Imagine evaluating two treatments for a severe illness using real-world hospital records. Treatment A shows a 40% mortality rate, while Treatment B shows a 10% mortality rate. A naive analysis would conclude that Treatment B is vastly superior.
But what if clinicians reserve Treatment A exclusively for patients on the brink of death, while prescribing Treatment B to those with mild, manageable symptoms? In this scenario, Treatment A's high mortality rate isn't necessarily a reflection of its failure; it is a reflection of the patients' baseline risk.
This is the core mechanism of confounding by indication (often manifesting as treatment selection bias or channeling bias). It is one of the most pervasive challenges in observational health research, arising whenever the clinical reason for prescribing a treatment is systematically linked to the patient's likelihood of experiencing the outcome.
The Broken Coin Flip
In a randomized controlled trial (RCT), the link between a patient's health status and their treatment assignment is intentionally broken by a coin flip. Because assignment is random, both visible and hidden patient characteristics balance out across groups.
In observational data, however, treatments are assigned by human decisions. Clinicians use their best judgment, balancing a patient's age, symptom severity, comorbidities, and prognosis to select a therapy. The "indication"—the very reason a drug or surgery is chosen—forges a powerful, non-random link between the treatment and the baseline risk of the patient.
Randomized Controlled Trial:
Patient Baseline Risk - - - (Broken Link) - - -> Treatment Assignment ---> Outcome
Observational Study:
Patient Baseline Risk -------------------------> Treatment Assignment ---> Outcome
\ /
\-----------------------------------------------------------------/
When researchers attempt to evaluate the real-world effectiveness of a drug retrospectively, this backdoor path distorts the truth. As Tianhong Zhang et al. observe in Cognitive functions following initiation of antipsychotic medication..., naturalistic clinical settings inherently bias treatment groups because individuals with more severe baseline symptoms are disproportionately medicated. Without accounting for this, the treatment can falsely appear to cause poor outcomes, when it is simply acting as a marker for initial clinical severity.
Two Sides of the Selection Bias Coin
Confounding by indication typically distorts real-world evidence in one of two directions: making a necessary treatment look harmful, or making an aggressive intervention look miraculously effective.
1. The Illusion of Harm (Sicker Patients Get More Treatment)
When a condition worsens, doctors escalate care. In a retrospective study, this creates a spurious correlation between intensive treatments and adverse outcomes. For example, in an analysis of sepsis patients, Jingjing Meng et al. found a striking association between prescription opioid use and increased mortality in Prescription opioids are associated with higher mortality in patients diagnosed with sepsis.... Decoupling whether the medication itself drove the mortality or whether it merely served as a proxy for severe, late-stage physiological distress is exceptionally difficult without objective, time-resolved measures of pain and baseline acuity.
2. The Healthy User Effect (Fitter Patients Get Major Surgery)
Conversely, selection bias can make an intervention look better than it is. Major surgeries carry immediate risks, so surgeons selectively operate on patients who are robust enough to survive the procedure.
In Long-Term Survival in Men With High Grade Prostate Cancer..., Ashutosh Tewari et al. encountered this when comparing radical prostatectomy to radiation and conservative management. Healthier, younger men were systematically channeled toward surgery. Similarly, the authors of Total hip arthroplasty versus hemiarthroplasty for independently mobile older adults with intracapsular hip fractures note that superior survival rates in observational cohorts often reflect the initial fitness of the patients selected for complex joint replacements rather than the isolated efficacy of the surgical technique itself.
Channeling Bias
A closely related variant is channeling bias, where a patient's specific comorbidity profile dictates drug choice. As Yee-Suk Kim et al. demonstrate in Comparative safety and effectiveness of alendronate versus raloxifene in women with osteoporosis, doctors actively avoid prescribing certain medications to patients with known risk factors (such as avoiding clot-promoting drugs in patients prone to thrombosis). This systematic steering creates treatment groups with fundamentally lopsided baseline safety profiles.
Untangling the Web: Methodologies for Causal Inference
To extract reliable insights from real-world data, researchers use causal inference frameworks designed to mathematically reconstruct the "coin flip" of an RCT.
Propensity Score Matching (PSM)
Instead of comparing raw treatment groups, researchers calculate a patient's propensity score: the probability that they would receive a specific treatment based on all their observed baseline characteristics (age, lab values, comorbidities).
Once these scores are calculated, researchers pair a treated patient with an untreated patient who had the exact same probability of receiving the treatment. For instance, Ke Su utilized this approach in Gamma knife radiosurgery versus transcatheter arterial chemoembolization... to balance baseline health profiles, ensuring that the compared groups were clinically interchangeable before evaluating survival outcomes. The same balancing act was crucial in the Head-to-head comparison of azvudine and nirmatrelvir/ritonavir for the hospitalized patients with COVID-19... study to isolate true therapeutic efficacy from physician prescribing preferences.
Target Trial Emulation
A more comprehensive conceptual framework is target trial emulation, which applies the rigorous design principles of a prospective RCT to the analysis of retrospective data. By explicitly defining inclusion criteria, treatment strategies, and the exact "Time Zero" (the moment treatment eligibility begins), researchers prevent common observational traps like immortal time bias.
By applying this framework, Camilla Pegram et al. showed in Target trial emulation: Do antimicrobials or gastrointestinal nutraceuticals prescribed at first presentation for acute diarrhoea cause a better clinical outcome in dogs... how accounting for initial symptom severity prevents researchers from misattributing natural recovery patterns to medical interventions.
Machine Learning-Based Causal Inference
When the clinical decision-making process involves complex, non-linear interactions between dozens of variables, traditional statistical matching falls short. Modern researchers leverage machine learning to model the treatment assignment mechanism more precisely.
In studies examining serious mental illness care, such as Revisiting diabetes risk of olanzapine versus aripiprazole... by Denis Agniel et al. and Antipsychotics and the risk of diabetes and death... by Sharon-Lise T. Normand et al., advanced machine learning algorithms were deployed to untangle complex prescribing patterns. By accurately capturing how clinicians selectively prescribe metabolically active drugs based on a patient's pre-existing risk profiles, these models successfully isolated the true pharmacological impacts from baseline vulnerabilities.
Summary
Observational health data is a reflection of clinical practice, and clinical practice is inherently non-random. Confounding by indication occurs because doctors treat patients based on need and risk. Understanding why a drug was prescribed is an absolute prerequisite to understanding what that drug actually achieved. By using advanced causal inference techniques like propensity matching and target trial emulation, researchers can mathematically control for these clinical intentions, transforming messy real-world data into reliable, actionable evidence.