Learn Plain-English guides to statistical analysis, method validation, quality control and diagnostic evaluation — choosing the right procedure, the common mistakes, and results that stand up.

Getting started

• Which CLSI EP protocol do you need? A method validation roadmap
  A map of the EP protocols by the question each one answers. Start here and follow the links to the detail.
• Validation vs verification in the laboratory
  Validating a method from scratch and verifying a manufacturer’s claims are different jobs of very different size. How to tell which one your assay requires.
• Trueness vs precision vs accuracy
  Three words everyday language blurs into one. What each means, and why accuracy is not a synonym for trueness.

Precision & verification

• Establishing precision: repeatability and reproducibility (EP05-A3)
  The nested study design, the variance components it separates, and how precision changes across the measuring range.
• What is measurement system analysis (MSA)?
  What MSA covers, the two traditions the term belongs to, and which study answers which part of it.
• Within-run, between-run and intermediate precision
  A hierarchy of variance components, not interchangeable numbers — and why variances add but standard deviations do not.
• Verifying manufacturer precision claims (EP15-A3)
  Establish or verify? Which of EP05 and EP15 your situation calls for, how to run the short verification study, and what to do when it fails.
• Rapid method evaluation before you commit (EP10-A3)
  A short screen catches gross precision, bias and linearity problems in a few days — before a full validation.
• Interference and recovery testing explained (EP07-Ed3)
  Putting a number on the shift an interferent causes, and judging it — like recovery — against an allowable limit.
• Robustness testing: varying several factors at once
  Why one factor at a time misses interactions, and how a factorial design and a multi-factor model find them.

Detection & measuring range

• LoB, LoD and LoQ explained (EP17-A2)
  Limit of blank, detection and quantitation — what each means, how they relate, how to estimate them, and how to read the LoQ off a precision profile.
• Probit analysis for the limit of detection
  For detected / not-detected assays, the detection limit is the concentration detected 95% of the time — found by probit regression.
• Fitting a calibration curve
  Choosing the model, why weighting decides the low end of the curve, and why a high r² is not evidence the fit is right.
• Assessing linearity of a measurement procedure (EP06-Ed2)
  How to establish the linear measuring interval, and why a correlation coefficient is the wrong tool for it.
• Analytical measurement range, measuring interval and reportable range
  Three limits — the LoQ, the range you measure directly, and the range you report after dilution — and why they must not be collapsed into one.

Method comparison & agreement

• Why correlation is the wrong statistic for method comparison
  The most reported number in a comparison study is blind to bias and dependent on your sample range. What to report instead.
• Deming, Passing–Bablok or OLS regression?
  A decision guide to choosing a regression for method comparison — with a worked example and a downloadable workbook.
• Bland–Altman explained: limits of agreement
  How the difference plot works, what the limits of agreement mean, and when to use it rather than regression.
• Estimating bias at a medical decision point
  The practical output of a method comparison is the bias where clinical decisions turn — how to estimate it, bound it, and judge it.
• How many samples do you need for a method comparison study?
  Not a single number — what determines it, and why the spread of concentrations matters as much as the count.
• Commutability of reference materials and processed samples explained (EP14-A3)
  A reference material is only useful if it behaves like a patient sample across methods. What commutability means, and how it is assessed.

Total error & performance goals

• Total analytical error explained (EP21-A)
  How imprecision and bias combine into the total error of a single result, and how to judge it against an allowable limit.
• Allowable total error and performance specifications
  The standard every acceptance judgement is made against — the Milan hierarchy, biological variation, regulatory limits and state of the art, and how to choose among them.
• Estimating measurement uncertainty, top-down
  ISO 15189 asks for it — how to build it from the imprecision and bias a validation already produced, rather than a component-by-component budget.
• Total error or measurement uncertainty?
  Both combine bias and imprecision but answer different questions. How the two frameworks differ, and when each is the right one.

Regulation & accreditation

• What goes in a validation report
  What each performance characteristic owes the document, what to fix before the study runs, and the exclusions inspectors look for first.
• Method validation under ICH Q2(R2)
  The performance characteristics ICH asks you to demonstrate, the study behind each, and how the terminology maps onto the clinical equivalents.
• What CLIA requires before you report a result
  The performance characteristics you are answerable for, which of the two lists applies to you, and where the acceptance limits have to come from.
• CAP accreditation and AMR verification
  The checklist requirements that repeat every six months — what AMR verification asks for, and when a separate study is not needed.
• Comparing instruments within one laboratory (EP31)
  Two analysers running the same method should agree. Why verifying that is not a method comparison, and what to do when they do not.
• Performance studies for an IVD 510(k)
  What FDA recommends for reporting diagnostic performance, and the four analysis practices it calls statistically inappropriate.

Reference intervals

• Choosing a reference interval method (EP28-A3C)
  Parametric, non-parametric, robust, bootstrap, Harrell–Davis — which to use, by sample size and distribution.
• Transferring and verifying a reference interval (EP28-A3C)
  Adopt an existing interval and confirm it holds in your population with about 20 individuals, rather than establishing from scratch.
• When to partition a reference interval
  Separate intervals by sex or age — how the Harris–Boyd criteria decide, and the sample-size cost of splitting.
• Handling outliers and skew in reference interval data
  Reference limits live in the tails, where a stray value or untreated skew does the most damage. How to screen and handle the shape.

Diagnostic performance

• ROC curves and AUC explained (EP24-A2)
  What a ROC curve shows, what the area under it means, how to compare two tests, and how to choose a cut-off.
• Comparing two diagnostic tests: DeLong’s test for the AUC
  Two tests on the same subjects give correlated AUCs. Why that needs a paired comparison, and how DeLong’s test handles it.
• Choosing a cut-off for a diagnostic test
  Youden’s index, cost- and prevalence-weighted thresholds, and why the “optimal” cut-off from your own data is optimistic.
• Sensitivity, specificity and predictive values explained
  Two numbers describe the test, two describe your population — and why a positive result means less than most people think.
• Positive and negative likelihood ratios explained
  One prevalence-independent number that turns a patient’s pre-test odds into post-test odds — the bridge sensitivity and predictive values leave out.
• How many samples for a diagnostic accuracy study?
  Sensitivity and specificity are proportions with wide intervals at small n — how to size each arm by the interval width you need.
• Positive and negative percent agreement (PPA, NPA) vs sensitivity and specificity
  When there is no gold standard, sensitivity and specificity are the wrong words. What percent agreement is, and when to use it.
• Evaluating a qualitative test (EP12-A2)
  Study design around the cut-off, accuracy against a reference or comparator, and the C5–C95 interval that captures qualitative precision.
• Cohen’s kappa and weighted kappa explained
  Agreement corrected for chance, when to weight it, and the prevalence paradox that pulls kappa down even when raw agreement is high.

Survival analysis

• Kaplan–Meier survival curves and censoring
  How the estimator handles censored time-to-event data, and how to read median survival, survival at a time, and confidence bands.
• Comparing survival between groups: log-rank and weighted tests
  Testing whether survival curves differ under censoring, and how the Wilcoxon and Fleming–Harrington weightings change what you detect.
• Cox proportional hazards regression and hazard ratios
  Reading a hazard ratio and its interval, why interactions make it conditional, and what the proportional-hazards assumption requires.

Statistics fundamentals

• Choosing the right statistical test
  What you are asking, what kind of measurement you have, and whether the observations are paired — three questions that narrow it to one test.
• Confidence intervals and p-values
  What a p-value does and does not mean, and why only the interval separates “too small to matter” from “we could not tell”.
• Student’s, Welch’s or paired? Choosing a t-test
  Why Welch’s is the better default, what pairing buys you, and the estimate you should be reporting instead of the p-value.
• Testing normality — and what to do when it fails
  The test raises the alarm at sample sizes where it does not matter and stays quiet where it does. What to look at instead.
• Non-parametric tests: what they assume, and what they cost
  About 5% of the power under ideal conditions and an advantage otherwise — but they are not the assumption-free methods they are sold as.
• How many samples do you need to compare two groups?
  The quantities that fix it, why halving the detectable difference costs four times the subjects, and why post-hoc power is not worth computing.
• Chi-square, Fisher exact or McNemar?
  Choosing a test for a table of counts — independent or paired classifications — and which effect measure belongs beside it.

ANOVA & multiple comparisons

• One-way, two-way and multi-factor ANOVA
  What ANOVA partitions, why an interaction changes what a main effect means, and when the order of terms matters.
• Choosing a multiple comparison procedure
  Tukey-Kramer, Dunnett, Hsu or Scheffé — the structure of the comparisons you need selects it, and asking for fewer buys tighter intervals.

Regression & multivariate

• Building a multiple regression model
  Coefficients as adjusted effects, collinearity and VIF, and comparing models with adjusted R², AIC and BIC rather than R².
• Reading residual plots: what each pattern means
  Curvature, funnelling, outliers and influence — the patterns a high R² will never show you.
• Logistic regression and odds ratios explained
  Why a yes/no outcome needs a different model, how to read an odds ratio and its interval, and why it is not a risk ratio.
• Pearson, Spearman or Kendall?
  Linear association or merely monotonic, when ranks are the right scale, and the scatter plots that decide it.
• Principal component analysis explained
  Scores, loadings and eigenvalues, how to read a biplot, how many components to keep, and how PCA differs from factor analysis.
• Factor analysis explained
  The common factor model, communality and uniqueness, how many factors to keep, and what rotation does and does not change.