def test_get_mc0_precision():
std_theta = 1
var_epsilon = 1
n_students = 20
n_classes = 10
errors = np.zeros(n_classes)
precisions = np.zeros(n_classes)
for class_ in range(n_classes):
mu_1 = np.random.normal()
mu_0 = np.random.normal()
df = construct_class_df(0, 0, n_students, std_theta, mu_0, mu_1, var_epsilon)
collapsed = calculate_va_continuous.collapse(df)
collapsed['m_c1'] = calculate_va_continuous.get_mc1(df)
mc_0 = calculate_va_continuous.get_mc0(collapsed)
errors[class_] = mc_0 - mu_0
precisions[class_] = calculate_va_continuous.get_mc0_precision(std_theta**2, var_epsilon, df['continuous var'].values)
check_calibration(errors, precisions)
def test_get_mc0():
var_epsilon = abs(np.random.normal())
std_theta = abs(np.random.normal())
n_students = 25
n_classes = 1000
errors = np.zeros(n_classes)
for class_ in range(n_classes):
mu_1 = np.random.normal()
mu_0 = np.random.normal()
df = construct_class_df(0, 0, n_students, std_theta, mu_0, mu_1, var_epsilon)
df = calculate_va_continuous.collapse(df)
df['m_c1'] = mu_1 # Just assume this is correctly estimated
mc_0 = calculate_va_continuous.get_mc0(df)
errors[class_] = mc_0 - mu_0
error_se = (np.var(errors) / n_classes)**(.5)
mean_error = np.mean(errors)
assert mean_error > -2 * error_se
assert mean_error < 2 * error_se