An indicator variable encodes whether a condition holds using two values, conventionally 1 and 0. It converts a continuous or categorical question into a binary one: for example, whether in-game purchases exceed $120. section 8.3
In causal analysis, a binary treatment indicator represents alternative interventions such as treatment versus no treatment. In linear regression, its coefficient can represent the average treatment effect when the model and adjustment assumptions are appropriate. section 11.4.1
Dummy variables also support workflow checks rather than only primary modeling:
- Add a random common cause to test whether estimated effects remain stable; a non-confounding variable should not materially change the ATE, although it can add noise. section 11.5.2
- Replace treatment with a permuted placebo treatment; the expected effect is zero, so a changed estimate can expose estimator or identification problems. section 11.5.3
- Generate a synthetic outcome from covariates; its treatment effect should be zero, providing another check on causal assumptions. section 11.5.4
Listing 11.17 Replacing the treatment variable with a dummy variable
identified_estimand.set_identifier_method("backdoor")
res_placebo = model.refute_estimate(
identified_estimand, #1
causal_estimate_ipw, #1
method_name="placebo_treatment_refuter", #1
placebo_type="permute", #1
num_simulations=100 #1
)
print(res_placebo)
A Python-style listing shows how to permute treatment, run repeated placebo simulations, and inspect the estimated and new effects. section 11.5.3
