The FlowScript language

FlowScript is the language you write in the Lab. A FlowScript document is a list of typed blocks (a study, a cohort, an estimate, a survival curve, a task) whose attributes hold numbers with units and intervals, text, references to other blocks and derived values. One or more view lines then say which figures to draw from it. Because every block has a known kind, the Lab can check the document as you type: that each count is a whole number that can exist, that a flow balances at every step, that a stated rate or p-value matches the counts beside it, that a unit is real and of the right kind, that an interval contains its own estimate, that every reference points at something, and that your own stated rules hold. This page is the complete reference: the syntax, every kind of value, units, uncertainty, the expression language, the check block, view lines, all 125 block kinds with every attribute, and every error and warning the Lab can show you, with what to do about each.

At a glance

A first document

// A two-arm trial. Edit any n and the Lab checks the flow again.
study trial {
  design: "Randomised, parallel-group"
  registry: "NCT00000000"
}

cohort screened { n: 1502 source "Screening log, May 2024" }

cohort randomised {
  n: 1200
  from: screened
  excluded: { ineligible: 238, declined: 64 }
}

arm placebo { n: 600, from: randomised, intervention: "Placebo" }
arm drug_x  { n: 600, from: randomised, intervention: "Drug X 50 mg daily" }

outcome death_30d {
  metric: "30-day all-cause mortality"
  control:   { events: 72, n: 600, rate: = events / n * 100 }
  treatment: { events: 48, n: 600, rate: 8.0% }
  test: "chi-squared"
  p: 0.021
}

check arms_balance {
  that: = sum(kind:arm, "n") == randomised.n
  else: "The arms hold {sum(kind:arm, \"n\")} people, not {randomised.n}."
}

view flow: consort(trial)

This document compiles with no issues. Change excluded: { ineligible: 238, declined: 64 } to { ineligible: 238 } and the Lab reports:

Arithmetic inconsistency: screened.n (1502) − sum(excluded) (238) = 1264, but randomised.n is 1200.

Change p: 0.021 to p: 0.9 and it warns that the p-value is not what a chi-squared test gives for 72/600 against 48/600 ("that is 0.021 or 0.027", with and without the continuity correction).

How a document is built

Blocks

A block is a kind, an optional id, and a body in braces:

cohort screened { n: 1502 }

cohort eligible {
  n: 1240
  from: screened
  excluded: { ineligible: 198, declined: 64 }
}

Identifiers

An identifier starts with a letter or _, then letters, digits, _ and -. Kinds, ids, attribute names and bare words are all identifiers.

Tip: Use only letters, digits and _ in block ids. A hyphen is legal in an id, but an expression reads arm-a.n as "arm minus a.n", and renaming with F2 refuses ids with hyphens.

Attributes

An attribute is a name, a colon and a value: n: 1240. Attributes are separated by new lines; commas between them are optional, so { n: 600, from: randomised } and the same on two lines are identical. Each kind declares the attributes it understands. An attribute it does not know is kept as data and warned about: "Attribute "x" is not part of the study schema — it will be carried through as freeform data." A required attribute that is missing is an error: "outcome is missing required attribute metric."

Nesting

Blocks may contain blocks:

survival os {
  time_unit: "months"
  group drug  { label: "Drug X",  times: [3, 8, 12, 20], events: [1, 0, 1, 0] }
  group usual { label: "Usual care", times: [2, 5, 9, 14], events: [1, 1, 0, 1] }
}

Nesting never makes a document invalid. Some kinds list their typical children (a study holds cohort, arm, outcome, view and note); putting something else inside one gives a warning such as "task is not a typical child of study (allowed: cohort, arm, outcome, view, note)." For several figures, nesting carries meaning (a group inside a survival is a curve; inside a table_one it is a column), so check Lab views and renderers for the shape each figure expects.

Comments

// A line comment runs to the end of the line.
/* A block comment
   can span lines. */
cohort a { n: 10 } // a comment may follow code

An unclosed /* is an error ("Unterminated /* … */ block comment") that stops the whole document being read.

Your own kinds

The vocabulary is open. factor laser_power { level: 3 } is a valid block even though factor is not a registered kind: it is not checked against any schema, its numbers can still be referenced and used in expressions, and nothing warns about its attributes. Registered kinds are the ones the Lab's figures, checks and reporting standards understand.

The only alias is captable, which is read as cap_table.

Values

What each attribute type accepts

The block reference gives each attribute a type:

Allowed words are matched without regard to case or punctuation, so very_low, very-low and Very Low all match very_low, and judgement: low and judgement: "low" are the same.

References

A reference points from one block to another.

A bare word in an attribute that is not declared as a reference (mode: web, arm: treatment) is never treated as a reference, so it is never reported as missing.

Duplicate ids are an error: "Duplicate id "s" — every block id must be unique." The first block keeps the id; references resolve to it.

Working with references in the editor

• F12 on an id jumps to its declaration; ⇧F12 lists every use.
• F2 renames an id in every place it is used: its declaration, references, derived values, selectors and view arguments, but not inside strings or comments. The rename refuses an id that is already taken, one that is not a valid id, or one that is not declared, with the reason.
• Hovering an id shows the block and up to six of its numbers.

Sources

Any attribute value may be followed by where it came from. The source is recorded with the number; it shows in the editor's hover, the Provenance pane and the submission bundle.

cohort screened { n: 1502 source "Screening log, exported 2024-05-03" }

dataset trial_data { path: "trial.csv", format: csv, rows: 240 }
distribution age {
  mean: 61.4 source { dataset: "trial.csv", column: "age", rows: 240, aggregation: "mean" }
}

A number with no source is recorded as typed by hand. The Provenance pane lists those so you can decide which need one.

Units

A unit is written straight after a number, on the same line: dose: 5 mg/kg/day. The Lab parses it, knows what kind of quantity it is, converts between compatible units when it compares numbers, and refuses to compare incompatible ones.

Writing a unit

Three words that can follow a value are never units: source, se and to. A unit is only taken when what follows it can end a value (the end of the line, a comma, a closing bracket, an interval, ±, @, a comment, source or se), which is what keeps { x: 330 y: 230 } reading as two attributes.

Warning: In an attribute value, write per 100000 or per 100 000, not per 100,000: the comma ends the value. Hyphenated unit names such as person-years cannot follow a number in an attribute; write py (or pyrs) for person-years. Write the US fluid ounce as floz, not fl oz.

Supported units

Word spellings of three or more letters may differ in case (Days, MOL). Short symbols keep their case: mM is millimolar and Mm megametre.

Some conventions to know: a month is one twelfth of a 365.25-day year, so 3 mo is 13.04 weeks, not 12; mmHg and torr differ in the eighth significant figure; a temperature offset applies only to an absolute temperature, so °C/min is a rate of change; counting labels are dimensionless but are not interchangeable (events/participants is a rate, while events + participants is refused); and Eq and IU cannot be converted to moles or mass, because that needs a substance-specific factor.

Refused symbols

These are warnings on the attribute ("Unit x on key is not recognised: … The number is kept; the unit is carried through as text."). The number is kept; the unit is just not checked.

Plausibility checks

Some values are physically impossible and are errors: a negative count of a counting unit, a duration below zero, a temperature below absolute zero, a proportion or probability outside 0 to 1, a p-value outside 0 to 1. Others are merely implausible and are warnings, against these ranges:

A negative value of a quantity that cannot be negative (a pressure, a mass) is also a warning: "-4.2 mmHg is negative, and pressure is not. If this is a change or a difference, name it as one."

A percentage gets three extra checks, but only in attributes declared as percent or estimate: above 100 ("150% exceeds 100. A percentage of a whole cannot; a relative change can — label it as a change if that is what this is."), below zero, and a note when a tiny percentage looks like a proportion ("0.1% is 0.001 of the whole — one part in 1000. If this is a proportion, write 10% or change the unit to proportion."). A % on any other attribute is never questioned.

Tip: For a difference measure such as a mean difference in blood pressure, put the unit in the block's unit: attribute (estimate e { measure: md, value: -4.2 [-7.3, -1.1], unit: "mmHg" }) rather than on the number, so the blood-pressure plausibility range does not apply to a change.

Comparing numbers in different units

Wherever the Lab adds or compares numbers (a cohort flow, an arm sum, an expression), it uses one rule:

1. If no number carries a unit, the numbers are compared as written.
2. If every unit present is %, they are also compared as written.
3. Otherwise every number must be the same kind of quantity, or the comparison is refused with a unit error and no arithmetic claim is made.
4. Compatible numbers are converted to the first one's unit, so 1500 participants minus 300 participants is 1200, and 128 per 1000 reconciles with 12.8%.
5. A plain number beside a counting label is read as that count. A plain number beside a ratio unit is refused rather than guessed, because 0.12 beside % could mean 0.12% or 12%: "Ambiguous scale: … Give … a unit".

Uncertainty

An estimate can carry its uncertainty in the same value. The Lab fills in the missing half (the standard error from an interval, or the interval from a standard error), on the right scale, and checks what cannot be true.

The notations

A unit may follow the point estimate: 4.2 mmHg [1.1, 7.3]. Units on the bounds are accepted and ignored. The whole notation must be on one line.

Ratio measures and the log scale

When a block says its measure is a ratio (rr, or, hr, irr, rrr, ratio, gmr, sir or smr) or says scale: log, its intervals are treated on the log scale: the standard error is the standard error of the log ratio, the interval is asymmetric, and every bound must be above zero. A ratio of 0 or below is an error: "a ratio measure cannot be 0 — a risk, odds or hazard ratio is strictly positive. Set scale: identity if this is a difference."

Two ways to write an estimate

The estimate kinds (estimate, effect, contrast, meta, subgroup) accept either one value or separate parts:

estimate hr_os { measure: hr, value: 0.68 [0.51, 0.90] }
estimate hr_pfs { measure: hr, point: 0.71, lo: 0.58, hi: 0.87, p: 0.001 }

Both are checked the same way. Two differences matter:

• An estimate written with separate point, lo, hi or se is also checked against its own p:. If the interval excludes no effect but the p-value says not significant (or the reverse), you get "the interval and the p-value disagree … One of the two was copied from a different analysis." A small band around α is allowed for rounding.
• Expressions see the interval of an inline value: but not of separate parts. Using sep.point in an expression gives a note that the result "was computed without the uncertainty of sep.point", with the advice to write the estimate inline.

Writing both an inline value: and separate parts that complete an estimate (a point with lo/hi or se), and having the two points disagree, is a warning: "twice states its estimate twice and the two disagree: value says … and point says …. Keep one of them".

Use level: 0.95 (a fraction) on the block when you need a coverage level with separate parts; level: 95 is an error because the attribute takes a fraction from 0 to 1.

What is checked in an estimate

When df is given on the block, intervals use the t distribution rather than the normal.

Expressions (derived values)

Any attribute value that starts with = is an expression, computed from the rest of the document every time it changes:

cohort randomised { n: 1200 }
cohort analysed {
  n: = randomised.n - excluded.withdrew
  from: randomised
  excluded: { withdrew: 40 }
}

outcome death_30d {
  metric: "30-day mortality"
  control: { events: 72, n: 600, rate: = events / n * 100 }
}

funnel_step trials { n: 2400, title: "Started a trial" }
funnel_step paid   { n: 312, from: trials, title: "Paid" }
kpi conversion { title: "Trial to paid", baseline: = pct(paid.n, trials.n) }

A derived number cannot go stale: when a count is corrected, everything computed from it follows.

Where an expression can go

• Any attribute of any block, and any entry one level inside an object (control: { rate: = … }).
• The expression runs to the end of the line, a comment, a comma outside brackets, or a closing brace, so { a: = x + 1, b: 2 } works on one line. Without the comma, everything up to the end of the line would be read as part of the expression.
• An expression must produce a number, except in check's that:, which must produce true or false. A true/false attribute anywhere else cannot be derived: "risk_table holds a truth value, but nothing reads the answer to a condition written there … Write the answer out, or state the condition as check … { that: = … }".
• Expressions are evaluated in dependency order, so they may refer to each other in any order. A loop is an error: "Circular definition: b.n → c.n → b.n. One of these has to be written out as a number."

Names

A path that names nothing is an error: "nope.n does not resolve to a value — check the block id and attribute name."

Operators

From loosest to tightest binding:

There is no truthiness: a number where true or false is needed is an error ("The condition of if must be true or false, got a number — this language has no truthiness, write the comparison out."). Juxtaposition is never multiplication: 2 randomised.n is an error ("Unexpected randomised after the end of the expression."), not "twice randomised"; write 2 * randomised.n.

Numbers and units inside expressions

Arithmetic in expressions is exact for ordinary decimals: 0.1 + 0.2 is exactly 0.3, not 0.30000000000000004.

Built-in functions

The wrong number of arguments is an error ("round takes 1–2 arguments, got 3."), as is a name that is not a function ("foo is not a built-in function. Available: sum, mean, …").

Aggregates over blocks

sum, mean, median, min, max, count, sd and se can run over a set of blocks instead of a list of values. Write a selector, a comma, and the attribute name in quotes:

cohort randomised { n: 1200 }
arm placebo { n: 600, from: randomised }
arm drug_x  { n: 600, from: randomised }

check arms { that: = sum(kind:arm, "n") == randomised.n }
kpi large_arms { title: "Arms over 500", baseline: = count(where(kind:arm, n > 500)) }

A where test compares a field with a literal (n > 500, label == "control", "role" == "primary"), and tests can be combined with and, or, not and parentheses. The field may be kind, id, or any attribute (dotted for an object entry, such as control.events). The right-hand side must be a literal number, text, true, false or null.

Blocks that lack the field are skipped. A sum over a selector that matches nothing is 0 with a warning ("Nothing matched over kind:sharehodler with a shares, so this total is 0. Check the selector."); every other aggregate over nothing is an error ("mean needs at least one value; nothing matched over kind:widget."). The field must be quoted ("The field of an aggregate must be a quoted name, e.g. sum(kind:arm, "n").") and every aggregate except count needs one. Values in different units are converted when they are the same kind of quantity; values that cannot be reconciled are left out of the total with a warning saying so.

Uncertainty through expressions

An expression that uses an inline estimate (value: 0.68 [0.51, 0.90]) carries its uncertainty through +, -, *, /, ^, ln, exp, sum and mean. For example, md1.value + md2.value with intervals [1.1, 7.3] and [0.5, 3.5] gives 6.2 with its own interval. A plain number you write (the 2 in x * 2) counts as exact.

Other operations keep the point estimate and drop the interval, with a note saying so ("median has no uncertainty algebra, so the result is a point estimate and the interval stops here."). Estimates on different scales cannot be combined: adding a ratio to a difference, or computing 1 - hr.value, is refused ("Cannot subtract these two estimates — one is a ratio measure (0.68) and the other is the plain number 1 … state it as an estimate of its own."). Aggregates over a selector use the numbers alone and carry no interval.

Limits

Checks

A check block states a rule about your document. The Lab runs it after everything else is computed and reports a violation on the status badge like any other issue.

cap_table acme { title: "Acme Ltd", currency: "GBP" }
shareholder founders { name: "Founders", shares: 7000000 }
shareholder seed     { name: "Seed investors", shares: 2000000 }
option_pool pool     { shares: 1000000 }

check fully_diluted {
  title: "Fully diluted share count"
  that: = sum(kind:shareholder, "shares") + pool.shares == 10000000
  else: "Shares and pool total {sum(kind:shareholder, \"shares\") + pool.shares}, not 10,000,000."
  severity: error
  about: acme
}

This document passes. Change the seed shareholder to shares: 1500000 and the badge shows the error "Check fully_diluted: Shares and pool total 9500000, not 10,000,000." Without else, the Lab prints the condition and the numbers it compared, for example "Check fully_diluted: sum(kind:shareholder, "shares") + pool.shares == 10000000 is false — …".

What makes a check fail to run, always reported as an error whatever its severity:

A selector that matches nothing makes a sum 0 and warns, which is usually why a check you believe in comes back false.

The arithmetic the Lab checks

Some rules are built in for specific kinds. They are errors when a number cannot exist and warnings when two numbers disagree.

The arm and outcome effect checks compare intervals, not points, so an adjusted estimate that differs from the crude one is only reported when it excludes it. Tests other than chi-squared and Fisher (a t-test, a log-rank test, a mixed model) are not recomputed.

The Reporting standards checks add many more rules (the ICER follows from costs and effects, a 2×2 table adds up, and so on) when you score a document against a checklist.

View lines

A view is one figure drawn from the document.

The one-line form

view flow: consort(trial)
view survival: km(os)
view timeline: gantt
view rob: rob_traffic_light(rob2)

view, then the view's name, a colon, a renderer, and optionally arguments in parentheses separated by commas. Arguments may be ids, quoted strings or numbers. Most renderers take the id of the block to draw; the risk-of-bias figures take an instrument name instead. A view line can sit at the top level or inside a block (inside a block it must have a name).

The block form

view figure_1 {
  renderer: "forest"
  args: ["pooled"]
  title: "Figure 1"
  caption: "Risk of the primary outcome, random-effects meta-analysis."
  alt: "Forest plot of eleven trials; pooled risk ratio 0.78."
  width: 183 mm
  publish: { journal: "nature", column: double }
}

The block form takes the same renderer and arguments plus presentation and publishing settings (title, caption, alt, width, height, theme, palette, publish, type_pt, line_pt, guideline). Write renderer: and each argument as quoted strings: a bare word in renderer: is not read, and the view then falls back to the CONSORT renderer. A block-form view without renderer: is an error ("view is missing required attribute renderer.").

What is checked in a view line

Two views with the same name are a duplicate-id error ("Duplicate id "a" — every block id must be unique."); until you rename one, the second is listed as a-2 (then a-3 and so on). Several unnamed views are listed as untitled, untitled-2 and so on.

Renderers

The Lab has 36 renderers, most with aliases; there are 101 accepted names in all. Each renderer, the blocks it is built from, its argument and its messages are documented on Lab views and renderers.

Note the two near-collisions: funnel is the business conversion funnel (the meta-analysis funnel plot is funnel_plot), and timeline is an alias of the clinical case timeline (the project timeline is gantt).

A note whose text contains "Illustrative values" or "(illustrative)" is drawn in a band below every view.

Block kinds reference

Each kind below lists its attributes with their type, the rules the Lab enforces, and what they mean. "Takes an id" means a block without one is warned about. "Typical children" are the kinds that nest inside it without a warning. "Read by" names the views that draw it. A rule breach is an error unless noted; the messages are listed under Messages reference.

Shared word lists

Several attributes take a word from one of these lists. Case and punctuation are ignored when matching.

Blocks that do not need an id

note, panel, view, caption, legend, alt, licence and the drawing primitives (shape, label_2d, arrow_2d, annotation, box3d, sphere3d, cylinder3d, plane3d, edge3d, label3d). Every other registered kind takes an id.

Required attributes

Only these attributes are required; leaving one out is an error. Kinds not listed have no required attributes, and the reporting standards say in their own words what is missing.

Core blocks

Four kinds belong to no domain: notes, layout panels, your own rules (check) and views.

note — Annotation. Free-form annotation attached to another block. The id is optional. Drawn below any view when its text marks the values as illustrative.

panel — Layout panel. Composes a named view into a multi-panel figure — grid position, spans, shared axes, panel letter. The id is optional. Read by the panel view.

check — Invariant. A rule you state about this document, the message to show when it is broken, and how loudly to say it. See Checks. Takes an id.

view — View. One figure drawn from the document. Usually written on one line as view name: renderer(args); see View lines. The id is optional.

Clinical trials, protocols and case reports

Trial flows, outcomes, case timelines and protocol schedules. The Lab checks the arithmetic of cohort and arm blocks and the rates, p-values and effects of outcome blocks (see The arithmetic the Lab checks).

study — Study. The study a figure reports: its design, place, period and registration. Takes an id. Typical children: cohort, arm, outcome, view, note. Read by the consort, rob_traffic_light, rob_summary_bar, appraisal_table and intervention_table views.

cohort — Cohort. One box of a participant flow: how many people, which box they came from, and why the others left. Takes an id. Read by the consort, prisma and experiment_flow views.

arm — Trial arm. One trial arm: how many were allocated to it and what they received. Takes an id. Read by the consort, experiment_flow, intervention_table and spirit_schedule views.

outcome — Outcome. One outcome, with each group's counts, the test, the p-value and the effect estimates. Takes an id. Read by the grade_sof, experiment_flow and spirit_schedule views.

timeline — Case timeline. The episode of care laid out against time (CARE item 7) — a title and the origin every event's time is counted from. Takes an id. Read by the clinical_timeline view.

timeline_event — Timeline event. One dated event in a clinical case — when it happened, what happened, and the phase of care it belongs to. Takes an id. Read by the clinical_timeline view.

spirit_schedule — SPIRIT schedule. The SPIRIT 2013 item 13 figure — the study periods and time points a protocol's visits are scheduled against. Takes an id. Read by the spirit_schedule view.

spirit_row — SPIRIT schedule row. One row of the SPIRIT schedule — an enrolment step, an intervention or an assessment, and the time points it happens at. Takes an id. Read by the spirit_schedule view.

Statistics

Estimates and their intervals, pooled analyses, time-to-event data, prediction-model performance, raw distributions and study design. The estimate kinds (estimate, effect, contrast, meta, subgroup) share one shape; see Uncertainty.

estimate — Effect estimate. One effect size with its interval, measure type and scale — the atom every forest plot is built from. Takes an id. Read by the forest, funnel_plot and grade_sof views.

effect — Effect size. An effect size stated on its own — the same shape as estimate, for documents that prefer the word. Takes an id. Read by the forest and funnel_plot views.

contrast — Contrast. A comparison between two named groups, carrying the contrast estimate and the test behind it. Takes an id. Read by the forest and funnel_plot views.

meta — Meta-analysis. A pooled estimate over study-level estimates, with its heterogeneity and small-study diagnostics. Takes an id. Typical children: estimate, effect, subgroup, study, note, view. Read by the forest and funnel_plot views.

subgroup — Subgroup. A study-level or participant-level subgroup estimate, with the interaction test that licenses it. Takes an id. Read by the forest view.

survival — Survival. A time-to-event curve — times, events, at-risk ticks per group, plus the hazard ratio. Takes an id. Typical children: group, series, estimate, effect, note, view. Read by the km view.

roc — ROC curve. A receiver-operating-characteristic curve with its AUC, thresholds and validation provenance. Takes an id. Read by the roc view.

calibration — Calibration. Predicted-versus-observed risk with intercept, slope and Brier score — the other half of discrimination. Takes an id. Read by the calibration view.

decision_curve — Decision curve. Net benefit across threshold probabilities, against treat-all and treat-none. Takes an id. Read by the decision_curve view.

diagnostic — Diagnostic 2×2. A diagnostic-accuracy 2×2 table and everything derived from it — sensitivity through likelihood ratios. Takes an id.

distribution — Distribution. Raw values for a box, violin, raincloud or histogram — plus the summary statistics they imply. Takes an id. Typical children: group, series, note. Read by the raincloud and table_one views.

series — Series. A named run of paired values with its error bars, encoding and redundant non-colour channel. Takes an id. Read by the roc, calibration, decision_curve, control_chart and table_one views.

group — Group. One group's raw values or survival row — the unit a box plot, violin or Kaplan–Meier curve repeats. Takes an id. Read by the km, raincloud and table_one views.

dose_response — Dose–response. Doses against responses with the fitted curve's parameters — EC50, Hill slope, plateaux. Takes an id.

power — Power. A power calculation: the effect assumed, the test, and the design that follows from them. Takes an id. Read by the experiment_flow view.

sample_size — Sample size. The planned and achieved sample size, and the assumption the plan rested on. Takes an id.

table_one — Baseline table. The baseline-characteristics table — groups, variables, and how each row is summarised. Takes an id. Typical children: group, distribution, note, view. Read by the table_one view.

control_chart — Control or run chart. A statistical-process-control or run chart's own parameters — baseline, centre line, limits, when the intervention started, and the rules it is read against. Takes an id. Read by the control_chart view.

Evidence appraisal and reporting

Risk of bias, GRADE, reporting standards, and the narrative "reporting homes" the checklists on the Reporting standards page read and write.

bias_domain — Bias domain. One domain of a risk-of-bias tool — RoB 2, ROBINS-I, QUADAS-2 — as a declared column. Takes an id. Read by the rob_traffic_light and rob_summary_bar views.

bias_assessment — Bias assessment. One study × domain risk-of-bias judgement with its supporting quotation and predicted direction. Takes an id. Read by the rob_traffic_light and rob_summary_bar views.

grade_outcome — GRADE outcome. A GRADE certainty rating with each domain's downgrade and the footnote reasons behind it. Takes an id. Read by the grade_sof view.

sof_row — Summary-of-findings row. One row of a GRADE Summary of Findings table — absolute risks, relative effect, certainty. Takes an id. Read by the grade_sof view.

guideline — Reporting guideline. Which reporting standard this figure claims to follow, at which version, and where the checklist is. Takes an id.

animal_study — Animal-study report. The statements an animal-research report has to make and a figure has no number for — experimental unit, randomisation, blinding, housing, welfare, ethics. Takes an id. Read by the experiment_flow view.

qi_project — Quality-improvement project. A quality-improvement report — the local problem, the rationale, the aim, the context, the intervention, and how attribution was studied. Takes an id.

qual_study — Qualitative-study report. The reporting home COREQ and SRQR share — the research team and its reflexivity, the sampling, the interview conduct, saturation, coding and trustworthiness. Takes an id. Read by the theme_map view.

tidier_intervention — Intervention description (TIDieR). One intervention described so another team could deliver it — what, why, by whom, how, where, when, how much, and how faithfully. Takes an id. Read by the intervention_table view.

appraisal — Appraisal protocol. What a risk-of-bias appraisal states once rather than per study — which tool at which version, who assessed, and how disagreements were settled. Takes an id. Read by the appraisal_table view.

signalling — Signalling answers. One assessment's recorded answers to its tool's signalling questions — what turns an asserted judgement into a re-runnable one. Takes an id.

applicability — Applicability concern. A per-study, per-domain judgement of concern about applicability — a different question from risk of bias, asked of the same pair. Takes an id.

confounder — Confounding domain. One confounding domain a non-randomised comparison was appraised against, and whether it was pre-specified, measured and controlled for. Takes an id.

amstar_item — AMSTAR 2 item. One answered item of an AMSTAR 2 appraisal form — the item number, the answer, and the note behind it. Takes an id. Read by the appraisal_table view.

Health economics

Cost-effectiveness models, strategies, ICERs and probabilistic sensitivity analyses.

econ_model — Cost-effectiveness model. The frame of a cost-effectiveness analysis: perspective, currency, time horizon and discount rate. Takes an id. Typical children: intervention, view, note. Read by the ce_plane view.

intervention — Intervention. One intervention's per-patient cost and QALYs, and what it is compared with. Takes an id. Read by the ce_plane view.

economic_strategy — Economic strategy. One strategy on the cost-effectiveness plane, with its costs, effects and frontier status. Takes an id. Read by the ce_plane view.

icer — ICER. An incremental cost-effectiveness ratio with the threshold and verdict that make it mean something. Takes an id. Read by the ce_plane view.

psa — Probabilistic sensitivity analysis. A probabilistic sensitivity analysis — draws, seed, and the acceptability curve they produce. Takes an id. Read by the ce_plane view.

econ_report — Economic-evaluation report. The narrative half of an economic evaluation — population, setting, how outcomes were valued, model structure, discounting, equity, engagement. Takes an id.

Data binding

Columns of a bound dataset. With rows bound on the Data bench, a column name the data does not have is an error (see The Data and Statistics benches).

column — Data column. One column of a bound dataset — its type, unit, role in the figure, and how it was derived. Takes an id.