def calc_error(self, truth, obs, ignore_nan):
"""
Calculate the error of the obs compared to the truth.
Parameters
----------
truth : array
Numeric NumPy array of gold standard data
obs : array
Numeric NumPy array of observed values
ignore_nan : bool
Should NaN's be ignored in the calculation?
Returns
-------
error : float
The error
Notes
-----
truth and obs must have equal length. The ith element in obs is a measurment or observation of the
ith element in truth.
"""
errors = []
for i in range(0, len(truth)):
if ignore_nan == True:
if utilities.is_nan(truth[i]) != True and utilities.is_nan(obs[i]) != True:
errors.append(self.dist(truth[i], obs[i]))
else:
errors.append(self.dist(truth[i], obs[i]))
if len(errors) == 0:
error = np.nan
else:
error = self.summarize(errors)
return error
def evaluate_estimates(out_file, key, est_dir, est_y, gold_standard_file, y):
"""
Evaluate estimates stored in a directory of csvs against the gold standard.
Parameters
----------
out_file : string
A path to a csv in which to store summary error metrics
key : list of strings
A list of strings that correspond to columns in the estimates files and in the gold standard files
to serve as the key in merging these files. If evaluates_estimates is being run from
the command line, then key must be enclosed in double quotes (e.g. \"['iso3','year']\")
est_dir : string
The path to a directory with the csvs of the estimates. All csvs in this directory will be assumed to contain
estimates. The path should end with a /
est_y : string
The name of the column in the estimate files that holds the predictions from the model
y : string
The name of the column in the gold standard file that holds the response variable. This should also
be the name of the column in the estimates files that holds the knocked out and noised response variable.
gold_standard_file : string
The path to the gold standard file
Notes
-----
see parse_filename for a guide to the naming convention of the predicted response variables
"""
model_design_vars = {}
data = utilities.read(gold_standard_file)
data = utilities.add_unique_id(data, key, 'unique_id_for_join_by')
files = os.listdir(est_dir)
for file in files:
file_key = parse_filename(file)
path = est_dir + file
new_data = utilities.read(path)
# rename variable
names = []
for name in new_data.dtype.names:
if name == est_y:
est_y_name = est_y + '_' + str(file_key['model']) + '_' + str(file_key['design']) + '_' + str(file_key['rep'])
names.append(est_y_name)
elif name == y:
y_name = y + '_' + str(file_key['model']) + '_' + str(file_key['design']) + '_' + str(file_key['rep'])
names.append(y_name)
else:
names.append(name)
new_data.dtype.names = tuple(names)
# collect up variables corresponding to a certain model and design
if model_design_vars.has_key(file_key['model']) == False:
model_design_vars[file_key['model']] = {}
if model_design_vars[file_key['model']].has_key(file_key['design']) == False:
model_design_vars[file_key['model']][file_key['design']] = []
model_design_vars[file_key['model']][file_key['design']].append(est_y_name)
new_data = utilities.add_unique_id(new_data, key, 'unique_id_for_join_by')
new_data = new_data[['unique_id_for_join_by', est_y_name, y_name]]
# http://stackoverflow.com/questions/2774949/merging-indexed-array-in-python
data = numpy.lib.recfunctions.join_by('unique_id_for_join_by', data, new_data)
# this would write a file with the gold standard and all the predictions
#utilities.write(out_predictions_file, data)
data = numpy.lib.recfunctions.drop_fields(data, 'unique_id_for_join_by')
model_design_errors = {}
for model in model_design_vars.keys():
model_design_errors[model] = {}
for design in model_design_vars[model].keys():
model_design_errors[model][design] = {}
truth = []
obs = []
for var in model_design_vars[model][design]:
y_var = var.replace(est_y, y)
for i in range(0, len(data[y])):
if utilities.is_nan(data[y_var][i]) == True:
pdb.set_trace()
truth.append(data[y][i])
obs.append(data[var][i])
truth = np.array(truth)
obs = np.array(obs)
model_design_errors[model][design] = {}
errors = errormetrics.get_error_metrics()
for error in errors:
error_str = 'errormetrics.' + error + '()'
error_class = eval(error_str)
model_design_errors[model][design][error] = error_class.calc_error(truth, obs, True)
errors = []
for model in model_design_errors.keys():
for design in model_design_errors[model].keys():
for error in model_design_errors[model][design].keys():
errors.append(error)
errors = np.unique(errors)
writer = csv.writer(open(out_file, 'wb'))
fieldnames = ['model','design'] + errors.tolist()
writer.writerow(fieldnames)
for model in model_design_errors.keys():
for design in model_design_errors[model].keys():
row = [model, design]
for error in errors:
if model_design_errors[model][design].has_key(error) == True:
row.append(model_design_errors[model][design][error])
else:
row.append('')
writer.writerow(row)
writer = []
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
def simulate_data(out_dir, y, se, gold_standard_file, design_file):
"""
Simulate data by knocking out and adding noise to a gold standard file. How data is knocked out and
how noise is added is determined by parameters specified in the design file.
Parameters
----------
out_dir : string
The path to a directory in which to output the noisy and knocked out data. The path should end with a /
y : string
The column name in the gold standard file corresponding to the response to be knocked out and noised up.
se : string
The column name in the gold standard file corresponding to the standard error of the response.
If se == '', then a se variable will be created named 'se' and filled with 0's. If noise is added, then
this se variable will be set to the standard error of the noise.
gold_standard_file : string
The path to a csv.
design_file : string
The path to a csv. If a knock out test is to be performed, there must be a column called knockerouters.
If noise is to be added, there must be a column called noisers. If there is a column called rep, then
each test will be repeated rep times. If no such column is provided, each test will only be run once.
All other columns are parameters for the knockerouter function or the noiser function. These two functions
must not not share column names for parameters. All column entries (not the header) must be enclosed in
double quotes. A string will be enclosed in single quotes and then double quotes (e.g. \"'USA'\"), whereas
a number or an array will be enclosed only in double quotes (e.g. \"2\", \"[1,2,3]\").
See Also
--------
utilities.read
"""
if os.path.isdir(out_dir) == False:
os.mkdir(out_dir)
gold_data = utilities.read(gold_standard_file)
if se == "":
gold_data = numpy.lib.recfunctions.append_fields(gold_data, "se", [0] * len(gold_data), "<f4")
se = "se"
reader = csv.reader(open(design_file))
on_header = True
index = 0
rep_index = np.nan
for row in reader:
if on_header == True:
header = row
on_header = False
for i in range(0, len(header)):
if header == "rep":
rep_index = i
else:
if utilities.is_nan(rep_index) == True:
reps = 1
else:
reps = int(row[rep_index])
for i in range(0, reps):
data = gold_data
row_dict = {}
for j, name in enumerate(header):
row_j = eval(row[j])
row_dict[name] = row_j
for func_collection in ["knockerouters", "noisers", "biasers"]:
if row_dict.has_key(func_collection) == True:
fun_str = func_collection + "." + row_dict[func_collection]
if func_collection == "biasers":
fun_str = fun_str + "(data, y"
else:
fun_str = fun_str + "(data, y, se"
get_args_str = "inspect.getargspec(" + func_collection + "." + row_dict[func_collection] + ")"
args = eval(get_args_str)[0]
for arg in args:
if (arg in ["data", "y", "se"]) == False:
fun_str = fun_str + ", row_dict['" + arg + "']"
fun_str = fun_str + ")"
data = eval(fun_str)
gold_standard_file = gold_standard_file.split("/")
gold_standard_file = gold_standard_file[len(gold_standard_file) - 1]
new_file = (
out_dir + "sim_" + gold_standard_file.replace(".csv", "") + "_" + str(index) + "_" + str(i) + ".csv"
)
utilities.write(new_file, data)
index = index + 1
