You interpret a mediation analysis result by finding the indirect effect, the a times b path, and checking whether its bootstrap confidence interval contains zero. If both ends of that interval sit on the same side of zero, the indirect effect is statistically supported. There is no p-value to look for, and the total effect c is a separate question from the one most people are actually asking.
This guide is written for anyone sitting in front of a PROCESS output window, a lavaan console, or an Mplus output who has run the model correctly and is stuck on what to do with what came back. It takes about ten minutes to read once and roughly two minutes to apply to your own output. As of 2026, the field’s standard is Andrew Hayes’s bootstrap approach, which retired the Baron and Kenny causal steps and the Sobel test that came with them.
Table of Contents
- 1What You Need
- 2Step-by-Step
- 31. Identify the variables and the mediation model you actually fitted
- 42. Read the direct, indirect, and total effects
- 53. Check whether the indirect effect is statistically supported
- 64. Compare the result with the research question and theory
- 75. How to interpret a mediation analysis result in APA 7
- 8Common Mistakes
- 9Frequently Asked Questions
- 10How do you interpret mediation analysis results?
- 11What are the statistics used to analyze mediation?
- 12How do you know if mediation is successful?
- 13What is the Baron-Kenny approach to mediation analysis?
- 14How many bootstrap resamples should I use, and which interval type?
- 15Do I need a significant total effect before testing mediation?
- 16Conclusion
What You Need
Before you can interpret a mediation analysis result you need five things in front of you. Missing any one of them is the most common reason people misread their own output.
First, the mediation model you specified, with the names of your variables written out: the independent variable X, the mediator M, the dependent variable Y, and any covariates you entered. If you cannot name the path you hypothesized before reading the output, you will find something.
Second, the output window itself, not a screenshot of the parts you liked. Third, the row in that window containing the indirect effect, its bootstrap interval bounds, and the resample count.
Fourth, a note of which software and version produced it, because the labels differ. PROCESS version 4 for SPSS, lavaan in R, JASP, Mplus, and SmartPLS 4 all estimate the same model and all name things differently. Fifth, and this is the one people skip, the design your data came from. Whether you measured X, M and Y at three separate time points decides whether your result is a causal claim or an association.
Step-by-Step
1. Identify the variables and the mediation model you actually fitted
Simple mediation output looks nothing like the output you get from a regression or a correlation, and mixing the two up wastes an afternoon. A correlation table gives you an r between each pair of variables. A mediation model gives you two regression equations run in sequence, one predicting M from X, one predicting Y from both X and M.
Read the model section at the top of your output first and confirm the variable names, the model number, the covariates, and the coding. PROCESS asks for a model number and defaults to model 4, which is simple mediation. Model 14 adds a moderator to the a path. If you meant model 4 and your output says 14, everything below this heading is irrelevant.
Note the missing-data handling too. PROCESS defaults to listwise deletion, so your effective N for the outcome equation can be smaller than the N in your first equation. That is worth a sentence in your methods, because it is exactly the kind of detail a reviewer checks.
2. Read the direct, indirect, and total effects
The path labels are old shorthand and they still show up in textbooks, so learn them once. Everything else follows from these five rows.
| Label | What it is | Formula | Where it lives in the output | Plain reading |
|---|---|---|---|---|
| a | Effect of X on M | M = i + aX + e | First regression equation, the row for X | Does the predictor move the mediator? |
| b | Effect of M on Y, controlling X | Y = i + c prime X + bM + e | Second equation, the row for M | Does the mediator move the outcome once X is held constant? |
| ab | Indirect effect | a multiplied by b | Effect section, row labelled indirect effect | How much of the X to Y effect travels through M |
| c prime | Direct effect | X coefficient in the outcome equation | Second equation, the row for X | What is left of the effect with M in the model |
| c | Total effect | c prime plus ab | Total effect model, the row for X | The overall X to Y relationship with M ignored |
Here is a worked example carried all the way to a results sentence. The model is trained stress predicting burnout through perceived lack of control, N = 342, three waves, PROCESS version 4 for SPSS, 10,000 bootstrap resamples.
The output gives a = 0.513 with a bootstrap interval of [0.331, 0.701], b = 0.708 with an interval of [0.479, 0.922], and an indirect effect of ab = 0.363 with a 95 percent bootstrap confidence interval of [0.224, 0.557]. The direct effect is c prime = -0.107, p = .341, so X has no leftover effect on burnout once control is in the model. The total effect is c = 0.257, p = .021.
Read that in words: training stress raised burnout, most of that effect ran through feeling unable to control anything, and almost none of it operated by any other route. The arithmetic checks out, 0.513 times 0.708 is 0.363, and -0.107 plus 0.363 is 0.256. That last figure lands a hair below the reported 0.257 because the two paths are rounded to three decimals before you add them, which is normal and not a sign of an error.
Look at the two R-squared values too. R-squared of 0.42 in the mediator equation means X explains 42 percent of the variance in perceived control. R-squared of 0.31 in the outcome equation means X and M together explain 31 percent of the variance in burnout. A low R-squared on the mediator is a warning sign, because it means a large part of the thing you are calling a mechanism is still unexplained.
3. Check whether the indirect effect is statistically supported
Judgment on the indirect effect rests on the bootstrap confidence interval for ab, and nothing else. In our worked example the interval runs from 0.224 to 0.557. Zero is not inside it, so the indirect effect is statistically supported at the .05 level.
If the interval had run from -0.180 to 0.310, zero would sit comfortably inside, and the honest conclusion is that the data do not distinguish an indirect effect from none. People find this hard to accept because both individual paths can be significant on their own. Significance of a and significance of b separately do not combine into significance of their product; the product has its own sampling distribution, which is why it needs resampling rather than a closed-form test.
That resampling is the bootstrap. The procedure draws thousands of resamples with replacement from your data, estimates ab in each one, and reads the 2.5th and 97.5th percentiles of that distribution as the interval bounds. Use 5,000 resamples as a floor and 10,000 when the computer will manage it. Set a random seed so your numbers reproduce exactly, and record it.
The interval type matters too. PROCESS offers percentile, bias-corrected, and BCa. The percentile interval is the default and the most conservative choice. Bias-corrected intervals can inflate the Type I error rate slightly, which means they occasionally declare an effect significant when it is not. Either is defensible; what is not defensible is failing to say which one you used.
Now the case that trips up almost everyone, and it is the one people search for most often: my indirect effect is significant but my total effect is not. Say a = 0.42 and b = -0.35, giving ab = -0.147 with a bootstrap interval of [-0.290, -0.031] that excludes zero. The direct effect is c prime = 0.28 and significant. The total effect, c prime plus ab, comes to 0.133 and its interval crosses zero.
Nothing is broken here. The two paths point in opposite directions, so they partly cancel before anyone gets to the total effect. This is called inconsistent mediation, and Zhao, Lynch and Chen describe it as incompatible. The older Baron and Kenny rules said a mediation cannot exist unless the total effect is significant first, which would have told you to discard a real finding. Hayes retired that requirement, and current practice does not make you check c before estimating ab.
One more warning about the direct effect. A significant direct effect tells you X still predicts Y with M in the model. It does not tell you a mediation occurred. Readers, reviewers, and sometimes your own supervisor treat a significant c prime as confirmation, and it is not. The mediation claim stands or falls on ab and its interval.
4. Compare the result with the research question and theory
Statistical support is not the same as support for your hypothesis. To finish interpreting a mediation analysis result, look at the direction and the magnitude of ab and ask whether they match the mechanism you proposed.
The signs of a and b give you a small typology, the one Zhao and colleagues formalized in 2010. When a and b share a sign, the indirect effect points the same way as the total effect and the mediation is complementary. When the total effect is negative but the indirect effect is positive, you have competitive mediation, where the paths oppose each other but the total still runs in the indirect direction. When the indirect effect runs opposite to the total effect, that is the inconsistent or incompatible case above. Same sign on both paths but a total effect near zero is your suppression pattern.
For magnitude, proportion mediated, ab divided by c, is the intuitive share of the total effect that travels through M. In the worked example, 0.363 over 0.257, that is 1.41, which is larger than 100 percent. Nothing has gone wrong. A proportion above one simply means the mediator transmits more effect than the total shows because the direct path is subtracting from it. The ratio becomes unstable and hard to interpret when c is small, which is another reason not to hang your argument on it.
A more portable number is the completely standardized indirect effect, where a and b are standardized before multiplying. It lets you say the indirect effect is, for example, 0.21 standard deviations, which a reader in a different field can compare against something in their own work. Report the unstandardized ab in the table and the standardized version in the sentence if you want both.
Finally, be honest about what the design supports. Shrout and Bolger laid out the causal conditions, and none of them are optional. X must precede M, M must precede Y, and you need three waves rather than three variables measured at once. You also need to rule out an unmeasured variable that affects both M and Y. A cross-sectional design can show that the paths are consistent with a mechanism, but it cannot establish one, and your write-up should say exactly that rather than slipping into causal verbs.
5. How to interpret a mediation analysis result in APA 7
Writing up a mediation analysis result takes two pieces: a table of coefficients and one or two sentences of text. The table comes first.
Use this as a template and replace the labels with your own variables. Report unstandardized coefficients in the main table, with the bootstrap intervals beneath, and put standardized betas in a separate column only if your field expects them.
| Path | Estimate | SE | 95 percent BC | t | p |
|---|---|---|---|---|---|
| a: X to M | 0.513 | 0.094 | [0.331, 0.701] | 5.46 | < .001 |
| b: M to Y, controlling X | 0.708 | 0.113 | [0.479, 0.922] | 6.27 | < .001 |
| ab: indirect effect | 0.363 | 0.085 | [0.224, 0.557] | 4.27 | < .001 |
| c prime: direct effect | -0.107 | 0.112 | [-0.328, 0.115] | -0.96 | .341 |
| c: total effect | 0.257 | 0.112 | [0.038, 0.481] | 2.30 | .021 |
Then write the sentence. Do not report a p-value for the indirect effect, because the bootstrap interval is the test. Do not call the result full or partial mediation.
A defensible version reads: A mediation analysis examined whether perceived lack of control accounted for the relationship between training stress and burnout in a sample of 342 employees. Perceived control was positively related to training stress, B = 0.513, SE = 0.094, 95 percent bootstrap CI [0.331, 0.701], and burnout was positively related to perceived control controlling for training stress, B = 0.708, SE = 0.113, 95 percent bootstrap CI [0.479, 0.922]. There was a significant indirect effect of training stress on burnout through perceived control, B = 0.363, SE = 0.085, 95 percent bootstrap CI [0.224, 0.557], 10,000 bootstrap resamples. The direct effect was not significant, B = -0.107, SE = 0.112, 95 percent bootstrap CI [-0.328, 0.115], and the total effect was significant, B = 0.257, SE = 0.112, 95 percent bootstrap CI [0.038, 0.481].
Add one sentence of interpretation and stop. Something like: Training stress was associated with higher burnout largely through employees’ sense of reduced control, a pattern consistent with the proposed mechanism but limited by the observational design.
The full versus partial label is the piece of old vocabulary to retire. It classified a result as partial when c prime stayed significant, and full when it went away, which makes the verdict a function of sample size rather than of the data’s structure. With a larger sample the direct effect tends to surface and the label flips. Report the sizes and directions of both paths and let the reader see the pattern.
Common Mistakes
Most interpretation errors fall into a small number of repeat patterns. Each one below is something reviewers flag, and each has a straightforward fix.
| Mistake | Why it is wrong | The fix |
|---|---|---|
| Using the Sobel test | It assumes the product of two coefficients is normally distributed, which it usually is not | Report the bootstrap interval; Hayes (2022) treats the Sobel test as retired |
| Reporting a p-value for ab | A bootstrap interval is the inference, not a separate test | Give ab, the interval bounds, and the resample count instead |
| Requiring a significant total effect first | Baron and Kenny (1986) made this a precondition; it discards real inconsistent mediation | Estimate ab regardless of c and discuss the pattern you get |
| Calling the result full or partial mediation | The label flips with sample size and describes nothing about mechanism | Describe the direction and magnitude of each path |
| Reading the wrong output row | The outcome equation has an X row and an M row; the effect block has the ab row | Scroll to the effect section and find the row labelled indirect effect |
| Omitting the resample count | A reader cannot judge whether the interval is trustworthy | State resamples, interval type, and seed in the methods or results |
| Dichotomizing the mediator | Collapsing a scale into two groups throws away variance and biases the b path | Keep M continuous, or justify the split with a principled reason |
| Adjusting for a variable on the causal path | Controlling for a consequence of M or Y removes part of the effect you are testing | Adjust only for pre-exposure covariates, and report both adjusted and unadjusted models |
Two of those deserve extra attention because they happen silently. Dichotomizing a mediator, splitting a five-item control scale at its midpoint for instance, reduces the b path toward zero and can make a real mechanism disappear.
Covariates are the other one, and the forum evidence is full of people whose mediation flipped after adding controls. Usually one of the new covariates is measured after X, which means it sits on the causal path and absorbs part of the effect. Another possibility is that the covariate is a collider. If you added baseline burnout to a study of training stress, you have conditioned on a consequence of stress, and that can manufacture an indirect effect that was not there before.
One more habit worth building: report the unmediated model too. Running the outcome equation without M in it, or simply noting the zero-order correlation between X and Y, gives reviewers the c they need to check your arithmetic.
Frequently Asked Questions
How do you interpret mediation analysis results?
Start with the indirect effect, the product of the a and b paths, and read its bootstrap confidence interval. If the interval excludes zero, the indirect effect is statistically supported. Then look at the direct effect c prime and the total effect c. Neither of those decides whether mediation occurred, and a significant direct effect is not evidence of a mechanism. Finish by checking the direction of the effect against your hypothesis, because a statistically supported indirect effect can still contradict the theory you proposed.
What are the statistics used to analyze mediation?
You report five path coefficients: a from X to M, b from M to Y controlling for X, the indirect effect ab, the direct effect c prime, and the total effect c, which equals c prime plus ab. Each comes with a standard error and a confidence interval. The key statistic is the bootstrap confidence interval for ab, which is the inference for mediation. The a and b paths are also read through their own intervals, and the two R-squared values tell you how much variance X explains in M and how much X and M explain in Y.
How do you know if mediation is successful?
Your data support mediation when the 95 percent bootstrap confidence interval for the indirect effect does not contain zero. That is the whole rule, and it applies whether or not your total effect was significant. Because the product of two coefficients has its own sampling distribution, significant a and significant b do not add up to a significant ab. Run 5,000 bootstrap resamples as a minimum, 10,000 where you can, and check the interval rather than reaching for a p-value. Support is statistical, so a separate question is whether the effect runs in the direction your theory predicted.
What is the Baron-Kenny approach to mediation analysis?
Baron and Kenny (1986) proposed four causal steps: a significant regression of M on X, a significant regression of Y on M, a significant regression of Y on X, and a drop in that last coefficient once M enters the model. It shaped three decades of teaching and is still the procedure many textbooks describe. Current practice following Hayes (2022) keeps the first regression but judges the indirect effect with a bootstrap interval, drops the requirement that the total effect be significant, and drops the full versus partial label that step four produced.
How many bootstrap resamples should I use, and which interval type?
Use 5,000 resamples at minimum and 10,000 for final analyses, since more draws give a more stable estimate of the interval tails. The percentile interval is the default in PROCESS and the conservative choice. Bias-corrected intervals can push the Type I error rate above your nominal level, which occasionally flags an effect that is not real, so if you use one, name it in the write-up. Set a random seed so the numbers reproduce, and record the seed along with the resample count.
Do I need a significant total effect before testing mediation?
No. The requirement came from Baron and Kenny’s causal steps and it is no longer part of accepted practice. It fails precisely when the two paths point in opposite directions: the indirect effect is real and significant, but the paths cancel and the total effect is small and non-significant. That pattern is called inconsistent or incompatible mediation. Estimate the indirect effect, report it with its interval, and describe the two paths separately so a reader can see the cancellation rather than a null result.
Conclusion
Do one thing first: find the row labelled indirect effect and read the two numbers at the end of it, the lower and upper bounds of its bootstrap confidence interval. If they sit on the same side of zero, the indirect effect is supported, and everything after that is about description rather than decision.
Then compare the direction of that effect with what your theory predicted, note the size of the direct effect and the total effect without reaching for the words full or partial, and write the sentence with the coefficient, the interval, and the resample count in it. Statistical mediation is a claim about a pattern in the data. Saying so plainly keeps you clear of the overreach reviewers are looking for.
