def main(model_group, location_set_version_id, cause_set_version_id,
launch_set_id):
print_log_message(
"Beginning NR modeling for model_group {}".format(model_group))
cache_dir = CONF.get_directory('db_cache')
read_file_cache_options = {
'block_rerun': True,
'cache_dir': cache_dir,
'force_rerun': False,
'cache_results': False
}
print_log_message("Preparing location hierarchy")
location_hierarchy = get_current_location_hierarchy(
location_set_version_id=location_set_version_id,
**read_file_cache_options)
cause_meta_df = get_current_cause_hierarchy(
cause_set_version_id=cause_set_version_id, **read_file_cache_options)
age_meta_df = get_ages(**read_file_cache_options)
print_log_message("Preparing model data")
model_df = get_model_data(model_group, location_hierarchy,
location_set_version_id, cause_meta_df)
print_log_message("Got {} rows of model data".format(len(model_df)))
if len(model_df) == 0:
print_log_message("Exiting...")
return
model_df = model_df.drop('deaths', axis=1)
code_system_cause_dict = get_code_system_cause_ids(model_df)
if (model_group.startswith("VR")) or (model_group.startswith("Cancer")):
print_log_message("Bringing back zeros (squaring) so noise reduction "
"knows to depress time series")
squarer = Squarer(cause_meta_df, age_meta_df)
model_df = squarer.get_computed_dataframe(model_df)
elif "HH_SURVEYS" in model_group:
model_df = square_dhs_data(model_df, cause_meta_df, age_meta_df,
location_hierarchy)
print_log_message(log_statistic(model_df))
print_log_message("Restricting model data to only existing cause_ids")
model_df = restrict_to_cause_ids(code_system_cause_dict, model_df)
print_log_message("Adding NR location info")
model_df = format_for_nr(model_df, location_hierarchy)
if model_group_is_run_by_cause(model_group):
run_phase_by_cause(model_df, model_group, launch_set_id)
else:
run_phase_by_model_group(model_df, model_group, launch_set_id)
print_log_message("Job complete. Exiting...")
def prep_restrictions(self, cause_set_version_id):
df = get_current_cause_hierarchy(
cause_set_version_id=cause_set_version_id,
**self.standard_cache_options
)
df = df[['acause', 'cause_id', 'male',
'female', 'yll_age_start', 'yll_age_end']]
return df
def prep_restrictions(self, cause_set_version_id):
"""Get all the needed restrictions data
Implemented by pulling from central get_causes function,
then subsetting to just the restriction info for each cause_id.
"""
df = get_current_cause_hierarchy(
cause_set_version_id=cause_set_version_id,
**self.standard_cache_options)
df = df[[
'acause', 'cause_id', 'male', 'female', 'yll_age_start',
'yll_age_end'
]]
return df
def drop_age_restricted_cols(df):
start = len(df)
age_meta_df = get_ages(force_rerun=False, block_rerun=True)
# secret causes in restrictions
cause_meta_df = get_current_cause_hierarchy(cause_set_id=4,
**{
'block_rerun': True,
'force_rerun': False
})
restrict_df = pd.read_csv(
"/homes/agesak/thesis/maps/injuries_overrides.csv")
restrict_df = add_cause_metadata(restrict_df,
add_cols='cause_id',
merge_col='acause',
cause_meta_df=cause_meta_df)
restrict_df["age_start_group"] = restrict_df["age_start_group"].fillna(0)
orig_cols = df.columns
df = add_age_metadata(
df,
add_cols=['age_group_years_start', 'age_group_years_end'],
age_meta_df=age_meta_df)
df = df.merge(restrict_df, on='cause_id', how='left')
# age_group_years_end is weird, 0-14 means age_group_years_end 15
too_young = df["age_group_years_end"] <= df["age_start_group"]
too_old = df["age_group_years_start"] > df["age_end_group"]
df = df[~(too_young | too_old)]
df = df[orig_cols]
end = len(df)
print_log_message(
f"dropping {start - end} cols that violate age restrictions")
return df
def run_phase(df,
csvid,
nid,
extract_type_id,
lsvid,
pop_run_id,
cmvid,
launch_set_id,
remove_decimal,
write_diagnostics=True):
read_file_cache_options = {
'block_rerun': True,
'cache_dir': CACHE_DIR,
'force_rerun': False,
'cache_results': False
}
iso3 = get_value_from_nid(nid,
'iso3',
extract_type_id=extract_type_id,
location_set_version_id=lsvid)
code_system_id = int(
get_value_from_nid(nid,
'code_system_id',
extract_type_id=extract_type_id))
data_type_id = get_value_from_nid(nid,
'data_type_id',
extract_type_id=extract_type_id)
cause_map = get_cause_map(code_map_version_id=cmvid,
**read_file_cache_options)
orig_deaths_sum = int(df['deaths'].sum())
if remove_decimal:
print_log_message("Removing decimal from code map")
cause_map['value'] = cause_map['value'].apply(
lambda x: x.replace(".", ""))
if needs_garbage_correction(iso3, data_type_id):
print_log_message("Correcting Garbage for {}".format(iso3))
orig_gc_sum = int(df.query('cause_id == 743')['deaths'].sum())
cause_meta_df = get_current_cause_hierarchy(cause_set_version_id=csvid,
**read_file_cache_options)
age_meta_df = get_ages(**read_file_cache_options)
loc_meta_df = get_current_location_hierarchy(
location_set_version_id=lsvid, **read_file_cache_options)
pop_meta_df = get_pop(pop_run_id=pop_run_id, **read_file_cache_options)
hiv_corrector = HIVCorrector(df,
iso3,
code_system_id,
pop_meta_df,
cause_meta_df,
loc_meta_df,
age_meta_df,
correct_garbage=True)
df = hiv_corrector.get_computed_dataframe()
after_gc_sum = int(df.query('cause_id == 743')['deaths'].sum())
after_deaths_sum = int(df['deaths'].sum())
print_log_message("""
Stage [gc deaths / total deaths]
Before GC correction [{gco} / {to}]
After GC correction [{gca} / {ta}]
""".format(gco=orig_gc_sum,
to=orig_deaths_sum,
gca=after_gc_sum,
ta=after_deaths_sum))
df = add_code_metadata(df, ['value', 'code_system_id'],
code_map=cause_map,
**read_file_cache_options)
assert (df['code_system_id'] == code_system_id).all(), "Variable code " \
"system id {} did not agree with all values of df code " \
"system id: \n{}".format(
code_system_id, df.loc[df['code_system_id'] != code_system_id])
print_log_message("Formatting data for redistribution")
df = format_age_groups(df)
# drop observations with 0 deaths
df = drop_zero_deaths(df)
# merge on redistribution location hierarchy
df = add_rd_locations(df, lsvid)
# fill in any missing stuff that may have come from rd hierarchy
df = fill_missing_df(df, verify_all=True)
df = add_split_group_id_column(df)
# final check to make sure we have all the necessary columns
df = format_columns_for_rd(df, code_system_id)
split_groups = list(df.split_group.unique())
parallel = len(split_groups) > 1
print_log_message("Submitting/Running split groups")
for split_group in split_groups:
# remove intermediate files from previous run
delete_split_group_output(nid, extract_type_id, split_group)
# save to file
split_df = df.loc[df['split_group'] == split_group]
write_split_group_input(split_df, nid, extract_type_id, split_group)
if parallel:
submit_split_group(nid, extract_type_id, split_group,
code_system_id, launch_set_id)
else:
worker_main(nid, extract_type_id, split_group, code_system_id)
if parallel:
print_log_message("Waiting for splits to complete...")
wait('claude_redistributionworker_{}'.format(nid), 30)
print_log_message("Done waiting. Appending them together")
df = read_append_split_groups(split_groups, nid, extract_type_id,
cause_map)
print_log_message("Done appending files - {} rows assembled".format(
len(df)))
df = revert_variables(df)
after_deaths_sum = int(df['deaths'].sum())
before_after_text = """
Before GC redistribution: {a}
After GC redistribution: {b}
""".format(a=orig_deaths_sum, b=after_deaths_sum)
diff = abs(orig_deaths_sum - after_deaths_sum)
diff_threshold = max(.02 * orig_deaths_sum, 5)
if not diff < diff_threshold:
raise AssertionError("Deaths not close.\n" + before_after_text)
else:
print_log_message(before_after_text)
return df
def get_computed_dataframe(self, df, location_meta_df):
"""Split value_column into detailed age and sex groups.
Applies a relative rate splitting algorithm with a K-multiplier that
adjusts for the specific population that the data to be split applies
to.
Arguments and Attributes:
df (pandas.DataFrame): must contain all columns needed to merge on
population:
['location_id', 'age_group_id', 'sex_id', 'year_id'].
Must be unique on id_cols.
id_cols (list): list of columns that must exist in df and identify
observations. Used to preserve df in every way except for
splitting value_column, age_group_id, and sex_id.
pop_run_id (int): which population version to use
cause_set_version_id (int): which cause set version id to use
value_column (str): must be a column in df that contains values
to be split
gbd_round_id (int): which gbd_round is it
gbd_team_for_ages (str): what gbd team to use to call the shared
function db_queries.get_demographics
Returns:
split_df (pandas.DataFrame): contains all the columns passed
in df, but all age_group_id values will be detailed, all
sex_ids will be detailed (1, 2), and val will be split
into these detailed ids.
"""
# set cache options
standard_cache_options = {
'force_rerun': False,
'block_rerun': True,
'cache_dir': "standard",
'cache_results': False
}
verbose = self.verbose
value_column = self.value_column
pop_run_id = self.pop_run_id
cause_set_version_id = self.cause_set_version_id
gbd_round_id = self.conf.get_id('gbd_round')
id_cols = self.id_cols
gbd_team_for_ages = self.gbd_team_for_ages
orig_val_sum = df[self.value_column].sum()
# pull in populations
# get relevant populations
if verbose:
print("[{}] Prepping population".format(str(datetime.now())))
locations_in_data = list(set(df.location_id))
mapping_to_country_location_id = get_country_level_location_id(
locations_in_data, location_meta_df)
# Map subnational to it's country
df = df.merge(mapping_to_country_location_id,
how='left',
on='location_id')
df.rename(columns={'location_id': 'orig_location_id'}, inplace=True)
df['location_id'] = df['country_location_id']
df.drop('country_location_id', axis=1, inplace=True)
country_locations_in_data = list(df['location_id'].unique())
years_in_data = list(set(df.year_id))
pop_df = get_pop(pop_run_id=pop_run_id, **standard_cache_options)
pop_df = pop_df.loc[
(pop_df['location_id'].isin(country_locations_in_data))
& (pop_df['year_id'].isin(years_in_data))]
# what columns identify population data
pop_id_cols = ['location_id', 'age_group_id', 'sex_id', 'year_id']
assert not pop_df[pop_id_cols].duplicated().any()
# pull causes table
if verbose:
print("[{}] Prepping cause metadata".format(str(datetime.now())))
cause_meta_df = get_current_cause_hierarchy(
cause_set_version_id=cause_set_version_id,
**standard_cache_options)
# pull age sex weights
if verbose:
print("[{}] Prepping age sex weights".format(str(datetime.now())))
dist_df = get_cause_age_sex_distributions(
distribution_set_version_id=self.distribution_set_version_id,
**standard_cache_options)
keep_cols = ['cause_id', 'age_group_id', 'sex_id', 'weight']
dist_df = dist_df[keep_cols]
# pull age detail map
if verbose:
print("[{}] Prepping age agg to detail "
"map".format(str(datetime.now())))
age_detail_map = getcache_age_aggregate_to_detail_map(
gbd_round_id=gbd_round_id, **standard_cache_options)
# create map from aggregate sex ids to detail sex ids
if verbose:
print("[{}] Prepping sex detail map".format(str(datetime.now())))
sex_detail_map = AgeSexSplitter.prep_sex_aggregate_to_detail_map()
detail_maps = {
'age_group_id': age_detail_map,
'sex_id': sex_detail_map
}
dist_causes = dist_df.cause_id.unique()
if verbose:
print("[{}] Prepping cause_id to weight cause "
"map".format(str(datetime.now())))
cause_to_weight_cause_map = \
AgeSexSplitter.prep_cause_to_weight_cause_map(
cause_meta_df, dist_causes)
val_to_dist_maps = {'cause_id': cause_to_weight_cause_map}
# which columns are to be split
split_cols = ['age_group_id', 'sex_id']
split_inform_cols = ['cause_id']
value_cols = [value_column]
if verbose:
print("[{}] Running RR splitting "
"algorithm".format(str(datetime.now())))
split_df = relative_rate_split(df,
pop_df,
dist_df,
detail_maps,
split_cols,
split_inform_cols,
pop_id_cols,
value_cols,
pop_val_name='population',
val_to_dist_map_dict=val_to_dist_maps,
verbose=verbose)
df.drop('location_id', axis=1, inplace=True)
df.rename(columns={'orig_location_id': 'location_id'}, inplace=True)
if self.collect_diagnostics:
# making this optional because of memory usage
self.diag_df = split_df.copy()
group_columns = list(df.columns)
group_columns.remove(value_column)
if verbose:
print("[{}] Collapsing result".format(str(datetime.now())))
split_df = split_df.groupby(group_columns,
as_index=False)[value_column].sum()
if verbose:
print("[{}] Asserting valid results".format(str(datetime.now())))
val_diff = abs(split_df[value_column].sum() - orig_val_sum)
if not np.allclose(split_df[value_column].sum(), orig_val_sum):
text = "Difference of {} {} from age sex " \
"splitting".format(val_diff, value_column)
raise AssertionError(text)
# check that all age group ids are good
good_age_group_ids = db_queries.get_demographics(
gbd_team_for_ages, gbd_round_id=gbd_round_id)['age_group_id']
bad = set(split_df.age_group_id) - set(good_age_group_ids)
if len(bad) > 0:
text = "Some age group ids still aggregate: {}".format(bad)
raise AssertionError(text)
# should be the same set of cause ids
assert set(split_df.cause_id) == set(df.cause_id)
return split_df
def run_phase(df,
csvid,
nid,
extract_type_id,
lsvid,
pop_run_id,
cmvid,
launch_set_id,
remove_decimal,
write_diagnostics=True):
"""String together processes for redistribution."""
# what to do about caching throughout the phase
read_file_cache_options = {
'block_rerun': True,
'cache_dir': CACHE_DIR,
'force_rerun': False,
'cache_results': False
}
# the iso3 of this data
iso3 = get_value_from_nid(nid,
'iso3',
extract_type_id=extract_type_id,
location_set_version_id=lsvid)
# the code system id
code_system_id = int(
get_value_from_nid(nid,
'code_system_id',
extract_type_id=extract_type_id))
# the data type
data_type_id = get_value_from_nid(nid,
'data_type_id',
extract_type_id=extract_type_id)
# cause map
cause_map = get_cause_map(code_map_version_id=cmvid,
**read_file_cache_options)
orig_deaths_sum = int(df['deaths'].sum())
if remove_decimal:
print_log_message("Removing decimal from code map")
cause_map['value'] = cause_map['value'].apply(
lambda x: x.replace(".", ""))
if needs_garbage_correction(iso3, data_type_id):
print_log_message("Correcting Garbage for {}".format(iso3))
orig_gc_sum = int(df.query('cause_id == 743')['deaths'].sum())
cause_meta_df = get_current_cause_hierarchy(cause_set_version_id=csvid,
**read_file_cache_options)
# get age group ids
age_meta_df = get_ages(**read_file_cache_options)
loc_meta_df = get_current_location_hierarchy(
location_set_version_id=lsvid, **read_file_cache_options)
pop_meta_df = get_pop(pop_run_id=pop_run_id, **read_file_cache_options)
# Move garbage to hiv first
hiv_corrector = HIVCorrector(df,
iso3,
code_system_id,
pop_meta_df,
cause_meta_df,
loc_meta_df,
age_meta_df,
correct_garbage=True)
df = hiv_corrector.get_computed_dataframe()
after_gc_sum = int(df.query('cause_id == 743')['deaths'].sum())
after_deaths_sum = int(df['deaths'].sum())
print_log_message("""
Stage [gc deaths / total deaths]
Before GC correction [{gco} / {to}]
After GC correction [{gca} / {ta}]
""".format(gco=orig_gc_sum,
to=orig_deaths_sum,
gca=after_gc_sum,
ta=after_deaths_sum))
df = add_code_metadata(df, ['value', 'code_system_id'],
code_map=cause_map,
**read_file_cache_options)
# recognizing that it is weird for code_system_id to come from two places,
# make sure they are consistent
assert (df['code_system_id'] == code_system_id).all(), "Variable code " \
"system id {} did not agree with all values of df code " \
"system id: \n{}".format(
code_system_id, df.loc[df['code_system_id'] != code_system_id])
print_log_message("Formatting data for redistribution")
# do we have all the packages we need?
# verify_packages(df)
# format age groups to match package parameters
df = format_age_groups(df)
# drop observations with 0 deaths
df = drop_zero_deaths(df)
# merge on redistribution location hierarchy
df = add_rd_locations(df, lsvid)
# fill in any missing stuff that may have come from rd hierarchy
df = fill_missing_df(df, verify_all=True)
# create split groups
# NO SPLIT GROUP NEEDED
df = add_split_group_id_column(df)
# final check to make sure we have all the necessary columns
df = format_columns_for_rd(df, code_system_id)
split_groups = list(df.split_group.unique())
parallel = len(split_groups) > 1
print_log_message("Submitting/Running split groups")
for split_group in split_groups:
# remove intermediate files from previous run
delete_split_group_output(nid, extract_type_id, split_group)
# save to file
split_df = df.loc[df['split_group'] == split_group]
write_split_group_input(split_df, nid, extract_type_id, split_group)
# submit jobs or just run them here
if parallel:
submit_split_group(nid, extract_type_id, split_group,
code_system_id, launch_set_id)
else:
worker_main(nid, extract_type_id, split_group, code_system_id)
if parallel:
print_log_message("Waiting for splits to complete...")
# wait until all jobs for a given nid have completed
# eventually need logic for files not being present
wait('claude_redistributionworker_{}'.format(nid), 30)
# This seems to be necessary to wait for files
print_log_message("Done waiting. Appending them together")
# append split groups together
df = read_append_split_groups(split_groups, nid, extract_type_id,
cause_map)
print_log_message("Done appending files - {} rows assembled".format(
len(df)))
df = revert_variables(df)
after_deaths_sum = int(df['deaths'].sum())
before_after_text = """
Before GC redistribution: {a}
After GC redistribution: {b}
""".format(a=orig_deaths_sum, b=after_deaths_sum)
diff = abs(orig_deaths_sum - after_deaths_sum)
# bad if change 2% or 5 deaths, whichever is greater
# (somewhat arbitrary, just trying to avoid annoying/non-issue failures)
diff_threshold = max(.02 * orig_deaths_sum, 5)
if not diff < diff_threshold:
raise AssertionError("Deaths not close.\n" + before_after_text)
else:
print_log_message(before_after_text)
return df
def run_phase(df, nid, extract_type_id, pop_run_id, cause_set_version_id,
location_set_version_id):
"""Run the full phase, chaining together computational elements."""
cache_dir = CONF.get_directory('FILEPATH')
orig_deaths = df['deaths'].sum()
standard_cache_options = {
'force_rerun': False,
'block_rerun': True,
'cache_dir': cache_dir,
'cache_results': False
}
code_system_id = get_value_from_nid(nid,
'code_system_id',
extract_type_id=extract_type_id)
# this queries the database, maybe should be passed in directly
code_system = get_code_system_from_id(code_system_id)
source = get_value_from_nid(nid, 'source', extract_type_id=extract_type_id)
data_type_id = get_value_from_nid(
nid,
'data_type_id',
extract_type_id=extract_type_id,
location_set_version_id=location_set_version_id)
# get cause hierarchy
cause_meta_df = get_current_cause_hierarchy(
cause_set_version_id=cause_set_version_id, **standard_cache_options)
is_vr = data_type_id in [9, 10]
if not skip_hiv_correction(source) and is_vr:
# get location hierarchy
loc_meta_df = get_current_location_hierarchy(
location_set_version_id=location_set_version_id,
**standard_cache_options)
# get population
pop_meta_df = get_pop(pop_run_id=pop_run_id, **standard_cache_options)
# get age metadata
age_meta_df = get_ages(**standard_cache_options)
# get the country
iso3 = get_value_from_nid(
nid,
'iso3',
extract_type_id=extract_type_id,
location_set_version_id=location_set_version_id)
assert pd.notnull(iso3), "Could not find iso3 for nid {}, " \
"extract_type_id {}".format(nid, extract_type_id)
hiv_corrector = HIVCorrector(df,
iso3,
code_system_id,
pop_meta_df,
cause_meta_df,
loc_meta_df,
age_meta_df,
correct_garbage=False)
print_log_message("Running hiv correction for iso3 {}".format(iso3))
df = hiv_corrector.get_computed_dataframe()
if needs_injury_redistribution(source):
print_log_message("Correcting injuries")
if not 'loc_meta_df' in vars():
# get location hierarchy
loc_meta_df = get_current_location_hierarchy(
location_set_version_id=location_set_version_id,
**standard_cache_options)
injury_redistributor = InjuryRedistributor(df, loc_meta_df,
cause_meta_df)
df = injury_redistributor.get_computed_dataframe()
df = combine_with_rd_raw(df, nid, extract_type_id, location_set_version_id)
val_cols = ['deaths', 'deaths_raw', 'deaths_corr', 'deaths_rd']
# run china VR rescaling
if needs_subnational_rescale(source):
china_rescaler = ChinaHospitalUrbanicityRescaler()
df = china_rescaler.get_computed_dataframe(df)
if needs_strata_collapse(source):
# set site id to blank site id and collapse
df['site_id'] = 2
group_cols = list(set(df.columns) - set(val_cols))
df = df.groupby(group_cols, as_index=False)[val_cols].sum()
if is_vr:
# drop if deaths are 0 across all current deaths columns
df = df.loc[df[val_cols].sum(axis=1) != 0]
# restrict causes based on code system
print_log_message("Running bridge mapper")
bridge_mapper = BridgeMapper(source, cause_meta_df, code_system)
df = bridge_mapper.get_computed_dataframe(df)
# run recodes based on expert opinion
print_log_message("Enforcing some very hard priors (expert opinion)")
expert_opinion_recoder = Recoder(cause_meta_df, source, code_system_id,
data_type_id)
df = expert_opinion_recoder.get_computed_dataframe(df)
end_deaths = df['deaths'].sum()
print_log_message("Checking no large loss or gain of deaths")
if abs(orig_deaths - end_deaths) >= (.1 * end_deaths):
diff = round(abs(orig_deaths - end_deaths), 2)
old = round(abs(orig_deaths))
new = round(abs(end_deaths))
raise AssertionError("Change of {} deaths [{}] to [{}]".format(
diff, old, new))
return df
def run_phase(df, nid, extract_type_id, env_run_id, pop_run_id,
location_set_version_id, cause_set_version_id):
cache_dir = CONF.get_directory('db_cache')
source = get_value_from_nid(
nid,
'source',
extract_type_id=extract_type_id,
location_set_version_id=location_set_version_id)
data_type_id = get_value_from_nid(
nid,
'data_type_id',
extract_type_id=extract_type_id,
location_set_version_id=location_set_version_id)
iso3 = get_value_from_nid(nid,
'iso3',
extract_type_id=extract_type_id,
location_set_version_id=location_set_version_id)
standard_cache_options = {
'force_rerun': False,
'block_rerun': True,
'cache_dir': cache_dir,
'cache_results': False
}
# ************************************************************
# Get cached metadata
# ************************************************************
print_log_message("Getting cached db resources")
location_hierarchy = get_current_location_hierarchy(
location_set_version_id=location_set_version_id,
**standard_cache_options)
pop_df = get_pop(pop_run_id=pop_run_id, **standard_cache_options)
env_df = get_env(env_run_id=env_run_id, **standard_cache_options)
age_weight_df = get_age_weights(**standard_cache_options)
cause_meta_df = get_current_cause_hierarchy(
cause_set_version_id=cause_set_version_id, **standard_cache_options)
age_meta_df = get_ages(**standard_cache_options)
# ************************************************************
# RAKING
# ************************************************************
# Rake if appropriate based on this logic
if ((data_type_id in [8, 9, 10] and (source != 'Other_Maternal'))
or source in MATERNAL_NR_SOURCES):
if source not in NOT_RAKED_SOURCES:
print_log_message("Raking sub national estimates")
raker = Raker(df, source)
df = raker.get_computed_dataframe(location_hierarchy)
# for the Other_Maternal source we only rake household surveys
elif source == "Other_Maternal":
model_groups = get_datasets(nid,
extract_type_id,
block_rerun=True,
force_rerun=False).model_group.unique()
assert len(model_groups) == 1
model_group = model_groups[0]
if "HH_SURVEYS" in model_group:
if model_group == 'MATERNAL-HH_SURVEYS-IND':
print_log_message("Raking sub national estimates," \
" applying double raking for India Maternal"
)
raker = Raker(df, source, double=True)
df = raker.get_computed_dataframe(location_hierarchy)
else:
print_log_message("Raking sub national estimates")
raker = Raker(df, source)
df = raker.get_computed_dataframe(location_hierarchy)
# ************************************************************
# DROP ZERO SAMPLE SIZE AND RESTRICTED AGE/SEX DATA
# ************************************************************
# data with zero sample size is almost certaintly some anomolous result
# of a program generating data it shouldn't have, and it shouldn't be
# included in codem models. Was probably already dropped, anyway, before
# running noise reduction.
df = df.query('sample_size != 0')
# uploading data before 1980 is a waste of space because neither codem
# nor codviz use it
df = df.loc[df['year_id'] >= 1980]
print_log_message("Enforcing age sex restrictions")
# this actually drops data from the dataframe if it violates age/sex
# restrictions (e.g. male maternity disorders)
df = enforce_asr(df, cause_meta_df, age_meta_df)
# ************************************************************
# FIT EACH DRAW TO NON-ZERO FLOOR
# ************************************************************
print_log_message("Fitting to non-zero floor...")
nonzero_floorer = NonZeroFloorer(df)
df = nonzero_floorer.get_computed_dataframe(pop_df, env_df, cause_meta_df)
# ************************************************************
# AGE AGGREGATION
# ************************************************************
print_log_message("Creating age standardized and all ages groups")
age_aggregator = AgeAggregator(df, pop_df, env_df, age_weight_df)
df = age_aggregator.get_computed_dataframe()
# ************************************************************
# Make CODEm and CoDViz metrics for uncertainty
# ************************************************************
# columns that should be present in the phase output
final_cols = [
'age_group_id', 'cause_id', 'cf_corr', 'cf_final', 'cf_raw', 'cf_rd',
'extract_type_id', 'location_id', 'nid', 'sample_size', 'sex_id',
'site_id', 'year_id'
]
# Use draws to make metrics for uncertainty to
# be used by CODEm and CoDViz
# also creates cf_final from mean of draws
print_log_message("Making metrics for CODEm and CoDViz")
if dataset_has_redistribution_variance(data_type_id, source):
df = RedistributionVarianceEstimator.make_codem_codviz_metrics(
df, pop_df)
final_cols += [
'cf_final_high_rd', 'cf_final_low_rd', 'variance_rd_log_dr',
'variance_rd_logit_cf'
]
# we did this in the old code-- no cfs over 1 nor below 0
for cf_col in ['cf_final', 'cf_rd', 'cf_raw', 'cf_corr']:
df.loc[df[cf_col] > 1, cf_col] = 1
df.loc[df[cf_col] < 0, cf_col] = 0
df = df[final_cols]
return df
def run_phase(df, nid, extract_type_id, env_run_id, pop_run_id,
location_set_version_id, cause_set_version_id):
cache_dir = CONF.get_directory('db_cache')
source = get_value_from_nid(
nid,
'source',
extract_type_id=extract_type_id,
location_set_version_id=location_set_version_id)
data_type_id = get_value_from_nid(
nid,
'data_type_id',
extract_type_id=extract_type_id,
location_set_version_id=location_set_version_id)
standard_cache_options = {
'force_rerun': False,
'block_rerun': True,
'cache_dir': cache_dir,
'cache_results': False
}
# ************************************************************
# Get cached metadata
# ************************************************************
print_log_message("Getting cached db resources")
location_hierarchy = get_current_location_hierarchy(
location_set_version_id=location_set_version_id,
**standard_cache_options)
pop_df = get_pop(pop_run_id=pop_run_id, **standard_cache_options)
env_df = get_env(env_run_id=env_run_id, **standard_cache_options)
age_weight_df = get_age_weights(**standard_cache_options)
cause_meta_df = get_current_cause_hierarchy(
cause_set_version_id=cause_set_version_id, **standard_cache_options)
age_meta_df = get_ages(**standard_cache_options)
# ************************************************************
# RAKING
# ************************************************************
if ((data_type_id in [8, 9, 10] and (source != "Other_Maternal"))
or source in MATERNAL_NR_SOURCES):
if source not in NOT_RAKED_SOURCES:
print_log_message("Raking sub national estimates")
raker = Raker(df, source)
df = raker.get_computed_dataframe(location_hierarchy)
# for the Other_Maternal source we only rake household surveys
elif source == "Other_Maternal":
model_groups = get_datasets(nid,
extract_type_id,
block_rerun=True,
force_rerun=False).model_group.unique()
assert len(model_groups) == 1
model_group = model_groups[0]
if "HH_SURVEYS" in model_group:
print_log_message("Raking sub national estimates")
raker = Raker(df, source)
df = raker.get_computed_dataframe(location_hierarchy)
# ************************************************************
# DROP ZERO SAMPLE SIZE AND RESTRICTED AGE/SEX DATA
# ************************************************************
df = df.query('sample_size != 0')
df = df.loc[df['year_id'] >= 1980]
print_log_message("Enforcing age sex restrictions")
df = enforce_asr(df, cause_meta_df, age_meta_df)
# ************************************************************
# FIT EACH DRAW TO NON-ZERO FLOOR
# ************************************************************
print_log_message("Fitting to non-zero floor...")
nonzero_floorer = NonZeroFloorer(df)
df = nonzero_floorer.get_computed_dataframe(pop_df, env_df, cause_meta_df)
# ************************************************************
# AGE AGGREGATION
# ************************************************************
print_log_message("Creating age standardized and all ages groups")
age_aggregator = AgeAggregator(df, pop_df, env_df, age_weight_df)
df = age_aggregator.get_computed_dataframe()
# ************************************************************
# Make CODEm and CoDViz metrics for uncertainty
# ************************************************************
# columns that should be present in the phase output
final_cols = [
'age_group_id', 'cause_id', 'cf_corr', 'cf_final', 'cf_raw', 'cf_rd',
'extract_type_id', 'location_id', 'nid', 'sample_size', 'sex_id',
'site_id', 'year_id'
]
print_log_message("Making metrics for CODEm and CoDViz")
if dataset_has_redistribution_variance(data_type_id, source):
df = RedistributionVarianceEstimator.make_codem_codviz_metrics(
df, pop_df)
final_cols += [
'cf_final_high_rd', 'cf_final_low_rd', 'variance_rd_log_dr',
'variance_rd_logit_cf'
]
for cf_col in ['cf_final', 'cf_rd', 'cf_raw', 'cf_corr']:
df.loc[df[cf_col] > 1, cf_col] = 1
df.loc[df[cf_col] < 0, cf_col] = 0
df = df[final_cols]
return df
def run_phase(df, nid, extract_type_id, pop_run_id, cause_set_version_id,
location_set_version_id):
"""Run the full phase, chaining together computational elements."""
# get filepaths
cache_dir = CONF.get_directory('db_cache')
orig_deaths = df['deaths'].sum()
standard_cache_options = {
'force_rerun': False,
'block_rerun': True,
'cache_dir': cache_dir,
'cache_results': False
}
code_system_id = get_value_from_nid(nid,
'code_system_id',
extract_type_id=extract_type_id)
code_system = get_code_system_from_id(code_system_id,
**standard_cache_options)
source = get_value_from_nid(nid, 'source', extract_type_id=extract_type_id)
data_type_id = get_value_from_nid(
nid,
'data_type_id',
extract_type_id=extract_type_id,
location_set_version_id=location_set_version_id)
# get cause hierarchy
cause_meta_df = get_current_cause_hierarchy(
cause_set_version_id=cause_set_version_id, **standard_cache_options)
is_vr = data_type_id in [9, 10]
# run hiv correction on VR, but not Other_Maternal
# countries to correct will be further pruned by the master cause
# selections csv in the hiv corrector class
if not skip_hiv_correction(source) and is_vr:
# get location hierarchy
loc_meta_df = get_current_location_hierarchy(
location_set_version_id=location_set_version_id,
**standard_cache_options)
# get population
pop_meta_df = get_pop(pop_run_id=pop_run_id, **standard_cache_options)
# get age metadata
age_meta_df = get_ages(**standard_cache_options)
# get the country
iso3 = get_value_from_nid(
nid,
'iso3',
extract_type_id=extract_type_id,
location_set_version_id=location_set_version_id)
assert pd.notnull(iso3), "Could not find iso3 for nid {}, " \
"extract_type_id {}".format(nid, extract_type_id)
hiv_corrector = HIVCorrector(df,
iso3,
code_system_id,
pop_meta_df,
cause_meta_df,
loc_meta_df,
age_meta_df,
correct_garbage=False)
print_log_message("Running hiv correction for iso3 {}".format(iso3))
df = hiv_corrector.get_computed_dataframe()
if needs_injury_redistribution(source):
print_log_message("Correcting injuries")
if not 'loc_meta_df' in vars():
# get location hierarchy
loc_meta_df = get_current_location_hierarchy(
location_set_version_id=location_set_version_id,
**standard_cache_options)
injury_redistributor = InjuryRedistributor(df, loc_meta_df,
cause_meta_df)
df = injury_redistributor.get_computed_dataframe()
# apply redistribution of LRI to tb in under 15, non-neonatal ages based
# on location/year specific proportions
print_log_message(
"Applying special redistribution of LRI to TB in under 15")
lri_tb_redistributor = LRIRedistributor(df, cause_meta_df)
df = lri_tb_redistributor.get_computed_dataframe()
# merge in raw and rd here because recodes and bridge mapping should
# also apply to the causes that are in previous phases (raw deaths for
# secret codes need to be moved up to their parent cause, for example)
df = combine_with_rd_raw(df, nid, extract_type_id, location_set_version_id)
val_cols = ['deaths', 'deaths_raw', 'deaths_corr', 'deaths_rd']
# run china VR rescaling
if needs_subnational_rescale(source):
china_rescaler = ChinaHospitalUrbanicityRescaler()
df = china_rescaler.get_computed_dataframe(df)
if needs_strata_collapse(source):
# set site id to blank site id and collapse
df['site_id'] = 2
group_cols = list(set(df.columns) - set(val_cols))
df = df.groupby(group_cols, as_index=False)[val_cols].sum()
if is_vr:
# drop if deaths are 0 across all current deaths columns
df = df.loc[df[val_cols].sum(axis=1) != 0]
# restrict causes based on code system
print_log_message("Running bridge mapper")
bridge_mapper = BridgeMapper(source, cause_meta_df, code_system)
df = bridge_mapper.get_computed_dataframe(df)
# run recodes based on expert opinion
print_log_message("Enforcing some very hard priors (expert opinion)")
expert_opinion_recoder = Recoder(cause_meta_df, source, code_system_id,
data_type_id)
df = expert_opinion_recoder.get_computed_dataframe(df)
end_deaths = df['deaths'].sum()
print_log_message("Checking no large loss or gain of deaths")
if abs(orig_deaths - end_deaths) >= (.1 * end_deaths):
diff = round(abs(orig_deaths - end_deaths), 2)
old = round(abs(orig_deaths))
new = round(abs(end_deaths))
raise AssertionError("Change of {} deaths [{}] to [{}]".format(
diff, old, new))
return df