Жазу · Бастапқы
Build tables that readers can interpret independently
Build tables with clear purposes, units, summaries and denominators.
11 мин оқу · 2026 ж. 26 қаз.
Бұл бет әлі аударылмаған, сондықтан ағылшын нұсқасын оқып отырсыз.
Негізгі тұжырымдар
- Use exact values where they help.
- Define uncertainty and missing entries.
- Audit tables against analysis output.
Мазмұны
Build tables with clear purposes, units, summaries and denominators.
A table is useful when readers need exact values, several related measures or a structured comparison. A figure is usually better for immediately seeing a pattern. Good papers use each deliberately. A table should help someone evaluate the evidence without guessing what a column means or searching several pages for its denominator.
This tutorial shows how to plan a table, report descriptive and inferential statistics, organise evidence from review papers and audit the final output. The worked controller table uses synthetic data supplied with this collection. Other examples are explicitly hypothetical.
Give every table a clear job
Before creating columns, finish the sentence “The reader needs this table to…” Possible answers include understanding the sample, comparing primary outcomes, evaluating a regression model, identifying materials or examining differences among studies.
A table that tries to perform all these jobs usually becomes too wide and too difficult to interpret. Separate a participant description from the primary result when the distinction helps. Move lengthy technical details to a supplement if needed, but keep information essential to evaluating the main claim in the main paper or an immediately accessible location.
In a review article, an evidence table should support synthesis rather than become a catalogue of abstracts. Include dimensions that explain agreement and disagreement: population, design, measurement, timing, comparison, effect estimate, uncertainty and limitations. The appropriate columns follow the review question.
In an engineering paper, a benchmark table may need hardware, dataset, evaluation split and resource limits as well as accuracy. Reporting a performance number without its evaluation conditions can make an apparently simple ranking scientifically misleading.
Design the reader's route through the table
Use a concise descriptive title. Put the main comparison near the beginning, arrange rows in a meaningful order and put units in column headings. Keep category names consistent with Methods and figures. “Control”, “reference” and “baseline” should not silently refer to the same condition in different parts of the paper.
Prefer a small number of well defined columns to many abbreviated ones. If a heading requires several lines of explanation, ask whether the table should be divided. Group related rows with clear text labels rather than relying only on colour or spacing.
Distinguish information types. A descriptive table reports what was observed in the sample. An inferential table reports an estimated effect or model parameter with uncertainty. A materials table identifies resources. These functions can coexist when carefully organised, but a reader should never have to infer which kind of number a column contains.
Name each summary and denominator
For a continuous variable, say whether the entry is a mean with SD, median with IQR, range or another statistic. “Mean ± error” is inadequate. For a categorical variable, show a count and percentage with a clear denominator rule. If denominators differ by row because of missing responses, state the available count for each row.
A hypothetical survey table might report “Access to a quiet study space: 18/30 respondents, 60%”. If only 30 of 40 recruited students answered that item, add that context. Do not put 60% beside a heading of n = 40 without explaining that the item denominator differs.
A weighted survey percentage should be labelled as weighted and accompanied by the relevant unweighted sample size. A count of positive responses cannot necessarily be reconstructed by multiplying a weighted percentage by the unweighted n. The notes should describe enough of the weighting and design to avoid that mistake.
For laboratory and engineering studies, define what n counts. Twelve repeated readings from three specimens are not twelve independent specimens. State the number of independent units and the handling of technical repeats. A single unexplained n in a table title can conceal a major design problem.
Work through a complete result table
The supplied dataset contains twelve synthetic layout cases, each with positioning errors for controllers A and B. The purpose of the table below is to report the paired comparison, not to validate a real controller.
Table 1. Positioning errors and paired differences in twelve synthetic layout cases.
| Quantity | Estimate | Interpretation |
|---|---|---|
| Independent cases assumed for the teaching calculation | 12 | Each case contains one A and one B value |
| Controller A mean error | 20.33 mm | Arithmetic mean across cases |
| Controller A SD | 2.84 mm | Spread of A values across cases |
| Controller B mean error | 16.75 mm | Arithmetic mean across cases |
| Controller B SD | 2.38 mm | Spread of B values across cases |
| Mean A minus B difference | 3.58 mm | Average within case reduction in error |
| SD of within case differences | 1.00 mm | Spread of the twelve differences |
| 95% confidence interval for mean difference | 2.95 to 4.22 mm | t based interval with eleven degrees of freedom |
Table note: All values are constructed for teaching. Differences were calculated within case as A minus B, so positive values indicate lower B error. All twelve cases are included; no values are missing. SD means sample standard deviation. Calculations used unrounded values before display rounding. The interval illustrates a conventional paired t calculation under its assumptions and provides no empirical validation of a navigation system.
The note does several jobs. It defines the subtraction, the source of n, the missingness rule, the summary and the interpretive boundary. Readers do not have to guess whether the interval belongs to A, B or their difference.
The interval was not obtained by subtracting the endpoints of two unrelated group intervals. Pairing affects the variance of the difference. A table that lists only group summaries should not encourage readers to reconstruct uncertainty through an invalid shortcut.
Distinguish estimates from tests
An effect estimate describes a magnitude, such as a mean difference, risk ratio or regression coefficient. A confidence interval describes uncertainty under the stated method and assumptions. A p value relates the observed test statistic to a specified null model. These quantities have different roles.
If p values appear, specify whether they are raw or adjusted, name the adjustment and identify the comparison family. Avoid unexplained asterisks. A star may attract attention while hiding whether the effect is large enough to matter or too imprecise to guide a decision.
A regression table should define the outcome, model type, units, reference categories, covariates and analysis sample. For example, a coefficient of 2 is uninterpretable if readers do not know whether the predictor unit is one hour, ten hours or one SD. For logistic regression, distinguish log odds coefficients from exponentiated odds ratios.
Do not add a significance test to every row of a descriptive table simply because software offers it. The need for a comparison follows the design and research question. A table of sample characteristics can describe who was studied without becoming a collection of unplanned tests.
Use precision that reflects the measurements
Keep full precision during calculation and round only the displayed output. Comparable entries should use consistent precision unless their scale requires otherwise. Too many digits imply measurement accuracy or certainty that the study may not support. Too few digits can conceal important differences or round a small p value to zero.
A hypothetical mass measurement recorded to the nearest 0.1 g does not usually need a sample mean presented to eight decimal places. However, an intermediate calculation should retain precision until the final reporting decision. Rounding every observation before calculating a small difference can alter that difference unnecessarily.
Use explicit labels for missing, not applicable and below detection values. A blank cell is ambiguous. Zero should mean an observed or calculated zero, not “not measured”. If a value falls below an assay's detection threshold, state the threshold and how it was handled in analysis.
For intervals that include negative values, write the limits unambiguously, such as “negative 1.2 to 3.4”, or use a clear signed numerical convention. In ratio tables, remember that one is the null value. Do not use the same interpretation for all columns merely because each contains an interval.
Build review tables for comparison and reasoning
A weak review table has columns for author, year and a long conclusion copied from each abstract. It can become a list without synthesis. A better table uses the same questions for every study and preserves differences that matter.
For a hypothetical review of classroom feedback, consider columns for study design, learner group, feedback method, comparator, outcome timing, main estimate and key limitation. A study of immediate quiz scores and a study of performance a year later do not automatically measure the same outcome, even if both conclude that feedback “works”.
For a hypothetical review of battery cooling methods, include the operating conditions, ambient temperature, heat load, performance metric, energy cost and whether evidence comes from simulation or physical experiments. A lower peak temperature under an easier workload should not be ranked as superior without qualification.
Link every study row to its actual citation, and distinguish “not reported” from an observed absence. If a paper does not mention randomisation, do not automatically write “not randomised” unless other information supports that conclusion. If several reports describe the same study, prevent duplicate counting by using a study identifier and linking its associated reports.
An evidence table can also distinguish the review authors' appraisal from the original authors' claims. Put a concise risk of bias or limitation judgement in a separate column and explain the assessment method. Do not invent a numerical quality score merely because the table looks incomplete without one.
Make tables accessible on the website and in a PDF
Use real text tables rather than screenshots of spreadsheet cells. A screen reader needs programmatic relationships between headers and data cells. W3C guidance explains header cells and the use of scope attributes where direction would otherwise be ambiguous. S407
As an author, provide clear headings, simple structure and descriptive notes. Avoid merged cells when a simpler layout conveys the information. Make sure a narrow screen can still reveal the relationship between each value and its heading. A very wide table may need a responsive layout or a downloadable supplementary version.
For a PDF, check page breaks, repeated headers, line wrapping and whether notes remain close to their tables. Do not shrink the font until a large table technically fits but is unreadable. General manuscript guidance also emphasises clear table titles, explanations and appropriate notes. ICMJE is one medical publishing example; follow the destination's requirements for your field. S01
Audit the table against the analysis
Generate tables from analysis outputs where possible. Manual copying creates opportunities for old estimates and new intervals to become mixed. Keep the extraction or calculation script and record which dataset and analysis version produced the table.
Perform a deliberate audit:
- Recalculate totals and check denominators against the data.
- Confirm units, effect direction and reference categories.
- Check that intervals belong to the displayed estimates.
- Compare quoted values with the abstract, Results and figures.
- Verify all source citations in review tables.
- Inspect the exported table at its intended reading size.
If a result changes, update every occurrence rather than correcting one cell. A manuscript that gives 3.58 mm in the table and 3.85 mm in the abstract creates uncertainty about which analysis is authoritative.
Practice and model answer
Task: A table lists “Group A: 12.4 ± 1.2; Group B: 10.8 ± 1.0; p = 0.04”. What is missing?
Model answer: The variable and unit, group definitions, sample sizes, meaning of the ± quantity, independent or paired design, statistical test, direction and size of the intended contrast, confidence interval, missingness rule and any multiplicity adjustment. The displayed values are hypothetical and insufficient to reconstruct those details reliably.
Task: An evidence table marks a method “absent” whenever the source article does not describe it. Is this defensible?
Model answer: Usually the appropriate label is “not reported” unless there is evidence that the method was not used. Reporting completeness and actual conduct are related but distinct. Explain the appraisal rule and avoid filling gaps with assumptions.
Use Template 11 to build table notes and check whether a reader can interpret the table independently.