def test_run_SS(baseline, param_updates, filename, dask_client):
# Test SS.run_SS function. Provide inputs to function and
# ensure that output returned matches what it has been before.
SS.ENFORCE_SOLUTION_CHECKS = True #False
# if running reform, then need to solve baseline first to get values
if baseline is False:
p_base = Specifications(output_base=constants.BASELINE_DIR,
baseline_dir=constants.BASELINE_DIR,
baseline=True,
num_workers=NUM_WORKERS)
p_base.update_specifications(param_updates)
if p_base.use_zeta:
p_base.update_specifications({
'initial_guess_r_SS': 0.10,
'initial_guess_TR_SS': 0.02
})
p_base.baseline_spending = False
base_ss_outputs = SS.run_SS(p_base, client=dask_client)
utils.mkdirs(os.path.join(constants.BASELINE_DIR, "SS"))
ss_dir = os.path.join(constants.BASELINE_DIR, "SS", "SS_vars.pkl")
with open(ss_dir, "wb") as f:
pickle.dump(base_ss_outputs, f)
# now run specification for test
p = Specifications(baseline=baseline, num_workers=NUM_WORKERS)
p.update_specifications(param_updates)
test_dict = SS.run_SS(p, client=dask_client)
expected_dict = utils.safe_read_pickle(
os.path.join(CUR_PATH, 'test_io_data', filename))
for k, v in expected_dict.items():
print('Checking item = ', k)
assert (np.allclose(test_dict[k], v, atol=1e-06))
def test_run_TPI(baseline, param_updates, filename, tmp_path, dask_client):
'''
Test TPI.run_TPI function. Provide inputs to function and
ensure that output returned matches what it has been before.
'''
baseline_dir = os.path.join(CUR_PATH, 'baseline')
if baseline:
output_base = baseline_dir
else:
output_base = os.path.join(CUR_PATH, 'reform')
p = Specifications(baseline=baseline,
baseline_dir=baseline_dir,
output_base=output_base,
num_workers=NUM_WORKERS)
test_params = TEST_PARAM_DICT.copy()
test_params.update(param_updates)
p.update_specifications(test_params)
p.maxiter = 2 # this test runs through just two iterations
# Need to run SS first to get results
SS.ENFORCE_SOLUTION_CHECKS = False
ss_outputs = SS.run_SS(p, client=dask_client)
if p.baseline:
utils.mkdirs(os.path.join(p.baseline_dir, "SS"))
ss_dir = os.path.join(p.baseline_dir, "SS", "SS_vars.pkl")
with open(ss_dir, "wb") as f:
pickle.dump(ss_outputs, f)
else:
utils.mkdirs(os.path.join(p.output_base, "SS"))
ss_dir = os.path.join(p.output_base, "SS", "SS_vars.pkl")
with open(ss_dir, "wb") as f:
pickle.dump(ss_outputs, f)
TPI.ENFORCE_SOLUTION_CHECKS = False
test_dict = TPI.run_TPI(p, client=dask_client)
expected_dict = utils.safe_read_pickle(filename)
for k, v in expected_dict.items():
print('Max diff in ', k, ' = ')
try:
print(np.absolute(test_dict[k][:p.T] - v[:p.T]).max())
except ValueError:
print(np.absolute(test_dict[k][:p.T, :, :] - v[:p.T, :, :]).max())
for k, v in expected_dict.items():
try:
assert (np.allclose(test_dict[k][:p.T],
v[:p.T],
rtol=1e-04,
atol=1e-04))
except ValueError:
assert (np.allclose(test_dict[k][:p.T, :, :],
v[:p.T, :, :],
rtol=1e-04,
atol=1e-04))
def compute_se(beta_hat, W, K, p, h=0.01, client=None):
"""
Function to compute standard errors for the SMM estimator.
Args:
beta_hat (array-like): estimates of beta parameters
W (Numpy array): weighting matrix
K (int): number of moments
p (OG-USA Specifications object): model parameters
h (scalar): percentage to move parameters for numerical derivatives
client (Dask Client object): Dask client
Returns:
beta_se (array-like): standard errors for beta estimates
VCV_params (Numpy array): VCV matrix for parameter estimates
"""
# compute numerical derivatives that will need for SE's
model_moments_low = np.zeros((p.J, K))
model_moments_high = np.zeros((p.J, K))
beta_low = beta_hat
beta_high = beta_hat
for i in range(len(beta_hat)):
# compute moments with downward change in param
beta_low[i] = beta_hat[i] * (1 + h)
p.beta = beta_low
ss_output = ss_output = SS.run_SS(p, client=client)
model_moments_low[i, :] = calc_moments(ss_output, p)
# compute moments with upward change in param
beta_high[i] = beta_hat[i] * (1 - h)
p.beta = beta_low
ss_output = ss_output = SS.run_SS(p, client=client)
model_moments_high[i, :] = calc_moments(ss_output, p)
deriv_moments = (model_moments_high - model_moments_low).T / (2 * h *
beta_hat)
VCV_params = np.linalg.inv(
np.dot(np.dot(deriv_moments.T, W), deriv_moments))
beta_se = (np.diag(VCV_params))**(1 / 2)
return beta_se, VCV_params
def test_constant_demographics_TPI(dask_client):
'''
This tests solves the model under the assumption of constant
demographics, a balanced budget, and tax functions that do not vary
over time.
In this case, given how initial guesses for the time
path are made, the time path should be solved for on the first
iteration and the values all along the time path should equal their
steady-state values.
'''
# Create output directory structure
spec = Specifications(output_base=CUR_PATH,
baseline_dir=OUTPUT_DIR,
baseline=True,
num_workers=NUM_WORKERS)
og_spec = {
'constant_demographics': True,
'budget_balance': True,
'zero_taxes': True,
'maxiter': 2,
'r_gov_shift': 0.0,
'zeta_D': [0.0, 0.0],
'zeta_K': [0.0, 0.0],
'debt_ratio_ss': 1.0,
'initial_foreign_debt_ratio': 0.0,
'start_year': 2019,
'cit_rate': [0.0],
'PIA_rate_bkt_1': 0.0,
'PIA_rate_bkt_2': 0.0,
'PIA_rate_bkt_3': 0.0,
'eta':
(spec.omega_SS.reshape(spec.S, 1) * spec.lambdas.reshape(1, spec.J))
}
spec.update_specifications(og_spec)
spec.etr_params = np.zeros(
(spec.T + spec.S, spec.S, spec.etr_params.shape[2]))
spec.mtrx_params = np.zeros(
(spec.T + spec.S, spec.S, spec.mtrx_params.shape[2]))
spec.mtry_params = np.zeros(
(spec.T + spec.S, spec.S, spec.mtry_params.shape[2]))
# Run SS
ss_outputs = SS.run_SS(spec, client=dask_client)
# save SS results
utils.mkdirs(os.path.join(OUTPUT_DIR, "SS"))
ss_dir = os.path.join(OUTPUT_DIR, "SS", "SS_vars.pkl")
with open(ss_dir, "wb") as f:
pickle.dump(ss_outputs, f)
# Run TPI
tpi_output = TPI.run_TPI(spec, client=dask_client)
assert (np.allclose(tpi_output['bmat_splus1'][:spec.T, :, :],
ss_outputs['bssmat_splus1']))
def minstat(beta_guesses, *args):
"""
This function generates the weighted sum of squared differences
between the model and data moments.
Args:
beta_guesses (array-like): a vector of length J with the betas
args (tuple): length 6 tuple, variables needed for minimizer
Returns:
distance (scalar): weighted, squared deviation between data and
model moments
"""
# unpack args tuple
data_moments, W, p, client = args
# Update beta in parameters object with beta guesses
p.beta = beta_guesses
# Solve model SS
print("Baseline = ", p.baseline)
ss_output = SS.run_SS(p, client=client)
# Compute moments from model SS
model_moments = calc_moments(ss_output, p)
# distance with levels
distance = np.dot(
np.dot((np.array(model_moments) - np.array(data_moments)).T, W),
np.array(model_moments) - np.array(data_moments),
)
print("DATA and MODEL DISTANCE: ", distance)
return distance
def run_model(meta_param_dict, adjustment):
"""
Initializes classes from OG-USA that compute the model under
different policies. Then calls function get output objects.
"""
print("Meta_param_dict = ", meta_param_dict)
print("adjustment dict = ", adjustment)
meta_params = MetaParams()
meta_params.adjust(meta_param_dict)
if meta_params.data_source == "PUF":
data = retrieve_puf(AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY)
# set name of cached baseline file in case use below
cached_pickle = "TxFuncEst_baseline_PUF.pkl"
else:
data = "cps"
# set name of cached baseline file in case use below
cached_pickle = "TxFuncEst_baseline_CPS.pkl"
# Get TC params adjustments
iit_mods = convert_policy_adjustment(
adjustment["Tax-Calculator Parameters"]
)
# Create output directory structure
base_dir = os.path.join(CUR_DIR, BASELINE_DIR)
reform_dir = os.path.join(CUR_DIR, REFORM_DIR)
dirs = [base_dir, reform_dir]
for _dir in dirs:
utils.mkdirs(_dir)
# Dask parmeters
client = Client()
num_workers = 5
# TODO: Swap to these parameters when able to specify tax function
# and model workers separately
# num_workers_txf = 5
# num_workers_mod = 6
# whether to estimate tax functions from microdata
run_micro = True
time_path = meta_param_dict["time_path"][0]["value"]
# filter out OG-USA params that will not change between baseline and
# reform runs (these are the non-policy parameters)
filtered_ogusa_params = {}
constant_param_set = {
"frisch",
"beta_annual",
"sigma",
"g_y_annual",
"gamma",
"epsilon",
"Z",
"delta_annual",
"small_open",
"world_int_rate",
"initial_debt_ratio",
"initial_foreign_debt_ratio",
"zeta_D",
"zeta_K",
"tG1",
"tG2",
"rho_G",
"debt_ratio_ss",
"budget_balance",
}
filtered_ogusa_params = OrderedDict()
for k, v in adjustment["OG-USA Parameters"].items():
if k in constant_param_set:
filtered_ogusa_params[k] = v
# Solve baseline model
start_year = meta_param_dict["year"][0]["value"]
if start_year == 2020:
OGPATH = inspect.getfile(SS)
OGDIR = os.path.dirname(OGPATH)
tax_func_path = None # os.path.join(OGDIR, 'data', 'tax_functions',
# cached_pickle)
run_micro_baseline = True
else:
tax_func_path = None
run_micro_baseline = True
base_spec = {
**{
"start_year": start_year,
"tax_func_type": "DEP",
"age_specific": False,
},
**filtered_ogusa_params,
}
base_params = Specifications(
output_base=base_dir,
baseline_dir=base_dir,
baseline=True,
num_workers=num_workers,
)
base_params.update_specifications(
json.load(
open(
os.path.join(
"..", "..", "ogusa", "ogusa_default_parameters.json"
)
)
)
)
base_params.update_specifications(base_spec)
BW = TC_LAST_YEAR - start_year + 1
base_params.BW = BW
# Will need to figure out how to handle default tax functions here
# For now, estimating tax functions even for baseline
c_base = Calibration(
base_params,
iit_reform={},
estimate_tax_functions=True,
data=data,
client=client,
)
# update tax function parameters in Specifications Object
d_base = c_base.get_dict()
# additional parameters to change
updated_txfunc_params = {
"etr_params": d_base["etr_params"],
"mtrx_params": d_base["mtrx_params"],
"mtry_params": d_base["mtry_params"],
"mean_income_data": d_base["mean_income_data"],
"frac_tax_payroll": d_base["frac_tax_payroll"],
}
base_params.update_specifications(updated_txfunc_params)
base_ss = SS.run_SS(base_params, client=client)
utils.mkdirs(os.path.join(base_dir, "SS"))
base_ss_dir = os.path.join(base_dir, "SS", "SS_vars.pkl")
with open(base_ss_dir, "wb") as f:
pickle.dump(base_ss, f)
if time_path:
base_tpi = TPI.run_TPI(base_params, client=client)
tpi_dir = os.path.join(base_dir, "TPI", "TPI_vars.pkl")
with open(tpi_dir, "wb") as f:
pickle.dump(base_tpi, f)
else:
base_tpi = None
# Solve reform model
reform_spec = base_spec
reform_spec.update(adjustment["OG-USA Parameters"])
reform_params = Specifications(
output_base=reform_dir,
baseline_dir=base_dir,
baseline=False,
num_workers=num_workers,
)
reform_params.update_specifications(
json.load(
open(
os.path.join(
"..", "..", "ogusa", "ogusa_default_parameters.json"
)
)
)
)
reform_params.update_specifications(reform_spec)
reform_params.BW = BW
c_reform = Calibration(
reform_params,
iit_reform=iit_mods,
estimate_tax_functions=True,
data=data,
client=client,
)
# update tax function parameters in Specifications Object
d_reform = c_reform.get_dict()
# additional parameters to change
updated_txfunc_params = {
"etr_params": d_reform["etr_params"],
"mtrx_params": d_reform["mtrx_params"],
"mtry_params": d_reform["mtry_params"],
"mean_income_data": d_reform["mean_income_data"],
"frac_tax_payroll": d_reform["frac_tax_payroll"],
}
reform_params.update_specifications(updated_txfunc_params)
reform_ss = SS.run_SS(reform_params, client=client)
utils.mkdirs(os.path.join(reform_dir, "SS"))
reform_ss_dir = os.path.join(reform_dir, "SS", "SS_vars.pkl")
with open(reform_ss_dir, "wb") as f:
pickle.dump(reform_ss, f)
if time_path:
reform_tpi = TPI.run_TPI(reform_params, client=client)
else:
reform_tpi = None
comp_dict = comp_output(
base_params,
base_ss,
reform_params,
reform_ss,
time_path,
base_tpi,
reform_tpi,
)
# Shut down client and make sure all of its references are
# cleaned up.
client.close()
del client
return comp_dict
def chi_estimate(
income_tax_params,
ss_params,
iterative_params,
chi_guesses,
baseline_dir="./OUTPUT",
):
"""
--------------------------------------------------------------------
This function calls others to obtain the data momements and then
runs the simulated method of moments estimation by calling the
minimization routine.
INPUTS:
income_tax_parameters = length 4 tuple, (analytical_mtrs, etr_params, mtrx_params, mtry_params)
ss_parameters = length 21 tuple, (J, S, T, BW, beta, sigma, alpha, Z, delta, ltilde, nu, g_y,\
g_n_ss, tau_payroll, retire, mean_income_data,\
h_wealth, p_wealth, m_wealth, b_ellipse, upsilon)
iterative_params = [2,] vector, vector with max iterations and tolerance
for SS solution
chi_guesses = [J+S,] vector, initial guesses of chi_b and chi_n stacked together
baseline_dir = string, path where baseline results located
OTHER FUNCTIONS AND FILES CALLED BY THIS FUNCTION:
wealth.compute_wealth_moments()
labor.labor_data_moments()
minstat()
OBJECTS CREATED WITHIN FUNCTION:
wealth_moments = [J+2,] array, wealth moments from data
labor_moments = [S,] array, labor moments from data
data_moments = [J+2+S,] array, wealth and labor moments stacked
bnds = [S+J,] array, bounds for parameter estimates
chi_guesses_flat = [J+S,] vector, initial guesses of chi_b and chi_n stacked
min_arg = length 6 tuple, variables needed for minimizer
est_output = dictionary, output from minimizer
chi_params = [J+S,] vector, parameters estimates for chi_b and chi_n stacked
objective_func_min = scalar, minimum of statistical objective function
OUTPUT:
./baseline_dir/Calibration/chi_estimation.pkl
RETURNS: chi_params
--------------------------------------------------------------------
"""
# unpack tuples of parameters
(
J,
S,
T,
BW,
beta,
sigma,
alpha,
Z,
delta,
ltilde,
nu,
g_y,
g_n_ss,
tau_payroll,
tau_bq,
rho,
omega_SS,
lambdas,
imm_rates,
e,
retire,
mean_income_data,
h_wealth,
p_wealth,
m_wealth,
b_ellipse,
upsilon,
) = ss_params
chi_b_guess, chi_n_guess = chi_guesses
flag_graphs = False
# specify bootstrap iterations
n = 10000
# Generate Wealth data moments
scf, data = wealth.get_wealth_data()
wealth_moments = wealth.compute_wealth_moments(scf, lambdas, J)
# Generate labor data moments
cps = labor.get_labor_data()
labor_moments = labor.compute_labor_moments(cps, S)
# combine moments
data_moments = list(wealth_moments.flatten()) + list(
labor_moments.flatten()
)
# determine weighting matrix
optimal_weight = False
if optimal_weight:
VCV_wealth_moments = wealth.VCV_moments(scf, n, lambdas, J)
VCV_labor_moments = labor.VCV_moments(cps, n, lambdas, S)
VCV_data_moments = np.zeros((J + 2 + S, J + 2 + S))
VCV_data_moments[: J + 2, : J + 2] = VCV_wealth_moments
VCV_data_moments[J + 2 :, J + 2 :] = VCV_labor_moments
W = np.linalg.inv(VCV_data_moments)
# np.savetxt('VCV_data_moments.csv',VCV_data_moments)
else:
W = np.identity(J + 2 + S)
# call minimizer
bnds = np.tile(
np.array([1e-12, None]), (S + J, 1)
) # Need (1e-12, None) S+J times
chi_guesses_flat = list(chi_b_guess.flatten()) + list(
chi_n_guess.flatten()
)
min_args = (
data_moments,
W,
income_tax_params,
ss_params,
iterative_params,
chi_guesses_flat,
baseline_dir,
)
# est_output = opt.minimize(minstat, chi_guesses_flat, args=(min_args), method="L-BFGS-B", bounds=bnds,
# tol=1e-15, options={'maxfun': 1, 'maxiter': 1, 'maxls': 1})
# est_output = opt.minimize(minstat, chi_guesses_flat, args=(min_args), method="L-BFGS-B", bounds=bnds,
# tol=1e-15)
# chi_params = est_output.x
# objective_func_min = est_output.fun
#
# # pickle output
# utils.mkdirs(os.path.join(baseline_dir, "Calibration"))
# est_dir = os.path.join(baseline_dir, "Calibration/chi_estimation.pkl")
# pickle.dump(est_output, open(est_dir, "wb"))
#
# # save data and model moments and min stat to csv
# # to then put in table of paper
chi_params = chi_guesses_flat
chi_b = chi_params[:J]
chi_n = chi_params[J:]
chi_params_list = (chi_b, chi_n)
ss_output = SS.run_SS(
income_tax_params,
ss_params,
iterative_params,
chi_params_list,
True,
baseline_dir,
)
model_moments = calc_moments(ss_output, omega_SS, lambdas, S, J)
# # make dataframe for results
# columns = ['data_moment', 'model_moment', 'minstat']
# moment_fit = pd.DataFrame(index=range(0,J+2+S), columns=columns)
# moment_fit = moment_fit.fillna(0) # with 0s rather than NaNs
# moment_fit['data_moment'] = data_moments
# moment_fit['model_moment'] = model_moments
# moment_fit['minstat'] = objective_func_min
# est_dir = os.path.join(baseline_dir, "Calibration/moment_results.pkl")s
# moment_fit.to_csv(est_dir)
# calculate std errors
h = 0.0001 # pct change in parameter
model_moments_low = np.zeros((len(chi_params), len(model_moments)))
model_moments_high = np.zeros((len(chi_params), len(model_moments)))
chi_params_low = chi_params
chi_params_high = chi_params
for i in range(len(chi_params)):
chi_params_low[i] = chi_params[i] * (1 + h)
chi_b = chi_params_low[:J]
chi_n = chi_params_low[J:]
chi_params_list = (chi_b, chi_n)
ss_output = SS.run_SS(
income_tax_params,
ss_params,
iterative_params,
chi_params_list,
True,
baseline_dir,
)
model_moments_low[i, :] = calc_moments(
ss_output, omega_SS, lambdas, S, J
)
chi_params_high[i] = chi_params[i] * (1 + h)
chi_b = chi_params_high[:J]
chi_n = chi_params_high[J:]
chi_params_list = (chi_b, chi_n)
ss_output = SS.run_SS(
income_tax_params,
ss_params,
iterative_params,
chi_params_list,
True,
baseline_dir,
)
model_moments_high[i, :] = calc_moments(
ss_output, omega_SS, lambdas, S, J
)
deriv_moments = (
np.asarray(model_moments_high) - np.asarray(model_moments_low)
).T / (2.0 * h * np.asarray(chi_params))
VCV_params = np.linalg.inv(
np.dot(np.dot(deriv_moments.T, W), deriv_moments)
)
std_errors_chi = (np.diag(VCV_params)) ** (1 / 2.0)
sd_dir = os.path.join(baseline_dir, "Calibration/chi_std_errors.pkl")
with open(sd_dir, "wb") as f:
pickle.dump(std_errors_chi, f)
np.savetxt("chi_std_errors.csv", std_errors_chi)
return chi_params
def minstat(chi_guesses, *args):
"""
--------------------------------------------------------------------
This function generates the weighted sum of squared differences
between the model and data moments.
INPUTS:
chi_guesses = [J+S,] vector, initial guesses of chi_b and chi_n stacked together
arg = length 6 tuple, variables needed for minimizer
OTHER FUNCTIONS AND FILES CALLED BY THIS FUNCTION:
SS.run_SS()
calc_moments()
OBJECTS CREATED WITHIN FUNCTION:
ss_output = dictionary, variables from SS of model
model_moments = [J+2+S,] array, moments from the model solution
distance = scalar, weighted, squared deviation between data and model moments
RETURNS: distance
--------------------------------------------------------------------
"""
(
data_moments,
W,
income_tax_params,
ss_params,
iterative_params,
chi_params,
baseline_dir,
) = args
(
J,
S,
T,
BW,
beta,
sigma,
alpha,
Z,
delta,
ltilde,
nu,
g_y,
g_n_ss,
tau_payroll,
tau_bq,
rho,
omega_SS,
lambdas,
imm_rates,
e,
retire,
mean_income_data,
h_wealth,
p_wealth,
m_wealth,
b_ellipse,
upsilon,
) = ss_params
chi_b = chi_guesses[:J]
chi_n = chi_guesses[J:]
chi_params = (chi_b, chi_n)
ss_output = SS.run_SS(
income_tax_params,
ss_params,
iterative_params,
chi_params,
True,
baseline_dir,
)
model_moments = calc_moments(ss_output, omega_SS, lambdas, S, J)
# distance with levels
distance = np.dot(
np.dot((np.array(model_moments) - np.array(data_moments)).T, W),
np.array(model_moments) - np.array(data_moments),
)
# distance = ((np.array(model_moments) - np.array(data_moments))**2).sum()
print("DATA and MODEL DISTANCE: ", distance)
# # distance with percentage diffs
# distance = (((model_moments - data_moments)/data_moments)**2).sum()
return distance
def runner(p, time_path=True, client=None):
'''
This function runs the OG-Core model, solving for the steady-state
and (optionally) the time path equilibrium.
Args:
p (Specifications object): model parameters
time_path (bool): whether to solve for the time path equilibrium
client (Dask client object): client
Returns:
None
'''
tick = time.time()
# Create output directory structure
ss_dir = os.path.join(p.output_base, "SS")
tpi_dir = os.path.join(p.output_base, "TPI")
dirs = [ss_dir, tpi_dir]
for _dir in dirs:
try:
print("making dir: ", _dir)
os.makedirs(_dir)
except OSError:
pass
print('In runner, baseline is ', p.baseline)
'''
------------------------------------------------------------------------
Run SS
------------------------------------------------------------------------
'''
ss_outputs = SS.run_SS(p, client=client)
'''
------------------------------------------------------------------------
Pickle SS results
------------------------------------------------------------------------
'''
utils.mkdirs(os.path.join(p.output_base, "SS"))
ss_dir = os.path.join(p.output_base, "SS", "SS_vars.pkl")
with open(ss_dir, "wb") as f:
pickle.dump(ss_outputs, f)
print('JUST SAVED SS output to ', ss_dir)
# Save pickle with parameter values for the run
param_dir = os.path.join(p.output_base, "model_params.pkl")
with open(param_dir, "wb") as f:
cloudpickle.dump((p), f)
if time_path:
'''
------------------------------------------------------------------------
Run the TPI simulation
------------------------------------------------------------------------
'''
tpi_output = TPI.run_TPI(p, client=client)
'''
------------------------------------------------------------------------
Pickle TPI results
------------------------------------------------------------------------
'''
tpi_dir = os.path.join(p.output_base, "TPI")
utils.mkdirs(tpi_dir)
tpi_vars = os.path.join(tpi_dir, "TPI_vars.pkl")
with open(tpi_vars, "wb") as f:
pickle.dump(tpi_output, f)
print("Time path iteration complete.")
print("It took {0} seconds to get that part done.".format(time.time() -
tick))