Guide · foundations
Identification comes before estimation
Estimation asks: given that this quantity is recoverable, what is it, and how precisely? Identification asks the prior question — is it recoverable at all? The two get confused constantly, because every estimator answers the second question whether or not the first one was ever asked.
The distinction
More data does not fix a missing route
An estimation problem gets better with sample size. Wide interval, noisy answer, unstable fit — collect more rows and it improves.
An identification problem does not. If a common cause of treatment and outcome was never measured, then a million rows gives you the same wrong number with a much tighter interval around it. The bias does not shrink; only the honesty does.
This is the case worth being frightened of, because it looks exactly like
success. Below is a real fit from this build, on a graph where the confounder
Z exists but was not observed.
A confident, precise, wrong answer
The truth is 2. The estimate
is off by +0.726, against a standard error
of 0.017 — the bias dwarfs the stated
uncertainty, and nothing in the output says so. Graph:
X -> Y, Z -> X, Z -> Y, measured
X, Y.
What axiom does with this graph
Status
downgraded, route
backdoor_unmeasured, and it names what is missing:
Z — which nobody
recorded. The verdict arrives before the fit, from the graph alone.
The routes
Three ways in, and they need different things
Identification is not one criterion. It is a search over several, each licensed by a different feature of the graph. Knowing which one you are on tells you what would break it.
Back door
Block every non-causal path from treatment to outcome by conditioning on a set that contains no descendant of the treatment. The workhorse: if you measured the common causes, this is your route.
Needs: the confounders measured.
Breaks on: one you did not think of.
this build:
backdoor ·
adjust on Z
Front door
When the confounder is unmeasured but the whole effect passes through a mediator you did measure, the effect can be recovered in two stages — treatment to mediator, mediator to outcome — without ever seeing the confounder.
Needs: a fully mediating variable with no
unblocked back door of its own.
Breaks on: any direct path that skips the mediator.
worked end to end in the epidemiology example
Instrument
A variable that moves the treatment, has no other route to the outcome, and shares no cause with it. It gives you a lever: variation in the treatment that is unrelated to the confounding.
Needs: relevance and exclusion — the second is
untestable and is where the arguments happen.
Breaks on: a weak instrument, or any second path.
this build: 2SLS recovers 1.999 where naive OLS gives 2.347 · truth 2
And sometimes there is no route
On the graph
X -> Y, X <-> Y axiom returns
blocked:
“no back-door, front-door, or instrumental route exists in the graph”. That is not a limitation of the
library — no method recovers this effect from these variables, and a tool that
returned a number here would be inventing one.
The three verdicts
What “identified” actually means, and what it does not
axiom returns a status rather than a boolean, because the interesting cases are the ones in between.
The habit worth building is to run identification first, on
the graph, before any data is loaded. It costs seconds, it needs no sample, and when
it says blocked it has saved you the entire analysis and told you what to
go and get. When the depot tutorial does exactly this, the answer is that the panel on
hand cannot support the decision — and the rest of that tutorial is about designing the
experiment that can.
One more thing it does not mean
Identified here is not identified everywhere
Identification is a statement about recovering an effect in the population the data came from. It says nothing about whether that effect applies to the population you are about to decide for. Those are different questions with different criteria, and the second one is skipped far more often — it is the subject of why a model does not travel, four guides on.