What are the ARRIVE guidelines?

The ARRIVE guidelines are a reporting checklist for animal research — Animal Research: Reporting of In Vivo Experiments — published by the NC3Rs in 2010 and revised as ARRIVE 2.0 in 2020. ARRIVE 2.0 contains 21 items, split into an Essential 10 that every manuscript must report and a Recommended Set of 11 adding context to the study described.

ARRIVE is a reporting standard, not a design standard. The guidelines say what a reader needs to see in order to judge a study; they do not tell you how to run one. That distinction is the source of most misunderstanding about them, and it is worth holding onto through everything below.

What ARRIVE stands for, and who publishes it

ARRIVE is the acronym for Animal Research: Reporting of In Vivo Experiments. The guidelines are maintained by the NC3Rs — the UK National Centre for the Replacement, Refinement and Reduction of Animals in Research — and are hosted at arriveguidelines.org, where each item has its own explanatory page.

The original guidelines were published by Carol Kilkenny, William J. Browne, Innes C. Cuthill, Michael Emerson and Douglas G. Altman in PLOS Biology in 2010, under the title “Improving Bioscience Research Reporting: The ARRIVE Guidelines for Reporting Animal Research”. That version had 20 items. ARRIVE 2.0 was published in PLOS Biology in 2020 by Nathalie Percie du Sert and colleagues, with a companion Explanation and Elaboration article giving the rationale and worked examples for every item.

The motivating evidence was a survey of 271 published animal studies, co-funded by the NC3Rs and NIH/OLAW and reported by Kilkenny and colleagues in PLOS ONE in 2009. In that sample, only 12% reported random allocation of animals to experimental groups, and among the 35 papers that used qualitative outcome scores only 14% reported blinding. Only 59% stated the hypothesis or objective of the study together with the number and characteristics of the animals used. Those three numbers are the reason the checklist exists.

Common questions about the ARRIVE guidelines

What does ARRIVE stand for?

ARRIVE stands for Animal Research: Reporting of In Vivo Experiments, a reporting checklist for animal research rather than a design or ethics standard. The NC3Rs — the UK National Centre for the Replacement, Refinement and Reduction of Animals in Research — publishes and maintains it, first in PLOS Biology in 2010 and revised as ARRIVE 2.0 in the same journal in 2020.

Who is behind the ARRIVE guidelines — a journal, a funder, or a regulator?

The NC3Rs is behind ARRIVE: a UK government-funded scientific body, not a journal, publisher or regulator. It hosts the checklist and item-by-item explanations at arriveguidelines.org, and the guidelines themselves are published as ordinary peer-reviewed articles in PLOS Biology rather than issued as binding regulation, which is why enforcement depends on individual journals rather than on ARRIVE itself.

Should I report against ARRIVE 2.0 or the 2010 guidelines?

Report against ARRIVE 2.0, the 2020 revision the NC3Rs hosts as the current version. A manuscript written to ARRIVE 2.0’s Essential 10 also satisfies the 2010 list, because 2.0 kept substantially the same content and reorganised it by priority rather than adding a separate set of demands. The reverse does not hold: the 2010 flat list does not ask for the explicit fallback statements that make items 3 and 4 checkable.

The ARRIVE Essential 10

The Essential 10 are items 1 to 10 of ARRIVE 2.0: study design, sample size, inclusion and exclusion criteria, randomisation, blinding/masking, outcome measures, statistical methods, experimental animals, experimental procedures, and results. ARRIVE 2.0 describes this group as information that is “the basic minimum to include in a manuscript, as without this information, reviewers and readers cannot confidently assess the reliability of the findings presented.”

Several of the Essential 10 are more demanding than their one-word titles suggest, and the exact wording matters.

Item 1, study design, asks for the groups being compared including controls, and — sub-item 1b — for “the experimental unit (e.g. a single animal, litter, or cage of animals)”. Naming the experimental unit is a single clause that most manuscripts omit and that determines whether the statistics are defensible; see pseudoreplication.

Item 2, sample size, asks you to “specify the exact number of experimental units allocated to each group, and the total number in each experiment” and to “also indicate the total number of animals used”. Sub-item 2b asks how the sample size was decided, with details of any a priori calculation “if done” — the phrase permits an honest statement that no calculation was performed, which is a far better answer than a retrospective one. On why retrospective calculations do not help, see post hoc power. A specialist distinction the item’s wording elides: a formal a priori calculation needs an expected effect size, which is often genuinely unavailable before a first-in-kind experiment. ARRIVE 2.0’s explanation for item 2b answers that case directly — “If no power calculation was used to determine the sample size, state this explicitly and provide the reasoning that was used to decide on the sample size per group” — so what the item asks for is a stated rationale, not a number reverse-engineered from the result.

Item 3, inclusion and exclusion criteria, is among the least often satisfied items in the set and has three sub-items. Sub-item 3a asks for the criteria and whether they were set a priori, adding: “If no criteria were set, state this explicitly.” Sub-item 3b asks you to report any animals or data points not included in the analysis and why, adding: “If there were no exclusions, state so.” Sub-item 3c asks you to “report the exact value of n in each experimental group” for each analysis. Two of the three sub-items require a sentence even when the answer is nothing, which is why silence is non-compliance rather than a neutral absence.

Item 4, randomisation, asks whether randomisation was used and, if so, “the method used to generate the randomisation sequence”. Sub-item 4b asks for the strategy used to minimise confounders such as order of treatment, order of measurement, and cage or rack position — with the same explicit fallback: “If confounders were not controlled, state this explicitly.” Cage position is not a decorative detail; temperature, light and handling gradients within a rack are real sources of confounding. Simple randomisation is not the safe default it looks like, either. ARRIVE 2.0’s explanation for item 4 calls it “rarely appropriate”, because assigning every animal at once without regard to other variables cannot guarantee balanced groups, and warns that “In studies where sample sizes are small, simple randomisation may result in unbalanced groups” — the ordinary preclinical case. The alternatives it names are randomisation within blocks and stratified randomisation, which balance allocation across a nuisance variable such as cage location or a trait such as weight class.

Item 5, blinding/masking, asks you to “describe who was aware of the group allocation at the different stages of the experiment (during the allocation, the conduct of the experiment, the outcome assessment, and the data analysis)”. The item is written as four stages precisely because “the experimenter was blinded” is uninterpretable — blinded at which stage, and to what. Full blinding is not always achievable — a survival-surgery sham cannot be hidden from the surgeon performing it — but ARRIVE 2.0’s explanation for item 5 is unsentimental about how far that excuse stretches: “Welfare or safety reasons may prevent blinding of animal care staff but in most cases, blinding is possible.” Where it genuinely is not, the explanation asks for a reason rather than an omission: “If blinding was not used at any of the steps outlined in the table above, explicitly state this and provide the reason why blinding was not possible, or not considered.” The four-stage structure exists so that a manuscript can say blinding was infeasible during the procedure itself while still naming who was blinded during outcome assessment and analysis.

Item 6, outcome measures, asks for a clear definition of every outcome assessed and, for hypothesis-testing studies, identification of the primary outcome — the one that drove the sample size.

Item 7, statistical methods, asks for the methods used for each analysis “including software used”, and separately for how the assumptions of the statistical approach were assessed “and what was done if the assumptions were not met”. The second half of item 7 is what turns a vague methods paragraph into something checkable; see wrong statistical test.

Item 8, experimental animals, asks for species, strain and substrain, sex, age or developmental stage, and weight where relevant, plus provenance, health and immune status, genetic modification status, genotype, and any previous procedures. Substrain is named deliberately: C57BL/6J and C57BL/6N differ at the Nnt locus and behave differently in metabolic work, so “C57BL/6” alone does not satisfy item 8a.

Item 9, experimental procedures, asks for what was done, how, when and how often, where, and why — in enough detail for replication.

Item 10, results, asks for summary or descriptive statistics for each group “with a measure of variability where applicable (e.g. mean and SD, or median and range)”, and, where applicable, the effect size with a confidence interval.

The ARRIVE Recommended Set

The Recommended Set is items 11 to 21 of ARRIVE 2.0, described in the guidelines as the group that “adds context to the study described”: abstract (11), background (12), objectives (13), ethical statement (14), housing and husbandry (15), animal care and monitoring (16), interpretation and scientific implications (17), generalisability and translation (18), protocol registration (19), data access (20), and declaration of interests (21).

The label “recommended” causes a predictable error. These eleven items are not optional courtesies; they are the items that are usually enforced by other machinery — journal ethics policies cover item 14, data availability policies cover item 20, and competing-interests forms cover item 21. What ARRIVE 2.0 did was rank the 21 items by what a reader cannot do without, not by what a journal will chase you for.

Three of the Recommended Set carry more weight in preclinical work than their placement implies. Item 15, housing and husbandry, includes cage type, bedding, light cycle, temperature and number of cage companions — the variables that most often explain why a phenotype does not travel between laboratories. Item 18, generalisability and translation, asks how the findings relate to human biology, which is where a single-sex, single-strain, single-age study must state its limits rather than imply none; see mouse model relevance. Item 19, protocol registration, asks whether a protocol with the research question, design and analysis plan was prepared before the study and where it was registered — a question most in vivo papers still answer with silence.

What changed between the 2010 guidelines and ARRIVE 2.0

ARRIVE 2010 presented 20 items as a flat list, all nominally of equal weight. ARRIVE 2.0 kept substantially the same content but did three things to it.

It prioritised. Splitting 21 items into the Essential 10 and a Recommended Set was the central change, made after a structured consultation with researchers, journal editors and statisticians. A flat 20-item list gives an editor no basis for deciding what to insist on; a ranked list does.

It rewrote items to be checkable. ARRIVE 2.0 phrases items as instructions with explicit fallbacks — “if no criteria were set, state this explicitly”, “if there were no exclusions, state so”, “if confounders were not controlled, state this explicitly”. A reader can score those. “Describe inclusion criteria” cannot be scored, because an absent criterion and an unreported criterion look identical on the page.

It moved the experimental unit to item 1. In the 2010 version, the concept was diffused across design and statistics. In ARRIVE 2.0 it is sub-item 1b, before sample size, because every downstream number depends on it.

The ARRIVE 2.0 abstract is unusually candid about why a revision was needed at all: “Despite considerable levels of endorsement by funders and journals over the years, adherence to the guidelines has been inconsistent, and the anticipated improvements in the quality of reporting in animal research publications have not been achieved.”

The experimental unit decides most of the other items

The experimental unit is the smallest division of the experimental material that can be independently allocated to a treatment. ARRIVE 2.0 asks for it in sub-item 1b, and the answer propagates through items 2, 3c, 7 and 10.

If a treatment is applied to the drinking water of a cage of five mice, the cage is the experimental unit and n is the number of cages, not 25 mice across five cages. If a dam is dosed and pups are measured, the litter is the experimental unit for anything transmitted through the dam. If each animal receives its own injection, the animal is the unit — but if three tissue sections are quantified per animal, those sections are technical replicates and n remains the number of animals.

Getting this wrong inflates n by the number of measurements per animal, which shrinks standard errors and manufactures significance. It is a common statistical defect in in vivo manuscripts and the reason reviewers object that n is not independent. It is also invisible unless item 1b is stated, which is exactly why ARRIVE asks for it.

None of this requires discarding the extra measurements to fix, and ARRIVE 2.0’s explanation for item 7 names both acceptable routes: “The hierarchy can induce a clustering effect; for example, cage, litter or animal effects can occur where the outcomes measured for animals from the same cage/litter, or for cells from the same animal, are more similar to each other. This relationship has to be managed in the statistical analysis by including cage/litter/animal effects in the model or by aggregating the outcome measure to the cage/litter/animal level.” Entering cage, litter or animal as a random effect in a hierarchical model uses every section or pup while still returning a standard error appropriate to the true experimental unit; aggregating to one value per unit before analysis is simpler to describe in a methods paragraph. Either satisfies the item. What does not is a t-test or ANOVA run directly on the ungrouped measurements as though each one were an independent animal.

Common questions about sample size and the experimental unit

Is n=3 mice enough for a paper?

ARRIVE sets no minimum n — item 2b requires you to report how the sample size was decided, not to reach a particular number. Three animals per group can be defensible for a large, consistent effect with a stated rationale, but ARRIVE 2.0’s explanation closes the usual shortcut: “Pilot studies alone are unlikely to provide adequate data on variability for a power calculation for future experiments. Systematic reviews and previous studies are more appropriate sources of information on variability.” An “n=3” with no stated reasoning reads as arbitrary regardless of the true effect size.

Why did the reviewer say my sample size isn’t independent when I have plenty of data points?

ARRIVE calls the thing being objected to the experimental unit, and that — not the raw count of data points — is what the comment is about. If 30 tissue sections came from 6 animals, the sections are not independent observations — the true n is 6, not 30. ARRIVE sub-item 1b asks you to name the unit precisely so this is checkable before a reviewer has to raise it.

If I dose the drinking water for a cage of five mice, is n=5 or n=1?

The cage is the experimental unit, so n=1 for that cage. Because the treatment was delivered to the cage rather than to each mouse individually, the cage — not the mouse — is the experimental unit under ARRIVE sub-item 1b, and the study’s n is the number of cages, not the animals inside them. Housing five mice per treatment cage without replicating the cage itself gives an n of 1 no matter how many mice are counted.

How ARRIVE non-compliance is detected in a real manuscript

ARRIVE non-compliance is detected with a set of specific arithmetic and cross-reference checks run against the submitted text, not with a subjective reading of the prose. The following seven checks find most of it, and each can be run against a submitted manuscript in a few minutes.

Reconcile the animal arithmetic. Sum the group ns given in the methods, compare with the total number of animals stated under item 2a, and compare both with the ethics-approval numbers if given. A discrepancy that is not explained by a stated exclusion is an item 3b failure. In practice this check finds attrition that nobody hid deliberately and nobody reported either.

Cross-check n between the figure legends and the methods. Legends reading “n = 3” against a methods section describing six animals per group indicates either unreported exclusions or a change in the experimental unit between the bench and the figure. Item 3c exists to close this gap: it asks for the exact n “for each analysis”, not once for the study.

Search for the words “randomly”, “randomised” and “randomized” and read what follows. “Animals were randomly divided into four groups” names no method and therefore fails item 4a. Compliant text names the mechanism: a random number generator, a randomisation sequence prepared by someone not involved in dosing, block randomisation stratified by cage or by baseline weight. This check needs a disambiguation a simple text match misses: “randomly selected fields for imaging” or “concentrations were varied randomly during optimisation” contain the same word but describe a different decision than group allocation, so a hit still requires a human to confirm the sentence is actually about item 4 before scoring it either way.

Search for “blind” and check whether a stage is named. “The experimenter was blinded” fails item 5, which is written as four stages. A compliant sentence identifies who was unaware of allocation during allocation, conduct, outcome assessment and analysis, and it is entirely acceptable to state that blinding was not possible at a given stage.

Check whether an exclusion statement exists at all. Absence is the failure mode. Under items 3a and 3b, a manuscript with no exclusions must say so; the checkable question is whether the sentence is present, not whether animals were excluded.

Check whether the primary outcome is identified and whether it matches the sample size justification. Item 6b and item 2b are a matched pair. A manuscript reporting eleven outcomes with no primary named, and a power calculation based on an outcome that does not appear in the results, is a common and consequential mismatch — see multiple comparisons.

Check whether the statistical unit in the analysis matches the stated experimental unit. A t-test across 30 sections from 6 animals is an item 1b and item 7a failure simultaneously. Image-based endpoints are where this most often hides; see image quantification reporting.

The NC3Rs publishes a compliance questionnaire built from operationalised questions on the Essential 10, intended for journal staff, editors and peer reviewers rather than for authors. PerfectPaper runs the equivalent checks on a submitted manuscript before a journal does, reporting which Essential 10 items have no supporting sentence and quoting the sentence that would need to change; see in vivo rigour and preclinical peer review.

What a peer reviewer says when ARRIVE items are mishandled

Reviewer comments on ARRIVE failures are recognisable and repetitive. The phrasings below are the common ones in preclinical reports, and each maps to a numbered item.

“The manuscript does not state the experimental unit; if the cage was the unit of treatment, the reported n is inflated.” (item 1b)

“Please state how the sample size was determined. If no a priori power calculation was performed, say so.” (item 2b)

“No information is given on whether animals or data points were excluded. Please state exclusion criteria and report all exclusions with reasons, or confirm that there were none.” (items 3a, 3b)

“Animals are said to have been randomly assigned, but the method of sequence generation is not described.” (item 4a)

“It is unclear who was blinded and at which stage. Please specify whether outcome assessment and analysis were performed blind to allocation.” (item 5)

“A primary outcome is not identified, and it is therefore not possible to determine which comparisons were prespecified.” (item 6b)

“The statistical unit does not appear to match the experimental unit. Sections from the same animal are not independent observations.” (items 1b, 7a)

“The strain designation is incomplete. Please specify substrain and supplier.” (item 8a)

“Housing conditions, light cycle and group size are not reported, which limits assessment of reproducibility.” (item 15)

Two features of this list are worth noting. First, none of these comments is about the biology; they are all about whether the reader can evaluate the biology, which is why they arrive from statistical reviewers as often as from subject specialists. Second, most are unanswerable after the fact — an experiment run without randomisation cannot be retrospectively randomised, and the only honest response is to state the limitation. On writing that response, see how to write a response to reviewers.

Common questions about reviewer comments on ARRIVE items

Reviewer 2 says my sample size isn’t justified — what do I say back?

ARRIVE item 2b lets you answer this honestly: state plainly whether a power calculation was done. Its “if done” wording permits “no a priori calculation was performed” as a compliant answer; a retrospective calculation from your own observed effect is circular and does not satisfy the item. Name the method if one was used — if none was, say so rather than defending a number after the fact.

What do I write if I didn’t blind the outcome assessment?

ARRIVE item 5 asks who was aware of group allocation at each of four stages — allocation, conduct, outcome assessment and analysis — and its explanation asks you to “explicitly state this and provide the reason why blinding was not possible, or not considered” for any stage where blinding was not used. So name the stage and the reason; what fails the item is silence, not an admitted absence of blinding. Expect a reviewer to press hardest on this particular stage, because ARRIVE singles it out: “Blinding is especially important during outcome assessment, particularly if there is a subjective element.”

Do I need to list the substrain of my mice, or is the strain alone enough?

ARRIVE item 8a asks for strain and substrain, so list the substrain — specifically because substrains diverge genetically and phenotypically — C57BL/6J and C57BL/6N differ at the Nnt locus and behave differently in metabolic work — so “C57BL/6” alone leaves a reader unable to judge whether the result would replicate in a colony from a different supplier.

Whether ARRIVE has actually improved reporting

ARRIVE has improved reporting much less than its authors expected, and the evidence for that shortfall is unusually strong — two independent studies measured it directly rather than assuming it.

Baker, Lidster, Sottomayor and Amor examined journals that had endorsed ARRIVE and published “Two Years Later: Journals Are Not Yet Enforcing the ARRIVE Guidelines on Reporting Standards for Pre-Clinical Animal Studies” in PLOS Biology in 2014. Their finding is in the title: two years after endorsement, reporting of the core rigour items — randomisation, blinding, sample size justification — remained very low in the endorsing journals they sampled. Endorsement without enforcement moved almost nothing.

The strongest evidence is the IICARus trial — “A randomised controlled trial of an Intervention to Improve Compliance with the ARRIVE guidelines”, by Kaitlyn Hair, Malcolm R. Macleod and Emily S. Sena with the IICARus Collaboration, in Research Integrity and Peer Review in 2019. The trial randomised manuscripts submitted to PLOS ONE between an intervention arm, in which the journal asked authors to complete an ARRIVE checklist at submission, and standard editorial practice. Compliance with the checklist items was then scored in the manuscripts that went on to be published. Full compliance was essentially absent in both arms, overall compliance was low in both, and the intervention produced no broad improvement. The authors concluded that altering the editorial process to request a completed ARRIVE checklist is not enough to improve compliance, and that other approaches — such as more stringent editorial policies or a targeted approach on key quality items — may be required.

That result should change how an author uses ARRIVE. A checklist ticked at submission demonstrably does not produce compliant text, because ticking a box and writing the sentence are different acts and the trial separated them cleanly. What produces compliant text is checking the manuscript against the items — the seven checks in the section above — and adding the missing sentences.

The mechanism behind that failure is worth naming, because it tells you where to spend effort. A checklist question asks whether something is reported; the manuscript is what does or does not report it. An author who ran the study, knows the animals were allocated by a random number generator, and reads the question “did you randomise?” answers yes truthfully — and the methods section still says only “animals were randomly divided into four groups”, which fails item 4a because the method is not named. The checklist and the text drift apart in exactly the cases where the author’s memory is doing the work the sentence should do. Checking the text, item by item, is the only step that closes that gap, and it is a step that can be done by anyone reading the manuscript, including a coauthor who was not at the bench.

Common questions about whether ARRIVE is required

Is ARRIVE mandatory, or only recommended?

ARRIVE is endorsed by journals and funders rather than enforced as regulation, and endorsement alone has proved weak: the 2019 IICARus trial at PLOS ONE found that asking authors to complete an ARRIVE checklist at submission did not produce broadly compliant published manuscripts. Whether ARRIVE is “mandatory” for a given paper depends on the journal’s editorial process, not on whether a checklist box was ticked.

My journal’s author guidelines don’t mention ARRIVE at all — do I still need to follow it?

Following ARRIVE is worth doing regardless, because the Essential 10 describe what a reviewer needs to judge reliability, not a journal-specific hoop. A 2014 PLOS Biology audit found that journals which had formally endorsed ARRIVE were still not enforcing its core items two years later — a mention in the guidelines predicts little about whether a reviewer raises these points, and its absence predicts even less.

Does filling out the ARRIVE checklist at submission make my paper compliant?

No. The IICARus trial tested exactly this and found that requiring a completed checklist at submission produced no broad improvement in the text of the manuscripts that were later published. Completing a checklist and writing the compliant sentence are different acts: a checklist question can be answered “yes” from memory while the methods section still omits the specific detail the item asks for.

What ARRIVE does not cover

ARRIVE governs reporting of in vivo experiments in the life sciences. Several adjacent things fall outside it.

ARRIVE is not a design or ethics standard. The guidelines ask you to report whether randomisation was used; they do not require it, and they do not grant or substitute for ethical approval. Item 14 asks for an ethical statement, which is a reporting requirement about approvals you obtained elsewhere.

ARRIVE does not cover planning. The PREPARE guidelines, published by Norecopa in 2018, address planning animal experiments and are the upstream companion to ARRIVE’s downstream reporting focus.

ARRIVE does not cover in vitro work, clinical trials, or reviews. CONSORT covers randomised clinical trials, STROBE observational human studies, PRISMA systematic reviews, and ARRIVE animal experiments; MDAR and journal-specific reporting summaries sit across several of these. A study with a mouse cohort and a cell-culture arm needs ARRIVE for the former and something else for the latter.

ARRIVE does not settle whether a finding is interesting. Full compliance with the Essential 10 is compatible with a well-reported study nobody needs, and non-compliance is compatible with an important result. Reviewers who conflate the two produce a novelty objection when the actual objection was rigour, or the reverse.

ARRIVE does not make claims proportionate. An impeccably reported single-model experiment can still be written up as though it established a mechanism in humans; that is a separate defect, checked against the evidence actually presented. See overclaiming.

Using ARRIVE before you have data

Six ARRIVE items cannot be satisfied retrospectively, no matter how carefully the manuscript is written afterward: items 1b, 2b, 4a, 4b, 5 and 6b — experimental unit, sample size justification, randomisation method, confounder control, blinding stages, and primary outcome. Every one of them is a decision made before the first animal is dosed, and no amount of careful writing recovers a decision that was never made.

The practical consequence is that ARRIVE is most useful as a study-planning document that happens to be published as a reporting checklist. Writing the six clauses above into a protocol before the study starts takes perhaps an hour, satisfies item 19 if the protocol is registered, and removes the category of reviewer comment that has no good answer. Reading the checklist for the first time during manuscript preparation, which is when most researchers meet it, leaves those six items as limitations to be confessed rather than methods to be described.

Related

Pseudoreplication · In vivo rigour · Reviewer says n is not independent · Mouse model relevance · Preclinical peer review

Checked before submission by in vivo rigour, which reports which of the ARRIVE Essential 10 have no supporting sentence in the manuscript and names the sentence that would need to change.

Review my manuscript