Targeted minimum loss-based estimation

Update initial outcome predictions in a direction determined by treatment probabilities, then average the targeted contrasts.

What does the TMLE targeting step do?

TMLE starts with outcome predictions and fits a targeted update using observed outcomes and propensity scores. For the average treatment effect here, the update removes the sample’s average signed, propensity-weighted prediction error.

Try it in the experiment

  1. Inspect the initial predictions and correction left to make.
  2. Move the targeting slider and watch both prediction curves change.
  3. Apply the full update and inspect the remaining correction.

A solved targeting equation is an estimation property, not proof of a valid causal estimate. TMLE still needs confounding control, consistency, overlap, and adequate nuisance estimation under the required regularity conditions.