Biostatistics uses statistical reasoning to learn from biological and health-related data. Its gold standard is to connect the question, the independent unit, the study design and the uncertainty with a conclusion that the evidence can support. A correct calculation is one part of that chain.
This A–B subject root brings the existing study-design, statistics, epidemiology, survival-analysis and biological-data guides together. Begin with Statistics for the general foundations, Biology for living systems or Epidemiology for population-health investigation. The wider learning route returns to The Gold Standard of Education and How X Works.
Find your next biostatistics learning step
Choose a route or open the grouped chapter index. All 25 teaching chapters remain open. Chapter 15 uses invented seed-germination records to show why a pooled percentage can reverse both subgroup comparisons. The example teaches interpretation, not a real biological finding.
Full chapter index
Chapters 1–5: Define the biological study
Chapters 6–10: Build a defensible comparison
Chapters 11–15: Interpret evidence and uncertainty
Chapters 16–20: Respect the structure of the data
Chapters 21–25: Connect analysis with biological learning
Worked pooled-comparison case · Existing-guide directory · Sources
CHAPTER 1 OF 25 · ROUTE 1: Define the biological study
1. Start with the population and the question
Biostatistics connects statistical reasoning with living systems. A question about seed germination, repeated blood measurements, species surveys or gene expression needs a biological definition before it needs a test. Name the population, the observation period and the quantity you want to estimate.
An estimand is that target quantity, stated precisely enough that two readers could recognise the same question. A difference in mean growth after a fixed period is not the same target as a difference in the probability of surviving that period. The target determines which records must be collected and which comparison would be meaningful.
Penn State STAT 509: objectives and endpoints develops the connection between study objectives and endpoints. Use this root for the biological evidence chain; the Statistics root supplies the broader statistical foundations, and Epidemiology develops population-health investigation. No example here is a personal medical recommendation.
Continue through the existing guides: The Gold Standard of Statistics · The Gold Standard of Epidemiology · The Gold Standard of Biology
CHAPTER 2 OF 25 · ROUTE 1: Define the biological study
2. Define variables, endpoints and measurement rules
A variable can record a category, a count, a continuous measurement or a time until an event. Record its unit and the rule used to obtain it. “Growth” could mean height change, mass change or the number of new leaves. These are related biological outcomes, but they are not interchangeable measurements.
An endpoint also needs a time and an assessment rule. In a supplied seed dataset, “germinated by day 10” and “ever germinated during observation” can classify the same seed differently. Establish the intended definition before inspecting which one produces the most attractive result.
A data dictionary should preserve allowed values, units, detection limits and reasons for missing entries. The Experimental Condition and Experimental Factor guides connect these choices with reusable scientific descriptions. Naming a variable consistently helps comparison; it does not make two differently measured variables equivalent.
Continue through the existing guides: Why Science? | Experimental Condition Ontology and Reproducible Context · Why Science? | Experimental Factor Ontology and Comparable Experimental Variables · Why Science? | Statistics Ontology and Traceable Statistical Analysis
CHAPTER 3 OF 25 · ROUTE 1: Define the biological study
3. Identify the experimental unit before counting rows
The experimental unit is the entity independently assigned to an intervention in the design. It may be a person, a plant pot, a culture preparation or a whole site. The measured unit can be smaller: several leaves from one treated pot are several measurements, not necessarily several independently treated experimental units.
NC3Rs: experimental units explains why identifying that unit matters. Write down how assignment occurred, which entities share conditions and which measurements are repeated. A spreadsheet with hundreds of rows may still contain only a small number of independently assigned units.
Technical repeats help describe measurement variability. Biological replicates address variation between the biological entities represented. Neither label should be accepted without the actual design. A defensible report can state both the number of independent units and the number of measurements made within each.
Continue through the existing guides: Why Science? | Experimental Condition Ontology and Reproducible Context · Why Science? | MASC, Mixed-Effects Models and Disease-Associated Cell Populations · The Gold Standard of Biology
CHAPTER 4 OF 25 · ROUTE 1: Define the biological study
4. Separate random sampling from random allocation
Random sampling concerns which members of a population enter a sample. Random allocation concerns which condition a recruited unit receives. A study can use one without the other. Randomly assigning a convenient group of plants does not make that group representative of every habitat.
Penn State STAT 506: sampling principles provides the sampling foundation. Ask how the sampling frame was constructed, which units were inaccessible and whether selection probabilities or survey weights matter. A large sample drawn through a narrow channel can precisely describe that channel while missing the target population.
For a biodiversity example, many observations beside one footpath are not automatically a balanced survey of the surrounding landscape. Preserve site selection and observation effort. The biodiversity-survey guide develops how richness and detection questions depend on the way observations were obtained.
Continue through the existing guides: Why Mathematics? | Biodiversity Surveys, Species Richness and Capture–Recapture · How Science Works | Epidemiology — Populations, Risk, Rates, Study Design, Bias and Causal Inference in Health · The Gold Standard of Probability
CHAPTER 5 OF 25 · ROUTE 1: Define the biological study
5. Use randomisation to support a fair comparison
Random allocation makes treatment assignment a chance process specified by the design. It helps prevent systematic selection of favourable units into one condition and supports a probability model for comparing assigned groups. It does not guarantee identical groups in a particular realised sample.
NC3Rs: allocation and randomisation connects allocation with experimental design. The record should preserve the allocation unit and any restrictions, such as randomising within blocks. A random-number procedure cannot repair an incorrectly chosen unit or concealment that failed during recruitment.
Keep randomisation separate from every later step. Exclusions, missing outcomes and departures from assigned conditions can change what the final comparison answers. Plan how such events will be recorded rather than silently analysing only the records that remain convenient.
Continue through the existing guides: Why Mathematics? | Clinical Trials, Randomisation and Sample Size · Why Science? | Experimental Condition Ontology and Reproducible Context · How Science Works | How Humans Build, Test and Correct Knowledge About the World
CHAPTER 6 OF 25 · ROUTE 2: Build a defensible comparison
6. Connect controls, blocking and masking with specific biases
A control supplies a comparison appropriate to the question. Blocking groups units with a relevant shared feature before comparisons are made within or across those groups. Masking limits which people know an assignment while collecting, assessing or analysing outcomes.
NC3Rs: masking explains masking as a way to reduce bias. Ask which judgement could be influenced by knowledge of the condition. An assessor interpreting an ambiguous image faces a different problem from an automated measurement with a fixed algorithm; neither situation excuses incomplete documentation.
Controls do not all answer the same question. A measurement blank, a reference specimen and a comparison condition serve different purposes. Explain which alternative explanation each design feature addresses, and which alternatives still remain. Good design is a chain of justified choices, not a checklist of impressive labels.
Continue through the existing guides: Why Science? | Experimental Condition Ontology and Reproducible Context · Why Science? | Experimental Factor Ontology and Comparable Experimental Variables · Why Mathematics? | Clinical Trials, Randomisation and Sample Size
CHAPTER 7 OF 25 · ROUTE 2: Build a defensible comparison
7. Treat observational adjustment as an assumption-dependent argument
In observational records, groups can differ for reasons connected with both their exposure and their outcome. Such confounding can make a crude association misleading for a causal question. Stratification or modelling may help, but only when the variables, timing and assumptions fit the question.
CDC Field Epidemiology Manual: analysis and interpretation develops analysis plans and stratified interpretation. Before adjusting, describe the proposed causal relationships. A variable measured after exposure could be a mediator; conditioning on another variable could introduce selection bias. Adding every available column is not a universal route to a causal estimate.
Keep descriptive, predictive and causal statements distinct. A useful adjusted association can remain non-causal. The worked seed example later in this root demonstrates how different subgroup weights change a descriptive comparison; it does not establish that one seed collection causes better outcomes.
Continue through the existing guides: Why Mathematics? | Cochran–Mantel–Haenszel Stratification, Common Odds Ratios and Confounding · The Gold Standard of Epidemiology · How Science Works | Epidemiology — Populations, Risk, Rates, Study Design, Bias and Causal Inference in Health
CHAPTER 8 OF 25 · ROUTE 2: Build a defensible comparison
8. Describe distributions before reducing them to one number
A mean describes one feature of a distribution. It does not show a long tail, a mixture of groups or a few extreme values. A median, spread and a suitable display can reveal different aspects of the same biological measurements.
For repeated growth records, inspect variation between units as well as variation within a unit over time. An unusual value might be a genuine biological observation, a unit-conversion error or a recording problem. Investigate provenance before deciding to exclude it.
A summary should preserve the denominator and the observation scale. Counts, concentrations and proportions need different descriptions. Report what was measured and how many independent units contributed, so a reader can distinguish biological heterogeneity from a change in measurement or sampling.
Continue through the existing guides: The Gold Standard of Statistics · Why Science? | Statistics Ontology and Traceable Statistical Analysis · Why Science? | Experimental Condition Ontology and Reproducible Context
CHAPTER 9 OF 25 · ROUTE 2: Build a defensible comparison
9. Choose a probability model that matches the data-generating process
A model makes assumptions about how observed outcomes could arise. A binomial model, for example, uses a fixed number of trials with independent binary outcomes and a common success probability within the model. Seeds sharing a pot or belonging to different conditions may not satisfy that simple account.
NIST: confidence intervals for proportions discusses interval estimation for a proportion. The fraction observed is easy to calculate; the uncertainty attached to it depends on the sampling and model assumptions. A familiar formula should not erase clustering or selection.
Ask what the model leaves out. Overdispersion, dependence or changing probabilities can matter biologically. The Probability root helps develop distributions and conditional reasoning; the Biostatistics task is to connect those ideas with the actual units and conditions of a living-system study.
Continue through the existing guides: The Gold Standard of Probability · Why Mathematics? | Wilson Score Intervals, Binomial Proportions and Honest Uncertainty · The Gold Standard of Statistics
CHAPTER 10 OF 25 · ROUTE 2: Build a defensible comparison
10. Report effect size with its scale and baseline
An effect description should make the comparison understandable. A difference between two proportions is measured in percentage points; a ratio compares one proportion with another. Both can be mathematically correct while communicating different features.
In an invented example, proportions of 20% and 10% differ by 10 percentage points and have a ratio of 2. The relative increase from 10% to 20% is 100%, not 10%. Name the reference group and observation period so these statements cannot be confused.
Risk and odds also have different denominators. The CDC analysis guide explains measures suited to different study designs. Do not rename an odds ratio as a risk ratio or infer a population risk directly from a case–control sample selected by outcome.
Continue through the existing guides: Why Mathematics? | Comparing Percentages Fairly · Why Mathematics? | Medical Screening, Base Rates and Test Results · Why Mathematics? | Cochran–Mantel–Haenszel Stratification, Common Odds Ratios and Confounding
CHAPTER 11 OF 25 · ROUTE 3: Interpret evidence and uncertainty
11. Read a confidence interval as part of a procedure
A frequentist confidence interval is produced by a procedure whose coverage is assessed over repeated samples under stated assumptions. A realised 95% interval is not, in that interpretation, a statement that the fixed unknown parameter has a 95% probability of lying inside those particular bounds.
Penn State STAT 500: confidence intervals develops this interpretation. The interval describes uncertainty associated with a specified estimation procedure. It does not automatically account for systematic measurement error, a biased sample or a wrong model.
Distinguish uncertainty in a group mean from variation among individual organisms and from uncertainty in predicting a new individual. These intervals answer different questions. A narrow interval around a mean can coexist with wide biological variation among the units being studied.
Continue through the existing guides: Why Mathematics? | Wilson Score Intervals, Binomial Proportions and Honest Uncertainty · The Gold Standard of Statistics · How Regression Works | From Conditional Relationships and Least Squares to Coefficients, Residuals, Interactions, Diagnostics, Prediction and Causal Boundaries
CHAPTER 12 OF 25 · ROUTE 3: Interpret evidence and uncertainty
12. Use p-values without turning them into truth scores
A p-value measures how incompatible the observed result, or one more extreme under the test rule, is with a specified statistical model. It is not the probability that the null hypothesis is true, and it does not measure the size or biological importance of a difference.
ASA statement on p-values: context, process and purpose sets out these boundaries. A result just below a conventional threshold is not a different kind of fact from a result just above it. Report the estimate, uncertainty, design and analysis choices alongside any test result.
A large p-value does not prove equality or establish absence of a biologically meaningful difference. The data may be too imprecise for that conclusion. If the question concerns equivalence or an acceptable bound, specify that question and its margin in an appropriate design.
Continue through the existing guides: How Statistical Significance Works | From a Predefined Error Threshold and P-Value to Binary Test Decisions, Misinterpretation, Multiplicity and Better Scientific Judgement · The Gold Standard of Statistics · Why Mathematics? | Clinical Trials, Randomisation and Sample Size
CHAPTER 13 OF 25 · ROUTE 3: Interpret evidence and uncertainty
13. Plan sample size around a meaningful target
Sample-size planning connects the question with an effect worth detecting or a precision worth achieving. For a power calculation, the assumed effect, variability, significance rule and design all matter. A calculation is conditional on those inputs, not a guarantee that the planned study will produce a desired result.
NC3Rs: groups and sample size develops the connection between groups and sample size. Clustering, repeated measurements and losses to follow-up can change how much information the records contain. Counting every measurement as independent can make a plan appear more informative than it is.
Write down why the chosen precision or effect matters biologically. If the inputs are uncertain, examine more than one plausible scenario. A transparent sensitivity analysis is more useful than a single exact-looking sample size supported by an unexplained assumption.
Continue through the existing guides: Why Mathematics? | Clinical Trials, Randomisation and Sample Size · Why Mathematics? | Wilson Score Intervals, Binomial Proportions and Honest Uncertainty · The Gold Standard of Probability
CHAPTER 14 OF 25 · ROUTE 3: Interpret evidence and uncertainty
14. Use regression to answer a stated conditional question
Regression relates an outcome to predictors within a model. The outcome type matters: a continuous measurement, a binary endpoint and a count generally require different modelling choices. A coefficient also depends on the scale, coding and other variables included.
NIST: interpreting regression models supplies a regression foundation. Check fit and residual behaviour appropriate to the model, inspect influential observations and ask whether the functional relationship makes biological sense. A fitted line is a model summary, not a substitute for the underlying record.
Interactions allow a relationship to vary with another variable. If a seed-collection contrast differs by setting, one overall coefficient may conceal that variation. Interpretation should state the condition being compared; a predictive coefficient does not become causal simply because other predictors were included.
Continue through the existing guides: How Regression Works | From Conditional Relationships and Least Squares to Coefficients, Residuals, Interactions, Diagnostics, Prediction and Causal Boundaries · Why Mathematics? | Cochran–Mantel–Haenszel Stratification, Common Odds Ratios and Confounding · Why Science? | MASC, Mixed-Effects Models and Disease-Associated Cell Populations
CHAPTER 15 OF 25 · ROUTE 3: Interpret evidence and uncertainty
15. Work through a pooled comparison that reverses within subgroups
The following records are invented for teaching. Two seed collections, A and B, share one endpoint: germinated by a stipulated assessment time. Each seed is counted once in these summaries. Records come from favourable and challenging settings; no random allocation or causal experiment is assumed.
Within the favourable setting, B has the larger observed proportion: 19/20 = 95%, compared with A at 90/100 = 90%. Within the challenging setting, B is again larger: 20/100 = 20%, compared with A at 1/10 = 10%. Now combine the records without retaining their setting.
A contributes mostly favourable-setting observations, while B contributes mostly challenging-setting observations. The pooled comparison therefore mixes the settings in different proportions. A pooled percentage is not an unweighted average of its two subgroup percentages.
| Setting | A: germinated / total | A: proportion | B: germinated / total | B: proportion |
|---|---|---|---|---|
| Favourable | 90 / 100 | 90.00% | 19 / 20 | 95.00% |
| Challenging | 1 / 10 | 10.00% | 20 / 100 | 20.00% |
| Pooled | 91 / 110 | 82.73% | 39 / 120 | 32.50% |
The pooled values are A: (90 + 1)/(100 + 10) = 91/110 ≈ 82.73%, and B: (19 + 20)/(20 + 100) = 39/120 = 32.50%. The apparent pooled advantage for A coexists with B’s higher proportion within each setting. Each calculation describes its own mixture of settings.
Now define a hypothetical target mixture containing equal weights for the two settings. The standardised descriptive proportions are A: 0.5 × 0.90 + 0.5 × 0.10 = 0.50, and B: 0.5 × 0.95 + 0.5 × 0.20 = 0.575. Under that specified mixture, B exceeds A by 7.5 percentage points. Equal weighting is a teaching choice, not an automatically correct population.
A real target mixture needs justification. Stratification also does not remove every possible difference between the collections. These invented observations establish neither a causal effect nor a precise population ranking. The challenging-setting proportion for A comes from only ten seeds, and the arithmetic here does not quantify sampling uncertainty.
To test transfer, change only the number of records supplied by each setting while keeping its proportions fixed. Explain which pooled quantities change, which subgroup comparisons stay unchanged and what additional design information would be required for a causal claim.
Continue through the existing guides: Why Mathematics? | Cochran–Mantel–Haenszel Stratification, Common Odds Ratios and Confounding · Why Mathematics? | Comparing Percentages Fairly · The Gold Standard of Statistics
CHAPTER 16 OF 25 · ROUTE 4: Respect the structure of the data
16. Keep clustering and repeated observations in the analysis
Measurements from the same organism can resemble one another more than measurements from different organisms. Cells from one donor, leaves from one plant and repeated visits to one site have a structure that should remain visible in both the data and the uncertainty calculation.
Lazic: the problem of pseudoreplication in neuroscientific studies explains pseudoreplication: treating dependent observations as if they supplied independent replication can exaggerate the information in a study. The remedy begins with identifying the design, not merely selecting a more elaborate test after analysis.
A mixed-effects model or another design-appropriate method may represent some dependencies. Its suitability depends on the number and arrangement of independent groups and the assumptions made. Report donors and samples separately from cell counts, and retain time and grouping identifiers so the model can be checked.
Continue through the existing guides: Why Science? | MASC, Mixed-Effects Models and Disease-Associated Cell Populations · Why Science? | distinct, Full Distributions and Multi-Sample Single-Cell Evidence · Why Science? | Experimental Condition Ontology and Reproducible Context
CHAPTER 17 OF 25 · ROUTE 4: Respect the structure of the data
17. Distinguish an observed event time from a censored record
Time-to-event data include both the event definition and the time origin. Some units have not experienced the event when observation ends, or are last known to be event-free at an earlier point. Such a record can be right-censored rather than assigned an invented event time.
Penn State STAT 507: survival analysis introduces survival analysis. A Kaplan–Meier estimate uses the risk set at each event time. Censored units contribute information while under observation but are not counted as events at their censoring times.
Interpretation requires assumptions about censoring in the analysis setting. If observation ends for reasons connected with unobserved event risk, the usual estimate can be misleading. A survival probability, a hazard and a median event time are different summaries, and their units and conditioning should remain explicit.
Continue through the existing guides: Why Mathematics? | Kaplan–Meier Risk Sets, Censoring and Product-Limit Survival Curves · Why Mathematics? | Survival Analysis, Kaplan–Meier Curves and Censoring · The Gold Standard of Epidemiology
CHAPTER 18 OF 25 · ROUTE 4: Respect the structure of the data
18. Investigate missingness instead of disguising it
A blank value can represent an unrecorded measurement, a failed assay, withdrawal or a value below a detection limit. Those reasons have different implications. Preserve the reason where it is known and report how many records contribute to each analysis.
Sterne and colleagues: multiple imputation, potential and pitfalls discusses multiple imputation and its pitfalls. Imputation depends on a model and assumptions about missingness; it does not recover the actual unobserved values or automatically remove bias. A detection limit also needs an appropriate measurement account rather than casual replacement with zero.
Complete-case analysis changes the included population and may be biased under some missingness mechanisms. Compare available records and consider sensitivity to plausible alternatives. The conclusion should identify which uncertainty is addressed by the analysis and which missing-data assumptions remain untestable from the observed values alone.
Continue through the existing guides: Why Science? | OMOP CDM and Reproducible Observational Health Research · Why Science? | SDRF-Proteomics and Sample Metadata · Why Science? | Statistics Ontology and Traceable Statistical Analysis
CHAPTER 19 OF 25 · ROUTE 4: Respect the structure of the data
19. Match multiplicity control with the discovery question
Testing many hypotheses creates opportunities for false discoveries. The relevant family of comparisons should be defined by the scientific question and analysis plan, not by retaining only the comparisons that looked promising.
Benjamini and Hochberg: controlling the false discovery rate introduced a false-discovery-rate procedure under specified assumptions. False discovery rate concerns the expected proportion of false rejections among rejections, with the proportion defined as zero when there are none. It is not the probability that one particular reported gene is a false finding.
Family-wise error control and false-discovery-rate control address different error goals. Dependence between tests and the chosen procedure matter. Keep exploratory findings distinct from confirmatory claims, disclose the wider analysis and seek independent evidence where a biological interpretation depends on a selected signal.
Continue through the existing guides: How Multiple Testing Works | From Families of Hypotheses and Inflated False Positives to FWER, FDR, Adjusted Decisions, Power Trade-Offs and Transparent Discovery · Why Mathematics? | RNA Sequencing, Count Normalisation and Differential Expression · Why Science? | GWAS Catalog and Curated Variant–Trait Associations
CHAPTER 20 OF 25 · ROUTE 4: Respect the structure of the data
20. Make Bayesian assumptions and updating visible
A Bayesian analysis combines a prior distribution with a likelihood to obtain a posterior distribution. The prior describes uncertainty before the current evidence; the likelihood connects candidate parameter values with the observed data under the model.
NIST: prior distributions, likelihood and Bayesian updating explains this updating structure in a reliability setting. The mathematical relationship also provides a foundation for biological models, but the appropriate prior and likelihood need biological justification. A posterior probability is conditional on those choices.
A credible interval has a posterior-probability interpretation, unlike the repeated-sampling interpretation of a frequentist confidence interval. Examine sensitivity to defensible alternative priors and assess whether the model can reproduce relevant features of the data. A Bayesian calculation does not remove dependence, measurement bias or selection problems.
Continue through the existing guides: How Bayesian Inference Works | From Prior Knowledge and Likelihood to Posterior Distributions, Predictive Checks, Hierarchies, Computation and Better Decisions · Why Science? | BASS, Bayesian Cell-Type Clustering and Spatial-Domain Evidence · Why Science? | BayesSpace, Bayesian Spatial Clustering and Subspot-Resolution Evidence
CHAPTER 21 OF 25 · ROUTE 5: Connect analysis with biological learning
21. Validate prediction beyond the records used to fit it
A predictor can learn peculiarities of its training records rather than a relationship that transfers. Evaluating it on the same records used for fitting can give an overly favourable impression. Model selection and preprocessing are part of training, not exempt steps.
Penn State STAT 555: cross-validation develops cross-validation and the role of a further hold-out sample. Keep information from evaluation records out of feature selection and fitting. When measurements share a donor or site, splitting rows at random can leak group-specific information across the split.
Match evaluation to the intended use. A new hospital, new donor or later period may require a different validation design from a new measurement on an already observed individual. Report error, calibration and relevant subgroup performance rather than relying on a single attractive score.
Continue through the existing guides: Why Mathematics? | Medical Screening, Base Rates and Test Results · The Gold Standard of Bioinformatics · How Regression Works | From Conditional Relationships and Least Squares to Coefficients, Residuals, Interactions, Diagnostics, Prediction and Causal Boundaries
CHAPTER 22 OF 25 · ROUTE 5: Connect analysis with biological learning
22. Read high-dimensional biological data with their measurement model
Sequencing counts, cell-type proportions and spatial measurements are not interchangeable numeric tables. Library size, compositional constraints, batch structure and the number of independent donors can change what a comparison means. More measured features do not compensate for absent biological replication.
The RNA-sequencing, ALDEx2, scCODA and distinct guides develop different parts of this landscape. Read their assumptions before transferring a method to a new assay. A change in a relative abundance can reflect changes elsewhere in the composition; it need not establish an absolute increase in the named component.
NHGRI: genome-wide association studies describes genome-wide association studies. An associated marker is a starting point for biological investigation, not by itself proof that a particular variant causes the trait. Keep multiplicity, population structure, replication and mechanistic follow-up in the evidence chain.
Continue through the existing guides: Why Mathematics? | RNA Sequencing, Count Normalisation and Differential Expression · Why Science? | ALDEx2, Compositional Counts and Log-Ratio Evidence · Why Science? | scCODA, Compositional Models and Credible Cell-Type Abundance Changes · Why Science? | distinct, Full Distributions and Multi-Sample Single-Cell Evidence · Why Science? | GWAS Catalog and Curated Variant–Trait Associations · Why Science? | MISTy, Multiview Spatial Contexts and Explainable Marker-Dependency Evidence
CHAPTER 23 OF 25 · ROUTE 5: Connect analysis with biological learning
23. Preserve a reproducible path from records to conclusion
A reproducible analysis needs more than a final graph. Preserve variable definitions, sample identities, inclusion decisions, transformations, model specifications and the versions of the analysis materials. Record which choices were planned and which arose during exploration.
NIH: rigor and reproducibility connects rigor with reproducibility. A script can reproduce a calculation faithfully while the study remains biased or the biological claim exceeds the design. Reproducibility and validity are related goals, not synonyms.
The Statistics Ontology, Experimental Condition and sample-metadata guides help organise the record. A vocabulary term improves traceability only when it describes what was actually done. For shared health or genomic records, use authorised access and appropriate safeguards; this teaching root does not require uploading personal data.
Continue through the existing guides: Why Science? | Statistics Ontology and Traceable Statistical Analysis · Why Science? | Experimental Condition Ontology and Reproducible Context · Why Science? | Experimental Factor Ontology and Comparable Experimental Variables · Why Science? | SDRF-Proteomics and Sample Metadata · Why Science? | OMOP CDM and Reproducible Observational Health Research
CHAPTER 24 OF 25 · ROUTE 5: Connect analysis with biological learning
24. Read a biological claim through its denominator and design
Before accepting a headline, reconstruct the population, outcome, comparison, independent unit and uncertainty. Ask whether the stated effect is absolute or relative, which records were excluded and whether the claim concerns description, prediction or causation.
Then ask whether the conclusion is stronger than the evidence. A statistically detectable change may be biologically small. An imprecise result may still allow an important difference. A subgroup pattern selected after extensive searching may require further evaluation before it supports a general claim.
For a health-related report, these questions support evidence literacy rather than personal diagnosis or treatment decisions. Public Health and Epidemiology connect individual studies with population evidence; the screening guide develops why baseline prevalence matters when interpreting test results.
Continue through the existing guides: The Gold Standard of Public Health | Prevention That Reaches Communities · The Gold Standard of Epidemiology · Why Mathematics? | Medical Screening, Base Rates and Test Results · Why Science? | Vaccines, Immunity and Reading Health Evidence
CHAPTER 25 OF 25 · ROUTE 5: Connect analysis with biological learning
25. Teach transfer by changing the question, not just the numbers
Begin with a small supplied dataset whose units and definitions are clear. Ask learners to describe what each row represents before calculating a percentage. Then introduce one change: a different subgroup mixture, repeated observations from one unit or an outcome measured at a different time.
The seed case offers a useful transfer task. Keep every subgroup percentage fixed but alter the number of records contributed by each setting. Learners should be able to explain why the pooled comparison changes while the within-setting comparisons do not.
Finish with an evidence statement containing the target, design, estimate, uncertainty and boundary of the claim. The education standard is learning that transfers into life: reading a biological claim with enough care to identify its denominator, its assumptions and the next question worth asking.
Continue through the existing guides: The Gold Standard of Teaching Biology | Living Systems and Testable Evidence · The Gold Standard of Education | Learning That Transfers into Life · What is Education | Education, Public Health, Epidemiology and Laboratory Capability — How Learning Builds the Workforce Behind Population Health Intelligence · How X Works | How Things, Systems, Subjects and the World Work | eduKateSG
Existing-guide directory: follow the biological evidence chain
These connected articles develop the root’s main subtopics. Choose a focused guide for a method or dataset, or return to a neighbouring root for its wider subject. Each route retains its own scope and assumptions.
Foundations and neighbouring roots
| Existing guide | Use this branch to… |
|---|---|
| Statistics | Develop distributions, estimation and general statistical reasoning. |
| Probability | Study conditioning, chance and model assumptions. |
| Biology | Connect measurements with living systems. |
| Bioinformatics | Follow biological data and computational inference. |
| Epidemiology | Develop population-health design and interpretation. |
| How Epidemiology Works | Trace rates, bias and causal boundaries. |
| Public Health | Connect evidence with community prevention. |
Study design and fair comparisons
| Existing guide | Use this branch to… |
|---|---|
| Clinical trials, randomisation and sample size | Connect allocation with precision and design. |
| Medical screening and base rates | Distinguish test properties from an individual result. |
| Epidemic modelling | Connect rates, growth and population assumptions. |
| Biodiversity surveys | Study sampling, richness and capture–recapture. |
| Vaccines and reading health evidence | Practise interpreting a health-related evidence claim. |
| Stratification and confounding | Continue into common odds ratios and adjusted comparisons. |
| Comparing percentages fairly | Preserve denominators and reference groups. |
Uncertainty, models and repeated records
| Existing guide | Use this branch to… |
|---|---|
| Wilson intervals for proportions | Continue into interval estimation for binary records. |
| Kaplan–Meier risk sets | Read censoring and product-limit calculations. |
| Survival analysis | Develop time-to-event interpretation. |
| Statistical significance | Read p-values and decision thresholds carefully. |
| Multiple testing | Distinguish error-control goals. |
| Bayesian inference | Connect priors, likelihood and posterior checks. |
| Regression | Develop conditional relationships and model diagnostics. |
Omics, spatial data and biological discovery
| Existing guide | Use this branch to… |
|---|---|
| ALDEx2 | Read compositional counts and log-ratio evidence. |
| RNA sequencing | Connect normalisation with differential expression. |
| MASC | Study mixed effects and cell-population comparisons. |
| scCODA | Read compositional cell-type abundance models. |
| distinct | Compare full distributions across biological samples. |
| MISTy | Explore multiview spatial marker relationships. |
| BASS | Connect Bayesian models with spatial domains. |
| BayesSpace | Read spatial clustering and resolution assumptions. |
| GWAS Catalog | Trace curated variant–trait associations. |
Metadata and accountable analysis
| Existing guide | Use this branch to… |
|---|---|
| Statistics Ontology | Preserve traceable statistical analysis descriptions. |
| Experimental Condition Ontology | Record reproducible experimental context. |
| Experimental Factor Ontology | Describe comparable experimental variables. |
| OMOP CDM | Connect observational health records with reproducible research. |
| SDRF-Proteomics | Preserve sample and assay metadata. |
Study, work and teaching routes
| Existing guide | Use this branch to… |
|---|---|
| University courses for Statistics | Explore study pathways in statistical reasoning. |
| University courses for Public Health | Connect study with population health. |
| University courses for Biology | Develop biological foundations. |
| Clinical trials and biomedical translation | Connect research with accountable translation. |
| Education, epidemiology and laboratory capability | Connect learning with scientific capability. |
| Teaching Biology | Teach living systems through testable evidence. |
| The Gold Standard of Education | Carry reasoning into new situations. |
| How Science Works | Develop testable and correctable explanations. |
| How X Works | Choose the next subject branch. |
Sources and boundaries of the teaching examples
The sources below include university teaching materials, government scientific resources, research-design guidance and primary methodological papers. Their links sit beside the relevant chapters. The seed records, percentage comparison and equal-weight target mixture are original invented teaching examples, not observations from an eduKate study. This article teaches evidence interpretation and does not provide individual medical advice.
1. Penn State STAT 509: objectives and endpoints
3. Penn State STAT 506: sampling principles
4. NC3Rs: allocation and randomisation
6. CDC Field Epidemiology Manual: analysis and interpretation
7. NIST: confidence intervals for proportions
8. Penn State STAT 500: confidence intervals
9. ASA statement on p-values: context, process and purpose
10. NC3Rs: groups and sample size
11. NIST: interpreting regression models
12. Lazic: the problem of pseudoreplication in neuroscientific studies
13. Penn State STAT 507: survival analysis
14. Sterne and colleagues: multiple imputation, potential and pitfalls
15. Benjamini and Hochberg: controlling the false discovery rate
16. NIST: prior distributions, likelihood and Bayesian updating
17. Penn State STAT 555: cross-validation
18. NHGRI: genome-wide association studies
19. NIH: rigor and reproducibility
Carry five questions into the next biological claim: what was the target, what was the unit, how was the comparison formed, how uncertain is it and what conclusion remains justified? Continue through The Gold Standard of Education or choose another branch through How X Works.