supatuffpinkpuff
commented
1 year ago I'm encountering this issue, but I think I've determined a cause.
It clearly has to do with line 82, in find_pe_for_covar_set, in flame_dame_helpers.py binarized_df = pd.get_dummies(x_treated.loc[:, non_bool_cols].astype(str))
Based on the error message coming from pd.get_dummies, this seems to indicate that pd.get_dummies is being passed an empty dataframe. As to why that's happening, I think it has to do with the precise number of columns that are boolean/non boolean used for analysis. Since the code is iteratively removing columns, potentially including the non-boolean columns, at some point depending on the dataset, there might be an iteration where there are no non-boolean columns, in which case pd.get_dummies is passed an empty dataframe leading to this issue.
I believe a fix would involve adding an additional check before running pd.get_dummies to make sure that it's not being passed an empty dataframe. Perhaps adding
if len(non_bool_cols) > 0:
before each call of pd.get_dummies would suffice? Will do some more testing tomorrow.
I made some fake datasets as test cases, and found the following:
The below dataset has only boolean integer columns, and fails with this error on iteration 1. data = pd.DataFrame({'bool_integer_categories1':[1, 1, 0, 0, 1], 'treatment':[0, 0, 0, 1, 1], 'outcome':[5, 1, 2, 3, 4], 'bool_integer_categories2':[1, 0, 1, 0, 0]})
The below dataset has only non-boolean integer columns, and succeeds after two iterations. data = pd.DataFrame({'nonbool_integer_categories1':[100, 100, 10, 10, 1], 'treatment':[0, 0, 0, 1, 1], 'outcome':[5, 1, 2, 3, 4], 'nonbool_integer_categories2':[1, 2, 3, 1, 2]})
The below dataset has one non-boolean integer column and one boolean integer column, and fails on iteration 1. data = pd.DataFrame({'nonbool_integer_categories':[100, 10, 10, 50, 50], 'treatment':[0, 0, 0, 1, 1], 'outcome':[5, 1, 2, 3, 4], 'boolean_integer_column':[0, 0, 1, 1, 0]})
Adding another boolean integer column, still fails on iteration 1. data = pd.DataFrame({'nonbool_integer_categories':[100, 10, 10, 50, 50], 'treatment':[0, 0, 0, 1, 1], 'outcome':[5, 1, 2, 3, 4], 'boolean_integer_column':[0, 0, 1, 1, 0], 'boolean_int_col_2':[1, 0, 1, 0, 1]})
However, adding more non_boolean columns still fails, but survives another iteration to iteration 2. data = pd.DataFrame({'nonbool_integer_col_1':[100, 10, 10, 50, 50], 'nonbool_integer_col_2':[1, 1, 2, 2, 3], 'treatment':[0, 0, 0, 1, 1], 'outcome':[5, 1, 2, 3, 4], 'boolean_integer_column':[0, 0, 1, 1, 0]})
Code, Logs, and Traceback from the last example below: `def test_mixed_bool_more_cols_1():
data = pd.DataFrame({'nonbool_integer_col_1':[100, 10, 10, 50, 50],
'nonbool_integer_col_2':[1, 1, 2, 2, 3],
'treatment':[0, 0, 0, 1, 1],
'outcome':[5, 1, 2, 3, 4],
'boolean_integer_column':[0, 0, 1, 1, 0]})
# return data
model_flame = dame_flame.matching.FLAME(repeats=False, verbose=3, adaptive_weights='decisiontree')
model_flame.fit(holdout_data=data, treatment_column_name='treatment', outcome_column_name='outcome')
result_flame = model_flame.predict(data)
print('ATE:')
print(dame_flame.utils.post_processing.ATE(model_flame))
# Visualizing CATE of matched groups from FLAME.
group_size_treated = []
group_size_overall = []
cate_of_group = []
for group in model_flame.units_per_group:
# find len of just treated units
df_mmg = data.loc[group]
treated = df_mmg.loc[df_mmg["treatment"] == 1]
group_size_treated.append(len(treated))
group_size_overall.append(len(group))
cate_of_group.append(dame_flame.utils.post_processing.CATE(model_flame, group[0]))
plt.scatter(group_size_treated, cate_of_group, alpha=0.25, edgecolors='b')
plt.axhline(y=0.0, color='r', linestyle='-')
plt.xlabel('Number of Treatment units in group', fontsize=12)
plt.ylabel('Estimated Treatment Effect of Group', fontsize=12)
plt.title("Visualizing CATE of matched groups from FLAME", fontsize=14)
plt.savefig('interpretability.png')` Iteration number: 1 Number of matched groups formed in total: 0 Unmatched treated units: 2 out of a total of 2 treated units Unmatched control units: 3 out of a total of 3 control units Predictive error of covariates chosen this iteration: 0 Number of matches made in this iteration: 0 Number of matches made so far: 0 In this iteration, the covariates dropped are: set() Iteration number: 2 Number of matched groups formed in total: 0 Unmatched treated units: 2 out of a total of 2 treated units Unmatched control units: 3 out of a total of 3 control units Predictive error of covariates chosen this iteration: 0.0 Number of matches made in this iteration: 0 Number of matches made so far: 0 In this iteration, the covariates dropped are: nonbool_integer_col_2
Error
Traceback (most recent call last):
File "test_mixed_bool_more_cols_1", line 1, in
nehargupta
When running with the adaptive_weights='decisionTreeCV' parameter, in late iterations, sometimes the error 'no object to concatenate' appears. It seems like users who see this error can avoid it by using the early_stop_iterations parameter, or by using another learning method than decision trees, and possibly by binarizing columns before running the algorithm as well