What Is Mediation? How to Interpret a Mediator in Quantitative Research
What is mediation
Mediation analysis tests whether X affects Y through M. In this model, X is usually the independent variable, M is the mediator, and Y is the dependent variable.
Suppose you want to test whether service quality affects loyalty through satisfaction. Service quality is X, satisfaction is M, and loyalty is Y. Your research question is not only whether X changes Y, but also whether that effect passes through M.
The basic model has three paths:
- Path a: X → M.
- Path b: M → Y, after controlling for X.
- Path c': X → Y, after M has been added to the model.
The indirect effect is usually represented by the product a × b. The total effect can be described as c = c' + a × b. Here, c' is the direct effect of X on Y after M has been included in the model.
You can also read what is a mediator if you are mixing up the name of the variable with the mediation technique. The two concepts are closely related, but a mediator is a variable in the model, whereas mediation is the way you test the model's mechanism.
Why mediation matters in quantitative research
Mediation helps you answer a question about the mechanism of an effect. A relationship may exist at the overall level, but you still need to know why it appears. For example, employee training may improve performance because training increases competence, which then leads to better performance.
In a thesis, mediation is commonly used when the model has the form “X affects Y through M”. Your hypotheses might be:
- H1: Service quality has a positive effect on satisfaction.
- H2: Satisfaction has a positive effect on loyalty.
- H3: Service quality has a positive effect on loyalty.
- H4: Satisfaction mediates the relationship between service quality and loyalty.
H4 is the mediation hypothesis. A significant effect of X on Y alone is not enough to support H4. You must examine the indirect effect and its confidence interval separately. With bootstrapping, the indirect effect is considered significant when the confidence interval does not contain 0 (Preacher and Hayes, 2008).
Mediation also prevents you from making an overly broad interpretation of one regression coefficient. If X affects Y, you know that a relationship or model-based effect exists. When the indirect effect is significant, you have additional grounds to report that part of the mechanism runs through M. Statistical results still do not prove absolute causality when your design is a cross-sectional survey.
If your model includes a variable that changes the strength of a relationship, that is moderation, not mediation. You can read what is moderating and what is a moderator before choosing the analysis.
What counts as acceptable mediation
Mediation does not have one fixed cutoff like Cronbach's Alpha or AVE. The main decision is whether the indirect effect is statistically significant, usually based on a bootstrap confidence interval. You should also check whether the sign of the indirect effect matches your hypothesis.
The table below lists common checks. Read each row together with its source and the method you selected.
| Check | How to read the result | Criterion or interpretation | Source |
|---|---|---|---|
| Indirect effect | Indirect effect or Specific indirect effects | The bootstrap confidence interval does not contain 0 | (Preacher and Hayes, 2008) |
| Number of bootstrap samples | Bootstrap subsamples | 5,000 subsamples can be used in the PLS-SEM procedure | (Hair et al., 2022) |
| Direct effect | Direct effect or c' | Check the coefficient, sign, and p-value or confidence interval | (Preacher and Hayes, 2008) |
| Total effect | Total effect | Check the overall effect of X on Y | (Preacher and Hayes, 2008) |
| Distinguishing mediator and moderator | Hypothesis structure | A mediator explains a mechanism, while a moderator changes the strength or direction of an effect | (Baron and Kenny, 1986) |
| VIF in the PLS-SEM model | Inner VIF Values | VIF below 5 | (Hair et al., 2019) |
| R² of the dependent variable | R Square | 0,75 substantial, 0,50 moderate, 0,25 weak in PLS-SEM | (Hair et al., 2011) |
Do not conclude that “mediation is acceptable” only because the p-value for X → M is below 0,05. A significant path a is evidence to consider, but the mediation conclusion must be based on the indirect effect. If its confidence interval contains 0, you do not yet have enough evidence to conclude that the indirect effect is significant.
Two common labels are full mediation and partial mediation. Full mediation is usually used when the indirect effect is significant but the direct effect is no longer significant after M is added. Partial mediation is usually used when both the indirect effect and direct effect are significant. Present this classification carefully, because some more recent guidance prioritizes reporting the indirect effect and confidence interval directly rather than relying entirely on the traditional sequence of steps.
How to read mediation output
In SPSS, mediation is commonly run with the PROCESS macro. You then read the tables for the total effect, direct effect, and indirect effect. In SmartPLS, run bootstrapping and inspect Specific Indirect Effects, Direct Effects, Total Effects, and Path Coefficients.
The following is illustrative output, not the result of a real study. Assume that X is service quality, M is satisfaction, and Y is loyalty.
| Output table | Relationship | Original Sample or Effect | Bootstrapped t or t value | p-value | 95% Confidence Interval | Illustrative interpretation |
|---|---|---|---|---|---|---|
| Path Coefficients | X → M | 0,62 | 10,41 | < 0,001 | 0,50 to 0,73 | Positive and significant path a |
| Direct Effects | X → Y | 0,21 | 2,48 | 0,013 | 0,05 to 0,37 | Direct effect c' remains significant |
| Specific Indirect Effects | X → M → Y | 0,29 | 4,96 | < 0,001 | 0,19 to 0,41 | Confidence interval does not contain 0 |
| Total Effects | X → Y | 0,50 | 8,02 | < 0,001 | 0,38 to 0,62 | Positive and significant total effect |
For the Specific Indirect Effects row, the interval from 0,19 to 0,41 does not contain 0. In this illustrative output, the indirect effect is significant and positive. Because the direct effect remains significant at 0,21, you can describe the result as partial mediation, provided that this classification fits your theoretical framework and chosen convention.
With SPSS PROCESS, pay attention to the Indirect effect(s) of X on Y table. Focus on Effect, BootSE, LLCI, and ULCI. LLCI is the lower limit and ULCI is the upper limit of the confidence interval. When LLCI and ULCI are both positive or both negative, the interval does not cross 0. You do not need to rely only on the p-value of an individual path to conclude that the indirect effect is significant.
With SmartPLS 4, the usual process is to run Calculate > Bootstrapping, select the number of subsamples for your analysis design, and open Bootstrapping > Specific Indirect Effects. SmartPLS 3 and SmartPLS 4 differ in the location of some menus, so check the version you are using. In PLS-SEM, assessment should separate the measurement model from the structural model using a two-step approach (Anderson and Gerbing, 1988), after which you interpret the mediation paths.
Read a mediation result as a sequence. First, check the measurement model if you are using latent variables. Then check the X → M and M → Y paths, the indirect effect, direct effect, total effect, coefficient signs, and confidence interval. Finally, compare the results with the hypotheses instead of copying the entire output into Chapter 4.
What to do when mediation is not supported
If the indirect effect is not significant, work through the following checks.
Check the variables and hypothesis direction
Confirm that you selected the correct X, M, and Y. A common error is to place a moderator where the mediator should be, or to write that “M moderates X and Y” while running a mediation model. Check the model diagram, variable coding, and wording of the hypotheses before running the model again.
Check the data and scale
If an item was reverse-coded but you did not reverse code it, the path coefficient may change sign or become weaker. Check missing values, unusual response rows, and outliers. When removing data, record the criterion and the number of observations removed. Do not delete rows simply to make the p-value look better.
Check the measurement model
In SmartPLS, inspect outer loadings, CR, AVE, and HTMT before interpreting mediation. An outer loading from 0.7 upward is commonly used in PLS-SEM (Chin, 1998), while CR from 0.7 and AVE from 0.5 are commonly cited thresholds (Fornell and Larcker, 1981). If the scale is unstable, the indirect-effect conclusion has a weak foundation.
Run an appropriate bootstrap
If you are relying only on a Sobel test or on the p-value of each individual path, consider bootstrapping the indirect effect. In SPSS, PROCESS allows you to set the number of bootstrap samples and the confidence interval. In SmartPLS, use the bootstrapping results for Specific Indirect Effects. Do not repeatedly change the number of bootstrap samples just to obtain a significant result.
Keep the result if it remains unsupported
If the indirect effect is still not significant after reasonable checks, report that the result is not supported. You can discuss that the data from this sample do not provide enough evidence for the mediator's role. Removing the mediator from the results chapter simply because the hypothesis was not supported makes the report less transparent.
Distinguishing mediation, moderation, and a mediator
Mediation answers the question “Through what mechanism does X affect Y”. Moderation answers “When, or for which group, is the relationship between X and Y stronger or weaker”.
Satisfaction is a mediator if service quality increases satisfaction and satisfaction then increases loyalty. Age may be a moderator if the effect of service quality on loyalty differs across age groups or changes with age.
A mediator usually lies on the causal path: X → M → Y. A moderator usually appears through an interaction term, such as X × W → Y. Having a third variable in the model does not automatically make the model a mediation model. Look at the research question, hypothesis diagram, and the way the third variable enters the equation.
Some topics include both mechanisms. For example, service quality may affect loyalty through satisfaction, while the strength of that effect differs according to customer engagement. This may be a moderated mediation model, which is more complex than basic mediation. State clearly which variable is the mediator and which is the moderator.
You should also distinguish mediation from simply adding a control variable to a regression. A control variable is included to reduce the influence of other factors according to the research design. A mediator needs a theoretical argument and a separate hypothesis about the X → M → Y mechanism.
Common mistakes when reading mediation
The first mistake is concluding that mediation exists only because X → M has a p-value below 0,05. Read the indirect-effect row and the confidence interval for the indirect effect itself.
The second mistake is using “full mediation” mechanically. A nonsignificant direct effect supports a full-mediation interpretation in your model, but you still need to report the coefficient, confidence interval, and theoretical basis. If the direct and indirect effects have opposite signs, do not immediately label the result partial mediation without explaining the pattern.
The third mistake is confusing the p-value with effect size. An indirect effect may be statistically significant while remaining small in magnitude. Report the effect, confidence interval, and research context rather than writing only “there is an effect”.
The fourth mistake is mixing CB-SEM criteria into SmartPLS. CFI, TLI, RMSEA, and SRMR are commonly used fit indices in CB-SEM, with the source (Hu and Bentler, 1999). Do not place these indices in a PLS-SEM guide without a clear methodological reason.
The final mistake is changing the model after several runs without keeping a record. If you remove an item or change a path, create a tracking table for the data version, reason for the adjustment, and result after each run. This gives you something concrete to explain when your supervisor asks why the final model differs from the proposed model.
How to write this in your thesis
You can begin with: “Bootstrapping results show that the indirect effect of [X] on [Y] through [M] is [coefficient], with a 95% confidence interval from [LLCI] to [ULCI], which does not contain 0. Therefore, hypothesis [H] is [accepted or supported]. The direct effect of [X] on [Y] [remains or no longer remains] significant, so the mediating role is interpreted as [partial or full] according to the study's criterion.”
Frequently asked questions
What is mediation in SPSS
In SPSS, mediation tests whether M transmits part or all of the effect of X on Y. You can use the PROCESS macro to examine the direct effect, indirect effect, and total effect together with the bootstrap confidence interval.
What is the difference between full and partial mediation
Full mediation is usually used when the indirect effect is significant but the direct effect is no longer significant after the mediator is added. Partial mediation is usually used when the indirect effect is significant and the direct effect remains significant. Report the coefficients and confidence intervals, not only the name of the mediation type.
What indirect effect is acceptable
There is no fixed number that applies to every model. An indirect effect is considered significant when its bootstrap confidence interval does not contain 0 (Preacher and Hayes, 2008). Then check the sign, coefficient size, and consistency with the hypothesis.
Where does the mediator appear in the model diagram
The mediator sits between X and Y, forming the path X → M → Y. In the service quality, satisfaction, and loyalty example, satisfaction is the mediator. If a third variable changes the strength of the X → Y relationship, it is usually a moderator.
Can you run moderated mediation in SPSS
Yes, but this is more complex than basic mediation. Define the mediator, moderator, interaction term, and hypotheses before choosing a PROCESS model or another suitable procedure. Do not run a basic mediation model and call the result moderated mediation.
Should you delete a hypothesis when mediation is not significant
Do not delete the result simply because the hypothesis was not supported. Check the coding, data, scale, and model first. If the result remains nonsignificant, report honestly that the data do not provide enough evidence for the indirect effect and discuss possible explanations.
Open your SPSS or SmartPLS file, identify X, M, and Y, and find the Indirect Effects or Specific Indirect Effects row before writing the conclusion. If you need to run mediation on your .sav or .csv file and turn the results into thesis chapters, see M4 data analysis.