be used
in the modeling.
"""
LOG.info("Getting all raw inputs.")
self.asdr = ASDR(
demographics=self.demographics,
decomp_step=self.decomp_step,
gbd_round_id=self.gbd_round_id
).get_raw()
self.csmr = CSMR(
cause_id=self.csmr_cause_id,
demographics=self.demographics,
decomp_step=self.decomp_step,
gbd_round_id=self.gbd_round_id,
process_version_id=self.csmr_process_version_id
).get_raw()
self.data = CrosswalkVersion(
crosswalk_version_id=self.crosswalk_version_id,
exclude_outliers=self.exclude_outliers,
demographics=self.demographics,
conn_def=self.conn_def,
gbd_round_id=self.gbd_round_id
).get_raw()
self.covariate_data = [CovariateData(
covariate_id=c,
demographics=self.demographics,
decomp_step=self.decomp_step,
gbd_round_id=self.gbd_round_id
).get_raw() for c in self.country_covariate_id]
self.population = Population(
demographics=self.demographics,
decomp_step=self.decomp_step,
gbd_round_id=self.gbd_round_id
).get_population()
def configure_inputs_for_dismod(self, settings: SettingsConfig,
midpoint: bool = False,
mortality_year_reduction: int = 5):
"""
Modifies the inputs for DisMod based on model-specific settings.
Arguments
---------
settings
Settings for the model
mortality_year_reduction
number of years to decimate csmr and asdr
"""
self.data_eta = data_eta_from_settings(settings)
self.density = density_from_settings(settings)
self.nu = nu_from_settings(settings)
self.measures_to_exclude = measures_to_exclude_from_settings(settings)
# If we are constraining omega, then we want to hold out the data
# from the DisMod fit for ASDR (but never CSMR -- always want to fit
# CSMR).
data = self.data.configure_for_dismod(
measures_to_exclude=self.measures_to_exclude,
relabel_incidence=settings.model.relabel_incidence,
midpoint=midpoint,
)
asdr = self.asdr.configure_for_dismod(
hold_out=settings.model.constrain_omega)
csmr = self.csmr.configure_for_dismod(hold_out=0)
if settings.model.constrain_omega:
self.omega = calculate_omega(asdr=asdr, csmr=csmr)
else:
self.omega = None
if not csmr.empty:
csmr = decimate_years(
data=csmr, num_years=mortality_year_reduction)
if not asdr.empty:
asdr = decimate_years(
data=asdr, num_years=mortality_year_reduction)
self.dismod_data = pd.concat([data, asdr, csmr], axis=0, sort=True)
self.dismod_data.reset_index(drop=True, inplace=True)
self.dismod_data["density"] = self.dismod_data.measure.apply(
self.density.__getitem__)
self.dismod_data["eta"] = self.dismod_data.measure.apply(
self.data_eta.__getitem__)
self.dismod_data["nu"] = self.dismod_data.measure.apply(
self.nu.__getitem__)
# This makes the specs not just for the country covariate but adds on
# the sex and one covariates.
self.covariate_specs = CovariateSpecs(
country_covariates=settings.country_covariate,
study_covariates=settings.study_covariate
)
self.country_covariate_data = {c.covariate_id: c.configure_for_dismod(
pop_df=self.population.configure_for_dismod(),
loc_df=self.location_dag.df
) for c in self.covariate_data}
self.dismod_data = self.add_covariates_to_data(df=self.dismod_data)
self.dismod_data.loc[
self.dismod_data.hold_out.isnull(), 'hold_out'] = 0.
self.dismod_data.drop(['age_group_id'], inplace=True, axis=1)
return self
def prune_mortality_data(self, parent_location_id: int) -> pd.DataFrame:
"""
Remove mortality data for descendents that are not children of parent_location_id
from the configured dismod data before it gets filled into the dismod database.
"""
df = self.dismod_data.copy()
direct_children = self.location_dag.parent_children(parent_location_id)
direct_children = df.location_id.isin(direct_children)
mortality_measures = df.measure.isin([
IntegrandEnum.mtall.name, IntegrandEnum.mtspecific.name
])
remove_rows = ~direct_children & mortality_measures
df = df.loc[~remove_rows].copy()
return df
def add_covariates_to_data(self, df: pd.DataFrame) -> pd.DataFrame:
"""
Add on covariates to a data frame that has age_group_id, year_id
or time-age upper / lower, and location_id and sex_id. Adds both
country-level and study-level covariates.
"""
cov_dict_for_interpolation = {
c.name: self.country_covariate_data[c.covariate_id]
for c in self.covariate_specs.covariate_specs
if c.study_country == 'country'
}
df = self.interpolate_country_covariate_values(
df=df, cov_dict=cov_dict_for_interpolation)
df = self.transform_country_covariates(df=df)
df['s_sex'] = df.sex_id.map(
SEX_ID_TO_NAME).map(StudyCovConstants.SEX_COV_VALUE_MAP)
df['s_one'] = StudyCovConstants.ONE_COV_VALUE
return df
def to_gbd_avgint(self, parent_location_id: int, sex_id: int) -> pd.DataFrame:
"""
Converts the demographics of the model to the avgint table.
"""
LOG.info(f"Getting grid for the avgint table "
f"for parent location ID {parent_location_id} "
f"and sex_id {sex_id}.")
if self.drill_location_start is not None:
locations = self.demographics.drill_locations
else:
locations = self.location_dag.parent_children(parent_location_id)
grid = expand_grid({
'sex_id': [sex_id],
'location_id': locations,
'year_id': self.demographics.year_id,
'age_group_id': self.demographics.age_group_id
})
grid['time_lower'] = grid['year_id'].astype(int)
grid['time_upper'] = grid['year_id'] + 1.
grid = BaseInput(
gbd_round_id=self.gbd_round_id).convert_to_age_lower_upper(df=grid)
LOG.info("Adding covariates to avgint grid.")
grid = self.add_covariates_to_data(df=grid)
return grid
def interpolate_country_covariate_values(self, df: pd.DataFrame, cov_dict: Dict[Union[float, str], pd.DataFrame]):
"""
Interpolates the covariate values onto the data
so that the non-standard ages and years match up to meaningful
covariate values.
"""
LOG.info(f"Interpolating and merging the country covariates.")
interp_df = get_interpolated_covariate_values(
data_df=df,
covariate_dict=cov_dict,
population_df=self.population.configure_for_dismod()
)
return interp_df
def transform_country_covariates(self, df):
"""
Transforms the covariate data with the transformation ID.
:param df: (pd.DataFrame)
:return: self
"""
for c in self.covariate_specs.covariate_specs:
if c.study_country == 'country':
LOG.info(f"Transforming the data for country covariate "
f"{c.covariate_id}.")
df[c.name] = df[c.name].apply(
lambda x: COVARIATE_TRANSFORMS[c.transformation_id](x)
)
return df
def calculate_country_covariate_reference_values(
self, parent_location_id: int, sex_id: int) -> CovariateSpecs:
"""
Gets the country covariate reference value for a covariate ID and a
parent location ID. Also gets the maximum difference between the
reference value and covariate values observed.
Run this when you're going to make a DisMod AT database for a specific
parent location and sex ID.
:param: (int)
:param parent_location_id: (int)
:param sex_id: (int)
:return: List[CovariateSpec] list of the covariate specs with the
correct reference values and max diff.
"""
covariate_specs = copy(self.covariate_specs)
age_min = self.dismod_data.age_lower.min()
age_max = self.dismod_data.age_upper.max()
time_min = self.dismod_data.time_lower.min()
time_max = self.dismod_data.time_upper.max()
children = self.location_dag.children(parent_location_id)
for c in covariate_specs.covariate_specs:
if c.study_country == 'study':
if c.name == 's_sex':
c.reference = StudyCovConstants.SEX_COV_VALUE_MAP[
SEX_ID_TO_NAME[sex_id]]
c.max_difference = StudyCovConstants.MAX_DIFFERENCE_SEX_COV
elif c.name == 's_one':
c.reference = StudyCovConstants.ONE_COV_VALUE
c.max_difference = StudyCovConstants.MAX_DIFFERENCE_ONE_COV
else:
raise ValueError(f"The only two study covariates allowed are sex and one, you tried {c.name}.")
elif c.study_country == 'country':
LOG.info(f"Calculating the reference and max difference for country covariate {c.covariate_id}.")
cov_df = self.country_covariate_data[c.covariate_id]
parent_df = (
cov_df.loc[cov_df.location_id == parent_location_id].copy()
)
child_df = cov_df.loc[cov_df.location_id.isin(children)].copy()
all_loc_df = pd.concat([child_df, parent_df], axis=0)
# if there is no data for the parent location at all (which
# there should be provided by Central Comp)
# then we are going to set the reference value to 0.
if cov_df.empty:
reference_value = 0
max_difference = np.nan
else:
pop_df = self.population.configure_for_dismod()
pop_df = (
pop_df.loc[pop_df.location_id == parent_location_id].copy()
)
df_to_interp = pd.DataFrame({
'location_id': parent_location_id,
'sex_id': [sex_id],
'age_lower': [age_min], 'age_upper': [age_max],
'time_lower': [time_min], 'time_upper': [time_max]
})
reference_value = get_interpolated_covariate_values(
data_df=df_to_interp,
covariate_dict={c.name: parent_df},
population_df=pop_df
)[c.name].iloc[0]
max_difference = np.max(
np.abs(all_loc_df.mean_value - reference_value)
) + CascadeConstants.PRECISION_FOR_REFERENCE_VALUES
c.reference = reference_value
c.max_difference = max_difference
covariate_specs.create_covariate_list()
return covariate_specs
def reset_index(self, drop, inplace):
pass
def make_cascade_dag(model_version_id: int, dag: LocationDAG,
location_start: int, sex_start: int, split_sex: bool,
n_sim: int = 100, n_pool: int = 100, skip_configure: bool = False) -> List[_CascadeOperation]:
"""
Make a traditional cascade dag for a model version. Relies on a location DAG and a starting
point in the DAG for locations and sexes.
Parameters
----------
model_version_id
Model version ID
dag
A location DAG that specifies the location hierarchy
location_start
Where to start in the location hierarchy
sex_start
Which sex to start with, can be most detailed or both.
split_sex
Whether or not to split sex into most detailed. If not, then will just stay at 'both' sex.
n_sim
Number of simulations to do in sample simulate
n_pool
Number of multiprocessing pools to create during sample simulate
skip_configure
Don't configure inputs. Only do this if it's already been done.
Returns
-------
List of _CascadeOperation.
"""
tasks = []
sexes = [sex_start]
if SEX_ID_TO_NAME[sex_start] == 'Both':
if split_sex:
sexes = [
SEX_NAME_TO_ID['Female'],
SEX_NAME_TO_ID['Male']
]
top_level = root_fit(
model_version_id=model_version_id,
location_id=location_start, sex_id=sex_start,
child_locations=dag.children(location_start), child_sexes=sexes,
mulcov_stats=True,
skip_configure=skip_configure,
n_sim=n_sim, n_pool=n_pool,
ode_fit_strategy=True,
)
tasks += top_level
for sex in sexes:
for location1 in dag.children(location_start):
branch_or_leaf(
dag=dag, location_id=location1, sex=sex, model_version_id=model_version_id,
parent_location=location_start, parent_sex=sex_start,
n_sim=n_sim, n_pool=n_pool, upstream=[top_level[-1].command], tasks=tasks
)
tasks.append(Upload(
model_version_id=model_version_id,
fit=True, prior=True,
upstream_commands=[tasks[-1].command],
executor_parameters={
'm_mem_free': '50G'
}
))
return tasks
def construct_two_level_model(self, location_dag: LocationDAG, parent_location_id: int,
covariate_specs: CovariateSpecs,
weights: Optional[Dict[str, Var]] = None,
omega_df: Optional[pd.DataFrame] = None,
update_prior: Optional[Dict[str, Dict[str, np.ndarray]]] = None,
min_cv: Optional[Dict[str, Dict[str, float]]] = None,
update_mulcov_prior: Optional[Dict[Tuple[str, str, str], _Prior]] = None):
"""
Construct a Model object for a parent location and its children.
Parameters
----------
location_dag
Location DAG specifying the location hierarchy
parent_location_id
Parent location to build the model for
covariate_specs
covariate specifications, specifically will use covariate_specs.covariate_multipliers
weights
omega_df
data frame with omega values in it (other cause mortality)
update_prior
dictionary of dictionary for prior updates to rates
update_mulcov_prior
dictionary of mulcov prior updates
min_cv
dictionary (can be defaultdict) for minimum coefficient of variation
keyed by cascade level, then by rate
"""
children = location_dag.children(parent_location_id)
cascade_level = str(location_dag.depth(parent_location_id)) # min_cv lookup expects a string key
is_leaf = location_dag.is_leaf(parent_location_id)
if is_leaf:
cascade_level = MOST_DETAILED_CASCADE_LEVEL
model = Model(
nonzero_rates=self.settings.rate,
parent_location=parent_location_id,
child_location=children,
covariates=covariate_specs.covariate_list,
weights=weights
)
# First construct the rate grid, and update with prior
# information from a parent for value, dage, and dtime.
for smooth in self.settings.rate:
rate_grid = self.get_smoothing_grid(rate=smooth)
if update_prior is not None:
if smooth.rate in update_prior:
self.override_priors(rate_grid=rate_grid, update_dict=update_prior[smooth.rate])
if min_cv is not None:
self.apply_min_cv_to_prior_grid(
prior_grid=rate_grid.value, min_cv=min_cv[cascade_level][smooth.rate]
)
model.rate[smooth.rate] = rate_grid
# Second construct the covariate grids
for mulcov in covariate_specs.covariate_multipliers:
grid = smooth_grid_from_smoothing_form(
default_age_time=self.age_time_grid,
single_age_time=self.single_age_time_grid,
smooth=mulcov.grid_spec
)
if update_mulcov_prior is not None and (mulcov.group, *mulcov.key) in update_mulcov_prior:
ages = grid.ages
times = grid.times
for age, time in itertools.product(ages, times):
lb = grid.value[age, time].lower
ub = grid.value[age, time].upper
update_mulcov_prior[(mulcov.group, *mulcov.key)].lower = lb
update_mulcov_prior[(mulcov.group, *mulcov.key)].upper = ub
grid.value[age, time] = update_mulcov_prior[(mulcov.group, *mulcov.key)]
model[mulcov.group][mulcov.key] = grid
# Construct the random effect grids, based on the parent location
# specified.
if self.settings.random_effect:
random_effect_by_rate = defaultdict(list)
for smooth in self.settings.random_effect:
re_grid = smooth_grid_from_smoothing_form(
default_age_time=self.age_time_grid,
single_age_time=self.single_age_time_grid,
smooth=smooth
)
if not smooth.is_field_unset("location") and smooth.location in model.child_location:
location = smooth.location
else:
location = None
model.random_effect[(smooth.rate, location)] = re_grid
random_effect_by_rate[smooth.rate].append(location)
for rate_to_check, locations in random_effect_by_rate.items():
if locations != [None] and set(locations) != set(model.child_location):
raise RuntimeError(f"Random effect for {rate_to_check} does not have "
f"entries for all child locations, only {locations} "
f"instead of {model.child_location}.")
# Lastly, constrain omega for the parent and the random effects for the children.
if self.settings.model.constrain_omega:
LOG.info("Adding the omega constraint.")
if omega_df is None:
raise RuntimeError("Need an omega data frame in order to constrain omega.")
parent_omega = omega_df.loc[omega_df.location_id == parent_location_id].copy()
if parent_omega.empty:
raise RuntimeError(f"No omega values for location {parent_location_id}.")
omega = rectangular_data_to_var(gridded_data=parent_omega)
model.rate["omega"] = constraint_from_rectangular_data(
rate_var=omega,
default_age_time=self.age_time_grid
)
locations = set(omega_df.location_id.unique().tolist())
children_without_omega = set(children) - set(locations)
if children_without_omega:
LOG.warning(f"Children of {parent_location_id} missing omega {children_without_omega}"
f"so not including child omega constraints")
else:
for child in children:
child_omega = omega_df.loc[omega_df.location_id == child].copy()
assert not child_omega.empty
child_rate = rectangular_data_to_var(gridded_data=child_omega)
def child_effect(age, time):
return np.log(child_rate(age, time) / omega(age, time))
model.random_effect[("omega", child)] = constraint_from_rectangular_data(
rate_var=child_effect,
default_age_time=self.age_time_grid
)
return model