Merge pull request #105 from PSLmodels/fixes

Remove negative weights and aggregate records

.gitignore

@@ -11,3 +11,4 @@ tax_microdata_benchmarking/calibration
 !tax_microdata_benchmarking/storage/input/*.csv
 **demographics_2015.csv
 **puf_2015.csv
+*.DS_STORE

Makefile

@@ -12,7 +12,6 @@ flat-file:
 	python tax_microdata_benchmarking/create_taxcalc_input_variables.py
 	python tax_microdata_benchmarking/create_taxcalc_growth_factors.py
 	python tax_microdata_benchmarking/create_taxcalc_sampling_weights.py
-	python tax_microdata_benchmarking/create_all_datasets.py
 
 data: install flat-file test
 

app.py

@@ -0,0 +1,66 @@
+import streamlit as st
+
+from tax_microdata_benchmarking.utils.soi_replication import *
+from tax_microdata_benchmarking.storage import STORAGE_FOLDER
+from tax_microdata_benchmarking.datasets import *
+import pandas as pd
+
+INPUTS = STORAGE_FOLDER / "input"
+OUTPUTS = STORAGE_FOLDER / "output"
+
+
+@st.cache_resource
+def generate_comparsions():
+    tmd_2021 = pd.read_csv(OUTPUTS / "tmd_2021.csv")
+    soi_from_tmd_2021 = compare_soi_replication_to_soi(
+        tc_to_soi(tmd_2021, 2021), 2021
+    )
+    return soi_from_tmd_2021
+
+
+def soi_statistic_passes_quality_test(df):
+    # Relative error lower than this => OK
+    RELATIVE_ERROR_THRESHOLD = 0.05
+
+    # Absolute error lower than this for filer counts => OK
+    COUNT_ABSOLUTE_ERROR_THRESHOLD = 1e6
+
+    # Absolute error lower than this for aggregates => OK
+    AGGREGATE_ABSOLUTE_ERROR_THRESHOLD = 1e9
+
+    relative_error_ok = (
+        df["Absolute relative error"] < RELATIVE_ERROR_THRESHOLD
+    )
+    absolute_error_threshold = np.where(
+        df.Count,
+        COUNT_ABSOLUTE_ERROR_THRESHOLD,
+        AGGREGATE_ABSOLUTE_ERROR_THRESHOLD,
+    )
+    absolute_error_ok = df["Absolute error"] < absolute_error_threshold
+
+    return relative_error_ok | absolute_error_ok
+
+
+# 2021 datasets
+
+comparisons = generate_comparsions()
+
+st.title("SOI replication results")
+
+st.write(
+    """
+    This page shows the results of replicating the SOI dataset from the TMD-2021 output data file. It is sorted by absolute error."""
+)
+
+st.dataframe(comparisons.sort_values("Absolute error", ascending=False))
+
+import plotly.express as px
+
+histogram = px.histogram(
+    comparisons,
+    x="Absolute error",
+    marginal="rug",
+    title="Histogram of absolute relative errors",
+)
+
+st.plotly_chart(histogram)

docs/app.py

@@ -1 +1,105 @@
 import streamlit as st
+
+from tax_microdata_benchmarking.utils.soi_replication import *
+from tax_microdata_benchmarking.storage import STORAGE_FOLDER
+from tax_microdata_benchmarking.datasets import *
+import pandas as pd
+
+INPUTS = STORAGE_FOLDER / "input"
+OUTPUTS = STORAGE_FOLDER / "output"
+
+
+@st.cache_resource
+def generate_comparsions(use_original_weights: bool = False):
+    tmd_2021 = pd.read_csv(OUTPUTS / "tmd_2021.csv")
+    if use_original_weights:
+        tmd_2021.s006 = tmd_2021.s006_original
+    soi_from_tmd_2021 = compare_soi_replication_to_soi(
+        tc_to_soi(tmd_2021, 2021), 2021
+    )
+    return soi_from_tmd_2021
+
+
+def soi_statistic_passes_quality_test(df):
+    # Relative error lower than this => OK
+    RELATIVE_ERROR_THRESHOLD = 0.05
+
+    # Absolute error lower than this for filer counts => OK
+    COUNT_ABSOLUTE_ERROR_THRESHOLD = 1e6
+
+    # Absolute error lower than this for aggregates => OK
+    AGGREGATE_ABSOLUTE_ERROR_THRESHOLD = 1e9
+
+    relative_error_ok = (
+        df["Absolute relative error"] < RELATIVE_ERROR_THRESHOLD
+    )
+    absolute_error_threshold = np.where(
+        df.Count,
+        COUNT_ABSOLUTE_ERROR_THRESHOLD,
+        AGGREGATE_ABSOLUTE_ERROR_THRESHOLD,
+    )
+    absolute_error_ok = df["Absolute error"] < absolute_error_threshold
+
+    return relative_error_ok | absolute_error_ok
+
+
+# 2021 datasets
+
+comparisons = generate_comparsions()
+comparisons_original_weights = generate_comparsions(use_original_weights=True)
+comparisons["Original weight value"] = comparisons_original_weights["Value"]
+comparisons["Original weight error"] = comparisons_original_weights["Error"]
+comparisons["Improved under reweighting"] = (
+    comparisons["Absolute error"] < comparisons["Original weight error"].abs()
+)
+soi_subset = comparisons
+time_period = 2021
+
+soi_subset = soi_subset[soi_subset["Filing status"] == "All"]
+soi_subset = soi_subset[soi_subset["Taxable only"] == False]
+agi_level_targeted_variables = [
+    "adjusted_gross_income",
+    "count",
+]
+aggregate_level_targeted_variables = [
+    # "qualified_business_income_deduction",
+]
+soi_subset = soi_subset[
+    soi_subset.Variable.isin(agi_level_targeted_variables)
+    & (
+        (soi_subset["AGI lower bound"] != -np.inf)
+        | (soi_subset["AGI upper bound"] != np.inf)
+    )
+    | (
+        soi_subset.Variable.isin(aggregate_level_targeted_variables)
+        & (soi_subset["AGI lower bound"] == -np.inf)
+        & (soi_subset["AGI upper bound"] == np.inf)
+    )
+]
+
+comparisons["Targeted"] = False
+comparisons["Targeted"][soi_subset.index] = True
+
+soi_subset["Targeted"] = True
+
+st.title("SOI replication results")
+
+st.write(
+    """
+    This page shows the results of replicating the SOI dataset from the TMD-2021 output data file. It is sorted by absolute error."""
+)
+
+st.dataframe(comparisons.sort_values("Absolute error", ascending=False))
+
+import plotly.express as px
+
+histogram = px.histogram(
+    comparisons,
+    x="Absolute error",
+    marginal="rug",
+    title="Histogram of absolute relative errors",
+)
+
+st.plotly_chart(histogram)
+
+st.subheader("Targets included in reweighting")