BIOSTATISTICS • DATA METHODS & STATISTICAL COMMUNICATION

Documenting Assumptions & Limitations — Document assumptions, limitations, and sources of bias

Transparent reporting of analytic decisions transforms statistical results into credible, reproducible science.

Historical Context & Motivation

Every statistical model rests on a scaffold of assumptions—conditions that must hold for the analysis to produce valid inferences. Throughout the history of biostatistics, failures to document those assumptions have led to retracted papers, ineffective therapies, and eroded public trust. The practice of documenting assumptions and limitations evolved as the field recognized that transparent reporting is not an optional courtesy but a scientific obligation. Understanding how this norm developed helps explain why modern journals, regulatory agencies, and funding bodies now require explicit statements of analytic assumptions, data limitations, and potential sources of bias.

Early biostatistical work in the nineteenth and early twentieth centuries often presented results with little discussion of the conditions under which they held. Ronald A. Fisher's foundational work in the 1920s formalized assumptions like normality and independence, yet even Fisher seldom devoted space to discussing when those assumptions might fail. As biostatistics matured through the mid-twentieth century—spurred by landmark clinical trials and growing regulatory oversight—researchers began to appreciate that unstated assumptions could silently undermine entire studies.

1948
Streptomycin Trial
The British Medical Research Council conducted the first modern randomized controlled trial for streptomycin. While groundbreaking, the published report offered minimal discussion of assumptions underlying its statistical comparisons, setting a precedent that would later be critiqued.
1996
CONSORT Statement
The Consolidated Standards of Reporting Trials (CONSORT) guidelines were published, requiring trialists to report study limitations, potential biases, and the statistical assumptions underlying their analyses. This marked a paradigm shift in how clinical research was communicated.
2007
STROBE Guidelines
The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement extended transparent reporting requirements to observational study designs, explicitly mandating discussion of bias and confounding.
2015
Replication Crisis & Open Science
Large-scale replication efforts—particularly in psychology and biomedical science—revealed that many published findings could not be reproduced. Poor documentation of analytic assumptions and researcher degrees of freedom were identified as central contributors, catalyzing the open-science movement.
2020s
AI & Algorithmic Fairness
The rise of machine-learning models in health care intensified demands for documenting sources of bias in training data, feature selection, and model deployment. Model cards and data sheets became standard tools for transparent documentation.

This historical arc reveals a clear trajectory: from an era in which assumptions were implicit and limitations unstated, to the modern expectation that every biostatistical analysis must include a forthright account of what was assumed, what could go wrong, and how bias may have shaped the results. The central question this lesson addresses is: How do we systematically identify, categorize, and communicate the assumptions, limitations, and biases inherent in any biostatistical analysis?

Core Principles & Definitions

Before you can document assumptions and limitations effectively, you need a precise vocabulary. Three concepts sit at the core of this practice: assumptions, limitations, and sources of bias. Although they are related, each term captures a distinct dimension of analytic uncertainty. Understanding the differences—and the relationships—between them is essential for producing credible biostatistical reports.

1

Assumptions

Conditions that must hold for a statistical method to yield valid inferences. These include distributional assumptions (e.g., normality), structural assumptions (e.g., linearity), and design assumptions (e.g., random sampling). Violations may invalidate p-values, confidence intervals, or effect estimates.
2

Limitations

Constraints on the study's scope, generalizability, or precision that arise from design choices, resource constraints, or data availability. Examples include small sample size, restricted study population, and incomplete follow-up. Limitations are inherent rather than errors—they define the boundary conditions of the study's conclusions.
3

Sources of Bias

Systematic errors that cause estimated associations to deviate from the true value. The three canonical categories are selection bias (non-representative sampling), information bias (measurement error or misclassification), and confounding (unmeasured third variables that distort the exposure–outcome relationship).
4

Transparency Principle

The overarching ethical and methodological norm that all analytic decisions—variable selection, model specification, handling of missing data—should be explicitly stated and justified so that readers can assess validity and reproduce the analysis. This principle underpins reporting guidelines like CONSORT, STROBE, and PRISMA.
5

Sensitivity Analysis

A complementary practice in which the analyst systematically varies key assumptions to determine whether conclusions are robust or fragile. Documenting the results of sensitivity analyses provides empirical evidence about the consequences of assumption violations.
KEY TAKEAWAY
Think of a statistical analysis as a bridge. The assumptions are the load-bearing specifications (maximum weight, wind tolerance) that must be true for the bridge to stand. The limitations are the design constraints (how many lanes, maximum span length) that define what traffic the bridge can handle. The sources of bias are the hidden geological faults beneath the foundation—unseen forces that could cause the entire structure to shift. An engineer who fails to document any of these risks builds a structure that no one should trust, and the same is true for a biostatistician who omits them from a report.

Visual Explanation — The Assumption–Bias–Limitation Framework

The diagram below illustrates how assumptions, limitations, and sources of bias interact within a biostatistical analysis pipeline. Data collection, analytic modeling, and interpretation each introduce distinct categories of assumptions that, when violated, produce different forms of bias or limitation. Tracing these connections visually helps you construct a comprehensive documentation checklist for any study.

The framework traces the analysis pipeline from data collection through analytic modeling to interpretation. Each stage carries assumptions that, when violated, produce specific forms of bias or misinterpretation. The green box at the bottom represents the documentation process that should accompany every analysis.

Notice that the framework is organized around the three stages where assumptions enter an analysis. At the data collection stage, the analyst assumes the sample is representative, measurements are accurate, and follow-up is complete. Violations produce selection bias or information bias. At the analytic modeling stage, assumptions about distributional form, independence, and linearity determine whether parameter estimates and their standard errors are valid. Finally, at the interpretation stage, assumptions about causality and generalizability shape whether conclusions are warranted. Effective documentation addresses all three stages systematically.

Quantifying the Impact of Assumption Violations

While documenting assumptions is fundamentally a qualitative reporting practice, biostatisticians frequently quantify the potential impact of assumption violations through mathematical sensitivity analysis. Understanding the formulaic underpinnings helps you appreciate why certain assumptions matter more than others—and it equips you to provide numerical evidence in your documentation rather than vague disclaimers.

Bias Due to Unmeasured Confounding

One of the most common assumption violations in observational biostatistics is the presence of unmeasured confounding. The bias formula for confounding provides a way to express how far an observed association might deviate from the true causal effect when a confounder is not controlled.

CONFOUNDING BIAS FORMULA
Bias = RR_observed − RR_true = RR_true × (RR_CU × P_C − 1) / (RR_CU × P_C + (1 − P_C))
Where RR_observed is the unadjusted relative risk, RR_true is the causal relative risk absent confounding, RR_CU is the relative risk between the confounder and the outcome, and P_C is the prevalence of the confounder in the exposed group. This simplified formula shows that bias magnitude depends on both the confounder–outcome strength and the confounder prevalence.

E-Value for Sensitivity to Unmeasured Confounding

The E-value, introduced by VanderWeele and Ding in 2017, provides a more elegant approach. It quantifies the minimum strength of association (on the risk ratio scale) that an unmeasured confounder would need to have with both the treatment and the outcome to fully explain away an observed association.

E-VALUE
E-value = RR_observed + √(RR_observed × (RR_observed − 1))
Where RR_observed is the point estimate of the risk ratio (or its conversion for other measures). A large E-value means an implausibly strong unmeasured confounder would be required to nullify the finding—strengthening confidence in the result.

Quantifying Misclassification Bias

NON-DIFFERENTIAL MISCLASSIFICATION
OR_observed = OR_true × [(Se × Sp + (1 − Se)(1 − Sp)) / (Se + Sp − 1)]²
Where Se is sensitivity and Sp is specificity of the classification instrument. Non-differential misclassification of a binary exposure typically biases the odds ratio toward the null, making this formula useful for documenting the likely direction and magnitude of information bias.
📊 When to Include Quantitative Bias Analysis
Whenever a study relies on an assumption that is plausibly violated, quantitative sensitivity analysis should accompany qualitative acknowledgment. At minimum, report the E-value for unmeasured confounding in observational studies. For studies using imperfect diagnostic tests or self-reported exposures, calculate the expected bias from misclassification. This transforms a generic limitation statement ('unmeasured confounding may exist') into a rigorous and informative disclosure.

Taxonomy of Assumptions, Limitations & Biases

To document assumptions and limitations comprehensively, it helps to work from a structured taxonomy rather than relying on memory or intuition alone. The table below classifies the most common assumptions, their associated biases or consequences when violated, and the documentation strategies recommended for each. This classification covers the three pipeline stages identified in Section 3—data collection, analytic modeling, and interpretation—and adds a fourth dimension for reporting and communication.

Comprehensive taxonomy of assumptions, violations, and documentation strategies in biostatistics
CategorySpecific AssumptionConsequence of ViolationDocumentation Strategy
SamplingRandom or probability-based sampling from target populationSelection bias; limited external validityDescribe sampling frame, inclusion/exclusion criteria, and population to which results apply
MeasurementExposure and outcome measured without error (non-differential or differential)Information bias (misclassification); attenuation or exaggeration of effectReport measurement validity (sensitivity, specificity); quantify expected bias direction
Missing DataData are missing completely at random (MCAR) or missing at random (MAR)Biased estimates if data are missing not at random (MNAR); loss of powerReport missingness patterns; justify chosen imputation method; conduct complete-case sensitivity analysis
DistributionalOutcome or residuals follow a specified distribution (e.g., normal, Poisson)Invalid standard errors, p-values, and confidence intervalsPresent diagnostic plots (Q-Q, residual plots); use robust methods if violated
StructuralLinear (or specified functional) relationship between predictors and outcomeModel misspecification; biased coefficient estimatesTest linearity with splines or polynomial terms; report residual-vs-fitted plots
IndependenceObservations are independent of one anotherUnderestimated standard errors; inflated type I errorAcknowledge clustering (e.g., patients within hospitals); use mixed-effects or GEE models
No Unmeasured ConfoundingAll relevant confounders are measured and adjusted forResidual confounding; spurious or masked associationsList potential unmeasured confounders; compute E-value; conduct negative-control analyses
Causal InferenceStudy design supports causal interpretation (e.g., exchangeability, positivity, consistency)Overstatement of causal claims; policy errorsSpecify causal framework (DAG, potential outcomes); distinguish association from causation
The concentric-layers diagram shows that documentation should proceed from the innermost layer (data collection) outward through analytic modeling, interpretation, and reporting. Violations at inner layers propagate outward, compounding uncertainty.

Worked Example — Drafting an Assumptions & Limitations Section

Consider a hypothetical observational cohort study examining whether daily aspirin use is associated with reduced risk of colorectal cancer among adults aged 50–75. The study uses electronic health records from a single hospital system, follows patients for five years, and applies a multivariable Cox proportional hazards model to estimate the hazard ratio. Below, we walk through the systematic construction of an assumptions and limitations section using the four-layer framework.

Drafting an Assumptions & Limitations Statement for an Observational Cohort Study
1
Step 1 — Enumerate Data Collection AssumptionsBegin at the innermost layer. The study assumes that electronic health records accurately capture aspirin use (exposure) and colorectal cancer diagnoses (outcome). It also assumes that the hospital-based cohort is reasonably representative of the broader target population of adults aged 50–75. Document these assumptions explicitly: 'We assumed that prescription records and pharmacy dispensing data accurately reflect daily aspirin use; however, over-the-counter aspirin purchases may not be captured, leading to potential misclassification of exposure status.'
Documented: measurement assumption, direction of bias (likely non-differential misclassification biasing toward the null).
2
Step 2 — Specify Analytic Model AssumptionsThe Cox model assumes proportional hazards—that the hazard ratio for aspirin vs. non-aspirin remains constant over time. It also assumes that censoring is non-informative (i.e., patients who are lost to follow-up do not differ systematically from those who remain). Document each assumption and report how it was assessed: 'The proportional hazards assumption was evaluated using Schoenfeld residual tests (global p = 0.23) and log-log survival plots, both of which were consistent with proportionality.'
Documented: proportional hazards assumption tested and justified; non-informative censoring acknowledged as untestable but discussed.
3
Step 3 — Address Unmeasured ConfoundingAlthough the model adjusted for age, sex, BMI, smoking, family history, and NSAID use, unmeasured confounders such as dietary fiber intake, physical activity levels, and socioeconomic status may introduce residual confounding. Compute the E-value: if the adjusted HR is 0.72 (converted to RR ≈ 0.72 on the risk ratio scale), the E-value for the point estimate is approximately: E-value = 1/0.72 + √(1/0.72 × (1/0.72 − 1)) = 1.39 + √(1.39 × 0.39) = 1.39 + 0.74 ≈ 2.13. This means an unmeasured confounder would need to be associated with both aspirin use and colorectal cancer by a risk ratio of at least 2.13 to explain the observed association.
Documented: E-value = 2.13; listed plausible unmeasured confounders; assessed whether any could realistically reach this threshold.
4
Step 4 — State Limitations and ScopeMove to the interpretation layer. The single-hospital design limits generalizability to patients with similar demographics and health care access. The five-year follow-up may be insufficient to capture colorectal cancers with long latency periods. The retrospective design precludes causal claims. Write: 'This study was conducted within a single academic medical center, potentially limiting generalizability to community-based or rural populations. The observational design permits only associational inference; residual and unmeasured confounding cannot be excluded.'
Documented: generalizability constraints, temporal limitation, causal interpretation boundary.
5
Step 5 — Compile into a Structured ParagraphAssemble the documented items into a coherent limitations section organized by the four layers: 'Several limitations should be considered when interpreting these findings. First, aspirin exposure was ascertained from prescription records and may not capture over-the-counter use, potentially introducing non-differential misclassification that would bias the hazard ratio toward the null. Second, although the proportional hazards assumption was supported by Schoenfeld residual tests (p = 0.23), non-informative censoring could not be directly verified. Third, despite adjustment for six covariates, residual confounding by unmeasured factors such as dietary fiber and physical activity is possible; the E-value of 2.13 suggests that a moderately strong unmeasured confounder would be required to fully explain the observed association. Finally, this single-center study with five-year follow-up may not generalize to broader populations and cannot support causal conclusions.'
Final product: a structured, four-layer limitations paragraph with quantitative sensitivity evidence.

Best Practices vs. Common Pitfalls

Not all assumptions-and-limitations sections are created equal. Many published papers include perfunctory disclaimers—generic sentences that acknowledge limitations without specificity—while others provide richly detailed, actionable documentation that enhances the paper's credibility. The table below contrasts best practices with the most frequently encountered pitfalls, drawn from systematic reviews of published biostatistical work.

Best practices versus common pitfalls in documenting assumptions and limitations
Best PracticeCommon Pitfall
Name each assumption explicitly and cite the statistical test or diagnostic used to assess it (e.g., 'We tested normality with the Shapiro-Wilk test, p = 0.41').Vaguely state 'standard assumptions were met' without specifying which assumptions or how they were evaluated.
Describe the expected direction and magnitude of bias when an assumption is likely violated (e.g., 'Non-differential misclassification would bias the OR toward the null').Acknowledge a limitation exists but omit any discussion of its likely impact on results (e.g., 'Measurement error may be present').
Report quantitative sensitivity analyses such as E-values, bias-adjusted estimates, or results under alternative assumptions.Provide no quantitative evidence about robustness; rely entirely on qualitative statements.
Distinguish between limitations that affect internal validity (bias, confounding) and those that affect external validity (generalizability).Lump all limitations together without distinguishing their implications for different aspects of validity.
Discuss how limitations inform the interpretation of results—qualifying claims where uncertainty is highest.Isolate the limitations section from the discussion; conclusions are stated without referencing acknowledged weaknesses.
Acknowledge strengths alongside limitations to provide balanced context for the reader.Either omit strengths entirely or use them to dismiss limitations ('Despite these limitations, our large sample size ensures validity').
KEY TAKEAWAY
A well-written limitations section functions like a product safety label on a pharmaceutical drug: it doesn't diminish the value of the product, but it honestly communicates the conditions under which it should and should not be used. A study with candidly reported limitations is far more trustworthy than one with no limitations section at all—because the reader knows that no study is perfect, and silence on this front signals either naïveté or concealment.

Connection to Advanced Methods in Bias Analysis

Documenting assumptions and limitations is the qualitative foundation upon which more sophisticated quantitative bias analysis (QBA) methods are built. As you advance in biostatistics, you will encounter formal frameworks that transform qualitative limitations into probabilistic adjustments—allowing you to present bias-corrected estimates alongside conventional results. Understanding the progression from basic documentation to advanced QBA methods clarifies why rigorous documentation matters at every level of expertise.

Progression from basic documentation to advanced quantitative bias analysis
FeatureBasic Documentation (This Lesson)Advanced QBA Methods
GoalTransparently identify and describe assumptions, limitations, and bias sourcesQuantify the magnitude and direction of bias; produce bias-adjusted effect estimates
OutputNarrative limitations section with E-values and qualitative assessmentBias-adjusted point estimates, confidence intervals incorporating systematic error, probabilistic distributions of bias parameters
TechniquesAssumption enumeration, diagnostic tests (Schoenfeld, Q-Q), E-value computationMonte Carlo sensitivity analysis, Bayesian bias adjustment, multiple bias modeling, external validation
Prerequisite KnowledgeFoundational biostatistics: regression, study design, basic epidemiologyBayesian inference, simulation methods, causal inference frameworks (DAGs, potential outcomes)
Key ReferenceSTROBE, CONSORT checklists; VanderWeele & Ding (2017) for E-valuesLash, Fox, & Fink, Applying Quantitative Bias Analysis to Epidemiologic Data (2009, 2nd ed. 2021)

The key insight is that basic documentation and advanced QBA are not competing approaches but successive rungs on the same ladder. Every quantitative bias analysis begins with the exact skills taught in this lesson—identifying the relevant assumptions, assessing their plausibility, and specifying the direction of potential bias. As you encounter courses in advanced epidemiologic methods or Bayesian statistics, you will find that the quality of your quantitative bias analysis is only as good as the quality of your qualitative documentation. A Monte Carlo simulation that draws bias parameters from poorly justified prior distributions is no better than a vague limitation statement. Rigorous documentation, therefore, is not merely a beginner's exercise but the enduring foundation of credible statistical communication.

Practice Problems

PROBLEM 1CONCEPTUAL
A researcher publishes a logistic regression analyzing the association between smoking and lung cancer using hospital-based controls. The limitations section states: 'Our study has some limitations.' Identify at least three specific deficiencies in this documentation and explain what information should have been provided instead.
PROBLEM 2BASIC CALCULATION
A case-control study reports an odds ratio of 2.5 for the association between pesticide exposure and Parkinson's disease. Compute the E-value for this point estimate. (Use the formula: E-value = OR + √(OR × (OR − 1)), where OR is treated as an approximation of the risk ratio under the rare disease assumption.)
PROBLEM 3INTERMEDIATE
You are analyzing the effect of a new drug on blood pressure reduction using a linear mixed-effects model with patients nested within clinics. List the key assumptions of this model, describe one diagnostic test or plot for each assumption, and draft a brief paragraph documenting the assumptions for a methods section.
PROBLEM 4APPLIED
A public health agency uses a machine-learning model trained on electronic health records from urban academic medical centers to predict 30-day hospital readmission risk. The model will be deployed at rural community hospitals. Draft a comprehensive assumptions-and-limitations section (4–6 sentences) for the deployment report, addressing at least three distinct categories of bias or limitation.
PROBLEM 5CRITICAL THINKING
Some scholars argue that detailed documentation of limitations can undermine a paper's credibility by drawing attention to weaknesses, while others contend it enhances credibility. Construct an argument for the latter position, drawing on epistemological principles, the replication crisis, and the concept of the E-value. Then identify one scenario in which excessive limitation documentation could be genuinely counterproductive, and explain why.

Lesson Summary

Documenting assumptions, limitations, and sources of bias is a foundational skill in biostatistical communication that transforms raw analytic output into credible, actionable science. This lesson established that assumptions are the conditions required for valid inference (such as normality, independence, and linearity), limitations are inherent design constraints (such as small sample size or restricted generalizability), and biases are systematic errors including selection bias, information bias, and confounding. The four-layer framework (data collection → analytic model → interpretation → reporting) provides a systematic checklist for comprehensive documentation.

Quantitative tools such as the E-value and misclassification bias formulas convert qualitative concerns into numerical evidence about robustness. Best practices call for naming each assumption explicitly, describing the likely direction and magnitude of any bias, reporting sensitivity analyses, and distinguishing between threats to internal validity and external validity. Reporting guidelines like CONSORT and STROBE codify these expectations. Ultimately, transparent documentation is not an admission of weakness but a hallmark of scientific integrity—it builds the trust upon which evidence-based medicine depends.

Varsity Tutors • Biostatistics • Documenting Assumptions & Limitations