Ignorability and Unconfoundedness

You want to know whether a trading desk's new risk dashboard reduces losses. You can't run a randomized trial — traders chose whether to use it. So you compare users to non-users, controlling for experience and desk size. This comparison is only valid causally if, once you've controlled for the right variables, the remaining reason some traders adopted the dashboard and others didn't has nothing to do with how much they'd have lost either way. That "nothing left over that matters" condition is ignorability, and it cannot be checked directly from data — it has to be argued for.

An analogy: a fair lottery, after you look at the ticket stubs

Imagine a raffle where prize eligibility depends on ticket color, but ticket color was assigned by a machine that also (secretly) depended on the buyer's shoe size. If you don't know about the shoe-size link, comparing winners to losers looks unbiased — after all, color was "randomly" assigned. But it wasn't unconditionally random; it was random given shoe size. If shoe size also affects the outcome you care about, your comparison is contaminated unless you control for shoe size too. Ignorability says: once you've listed and controlled for every such hidden lever, whatever's left over in who-got-treated really is as good as a coin flip with respect to the outcome.

The assumption, one symbol at a time

Let $Y_i(1)$ and $Y_i(0)$ be trader $i$'s potential loss with and without the dashboard, $D_i$ whether they adopted it, and $X_i$ a vector of observed covariates (experience, desk, book size). Ignorability (also called unconfoundedness or conditional independence) states:

$$(Y_i(1), Y_i(0)) \perp D_i \mid X_i$$

In plain English: once you know a trader's covariates $X_i$, knowing whether they adopted the dashboard tells you nothing extra about what their loss would have been under either choice. Treatment assignment is "as good as random" within each slice of covariates — even though across the whole population it was clearly not random at all.

A companion condition, overlap (or positivity), requires

$$0 < P(D_i = 1 \mid X_i) < 1 \quad \text{for all } X_i,$$

meaning every covariate profile has some chance of appearing in both the treated and untreated groups — you can't estimate an effect for a type of trader who always adopts or never adopts, because there's no comparison group for them.

Worked example 1: when it plausibly holds

Suppose desk managers assign the dashboard based purely on a fixed rule: anyone with under 3 years' experience gets it (junior traders need the guardrails), everyone else doesn't. If loss potential depends on experience but, within an experience bracket, has no other systematic link to dashboard use, then controlling for years of experience ($X_i$ = experience) achieves ignorability. Comparing 2-year traders with the dashboard to other 2-year traders without it (there are none here — that's the catch) would be valid; the real issue is overlap fails, because the rule is deterministic in $X$. This is why a hard experience cutoff, while conceptually simple, actually breaks the assumption you need — see regression discontinuity design for how to still extract a valid comparison right at the cutoff.

Worked example 2: when it plausibly fails

Now suppose adoption instead depends on each trader's self-assessed risk appetite — cautious traders opt in, aggressive ones skip it, and this self-assessment is not fully captured by experience, desk, or book size. Two traders with identical $X_i$ (same experience, same desk, same book) can still differ systematically in $D_i$ because of unobserved risk appetite, and risk appetite plausibly also predicts loss directly. Now $(Y_i(1), Y_i(0)) \not\perp D_i \mid X_i$: ignorability fails, and a covariate-adjusted comparison will still be biased, because the unmeasured "hidden lever" (risk appetite) is a confounder — see confounding and the backdoor criterion for how such hidden common causes bias comparisons.

What this means in practice

Ignorability is the assumption every observational causal method — matching, propensity scores, regression adjustment — silently relies on. It is never testable directly, because it's a statement about potential outcomes, and you only ever observe one potential outcome per unit. What you can do is argue for it using domain knowledge (which variables plausibly drive both treatment and outcome, drawn out as a DAG — see directed acyclic graphs for causality), check overlap in the data, and run sensitivity analyses asking "how strong would a hidden confounder need to be to overturn my result?"