How to Report Statistics in AMA Style

Most statistics guides stop at the point where the software prints a result. The manuscript does not. Between the output table and the Results section sits a set of typographic conventions that a statistically literate reviewer reads as fluently as the numbers themselves, and a sentence that breaks them signals, fairly or not, that the author is unfamiliar with the literature they are writing for. This guide sets out the conventions of the American Medical Association (AMA) Manual of Style, then gives a finished, correctly formatted sentence for the five tests clinical researchers report most often. Every number in those sentences was computed from a public dataset; none was generated.

What "AMA style" means and who asks for it

The AMA Manual of Style, now in its 11th edition, is the editorial standard of the JAMA Network journals and the model for the instructions to authors at most clinical and surgical journals.¹ Its statistics chapter governs how a P value is written, how an interval is punctuated, which symbols are italicised, and how descriptive statistics are presented. The Publication Manual of the American Psychological Association (APA), now in its 7th edition, is the counterpart used by psychology, education, and many nursing and allied-health journals.² The two agree on what to report and disagree on how it looks on the page; the table near the end of this guide lists the differences that matter.

Two points before the rules. First, the journal's own instructions always win: if a target journal asks for something different, follow the journal. Second, a style manual tells you how to write a result, not which result to write. What belongs in a complete result, for any test, is the estimate, its precision, the test statistic with its degrees of freedom, the exact P value, and an effect size. The SAMPL guidelines, written for biomedical journals specifically, make the same list.³

P values

AMA style writes the letter as a capital italic P, drops the zero before the decimal point because a probability cannot exceed 1, and sets spaces around the operator: P = .03, P < .001.¹ The number of digits follows the value:

• Two digits when P is .01 or larger: P = .08, P = .72.
• Three digits when P is between .001 and .01: P = .006.
• P < .001 for anything smaller. The value is never written as P = .000, which is a rounding artefact rather than a probability.
• Keep three digits when rounding to two would carry the value across .05. A result of P = .046 stays P = .046; rounding it to P = .05 would change what the reader concludes.

Two habits the Manual rules out entirely. "NS" or "not significant" in place of a number tells the reader nothing about how far from the threshold the result fell, and "P < .05" when the exact value is known is the same omission in a different costume. Report the value the software gave you. The American Statistical Association's 2016 statement on P values makes the broader case: a P value does not measure the size of an effect or the importance of a result, so the effect estimate and its interval must sit beside it, never behind it.⁴

One note on software. StatsPlease prints every P value to three decimals (P = .075), so that a value near a threshold is never hidden by rounding. The AMA digit rule then trims it: P = .075 becomes P = .08 in the manuscript, while P = .006 and P < .001 are already in their final form. The output cards in this guide show the three-decimal engine output; the sentences beside them show the trimmed version you would submit.

Confidence intervals and descriptive statistics

An interval is written with the level, a comma, and the two bounds joined by a hyphen or en dash: 95% CI, 1.40–2.52. When a bound is negative, the dash becomes ambiguous, so the word "to" replaces it: 95% CI, −0.4 to 8.9.¹ Square brackets around an interval are APA, not AMA. Inside a parenthetical the elements are separated by semicolons, which is why the finished sentences below read "mean difference, 4.2 years; 95% CI, −0.4 to 8.9; P = .08".

The plus-minus sign is retired. A mean and its standard deviation are written as mean (SD), and inside a parenthetical that already uses round brackets the SD moves to square ones: 54.8 [17.4] years. The Manual's reason is that the symbol never says whether the quantity after it is a standard deviation, a standard error, or a confidence bound, and those three describe very different things.¹ When a variable is skewed, the median and interquartile range replace the mean and SD, written as median (IQR) with the two quartiles, not a single width: 3.54 (2.25–5.16) mm.

Percentages carry their numerator and denominator the first time they appear, because 68.1% of 304 patients and 68.1% of 22 patients are different findings: 207 of 304 (68.1%). Decimal places match the precision of the measurement. A blood pressure recorded in whole millimetres of mercury does not acquire two decimals because the software printed them, and a percentage from a sample of 50 does not need one.

Effect sizes and statistic symbols

Every inferential test is reported with an effect size, because the test statistic and its P value answer only whether a difference is detectable, not how large it is.³ The CONSORT statement's outcomes item asks for "the estimated effect size and its precision (such as 95% confidence interval)" for each primary and secondary outcome, and the same expectation now runs through the reporting guidelines for observational designs.⁵ Which effect size depends on the test: a mean difference with its interval and a standardised difference such as Cohen d for a t test, a rank-biserial r or a Hodges-Lehmann shift for a Mann-Whitney test, an odds ratio or risk ratio for a 2 × 2 table, η² for an analysis of variance, and the coefficient itself for a correlation.

Symbols follow a simple typographic rule. Latin-letter statistics are italicised: t, F, U, r, d, P, N, n. Greek letters are set upright: χ², η², ρ. Degrees of freedom sit in a subscript or in parentheses immediately after the statistic (F_2,854 or F[2, 854]; either is acceptable, so pick one and keep it). A capital N is the whole sample and a lowercase n is a subgroup. Coefficients and effect sizes keep their leading zero (d = 0.25, r = 0.65) because, unlike a P value, they are not confined below 1.

The cheat sheet: five tests, five finished sentences

Each row below gives the elements a complete report needs and a sentence written in AMA form. The sentences are not templates with invented numbers. Each was computed with scipy from a public clinical dataset, credited in the linked guide, and the linked guide walks through the same result element by element.

The first four sentences come from two cohorts: the Odense melanoma cohort of 205 patients treated surgically between 1962 and 1977,⁷ and a registry of 857 patients with acute coronary syndrome distributed as a teaching dataset with the moonBook R package.⁹ The correlation is from Vanderbilt's diabetes screening cohort.⁸ Here is the t test result exactly as the engine returns it, before the AMA digit rule trims the P value:

Example data: the melanoma cohort of Andersen, Borgan, Gill, and Keiding (1993), 205 patients treated at Odense University Hospital, Denmark, a public teaching dataset distributed with the R survival package. Computed with scipy, not generated.

Notice what the non-significant sentence does. It gives both means, so the reader can see the four-year gap the test was assessing; it gives the interval, which runs from a small deficit to an excess of nearly nine years, so the reader can see that the study did not rule out a real age difference; and it gives the exact P value rather than a verdict. "No significant difference" is not the same claim as "no difference", and the interval is what stops a reader from confusing them.⁶

The Methods paragraph

A reviewer reads the statistical methods before the results, and an incomplete Methods paragraph invites the criticism before a single number has been checked. The SAMPL guidelines list what it should contain: each test named alongside the question it answers, the assumption checks and how a violation was handled, the significance level and whether tests were two-sided, how missing data were treated, and the software with its version.³ Written out for the t test above:

"Continuous variables were summarised as mean (SD) when approximately normally distributed and as median (IQR) otherwise; normality was assessed with the Shapiro-Wilk test and equality of variances with the Levene test. Age at surgery was compared between patients with and without tumour ulceration using the Welch t test, which does not assume equal variances. Effect sizes are reported as Cohen d with 95% CIs for the mean difference. All tests were two-sided with α = .05. Analyses were performed in StatsPlease (version as stated), which computes results with scipy."

Each clause answers a question a reviewer would otherwise have to ask. Which test, and why that one. What was checked before it ran. What the effect size is. What the threshold was. What produced the numbers.

AMA vs APA: the differences that matter

The two manuals ask for the same content and set it differently. Submitting APA typography to a JAMA Network journal, or the reverse, is a formatting correction the copyeditor will make for you, and it is entirely mechanical once the differences are known.^1,2

StatsPlease writes its deterministic results sentence in AMA form by default; the Reporting Style picker in the Analyze flow switches the sentence to APA when the target journal asks for it. The computation underneath is the same either way.

Where ChatGPT gets this right, and where it does not

Ask ChatGPT or any other large language model to "write this result in AMA style" and it will usually get the typography right, because the conventions above appear in thousands of published papers it has read. What it cannot do is compute the number the sentence is about. Paste in a spreadsheet and ask for the t test, and the model produces a sentence with the correct shape wrapped around a statistic it never calculated, a P value that may or may not correspond to the data, and an interval that looks plausible because plausibility is what the model is optimised for. The formatting is real and the result is not.

The safe division of labour is the opposite one. Let a deterministic engine compute the result, so that the same CSV in R or SPSS returns the identical statistic, and then apply the style rules, by hand or by software that applies them mechanically. StatsPlease does both steps without a model in the loop: the numbers are computed by scipy and the sentence is assembled from a fixed template. The AI never computes a number.

Frequently asked questions

Does AMA style use a leading zero before a P value?

No. AMA style writes P = .03 and P < .001, with a capital italic P and no zero before the decimal point, because a P value cannot exceed 1. Coefficients, effect sizes, and other quantities that can exceed 1 keep their leading zero, so a Cohen d of 0.25 is written 0.25, not .25.

How many decimal places should a P value have in AMA style?

Two digits when P is .01 or larger (P = .08), three digits when P is between .001 and .01 (P = .006), and P < .001 for anything smaller. Keep three digits when rounding to two would move the value across .05, so P = .046 stays P = .046 rather than becoming P = .05. Never write P = .000, P = NS, or P < .05 when the exact value is known.

What is the difference between AMA and APA statistical reporting?

The numbers are identical; only the typography differs. AMA writes a capital italic P, a confidence interval as 95% CI, 1.40–2.52, and descriptives as mean (SD). APA writes a lowercase italic p, square brackets around the interval as 95% CI [1.40, 2.52], and the abbreviations M and SD. AMA is the house style of JAMA Network and most clinical and surgical journals; APA is used by psychology, education, and many nursing journals.

References

1. Christiansen SL, Iverson C, Flanagin A, et al. AMA Manual of Style: A Guide for Authors and Editors. 11th ed. New York, NY: Oxford University Press; 2020. https://doi.org/10.1093/jama/9780190246556.001.0001
2. American Psychological Association. Publication Manual of the American Psychological Association. 7th ed. Washington, DC: American Psychological Association; 2020. https://doi.org/10.1037/0000165-000
3. Lang TA, Altman DG. Basic statistical reporting for articles published in biomedical journals: the "Statistical Analyses and Methods in the Published Literature" or the SAMPL Guidelines. Int J Nurs Stud. 2015;52(1):5-9. https://doi.org/10.1016/j.ijnurstu.2014.09.006
4. Wasserstein RL, Lazar NA. The ASA Statement on p-Values: Context, Process, and Purpose. Am Stat. 2016;70(2):129-133. https://doi.org/10.1080/00031305.2016.1154108
5. Schulz KF, Altman DG, Moher D; CONSORT Group. CONSORT 2010 Statement: updated guidelines for reporting parallel group randomised trials. BMJ. 2010;340:c332. https://doi.org/10.1136/bmj.c332
6. Altman DG, Bland JM. Absence of evidence is not evidence of absence. BMJ. 1995;311(7003):485. https://doi.org/10.1136/bmj.311.7003.485
7. Andersen PK, Borgan Ø, Gill RD, Keiding N. Statistical Models Based on Counting Processes. New York, NY: Springer-Verlag; 1993.
8. Vanderbilt University Department of Biostatistics. Public teaching datasets. https://hbiostat.org/data
9. Moon KW. moonBook: Functions and Datasets for the Book by Keon-Woong Moon. R package; the acs dataset (857 patients with acute coronary syndrome). https://CRAN.R-project.org/package=moonBook