Why Reporting Precision Matters
A p-value is often the final number readers see from a hypothesis test, but its presentation affects what they can reasonably conclude. A value reported as 0.038 communicates more than “small”: it gives an approximate numerical measure of how often a result at least as extreme as the observed one would occur under the null model and the test procedure. Reporting too few digits can hide a comparison that matters; reporting many digits can suggest a level of precision that the analysis does not justify.
The goal is not to display every digit a computer prints. It is to report enough precision to describe the result honestly, make any comparison with a significance level clear, and avoid implying that a probability is exactly zero. The p-value must still be interpreted in context: it is calculated assuming the null hypothesis and the test’s conditions, and it is not the probability that the null hypothesis is true.
It is also important to distinguish reporting a numerical p-value from using an exact test. “Exact” can describe a test whose probability calculation uses the null distribution directly rather than a large-sample approximation. That does not mean its p-value must be printed with every available decimal place. A p-value from an exact test can still be rounded for reporting; conversely, many digits from software do not by themselves make a test exact.
Choosing How Many Digits to Report
A common convention is to report p-values to three decimal places when they are not extremely small, using an equals sign for a rounded numerical value. For example, a calculated value of 0.037846 might be reported as \(p=0.038\). Follow any reporting standard specified for the analysis, journal, or assignment. When a value is near the significance level, include enough digits to make the comparison unambiguous.
The number of displayed digits should not be confused with accuracy in the broader sense. A rounded p-value is an approximation to the computed value. Also, the test itself relies on its model and assumptions: extra digits cannot correct a poor design, an inappropriate model, or an inaccurate probability calculation. It is usually enough to report a sensible number of digits, state the direction and context of the finding, and make the decision rule clear when one is being used.
For a conventional rounding example, suppose the calculated value is \(0.037846\). To round to three decimal places, keep the digits through the thousandths place, \(0.037\), and inspect the next digit, \(8\). Since \(8\) is at least \(5\), increase the thousandths digit by one:
The reported value \(p=0.038\) means the computed p-value rounds to 0.038 at the chosen precision. It does not mean that the unrounded value is exactly 0.038. It also does not mean there is a 3.8% probability that the null hypothesis is true.
Worked Example: Rounding a Discrete-Test P-Value
In a hypothetical discrete test, the test procedure identifies three possible outcomes that are at least as extreme as the observed outcome. Under the null hypothesis, their probabilities are 0.018, 0.011, and 0.006. Find and report the p-value to three decimal places.
The p-value is the sum of the null probabilities for the outcomes included in the specified extreme set:
Thus the p-value is \(0.035\), and reporting \(p=0.035\) preserves the value to three decimal places. In this hypothetical example, the listed probabilities and the sum are exact as given, so no further rounding is needed. In an actual analysis, the reported digits should reflect the calculated value and the reporting standard being used.
The interpretation remains conditional: if the null hypothesis and test procedure are appropriate, a result at least as extreme as the observed result would occur with probability 0.035 under the null model. It is not a probability that the null hypothesis is correct.
Reporting Very Small P-Values
When a p-value is smaller than the smallest value shown by a chosen reporting format, use a less-than sign rather than rounding it down to zero. For instance, if values are ordinarily reported to three decimal places, a p-value of 0.0000374 is below 0.001. It can be reported as \(p<0.001\). If a more precise numerical value is useful and supported by the calculation, scientific notation can also communicate its size clearly.
An inequality is a bound, not an exact replacement for the numerical p-value. The report \(p<0.001\) says that the p-value is below 0.001; it does not say it equals 0.001, and it does not reveal its precise size below that bound. Where distinctions among very small p-values matter, reporting a suitable numerical value may be more informative, provided the method and precision justify it.
Worked Example: Reporting a Very Small Value Without Using Zero
A hypothetical test produces a computed p-value of 0.0000374. Suppose the usual reporting convention is three decimal places. Determine an appropriate report and, if useful, an alternative that gives more detail.
At three decimal places, rounding the number would produce 0.000. That display would be misleading because the computed value is positive and less than 0.001. Instead, compare the value directly with 0.001:
Therefore an appropriate report under the stated convention is \(p<0.001\). If more detail is needed and the calculation supports it, rewrite the same value in scientific notation:
The inequality communicates that the p-value falls below the reporting threshold; the scientific notation gives a more precise numerical approximation. Neither report means that the probability of the null hypothesis is 0.0000374. Both describe the probability of a result at least as extreme as the observed one under the null hypothesis and test procedure.
A displayed \(p=0.000\) usually results from rounding or from software that cannot display a value smaller than its numerical limit. It should not be treated as proof that the p-value is exactly zero. For a valid observed outcome with positive probability under the null model, the event used to calculate the p-value includes that outcome, so its probability cannot be less than the probability of the observed outcome itself. A computer display of zero can therefore indicate limited display precision or numerical underflow rather than a literal zero probability.
In practice, report the smallest bound supported by the output or calculation, such as \(p<0.001\), and provide scientific notation or additional digits if they are meaningful for the reader. Do not replace a very small positive value with \(p=0\).
Rounding Near a Significance Level
Rounding requires particular care when the p-value is close to the chosen significance level \(\alpha\). A decision rule compares the calculated p-value with \(\alpha\), using the unrounded value when available. Rounding should be used to communicate the result, not to change which side of the decision boundary the calculation falls on.
For example, when rounding to three decimal places, values in the interval from 0.0495 up to, but not including, 0.0505 ordinarily round to 0.050. This includes values both below and above 0.050. Therefore, a display of \(p=0.050\) alone may not show whether the unrounded value is below, equal to, or above \(\alpha=0.05\). This follows directly from the rounding rule: all values in that interval share the same rounded display, even though they are not all on the same side of the threshold.
Worked Example: A Rounded Value Near the Decision Boundary
A hypothetical test uses significance level \(\alpha=0.05\) and produces an unrounded p-value of 0.0496. Show how it rounds to three decimal places and decide how to report and compare it.
To round 0.0496 to three decimal places, keep the digits through the thousandths place, 0.049, and inspect the next digit, 6. Since \(6\) is at least \(5\), the thousandths digit increases by one:
A report of \(p=0.050\) on its own hides the comparison with \(\alpha=0.05\). Use the unrounded value for the comparison:
Under a rule that rejects \(H_0\) when \(p\leq\alpha\), this result meets the rejection rule. To make the comparison clear, report \(p=0.0496\) rather than only \(p=0.050\), if that precision is supported. The conclusion should still be expressed in the study context and should not claim that the result proves the alternative hypothesis.
The same principle applies on the other side of the boundary. If an unrounded p-value is 0.0504, it also rounds to 0.050 at three decimal places, but \(0.0504>0.05\). A decision based only on the rounded display could therefore differ from the decision based on the computed value. When the value is this close to \(\alpha\), give enough digits or state the comparison explicitly.
A Practical Reporting Sequence
A clear report separates the numerical result from its interpretation. First use the p-value calculated by the test, before rounding, to make any comparison with the stated significance level. Then choose a reporting format that is consistent with the required convention and that does not hide an important distinction. Finally describe what the result says about the data under the null model, in the context of the study.
Retain the unrounded p-value for comparisons with \(\alpha\) and for any later calculations.
Use a rounded numerical value for ordinary results and an inequality or scientific notation for very small results.
If rounding could conceal whether the value is above or below \(\alpha\), report more digits or give the comparison using the unrounded value.
Explain what results at least as extreme as the observed one would mean under the null model; do not describe the p-value as a probability that a hypothesis is true.
Common Mistakes and Why This Matters
P-values help readers judge how compatible observed data are with a null model, but a report can distort that information even when the calculation itself is correct. Precision is especially important when results are close to a threshold or when values are very small.
- Reporting a small value as zero. The reasoning “it rounds to 0.000, so the p-value is zero” confuses a display with the calculated probability. Report a bound such as \(p<0.001\), or a suitable numerical value, instead.
- Comparing only the rounded value with \(\alpha\). A value that rounds to 0.050 might be below or above 0.05. Compare using the unrounded calculation, and report enough digits to make the result clear.
- Giving excessive digits. A long string of digits can imply more useful precision than the test and its assumptions warrant. Follow a consistent reporting convention and include extra digits only when they help interpretation.
- Calling the p-value the chance that the null is true. The p-value is calculated under the null model; it does not give the probability that the null hypothesis is true.
- Treating “exact test” as a required number of decimal places. Exactness describes a method of probability calculation, not how many digits must be printed. The report should still use suitable precision.
Check Your Understanding
For each question, focus on what the displayed p-value does and does not communicate.
- A test produces an unrounded p-value of 0.02674. What is the result to three decimal places, and what does the result mean under the null model?
- A software display shows \(p=0.000\) after rounding. Give a more appropriate report if the reporting convention uses three decimal places.
- A test uses \(\alpha=0.05\), and the unrounded p-value is 0.0503. What does this value round to at three decimal places, and which value should be used for the comparison with \(\alpha\)?
- Explain why reporting \(p<0.001\) does not say that the p-value equals 0.001.
- In one sentence, distinguish an exact test from reporting a p-value with many decimal places.