--- module: 102-03 language: en chapter: 102 title: "Experimental Biomedical Methods, Measurement, Omics, and Reproducibility" module_title: "Experimental design, uncertainty, reproducibility, ethics, and laboratory safety" source_sha256: 08e96fbf4d5b2887a451dd32e80a5e89b9ca6272d8023597c4bdb8331ad324ab --- # Design, reproducibility, ethics, and safety ## Defining the comparison ### Link causal question to answerable comparison ### Define intervention, control, unit, outcome first ### Experimental unit: smallest independent assignment #### Cells in one treated dish are subsamples #### Wrong unit: pseudoreplication, false precision ## Randomisation and blinding ### Chance assignment balances baseline on average ### Blocking balances batch, sex, site ### Randomise processing order against drift ### Implement and record, not merely state #### Alternation and cage position are predictable ### Blinding separates assignment from handling #### Outcomes, labels, images, code can be blinded #### Phenotypes or adverse effects break blinding #### Automation inherits human thresholds ## Controls ### Negative control estimates background ### Positive control shows response is detectable #### Failure makes negative results uninterpretable ### Vehicle, sham, isotype, rescue answer different questions ### Active negative control can hide or create effects ## Designs and biological variation ### Factorial designs estimate interactions efficiently ### Dose-response shows shape and threshold ### Time course separates cause from consequence ### Crossover reduces between-subject variation #### Needs stable disease and adequate washout #### Carryover or irreversible effects violate it ### Design biological variation in from the start #### Intended inference sets relevant diversity #### Standardisation for precision, heterogeneity for reach ## Sample size and protocol ### Needs outcome, variance, effect, error rates ### Pilot effect sizes too imprecise for power ### Sequential designs stop by prespecified rules ### More cells cannot replace too few donors ### Protocols make tacit choices visible ### Preregistration time-stamps hypotheses and analyses #### Makes selective change detectable #### Report deviations with reasons ## Records and quality ### Provenance: identifiers, lots, calibration, versions ### Electronic notebooks need access control and export ### Keep raw data; clean through traceable scripts ### Quality assurance prevents error ### Quality control monitors each run #### Passing a narrow control does not validate method ### Set acceptance criteria before seeing groups #### Re-running one group's failures biases results ## Reproducibility and integrity ### Methods, materials, data, code, independent tests ### Unexplained direction reversal needs investigation ### Direct repeats conditions; conceptual tests the claim ### Fabrication, falsification, plagiarism ### Omission, duplicate publication, gift authorship ### Disclose and manage conflicts of interest ### Uniform adjustment acceptable if original kept #### Selective alteration misrepresents evidence ### Quantify raw data, not presentation images ## Human and animal research ethics ### Respect, benefit over harm, fair selection ### Uninformative studies are themselves unethical ### Consent: purpose, risks, alternatives, voluntariness ### Extra safeguards for vulnerability ### Unjust exclusion denies groups evidence ### Coding does not anonymise genomic data #### Re-identification grows with linkage and technology ### Replacement, reduction, refinement #### Reduce without underpowered waste #### Humane endpoints before death #### Justify species and model validity ## Laboratory safety ### Match containment to agent and procedure ### Aerosols, sharps, centrifugation, spills, waste ### Biological safety cabinets are not fume hoods ### Chemical controls follow hazard and exposure #### Substitution and enclosure beat PPE alone ### Incidents arise from interacting system weaknesses #### Near misses are signals if reporting is safe ### Practise emergency plans for use under stress ## Complete reporting ### Link question, design, conduct, analysis, limits ### Separate planned from exploratory work ### Report null and adverse findings ### Give effect sizes, uncertainty, exclusions ### Avoid causal language beyond the design ### Guidelines are checklists, not judgment ### Aim: an auditable chain