Перейти к содержанию

Написание статей · Начальный

Write results that show the size and uncertainty of the findings

Report findings with their magnitude, units, sample size and uncertainty.

11 мин чтения · 26 окт. 2026 г.

Эта страница ещё не переведена, поэтому вы читаете английскую версию.

Главное

  • Lead with the primary outcome.
  • Preserve denominators and pairing.
  • Report uncertain and exploratory findings transparently.
Содержание

Report findings with their magnitude, units, sample size and uncertainty.

Results explain what the study found. They should let readers trace a claim to the observations and analysis supporting it. A good Results section is neither a dump of software output nor a sequence of statements that something was “significant”. It gives the magnitude, direction, uncertainty and relevant context of the findings.

The organising question is: what evidence answers the aim, and what else must the reader know to interpret that evidence fairly? This includes planned outcomes that were inconclusive, failures, missing observations and analyses that complicate the preferred explanation.

Start with the sample or evidence that reached analysis

Before presenting an effect, show how the analysis sample was obtained. Report the number of units considered, eligible, included, excluded and analysed, with reasons for important losses. Distinguish people from visits, animals from cells, devices from runs, and studies from reports.

A denominator is the total relevant to a count or percentage. If 30 of 40 invited students respond, the response rate is 75%. If only 24 respondents complete a particular questionnaire, that analysis has a different denominator. Do not describe all results as n = 40 merely because forty invitations were sent.

For a review, explain the selection flow and number of studies contributing to each synthesis. Ten reports may describe fewer than ten independent studies. For engineering, report unsuccessful runs as well as successful ones. A failure to complete the task is often an outcome, not an inconvenience to delete.

Describe characteristics needed to understand the sample and comparisons. Keep descriptive reporting separate from claims that groups are equivalent. Similar looking means or large p values do not prove that potentially important group differences are absent.

Organise the section around the research questions

Use a clear sequence: sample and data quality, primary result, secondary results, exploratory analyses, and sensitivity or diagnostic findings that change interpretation. This order helps readers distinguish the main analysis from patterns noticed later. APA's quantitative standards explicitly distinguish primary, secondary and exploratory components. S211

A useful paragraph has a topic sentence, the evidence, and a reference to a relevant figure or table. A heading such as “Mean error was lower with controller B in the teaching dataset” is informative when the evidence supports it. A heading such as “Amazing performance improvement” replaces information with enthusiasm.

Choose detail according to its role. The text should explain the main comparison. Tables preserve exact values and denominators. Figures show patterns, distributions or relationships. Repeating every table entry in prose increases length without improving understanding.

Learn the components of a quantitative result

A quantitative comparison usually needs the groups or conditions, the outcome and units, the analysis sample size, an effect estimate, an interval or other justified uncertainty measure, and the relevant test or model when used.

Term What it tells the reader A common reporting error
Mean or median A summary of the observed values Reporting it without the group or scale
Standard deviation, SD Spread of observations around a mean Treating it as uncertainty in the mean
Standard error, SE Estimated sampling variability of an estimator Presenting it without naming the estimator
Confidence interval, CI Statistical precision through a specified interval procedure Describing it as containing 95% of observations
Effect estimate Magnitude of a comparison or relationship Replacing it with a p value
Reference category The group against which another is compared Reporting an odds ratio without naming the comparison

A 95% confidence interval is associated with a procedure's repeated sampling coverage under its assumptions. It is not the range containing 95% of the data, and a conventional frequentist interval is not automatically a 95% probability statement about a fixed parameter. NIST explains the distinction for a mean. S14

A p value describes a tail probability under a specified model and assumptions. It does not measure effect size, practical importance or the probability that the hypothesis is true. The ASA's official summary emphasises these limits. S13

Worked revision: the paired controller comparison

The following values come from the supplied synthetic dataset and are not evidence about a real controller.

Weak version:

Controller B was significantly better than controller A. All results were excellent and confirmed our hypothesis. Figure 1 shows the improvement.

Improved version:

All twelve constructed cases contained values for both controllers and were included in the paired analysis. Mean positioning error was 20.33 mm for controller A, with SD 2.84 mm, and 16.75 mm for controller B, with SD 2.38 mm. The mean within case difference, defined as A minus B, was 3.58 mm. Its t based 95% confidence interval was 2.95 to 4.22 mm. Every constructed pair had a lower B value, with individual differences ranging from 2 to 5 mm. The paired t statistic was 12.46 with eleven degrees of freedom, and the computed two sided p value was approximately 7.89 × 10⁻⁸. These calculations describe the teaching values under the stated model.

The revision explains inclusion, pairing, units, subtraction direction and uncertainty. It also distinguishes the means for each controller from the mean of paired differences. The difference's uncertainty depends on the paired differences, not on treating the controller groups as independent.

Paired controller measurements show individual layouts and the estimated average reduction with a confidence interval.

Example legend: Each line joins controller A and B values from one constructed case. Positive A minus B differences indicate lower error for B. The summary panel shows the mean paired difference and a t based 95% confidence interval. There are twelve paired cases. Values were created for instruction and are not measurements from physical or simulated robots.

A figure legend should be understandable without searching the Methods for the identity of each mark. It does not need to reproduce every procedural detail, but it must identify the observations, units and displayed uncertainty.

Worked revision: social science association

The following counts and estimates are invented solely to demonstrate reporting. They are not outputs from a supplied dataset.

Weak version:

Mentoring improved confidence, although the result was not significant. A trend was observed, suggesting that mentoring should be expanded.

Improved version:

Of 40 invited students, 30 completed the survey, giving a response rate of 75%. Twenty four respondents supplied the attendance, confidence and prior experience variables required by the specified analysis. In this constructed reporting example, an additional mentoring session was associated with a 2.4 point higher confidence score, with a 95% confidence interval from 0.6 points lower to 5.4 points higher. The estimate was adjusted for the recorded prior experience measure. The interval includes associations close to zero as well as a positive association that could be relevant on this scale. The analysis does not establish that attendance caused the difference.

This version avoids calling an uncertain estimate a “trend” merely because its p value might be near a conventional threshold. It names the analysis denominator and predictor increment. An actual paper would also report the score range, model details, missing data handling and supporting diagnostics. The fictional numbers teach sentence structure, not a recommended regression for every survey.

Notice that “adjusted” describes the model's inclusion of specified variables. It does not mean that every source of confounding has been removed. Results can identify the model used without claiming that the resulting coefficient has a causal interpretation.

Report binary outcomes with counts and the right scale

A risk difference compares probabilities by subtraction. A risk ratio compares them by division. An odds ratio compares odds, where odds equal probability divided by one minus probability. These measures answer related but distinct questions.

Consider a hypothetical engineering test with successful completion in 18 of 20 runs for one implementation and 15 of 20 for another. The observed completion percentages are 90% and 75%. The difference is 15 percentage points, while the ratio of percentages is 1.20. Calling this a “15% increase” is ambiguous. These descriptive calculations do not supply a confidence interval or account for pairing, shared environments or repeated runs. Those require the actual design and analysis.

State which category is the event and which is the reference. If failure is the event, the interpretation of a ratio changes. If runs are paired by environment, preserve that matching in the analysis and explanation.

Report findings that do not establish a clear effect

A large p value does not prove no effect, and an interval crossing zero does not imply every meaningful effect has been excluded. Read the interval in relation to scientifically relevant magnitudes. Greenland and colleagues explain why statistical interpretation requires attention to assumptions and plausible effect sizes. S205

A hypothetical estimate of 0.2 units with an interval from 0.1 units lower to 0.5 units higher is different from the same estimate with an interval from 8 units lower to 9 units higher. The first may be sufficiently precise for a particular practical question; the second may leave it unresolved. Whether either supports equivalence depends on a justified equivalence margin and a suitable analysis, not merely failure to reject a zero effect.

Avoid “there was no difference” unless the wording matches the actual evidence and inferential aim. A more informative sentence states the estimate and interval, then explains in Discussion what range of meaningful effects remains possible.

Write qualitative and review Results appropriately

Qualitative findings should connect interpretations to the material that supports them. A theme is an analytical pattern with a clear meaning, not merely a topic label such as “challenges”. Explain what the pattern reveals, show appropriate evidence such as consented quotations, and include important variation or counterexamples. Do not invent quotations or assign population prevalence to purposively selected interview material.

A fictional theme might be “Confidence involved knowing when to seek help.” A useful findings paragraph would explain the distinction between asking for guidance and expecting a complete answer, demonstrate that distinction using real excerpts in an actual study, and discuss cases that complicate it. Some qualitative traditions combine interpretation with findings; follow the chosen methodological approach.

For reviews, Results should show what evidence was included and what the synthesis establishes. Distinguish study characteristics, appraisal findings and the substantive synthesis. Counting papers with p < 0.05 is not a sufficient substitute for comparing estimates, designs and uncertainty. Articles 26 to 28 provide fuller guidance for these forms of evidence.

Include sensitivity analyses and important diagnostics

If a reasonable alternative analysis changes the interpretation, report the change. State what assumption was varied and why. For example, a conclusion may depend on how missing outcomes were handled or whether one influential independent unit is included. Influence concerns how strongly an observation affects an estimate, which is different from simply being numerically extreme.

A diagnostic result can also limit the analysis. Report important model fit problems and the response to them. Do not claim that assumptions were “confirmed” just because a diagnostic test did not cross a threshold. Some issues are better evaluated using plots, design knowledge and alternative models.

Perform a final numerical audit

Check every number against the final analysis output, including the abstract and figure captions. Confirm units, denominators, reference groups, subtraction direction, interval level, sample size and rounding. If software prints a p value as zero, use a sensible bound or scientific notation instead of claiming exact zero.

Practice: Rewrite “B increased success by 20%” using the hypothetical 18 of 20 versus 15 of 20 example. Model answer: “Successful completion occurred in 18 of 20 runs with B and 15 of 20 with A, corresponding to 90% and 75%. The observed difference was 15 percentage points and the ratio was 1.20.” Add an appropriate uncertainty analysis only after determining whether the runs were independent or paired. Use Template 11 to practise result sentences, legends and table notes.

Sources

S211: American Psychological Association. Quantitative Research Design (JARS Quant).

S14: NIST. Confidence limits for the mean

S13: American Statistical Association. 7 March 2016. ASA releases statement on statistical significance and p values. Official summary and six principles.

S205: Greenland S, Senn SJ, Rothman KJ, Carlin JB, Poole C, Goodman SN, Altman DG. 2016. Statistical tests, P values, confidence intervals, and power: a guide to misinterpretations. European Journal of Epidemiology 31:337 to 350.

Файлы

Войдите, чтобы скачать эти файлы.