Repository navigation
A dataset reform simulation's baseline loses inputs the reform neutralizes #632
Description
Problem
When a simulation built from a dataset applies a reform that neutralizes an input variable, its baseline loses that input as well. The baseline then reads the variable's default, so reform-vs-baseline comparisons describe the wrong baseline.
Simulation.__init__ builds the reform simulation from the dataset under the reform's system. Holder.set_input ignores a value for a neutralized variable (holders/holder.py:460). The baseline is then get_branch("baseline"), rebound to the baseline system by _bind_to_tax_benefit_system (simulations/simulation.py:405-413 in 3.32.27). So the branch inherits holders that never received the column. The rebinding makes the variable un-neutralized again, but there is no data left to read.
#613 fixes the same gap for situations: it builds the baseline from the situation under the baseline system. It keeps the old binding when there is no situation (a dataset, or populations given), so the dataset path still has the bug.
Reproduction (core 3.32.27; master 037960c has the same set_input and baseline code)
import warnings
import numpy as np
from policyengine_core.entities import build_entity
from policyengine_core.model_api import Variable, YEAR
from policyengine_core.simulations import Simulation
from policyengine_core.taxbenefitsystems import TaxBenefitSystem
Person = build_entity("person", "persons", "Person", is_person=True)
class person_id(Variable):
value_type = int
entity = Person
definition_period = YEAR
label = "Person ID"
class hours(Variable):
value_type = float
entity = Person
definition_period = YEAR
label = "Hours, an input"
class System(TaxBenefitSystem):
def __init__(self, reform=None):
super().__init__([Person])
self.add_variables(person_id, hours)
if reform is not None:
self.apply_reform_set(reform)
class Sim(Simulation):
default_tax_benefit_system = System
default_tax_benefit_system_instance = System()
class Data:
name = "repro"
data_format = "arrays"
time_period = "2024"
file_path = None
def load(self):
return {"person_id": np.array([1, 2]), "hours": np.array([10.0, 20.0])}
class NeutralizeHours:
def apply(self, system):
system.neutralize_variable("hours")
with warnings.catch_warnings():
warnings.simplefilter("ignore")
reformed = Sim(dataset=Data(), reform=NeutralizeHours())
print("plain simulation: ", Sim(dataset=Data()).calculate("hours", 2024).tolist())
print("reform: ", reformed.calculate("hours", 2024).tolist())
print("reform's baseline: ", reformed.baseline.calculate("hours", 2024).tolist())Output:
plain simulation: [10.0, 20.0]
reform: [0.0, 0.0]
reform's baseline: [0.0, 0.0]
The baseline should read [10.0, 20.0], as the plain simulation does.
Possible fix
Build a dataset reform simulation's baseline branch by loading the dataset under the baseline system, as #613 does for situations. Alternatively, keep the dataset's values for neutralized variables, without exposing them to the reform, so the baseline branch can read them.
Found while reviewing PolicyEngine/policyengine-us#10447. A reform that neutralizes unemployment_compensation_start_month makes both the reform and its baseline treat the column as unsupplied.
🤖 Generated with Claude Code
Activity
MaxGhenis commentedon Oct 26, 2026
Still reproduces on 3.33.3, the version policyengine.py bundles, and on 3.33.12, the latest. Each run used the script above, with output identical to 3.32.27's:
At 3.33.3:
simulations/simulation.py:786-790, thenbuild_from_datasetat:880). The script's system applies its reform in its constructor, as country systems do. Sohoursis already neutralized when the dataset loads, andHolder.set_inputwarns and stores nothing (holders/holder.py:603-607).simulations/simulation.py:912-921; 3.32.27 does the same at:404-413). But that branch is a clone of the reformed simulation (get_branch,:2934-2943), and the clone has nohoursinput to read.So the fix still needs the baseline to get the dataset's inputs for variables the reform neutralizes, for example by loading the dataset before neutralizing or by keeping the neutralized inputs.