My p Value Is .06: What It Means and What to Write

You look at the output, you see .060, and you do the arithmetic in your head: ten thousandths short. From there the temptation is single and strong, to write “marginally significant”, “approached significance” or “showed a trend”, because it feels like the honest way to save six months of work. It is the most expensive sentence in the paper: of all the ways to close out a borderline result, that one is the only one a reviewer reads as an admission that you would have taken .05 without blinking had it fallen your way. There is a lot you can legitimately say about a .060, and none of it needs an adverb.

What p = .060 actually tells you (and what .042 would)

A p value answers one narrow question: if the effect in the population were exactly zero and the rest of the model assumptions held, how often would data at least as extreme as yours turn up? A .060 answers that, under that assumption, data like yours or more extreme would appear in roughly one repetition in 17. It does not say there is a 6% chance the null is true, or a 94% chance your hypothesis is right, or that the effect is small. It says something about data under an assumption, and nothing about the probability of your hypothesis.

Two twin studies show the problem better than any definition. Both compare a four-week sleep program against a waitlist in undergraduates, with 64 per arm. The first returns t(126) = 1.90, p = .060, d = 0.34, 95% CI [-0.01, 0.69]. The second returns t(126) = 2.05, p = .042, d = 0.36, 95% CI [0.01, 0.71]. The entire distance between the two studies is two hundredths of d. One gets written up as an intervention that works, the other as one that does not.

Read the intervals instead of the two p values: [-0.01, 0.69] and [0.01, 0.71] say the same thing about the world. Both are compatible with an effect large enough to change practice and both are compatible with something negligible. Evidence is continuous; what is discrete is the decision you made when you cut it at a point, and that cut does not sort results into two different kinds of reality. How to read an interval like that, segment by segment, is in interpreting confidence intervals.

So a .060 is not a half-result or a broken one. It is a result with a wide interval, which is a specific, writable and considerably more informative thing than the adverb you were about to reach for.

The .05 threshold is arbitrary, and binary the moment you choose it

The cut-off came out of no proof. Fisher (1925) offered .05 as a convenient level —one in twenty— for deciding which results were worth returning to, and said plainly that a researcher could set a different one for a different problem. Particle physics works at five sigma. In psychology there is a serious proposal to move the bar to .005 for new findings (Benjamin et al., 2018). None of that turns .05 into a constant of nature; it is a professional agreement.

What the agreement does do is split the space in two, and that binary half of the story is not Fisher's: it comes from the Neyman-Pearson decision framework, where alpha is an error rate you commit to before seeing the data; choose it afterwards and it stops being the rate you claim to control. “Significant” only means something inside that commitment. “Marginally significant” lives on mixing the two frameworks: you use the decision rule for the headline and the evidence continuum for the exceptions, whichever suits the sentence you are writing.

There is a home test that settles it. Nobody writes “marginally non-significant” about a p of .042, and nobody says an effect is “trending toward null” when the value slides from .012 to .049. The vocabulary of the margin only ever travels in one direction, the one that favors the author's hypothesis. When a term shows up only when it helps, it has stopped describing evidence and become an argument wearing a description's clothes.

The arbitrariness of the threshold is a good reason not to hang your whole conclusion on it, and that debate is old and well put in the case against p value tyranny. The way out, though, is not to soften the threshold case by case and always in your own favor. It is to report the effect with its interval and let the reader see exactly what you are seeing.

The sentences you cannot write

“Trending toward significance”. A trend needs a series: repeated measurements, an increasing dose, something that moves. Your p is a fixed number computed from fixed data. It is not on its way anywhere and it will not reach .04 if you wait.

“Marginally significant”. There is no margin. A margin would be a lane of indecision agreed in advance, and you agreed to no such thing: you set a cut at .05. If you genuinely wanted an intermediate zone, you declare it beforehand with an explicit rule, and then the .060 falls inside a criterion rather than an excuse.

“Almost reached significance”, “fell just short”, “bordered on significance”. All three describe a motion that does not exist, and all three land the same way with a reviewer: if .060 counts as an almost-yes, then .05 was never your criterion.

“Non-significant, but in the predicted direction”. Describing the direction of the estimate is fine; leaning on it as support is not. If the true effect were exactly zero, half of all studies would produce a difference in the direction you predicted, by chance alone. Direction on its own, with no magnitude and no interval, does not separate your hypothesis from noise.

None of this is a personal failing. Pritschet, Powell and Horne (2016) analyzed psychology articles across four decades and found that calling values between .05 and .10 “marginally significant” became more frequent over time in the journals they examined. It is a habit of the field, which is exactly why reviewers have it tagged: when the phrase appears, what comes into question is not the phrase, it is whether the rest of the analysis was decided before or after the results were on screen.

The sentences you can write, word for word

In Results, no adjectives. Full statistic with its degrees of freedom, the exact p, and the effect size with its interval: “The program did not differ from the waitlist at the pre-specified threshold, t(126) = 1.90, p = .060, d = 0.34, 95% CI [-0.01, 0.69].” The same comparison run as an ANOVA reads F(1, 126) = 3.61, p = .060, partial eta squared = .028, and carries exactly as little drama. Italics, decimals and leading zeros for the p itself are covered in how to report p values in APA 7.

In the Discussion, everything the interval licenses. “The confidence interval is compatible with effects as large as d = 0.69, a magnitude that would matter in practice for a program like this, and also with no effect; the study does not distinguish between the two.” Any methods reviewer will sign that sentence, it says considerably more than “marginally significant”, and it promises nothing the data cannot carry.

You can also state what your design was able to see. With 64 per arm, a two-tailed alpha of .05 and 80% power, the smallest detectable effect sits near d = 0.50: your study was never built for an effect of 0.34. That is a sensitivity analysis, computed from the design rather than from the result, which is the other way round from observed power: that one is derived from the effect you got and adds nothing to the p value you already have. If you also want to guide whoever replicates, detecting a d of 0.34 at that same power would take roughly 136 per arm, more than double what you had; give that figure as a guide, with the caveat that an effect estimated with an interval this wide is an imprecise basis for planning. To get d and its interval straight from the statistic, use the effect size calculator.

What none of those sentences allows is a claim that there is no effect. That is a different test, equivalence or a Bayes factor, and it is set out in what to do when your results are not significant. With an interval reaching 0.69, that test will come back inconclusive anyway, and writing “inconclusive” is the correct move when inconclusive is the truth.

The other .060: the one that shows up after correction

There is a version of this that stings more, because the result was briefly on the right side of the line. You ran five planned comparisons, the critical one came back at p = .012, and after correcting for multiplicity the adjusted value landed on .060: Bonferroni multiplies by the five tests and gives exactly that, and Benjamini-Hochberg returns the same for the smallest of the set unless another comparison in the family pulls the adjustment down. The temptation changes shape but not substance: report the .012 in the text and park the correction in a footnote, or write “significant before correction” as though that were a category.

It is not one. The rule you committed to includes the family of tests and the adjustment method; your threshold was never .05 on its own, it was .05 after correction. What you write is short and goes in whole: “the comparison did not survive adjustment, p = .012, p_FDR = .060, d = 0.34, 95% CI [-0.01, 0.69].” The uncorrected value belongs in the table because it is information, not because it is a spare result to deploy when the other one lets you down.

Where a reviewer really digs is the family. If the preregistration listed five comparisons and the manuscript shows three, or if the correction covers one block of analyses but not the decisive one, the .060 becomes the smallest of your problems. And there is a reading worth keeping in view: with Benjamini-Hochberg the adjusted values depend on the whole set, so a .012 that turns into .060 is telling you something about your battery of tests, not only about that one effect.

Four ways to push .060 under the line, and why all four leave marks

Switch to one-tailed. The fastest: .060 becomes .030 by dividing by two. It is legitimate only if the directional hypothesis and the one-tailed test were fixed before you looked, and even then it belongs in the Method section. Choosing it after seeing which way the difference went is not a different test; it is setting the threshold with the data in view.

Collect more participants until it drops. If the stopping rule is “keep going until it crosses .05”, the false positive rate is no longer the 5% you believe you are controlling, and combined with two or three other flexible choices it climbs sharply (Simmons et al., 2011). Adding participants is perfectly legitimate; what contaminates the analysis is a stopping rule that depends on the result you want.

Drop an outlier. An exclusion rule invented on seeing the .060 is not a rule, it is a lever. If the extreme case is genuine and there are reasons to exclude it, the manuscript carries both analyses, with and without, and says which is primary and why.

Add a covariate that helps. Adjusting for something planned is analysis; adjusting for the first variable that pushes the p down is model selection guided by the outcome, and it leaves an obvious trace: the covariate appears in the critical model and nowhere else in the paper.

None of the four is invisible. Degrees of freedom that do not match the sample described in Method, a one-tailed test surfacing in one comparison out of ten, an orphan covariate: all of that is visible on a careful read, without access to your data. The clean route costs nothing: the pre-planned analysis runs as primary, with its .060 written plainly, and anything else you tried is labeled exploratory. A well-declared exploratory analysis costs you no credibility; one dressed up as confirmatory takes the whole paper's credibility with it.

What the reviewer will ask, and how to close it off

A borderline result triggers three questions, and all three are better answered inside your text before anyone asks them. What effect your design could detect. What is still compatible with your data. And whether the analysis that produced the .060 is the one you planned or the one left standing after you tried a few.

The third decides the manuscript. With a preregistration or a dated analysis plan, the .060 is simply a number and the conversation ends. Without one, the reviewer can only judge from indirect signals, and a “marginally significant” weighs far more there than it looks: it is the tell that the threshold was read with the calculator already running.

There is a real consolation here, and it is not “don't worry”. Your estimate does not vanish: d = 0.34 with its variance enters any future meta-analysis on exactly the same footing as a significant study's, because no threshold operates at that level and what gets combined is effects, not decisions. What determines whether your study contributes or disappears is not the .060; it is whether you reported the effect and the interval precisely enough for someone else to use them.

Frequently asked questions

Is p = .06 significant?

Not at an alpha of .05. A .060 sits above the threshold and the decision is negative, with no qualifiers. What you can report is where the estimate points and what the interval still allows, written with the numbers in view: t(126) = 1.90, p = .060, d = 0.34, 95% CI [-0.01, 0.69]. What you cannot do is call it significant with an adverb attached.

Can I say my result approached significance?

No, because nothing approaches anything. A p value is a fixed number computed from fixed data: the one you have does not move. A larger study would return a different p, higher or lower, but that would be new data rather than a trend in the value you already have. Write the exact value and the effect size with its interval instead, and describe the direction of the estimate without promising a destination: “the difference favored the intervention, but the interval includes zero”.

Can I switch to a one-tailed test to get under .05?

Only if the directional hypothesis and the one-tailed test were fixed before you saw the data and the Method section says so. Switching after seeing a .060 turns it into .030 by arithmetic rather than by evidence: you are dividing by two, not measuring better. It is one of the first things a reviewer checks when a single one-tailed comparison shows up among ten two-tailed ones.

Can I collect more participants until p drops below .05?

Not with that stopping rule, because checking the p and deciding whether to continue destroys the error control that gives the threshold its meaning. You can extend the sample if the additional N is decided without looking at the result, declared in the manuscript and analyzed in full; or if you planned a sequential design with corrected boundaries from the start. The dividing line is whether the stopping rule depends on the answer you want.

Why is the threshold .05 and not some other number?

Because Fisher (1925) offered it as a convenient level, one in twenty, and the field adopted it. Nothing in the mathematics privileges that number: particle physics works at five sigma, and there is a well-argued proposal to move the bar to .005 for new findings in the social sciences (Benjamin et al., 2018). Being arbitrary does not make it flexible after the fact; what gives it meaning is having fixed it beforehand.

How do I report p = .06 in APA 7?

Italic letter, no leading zero, exact value: p = .060. APA 7 accepts two or three decimals, and for a borderline result three is the better choice, because .060 tells the reader where you are and .06 does not. Never replace it with p > .05 when you have the exact figure, and always pair it with the full statistic, its degrees of freedom and the effect size with its confidence interval.

A .060 does not sink a paper; a badly written .060 does. Give the exact value, the effect with its interval and a sensitivity analysis, keep the adverbs out, and label the exploratory work as exploratory. If the problem in front of you is deciding which analysis runs as primary when the result landed on the line, that is what my statistical consulting is for.