def retractOnWorld(cogstate,worldname,proposition): """ INPUT: a triple (cognitive state, worldname, proposition) OUTPUT: array of worlds WHAT IT DOES: Definition 4 (i) """ result=[] world=getWorldByName(worldname,cogstate) for situation in sitgen(world): # s if Forceable(situation, proposition, cogstate): continue # s may not force P adding=False for basis in getAllBases(world,cogstate): # s' if not subset(situation,basis): continue # s is has to be a subset of s' Maximal=True for t in subsitgen(basis): if Forceable(situation, proposition, cogstate): continue # t may not force P if subset(situation,t): if situation != t: Maximal=False if not Maximal: continue # s should be a maximal subset of s' adding=True if adding: result.append(situation) return result
def Forceable(situation, proposition, cogstate):
"""
INPUT: a triple (situation, proposition, cognitive state)
OUTPUT: a Boolean
WHAT IT DOES: Decides whether the given situation forces
the proposition in the cognitive state.
"""
for world in cogstate:
if world[meta][US]:
if subset(situation, world):
if world[meta][name] not in proposition:
return False
return True
def retractOnState(cogstate,proposition):
"""
INPUT: a pair (cognitive state, proposition)
OUTPUT: cognitive state
WHAT IT DOES: Definition 4 (ii)
"""
result=[]
for world in cogstate:
newworld={} # do not shoot ourselves in the foot
newworld["values"]=dict(world["values"])
newworld["meta"]=dict(world["meta"])
addingToFS=False
if world["meta"]["US"]:
for biworld in cogstate:
if biworld["meta"]["FS"]:
biretract=retractOnWorld(cogstate,biworld["meta"]["name"],proposition)
for s in biretract:
if subset(s,world):
addingToFS=True
newworld["meta"]["FS"]=addingToFS
result.append(dict(newworld))
return result
def knock_out_leaving_surveys(data, y, se, key, year, survey_span, num_surveys, prop, survey_date):
"""
Knock out values of y and se in data so that the remaining data simulates a set
of surveys. More specifically, in each level of key multiple surveys are generated by
randomly selecting a year to conduct each survey and then marking that year and a number
of previous years determined by survey_span to leave in the dataset.
Parameters
----------
data : ndarray
A structured NumPy array. Should probably not have any missing values
or else strange behavior will ensue
y : string
The label for the response variable in data
se : string
The label for the standard error variable in data
key : string or list of strings
The labels for the variables in data that will define separate levels
for the knock out scheme.
year : string
The label for the year in data
survey_span : int
The number of years that each survey covers
num_surveys : int
The number of surveys in each country
prop : float
Proportion of countries to apply the knock out design. 0 <= prop <= 1
survey_date : string
The name of a variable to be added to data that contains the year each survey
was conducted so that surveys can be distinguished within a given level of the key.
This comes in handy if you want to specify correlated noise among surveys in your noiser.
Returns
-------
ko_data : ndarray
The same as data except with values of y and se knocked out
Notes
-----
In this framework, multiple observations of the same data points cannot be generated
"""
ko_data = copy.copy(data)
ko_data = utilities.add_unique_id(ko_data, key, 'unique_id_for_ko')
r = np.where(np.arange(1.,len(np.unique(ko_data['unique_id_for_ko']))+1.) <= len(np.unique(ko_data['unique_id_for_ko']))*prop, True, False)
if type(r.tolist()) != type(True):
random.shuffle(r)
else:
r = [r]
should_be_kept = {}
survey_date_dict = {}
for i, id in enumerate(np.unique(ko_data['unique_id_for_ko'])):
ko_data_i = utilities.subset(ko_data, 'unique_id_for_ko', id)
should_be_kept[id] = []
survey_date_dict[id] = []
if r[i] == True:
for s in range(0, num_surveys):
survey_year_index = random.choice(range(0,len(ko_data_i[year])))
for j in range(survey_year_index-survey_span, survey_year_index):
if (j in range(0, len(ko_data_i[year]))) == True:
should_be_kept[id].append(ko_data_i[year][j])
survey_date_dict[id].append(ko_data_i[year][survey_year_index])
else:
for j in range(0, len(ko_data_i[year])):
should_be_kept[id].append(ko_data_i[year][j])
survey_date_dict[id].append(np.nan)
survey_date_list = [np.nan]*len(ko_data[y])
for i in range(0, len(ko_data[y])):
id = ko_data['unique_id_for_ko'][i]
yr = ko_data[year][i]
for j, kept_yr in enumerate(should_be_kept[id]):
if kept_yr == yr:
survey_date_list[i] = survey_date_dict[id][j]
break
if utilities.is_nan(survey_date_list[i]) == True:
ko_data[y][i] = np.nan
ko_data[se][i] = np.nan
ko_data = numpy.lib.recfunctions.append_fields(ko_data, survey_date, np.array(survey_date_list))
ko_data = numpy.lib.recfunctions.drop_fields(ko_data, 'unique_id_for_ko')
return ko_data
def knock_out_cluster_unit(data, y, se, cluster='iso3', unit='year', prop=.2, design='random'):
"""
Within levels defined by the cluster variable, knock out a proportion of
units in data by replacing values of the variable y.
Parameters
----------
data : ndarray
A structured array.
y : string
A label of variable in data that corresponds to the response variable to be knocked out
se : string
A label of variable in data that corresponds to the standard error variable to be knocked out
cluster : string or list of strings
A field or list of fields in self.data (e.g. 'iso3' or \"['iso3','age']\"). The knock out scheme is applied separately to
levels defined by cluster.
unit : string
A field in self.data. The unit of the data to knock out. Unit should not have multiple values
with in levels of cluster.
proportion : float
The proportion of data to knock out.
design : string
If 'random', then a proportion of data is knocked out randomally.
If 'first', then the first proportion of data is knocked out and
analagously for last.
Examples
--------
>>> dtype = [('iso3','|S4'),('year','<i4'),('y','<f4'),('se','<f4')]
>>> data = np.array([('USA',1990,1,.1),('USA',1991,2,.2),('CAN',1990,3,.3),('CAN',1991,4,.4)], dtype=dtype)
>>> ko_data = knock_out_cluster_unit(data,'y','se','iso3','year',.5,'first')
>>> utilities.is_nan(ko_data['y'][0])
True
>>> utilities.is_nan(ko_data['y'][1])
False
>>> utilities.is_nan(ko_data['y'][2])
True
>>> utilities.is_nan(ko_data['y'][3])
False
>>> utilities.is_nan(ko_data['se'][0])
True
>>> utilities.is_nan(ko_data['se'][1])
False
>>> utilities.is_nan(ko_data['se'][2])
True
>>> utilities.is_nan(ko_data['se'][3])
False
# Check to see that original data has not been changed
>>> utilities.is_nan(data['y'][0])
False
"""
data = copy.copy(data)
data_cluster = {}
if cluster == '':
data_cluster[''] = data
else:
if len(cluster[0]) >= 2:
data = utilities.add_unique_id(data, cluster, 'knockerouters_unique_cluster_id')
cluster = 'knockerouters_unique_cluster_id'
for level in np.unique(data[cluster]):
data_cluster[level] = utilities.subset(data, cluster, level)
for key in data_cluster.keys():
candidates = []
for i, val in enumerate(data_cluster[key][y]):
if utilities.is_nan(val) == False:
candidates.append(i)
should_be_knocked_out = {}
r = np.where(np.arange(1.,len(candidates)+1.) <= len(candidates)*prop, True, False)
if type(r.tolist()) != type(True):
random.shuffle(r)
else:
r = [r]
for index, i in enumerate(candidates):
level = data_cluster[key][unit][i]
if design == 'random':
should_be_knocked_out[level] = r[index]
elif design == 'first':
should_be_knocked_out[level] = (float(i+1)/len(candidates)) <= prop
elif design == 'last':
should_be_knocked_out[level] = (float(i+1)/len(candidates)) >= (1-prop)
for i, level in enumerate(data[unit]):
if (level in should_be_knocked_out.keys()) == True:
if cluster == '':
if should_be_knocked_out[level] == True:
data[y][i] = np.nan
data[se][i] = np.nan
else:
if should_be_knocked_out[level] == True and data[cluster][i] == key:
data[y][i] = np.nan
data[se][i] = np.nan
if cluster == 'knockerouters_unique_cluster_id':
data = numpy.lib.recfunctions.drop_fields(data, 'knockerouters_unique_cluster_id')
return data
