R biostatistics workflow

R Biostatistics Workflow

Use R to answer the stated scientific question with the simplest defensible model. A small p-value cannot repair a confounded design, pseudoreplication, outcome switching, selective exclusion, or an incorrect experimental unit.

Route specialized analyses

• Randomization, blocking, controls, and prospective design: experimental-design.
• Bulk RNA-seq counts: bio-differential-expression-deseq2-basics.
• Single-cell models: single-cell-rna-qc and scvi-tools.
• Data cleaning/joins: r-bioscience-data-wrangling.
• Figures: bio-data-visualization-ggplot2-fundamentals.
• Reproducible execution and package state: rstudio-reproducible-analysis.

Do not replace a domain-specific count, survival, longitudinal, or omics method with a generic t.test() or lm() for convenience.

Define the estimand before code

Write down:

1. experimental unit and unit of inference;
2. outcome, scale, units, and measurement process;
3. exposure/treatment and exact comparison or contrast;
4. biological versus technical replication;
5. pairing, repeated measures, nesting, blocking, batches, plates, runs, centers, donors, and families;
6. covariates selected before seeing results and why;
7. primary versus exploratory outcomes;
8. missing-data, exclusion, outlier, and multiplicity plan;
9. effect measure and uncertainty to report;
10. assumptions that are scientific, not merely mathematical.

If these cannot be resolved, produce an analysis-readiness report rather than inventing a model.

Model routing

Data/question	Starting model	Critical checks
Continuous outcome, independent units	lm()	linearity, residual pattern, influential observations, heteroskedasticity, estimability
Binary outcome	glm(..., family = binomial())	event counts, separation, link scale, calibration, predicted probabilities
Counts/rates	Poisson or negative-binomial GLM	exposure/offset, overdispersion, zero process, independent units
Paired two-condition data	paired contrast/model	pairing IDs, within-unit correlation, missing pairs
Repeated/nested data	mixed model or marginal model	random-effect structure, enough groups, singular fits, time structure
Ordered/multicategory outcome	appropriate ordinal/multinomial model	proportional-odds or category assumptions, sparse cells
Time-to-event	survival model	time origin, censoring, competing risks, proportional hazards
Many genes/features	domain method + FDR	feature universe, normalization, dependence, contrast, effect size
Very small or assumption-poor sample	exact/permutation/robust sensitivity analysis	exchangeability, discreteness, low power; avoid exaggerated certainty

Nonparametric does not mean assumption-free. A permutation test still requires an exchangeability structure that respects pairing, blocks, or clusters.

Build the model explicitly

Use a data argument and explicit factor levels. Inspect the model matrix and intended contrasts before fitting.

analysis$condition <- relevel(factor(analysis$condition), ref = "control")
analysis$batch <- factor(analysis$batch)

fit <- lm(response ~ batch + condition, data = analysis, na.action = na.exclude)
model.matrix(fit) |> head()
summary(fit)
confint(fit)

For binary outcomes:

fit <- glm(
  response_binary ~ batch + condition,
  family = binomial(),
  data = analysis,
  na.action = na.exclude
)

odds_ratio <- exp(coef(fit))
odds_ratio_ci <- exp(confint(fit))

State whether an effect is on the response, log-odds, log-count, hazard, or transformed scale. Convert it only when the conversion is valid and show the scale used.

For repeated measures, use a reviewed mixed-model workflow only when the required package is installed and the grouping structure is supported by the data:

if (requireNamespace("lme4", quietly = TRUE)) {
  fit <- lme4::lmer(response ~ condition * time + (1 | subject_id), data = analysis)
  lme4::isSingular(fit)
}

Do not add a maximal random-effects structure blindly, and do not remove random effects merely to obtain significance. Report singularity, group counts, and sensitivity to defensible structures.

Diagnostics are model-specific

Check the assumptions that affect the chosen estimate and inference:

• residuals versus fitted values and covariates;
• influence and leverage, with biological context;
• distribution/dispersion appropriate to the outcome;
• linearity on the link scale where required;
• calibration or predictive checks when relevant;
• random-effect singularity and enough independent groups;
• sparse cells, complete/quasi separation, and estimability;
• batch-treatment confounding and extrapolation;
• independence at the experimental-unit level.

Do not use a normality test as an automatic switch between parametric and nonparametric tests. With small samples it has low power; with large samples it flags harmless deviations. Use design, residual diagnostics, effect robustness, and sensitivity analyses together.

Do not delete an outlier only because it changes the p-value. Verify data integrity, document the biological/technical reason, report analyses with and without any defensible exclusion, and preserve the original data.

Missing data

1. Quantify missingness by variable, group, batch, and time.
2. Distinguish not measured, failed QC, below detection, structural missingness, dropout, and unknown.
3. State what complete-case analysis assumes and how many experimental units it removes.
4. Use imputation only with a method that matches the analysis, includes relevant predictors, and propagates uncertainty.
5. Run sensitivity analysis when missing-not-at-random mechanisms are plausible.

Never replace missing biological measurements with zero by default.

Multiplicity and selective analysis

• Define the family of hypotheses and primary endpoints before correction.
• Use FDR for high-dimensional exploratory feature sets when appropriate; use family-wise procedures for confirmatory families when required.
• Report the unadjusted estimate, uncertainty, adjusted measure, and tested family.
• Do not try many reasonable analyses and present only the smallest p-value.
• Label subgroup, post-hoc, and data-driven analyses as exploratory and test interactions rather than comparing “significant here, not significant there.”

Effect-first reporting

For every primary result report:

• sample size and independent unit count per group;
• exclusions/missingness and reasons;
• exact model/formula, contrast, link, and reference levels;
• effect estimate with units/scale;
• confidence or credible interval;
• test statistic, degrees of freedom when applicable, and adjusted p-value/FDR;
• diagnostics and sensitivity analyses;
• software/package versions;
• limitations, especially confounding, small group counts, model dependence, and generalizability.

Use model-derived plots and raw-data displays together. Do not draw causal conclusions from observational associations without a defensible causal design.

Completion gates

• Scientific question, estimand, and experimental unit agree.
• Model matches outcome and dependence structure.
• Formula, factor levels, and contrast were inspected.
• Diagnostics and sensitivity analyses were run and retained.
• Multiplicity and missingness are explicit.
• Code reruns in a clean project environment.
• Results emphasize effect size and uncertainty, not only significance.
• A qualified statistician reviews consequential, regulatory, clinical, or borderline analyses.

Example prompts

• “Use $r-biostatistics-workflow to choose and justify an R model for this experiment. Start from the experimental unit, pairing, batches and outcome type; do not choose a test from p-values.”
• “Bu tekrarlı ölçüm deneyinde subject, batch ve time yapısını kontrol et; R formülünü, kontrastı, diagnostics ve sensitivity analizini açıkla.”
• “Audit this R analysis for pseudoreplication, missing-data bias, multiplicity, factor-level errors, selective exclusions, and effect-size reporting.”

Primary references

• R lm: https://stat.ethz.ch/R-manual/R-devel/library/stats/html/lm.html
• R glm: https://stat.ethz.ch/R-manual/R-devel/library/stats/html/glm.html
• R model formulas: https://stat.ethz.ch/R-manual/R-devel/library/stats/html/formula.html
• R model diagnostics/accessors: https://stat.ethz.ch/R-manual/R-devel/library/stats/html/lm.summaries.html
• lme4 vignettes: https://lme4.github.io/lme4/articles/