The Ethical Implications of Excluding “Messy” Patients: When Data Cleaning Becomes a Moral Decision
Executive Summary
Most applied modeling workflows include an unexamined step:
“We excluded patients with incomplete data.”
This sentence is often presented as a technical necessity.
It is not.
Excluding “messy” patients is an ethical decision that shapes:
- who the model represents,
- who benefits from predictions,
- and where uncertainty is allowed to exist.
This post argues that data exclusion is not value-neutral, and that ethical modeling requires making these choices explicit, justified, and reviewable. In applied health data, missingness and exclusion are often tied to workflow, severity, and access rather than random inconvenience (Rubin 1976; Little and Rubin 2019; Rajkomar et al. 2018).
Who Are “Messy” Patients?
In clinical and operational data, “messy” rarely means random.
Patients with incomplete, irregular, or inconsistent data are often those who:
- are critically ill,
- receive care under time pressure,
- move rapidly between units,
- have limited access to consistent follow-up,
- are treated in under-resourced settings.
Messiness is often a signal of vulnerability, not noise. In real-world clinical systems, the hardest patients to measure are often the very patients for whom decisions are most consequential (National Research Council 2010; Rose et al. 2023).
Exclusion Changes the Population You Are Modeling
When you drop patients with missing or irregular data, you are no longer modeling:
“Patients who present to the system.”
You are modeling:
“Patients who are easy to document under stable conditions.”
This distinction matters.
table(is.na(data$key_predictor), data$outcome)If exclusion is correlated with severity or outcome, then:
- estimates shift,
- uncertainty narrows artificially,
- generalization fails exactly where it is needed most.
Exclusion Is an Estimand Choice
Every analysis answers a question.
Excluding messy patients silently changes it.
Original question:
What is the effect or risk in the target population?
After exclusion:
What is the effect among well-documented patients under this workflow?
Both may be valid. Only one is usually claimed.
Ethical practice requires alignment between the estimand and the claim. Otherwise, an apparently technical exclusion silently changes the target population and the meaning of the result (Hernán and Robins 2020; Holland 1986).
Why “Data Quality” Is Often a Proxy for Privilege
Data completeness is not evenly distributed.
It often reflects:
- staffing ratios,
- institutional resources,
- patient advocacy,
- language access,
- time availability.
Models trained on “clean” data may:
- underperform in resource-limited settings,
- misestimate risk in high-acuity contexts,
- amplify existing disparities.
This is not bias introduced by the algorithm. It is bias introduced by who was allowed into the dataset—a form of representation bias that can precede any model fitting step (Suresh and Guttag 2021; Gianfrancesco et al. 2018).
The False Comfort of Clean Validation
Clean datasets validate clean models.
# Common but misleading
analysis_df <- data %>% drop_na()Performance metrics improve. Confidence intervals shrink. Reviewers relax.
But deployment happens in the messy world.
A model that only works for tidy cases is not robust. It is selectively competent. Apparent performance can improve precisely because difficult cases were removed before evaluation (Steyerberg 2019; Gianfrancesco et al. 2018).
Ethical Alternatives to Exclusion
Exclusion is sometimes unavoidable. It should never be silent.
Ethical alternatives include:
Modeling missingness explicitly
data <- data %>%
mutate(predictor_missing = is.na(predictor))This acknowledges that absence carries meaning. Missingness indicators do not solve the problem, but they make part of the selection process visible (Little and Rubin 2019; Rose et al. 2023).
Using uncertainty-aware methods
Bayesian models, joint models, and hierarchical approaches:
- widen uncertainty where data are sparse,
- prevent overconfidence in under-measured groups.
library(brms)
fit <- brm(
bf(outcome ~ predictor + predictor_missing + (1 | site)),
data = data,
family = bernoulli()
)Ethical modeling allows uncertainty to appear where reality is uncertain. That is one reason hierarchical and Bayesian workflows can be ethically preferable to silent deletion when data quality differs across sites or patient groups (Gelman et al. 2013; Little and Rubin 2019).
Sensitivity analysis instead of deletion
Rather than asking:
“Can we drop these patients?”
Ask:
“How sensitive are conclusions to including them?”
If conclusions collapse when messy patients are included, that is information, not failure. Sensitivity analysis is valuable precisely because it reveals how dependent an answer is on the analyst’s exclusions (Little and Rubin 2019; National Research Council 2010).
When Exclusion Is Ethically Defensible
There are cases where exclusion is appropriate:
- measurements are structurally incomparable,
- outcomes are undefined,
- inclusion would misrepresent the question.
Ethical exclusion requires:
- explicit rationale,
- quantification of who was excluded,
- discussion of how conclusions might differ otherwise.
summary(is.na(data))Silence is the ethical failure—not exclusion itself.
Messy Patients Are Often the Ones We Care About Most
In clinical settings, the highest-risk patients are often:
- the hardest to measure,
- the most inconsistently documented,
- the least represented in clean datasets.
Excluding them optimizes models for:
- convenience,
- publishability,
- internal validity,
at the expense of clinical relevance and justice. Exclusion can systematically shift the model toward patients who were easier to document, easier to follow, and easier to measure (Rajkomar et al. 2018; Obermeyer et al. 2019).
Documentation Is an Ethical Act
An audit-ready, ethically grounded analysis documents:
- how many patients were excluded,
- why exclusion occurred,
- how excluded patients differ,
- how inclusion might change results.
Example language:
Patients with incomplete documentation were not excluded by default. Where exclusion was necessary, we report the number and characteristics of excluded cases and interpret results as conditional on documented observations under the observed workflow.
This is not hedging. It is honesty, and it is a prerequisite for downstream trust, review, and fair deployment (Sendak et al. 2020; Barocas et al. 2023).
Ethics Does Not Mean “Include Everyone No Matter What”
Ethical modeling is not maximal inclusion. It is transparent trade-offs.
The ethical failure is not imperfection. It is pretending imperfection does not exist.
A Practical Ethical Checklist
Before excluding any patient, ask:
- Who is being excluded?
- Why are they missing data?
- How does exclusion change the estimand?
- Would conclusions change if they were included?
- Have we documented this clearly?
If you cannot answer these, pause.
DoDTR-trained models that exclude non-surviving prehospital casualties systematically underrepresent the most severe injuries — the model never sees the patients who died before reaching a treatment facility, creating optimistic performance estimates and a blind spot for the highest-acuity presentations that most need accurate decision support. This is not a technical data quality problem that better ETL will fix; it is an ethical choice about whose outcomes the model is designed to predict, embedded silently in an inclusion criterion. When that model is then used to support triage decisions in the field, it is being applied to exactly the population it was never trained on. The performance statistics reported in the validation study are accurate — and entirely irrelevant to the clinical context where the model will do the most harm.
Closing: Messiness Is Part of Reality
Clinical data is messy because clinical care is messy.
Excluding that mess may make models look better. It often makes them less just.
Ethical modeling does not eliminate uncertainty. It chooses where uncertainty is allowed to appear.
This post is part of the Missing Data Toolkit — a companion reference with data exclusion audit templates, estimand-aligned analysis frameworks, sensitivity analysis scaffolds, and ethical documentation checklists for registry exclusions.
Series Callout
This post is part of a broader Ethics in Trauma Registry Analysis Series:
- Opacity Is Sometimes Ethical: When Black Boxes Save Lives
- Accountability Without Interpretability: Who Owns a Model’s Decision?
- Bias Isn’t Always Where You Think It Is: Ethical Failure Modes in Registry Data
- Prediction vs Responsibility: Why Risk Scores Can Be Ethically Dangerous
- Human-in-the-Loop Is Not a Panacea (and Sometimes a Lie)
- The Ethical Implications of Excluding “Messy” Patients
- Missingness as a Fairness Issue in Machine Learning
- You Can’t Trust What You Don’t Track: AI Performance Monitoring in Clinical Systems
- From Weeks to Minutes: The Ethics of Automating CPG Compliance
- Ontology Is Not Optional: Semantic Infrastructure as Ethical Foundation
- What Responsible AI in Clinical Guidance Actually Requires
- Modernizing the DOD Trauma Registry: An Ethical and Technical Imperative