In the 1990s and early 2000s, observational studies consistently reported that hormone replacement therapy reduced the risk of coronary heart disease by 40 to 50 percent in postmenopausal women. The studies were well-designed by contemporary standards, controlling for age, smoking, diet, exercise, and numerous other confounders. Physicians began prescribing HRT for cardiovascular protection based on this evidence. Then in 2002, the Women's Health Initiative—a large randomized controlled trial—reported the opposite: HRT increased cardiovascular risk. The trial stopped early due to harm. This was not careless methodology or naive analysis. Investigators controlled for the confounders they could measure. The problem was what they could not measure: socioeconomic status, health-seeking behavior, dietary quality, and exercise engagement were concentrated among HRT users but not fully captured in the observational data. This residual confounding was sufficient to flip the sign of the conclusion entirely—from apparent benefit to actual harm.

Every method we have developed in causal inference—backdoor adjustment, double machine learning, causal forests, counterfactual reasoning—rests on a single, hidden foundation: the assumption of unconfoundedness. This assumption says that we have measured all the important confounders; that once we condition on the right variables, the remaining association between treatment and outcome is genuinely causal. When this assumption holds, our methods work. When it fails—even partially—our estimates become biased, sometimes drastically. The HRT story shows the real-world stakes: a recommendation reversal with consequences for millions of patients. An analyst who cannot identify the limits of observational evidence, quantify the threat of unmeasured confounding, and communicate those limits honestly is an analyst who cannot distinguish between a robust causal conclusion and one that will collapse under scrutiny. This chapter is about building that capability.

The unconfoundedness assumption—also called conditional ignorability or selection on observables—states that once you condition on a sufficient set of measured variables, treatment assignment is independent of potential outcomes. In other words: once you control for the right confounders, the association you observe between treatment and outcome is causal, not confounded. This is not a testable assumption. You cannot examine your data and prove it true. You can only argue that it is plausible based on domain knowledge, the structure of the causal graph, and the completeness of your measurement. The assumption does real work: without it, the backdoor adjustment formula produces a biased estimate; causal forests estimate contaminated effects; every method in the toolkit fails. It is the bedrock assumption, and it must hold—at least approximately—for the analysis to be credible.

Here is the hard truth: you rarely measure all the important confounders. Some confounders are known unknowns—you know they exist but cannot measure them, like a patient's unmeasured genetic predisposition. Others are unknown unknowns—you do not know they are confounders until after the analysis. The HRT case had both. Socioeconomic status and health-seeking behavior are not exotic variables; they are standard confounders. But in the observational data, they were only partially captured through proxies like education and income. The residual confounding—the confounding from unmeasured or partially measured variables—was enough to reverse the sign of the effect. The gap between what you measure and what actually confounds the relationship is not a statistical problem you can solve with a bigger sample or a more sophisticated algorithm. It is a structural problem. You cannot control for what you did not measure.

Confounding bias arises when a variable affects both treatment assignment and the outcome, creating a non-causal association between them. In causal graph terms, a confounder is a variable with arrows pointing to both treatment and outcome—a common cause. When you do not condition on this variable, a backdoor path opens: treatment ← confounder → outcome. Association flows through this path even if treatment has no causal effect. Adjusting for the confounder blocks the backdoor path and isolates the causal effect. But adjustment only works if you have measured the confounder and included it in your model. If the confounder is unmeasured, the backdoor path remains open, and bias persists. The magnitude of the bias depends on two things: how strongly the confounder predicts treatment, and how strongly it predicts the outcome. The HRT case is a vivid example: socioeconomic status strongly predicted who took HRT and strongly predicted cardiovascular outcomes. That dual association created substantial bias.

Confounding is not the only way selection bias enters observational data. There is another: collider bias. A collider is a variable that has arrows pointing to it from two or more other variables—a common effect. If you condition on a collider, you create a spurious association between its parents. In causal graph terms: if X and Y both point to Z, a collider, and you condition on Z, then X and Y become associated even if they are not causally related. This is the opposite of confounding. Confounding bias enters when you do not condition on a confounder; collider bias enters when you do condition on a collider. This distinction matters. Identifying confounders and colliders in your graph is the first step to knowing whether to adjust, not adjust, or proceed with caution. The HRT analysis adjusted appropriately for confounders but did not fall victim to collider bias—the core problem was unmeasured confounding, not improper conditioning.

Sensitivity analysis is the tool for answering this question: how strong must an unmeasured confounder be in order to overturn my causal conclusion? The answer is expressed as an E-value or a Cornfield-style bound. The E-value is a ratio that answers: what is the minimum prevalence ratio or odds ratio between the unmeasured confounder and both treatment and outcome that would be needed to explain away the observed association? A higher E-value means your conclusion is more robust to unmeasured confounding; you would need a very strong confounder to flip the sign of the effect. A lower E-value means your conclusion is fragile; even a weak confounder could overturn it. For example, if your estimated causal effect has an E-value of 3, then an unmeasured confounder would need to be 3 times more prevalent in the treatment group than the control group, and also increase the outcome risk 3-fold, just to move the estimate to the null. Sensitivity analysis does not prove that your conclusion is unconfounded; it quantifies what you do not know.

Several conditions heighten the threat of unmeasured confounding. First: large, unexplained gaps between observational and experimental evidence. If an RCT contradicts your observational estimate by more than random chance would allow, unmeasured confounding is the leading explanation. Second: effect sizes that are implausibly large or directionally surprising given theory. If you estimate a 50 percent harm reduction for an intervention with a weak biological mechanism, unmeasured confounding is more plausible than a true effect. Third: observational studies that consistently exceed the effect sizes of RCTs in the same domain. This pattern is a red flag for selection bias. Fourth: populations or settings where unmeasured confounders are known to be large—for example, studies of people who self-select into a treatment. When these conditions hold, a single observational study is not sufficient. Triangulation becomes essential: use multiple design approaches and ask whether they point to the same conclusion. Credibility increases when different methods, each with different assumptions and vulnerabilities, agree.

The gap between observational evidence and experimental truth exists because of unmeasured confounding—and the tools for managing it form a coherent whole. First, causal graphs help you identify confounders and colliders, telling you which variables to control. Second, backdoor adjustment and other conditional methods use those graphs to estimate causal effects, but only under the unconfoundedness assumption. Third, sensitivity analysis quantifies how plausible unmeasured confounding must be to threaten your conclusion. Fourth, triangulation with other designs—each carrying different hidden assumptions—builds confidence when observational methods alone are fragile. None of these tools makes unmeasured confounding go away. But together, they make the analysis honest: they force you to state what you are assuming, to measure how fragile those assumptions make your conclusion, and to acknowledge when a single observational study is too weak to stand alone. The HRT story exemplifies the full cycle: well-designed observational analysis, high initial credibility, large effect size suggesting robustness, then collision with experimental truth.

Observational causal inference works when you have measured the important confounders and your effect is not driven by subtle selection bias. It fails when unmeasured confounders are present. The hard part is knowing which case you are in. You cannot test unconfoundedness from the data alone; you can only argue for it on theoretical grounds. This asymmetry is why every analyst must develop two skills: first, the ability to use causal diagrams to identify and adjust for confounders; second, the ability to recognize when unmeasured confounding is plausible, estimate its magnitude using sensitivity analysis, and decide whether your conclusion is robust. The HRT case teaches both lessons. The observational studies were technically competent. The failure was not in method but in acknowledging the limits of what observational data can credibly show. An honest analyst would have computed E-values, shown that the effect size was surprisingly large, and recommended that the question be settled by experiment. That level of honesty requires not just statistical technique but intellectual humility about what you do not know.

The unconfoundedness assumption cannot be tested from observational data. It is a claim about what the world looks like—whether you have measured all the confounders that matter. You can defend this claim using domain knowledge, causal theory, the completeness of your measurement, and prior credibility. But you cannot prove it true by looking at the data. This is why every analyst must develop a second skill beyond estimation: the skill of recognizing when unmeasured confounding is plausible, measuring its threat using sensitivity analysis, and deciding whether your conclusion is robust enough to act on. The HRT case shows what happens when this skill is absent. Well-designed observational analysis, high initial credibility, substantial effect sizes—all of it collapsed when tested experimentally. A sensitivity analysis before the RCT would not have prevented the mistake, but it would have made the analyst honest about the uncertainty. It would have said: this effect is fragile; only a moderately strong unmeasured confounder is needed to reverse it; we should be cautious, and experiment if possible. That is the honesty this chapter teaches.