What Are Mediators? How to Read and Handle Mediating Variables
You open your SPSS output and see paths a, b, c, and c', together with an Indirect effect table. These symbols answer a specific question: does X affect Y directly, or does the effect pass through an intervening variable M? That is the role of a mediator, also called a mediating variable. If you are looking for a way to run the analysis for your topic, you can also read what is a mediator in the Statistics resources.
What Are Mediators
A mediator explains the mechanism through which an independent variable X affects a dependent variable Y. M sits between X and Y in the research model. For example, website quality may increase satisfaction, and satisfaction may then increase repurchase intention. In this model, satisfaction is the mediator.
You can picture the basic model as follows:
X → M → Y
Path a represents the effect of X on M. Path b represents the effect of M on Y after controlling for X. Path c' represents the direct effect of X on Y after M has been added to the model. The indirect effect is calculated as a × b. The total effect is commonly represented as c = c' + ab.
Suppose X is the online shopping experience, M is satisfaction, and Y is repurchase intention. You are not only asking whether X affects Y. You also want to know whether X increases repurchase intention because customers become more satisfied, or whether another direct effect remains.
The term “mediators” is used here because it is the form students commonly search for, but your thesis should use either “mediating variable” or “mediator” consistently. If your model contains several mediating variables, state clearly whether they are parallel mediators or serial mediators.
The Meaning of Mediators in Quantitative Research
A model containing only X and Y shows the level of association or effect between the two variables. Adding M allows you to explain the mechanism behind that relationship. This is why mediators often appear in models of consumer behaviour, technology acceptance, service quality, and human resource management.
A mediator helps you answer three questions. First, does X change M? Second, is M related to Y after X has been controlled for? Third, is the product a × b statistically different from zero?
When the indirect effect is significant, you have a basis for saying that M transmits part of the effect of X to Y. This wording is more careful than saying that “M is certainly the cause,” because cross-sectional survey data do not automatically establish a complete causal relationship.
You also need to distinguish a mediator from a moderator. A mediator answers “why” or “through what mechanism” X is related to Y. A moderator answers “when” or “for which group” the effect of X on Y is stronger. You can read what is a moderator and moderating if your model contains a moderating variable.
Regarding software, SPSS can run regression and PROCESS for mediation analysis. SmartPLS uses path coefficient and bootstrapping to test indirect effects in a PLS-SEM model. AMOS is suitable for CB-SEM, while R and JASP are suitable when you are willing to work with syntax or a different interface. Paying for an SPSS analysis service may save time in the short term, but you still need to understand the output when the committee asks questions. The greatest risk is receiving a conclusion that does not match your model, data coding, or hypotheses.
How Many Mediators Are Acceptable
A mediator does not have a single cutoff like Cronbach's Alpha. You need to examine the indirect effect and its confidence interval. With bootstrapping, the indirect effect is considered significant when the confidence interval does not contain 0 (Preacher et al., 2008).
The table below summarises the points commonly reported. Each row includes a source so that you do not copy an unexplained number into Chapter 4.
| Content to check | How to read it or reference threshold | Source |
|---|---|---|
Indirect effect a × b | The bootstrap confidence interval does not contain 0 | (Preacher et al., 2008) |
| Number of bootstrap samples | You can set 5,000 samples in PLS-SEM | (Hair et al., 2022) |
| Paths in the model | Examine the path coefficient, effect sign, and p-value or confidence interval | (Hair et al., 2022) |
| Distinguishing mediator and moderator | A mediator explains the mechanism transmitting an effect, while a moderator changes the strength or direction of the effect | (Baron and Kenny, 1986) |
| R² of the endogenous variable | 0.75 substantial, 0.50 moderate, 0.25 weak in PLS-SEM | (Hair et al., 2011) |
| VIF among predictor variables | Below 5 is the threshold commonly reported in PLS-SEM | (Hair et al., 2019) |
| Indirect effect in PROCESS | Report Effect, BootSE, BootLLCI, and BootULCI | (Preacher et al., 2008) |
The word “acceptable” needs to be understood in relation to the purpose of the test. You should not remove a mediator simply because its coefficient is small. If the confidence interval does not contain 0, the indirect effect may be significant even when the effect is weak. Report the coefficient, confidence interval, and effect direction rather than writing only “mediation exists.”
If the indirect effect is significant and the direct effect c' remains significant, the model is commonly described as partial mediation. If the indirect effect is significant but c' is no longer significant, many theses call it full mediation. Do not rely only on whether the total effect c is significant. The indirect effect is the part that needs to be tested directly with bootstrap.
How to Read Mediators in the Output
Suppose you run PROCESS in SPSS with X as app experience, M as trust, and Y as continuance intention. The table below is illustrative output, not the result of a real study. The column names are shortened versions of the tables students commonly see in PROCESS output.
| Relationship or index | Coeff | SE | LLCI | ULCI | p-value |
|---|---|---|---|---|---|
| X → M, path a | 0,542 | 0,081 | 0,382 | 0,702 | < 0,001 |
| M → Y, controlling for X, path b | 0,417 | 0,096 | 0,227 | 0,608 | < 0,001 |
| X → Y, direct effect c' | 0,186 | 0,078 | 0,032 | 0,340 | 0,018 |
| X → Y, total effect c | 0,412 | 0,071 | 0,272 | 0,552 | < 0,001 |
| Indirect effect X → M → Y | 0,226 | 0,061 | 0,115 | 0,357 | BootLLCI, BootULCI |
Read the output in order. First, the positive path a shows that X predicts M in a positive direction. Next, the positive path b shows that M still predicts Y after X has been entered into the equation. Finally, BootLLCI = 0,115 and BootULCI = 0,357 are both greater than 0, so the confidence interval does not contain 0. The indirect effect in this example is statistically significant.
In the Indirect effect table, PROCESS usually does not use a p-value as the main decision line. Look at BootLLCI and BootULCI. If one endpoint is negative and the other is positive, the confidence interval contains 0, so there is not enough evidence to conclude that the mediator transmits the effect.
For the regression tables, you also need to check R-sq, Model, Coefficients, and the number of observations. If you run SmartPLS, the table names may be Path Coefficients, Specific Indirect Effects, Total Effects, and Bootstrapping results. The reading logic is similar. Examine the indirect coefficient, confidence interval, and effect direction.
How to write this in your thesis
You can adapt the following sentence to your actual output: “The bootstrapping results show that the indirect effect of [X] on [Y] through [M] has a coefficient of [ab], and the confidence interval [BootLLCI; BootULCI] does not contain 0. Therefore, [M] has a mediating role in the relationship between [X] and [Y], with [partial/full] mediation where this classification is consistent with the direct effect results.”
Do not enter the illustrative numbers into your thesis. Open the correct output table, record the coefficients from your own .sav or .csv file, and compare them with the hypotheses established in Chapter 2.
What to Do When Mediators Do Not Meet the Criteria
If the confidence interval for the indirect effect contains 0, first check whether you selected the correct X, M, and Y variables and the correct hypothesised direction. A common error is to read the Total effect line and draw a conclusion about mediation, when the relevant line is Indirect effect.
Next, check the scale coding. An item written in the opposite direction that has not been recoded can reduce the M score or reverse the relationship. Check missing values, values outside the scale range, and unusual response patterns. If the data contain an outlier, you can read the guidance on outliers, but do not delete an observation simply to make the p-value look better.
If the M scale is not yet sound, go back and check Cronbach's Alpha, Corrected Item-Total Correlation, EFA, or the measurement model. When an item is unsuitable in terms of both content and statistics, you may consider removing it, rerunning the model, and recording each processing round. The scale should be defended with a theoretical argument, not only by the fact that the coefficient increased.
If the mediator is not significant, retain that result when the data and analysis procedure are valid. You may check an alternative model supported by the theory, such as a direct effect or a different mediator, but you should not test dozens of models and report only the one with a small p-value.
In SmartPLS, check Outer Loadings, Construct Reliability and Validity, HTMT, VIF, Path Coefficients, and Specific Indirect Effects. The thresholds for CR, AVE, and HTMT need to be interpreted in the PLS-SEM context. You can compare them with (Fornell and Larcker, 1981) and (Henseler et al., 2015), respectively. CFI, TLI, RMSEA, and related fit indices belong to CB-SEM and should not be included in the SmartPLS model assessment section.
Distinguishing Mediators from Moderators
These two concepts are easy to confuse because both are added to a regression model. A mediator lies on the pathway through which an effect is transmitted. A moderator creates a condition that makes the relationship between X and Y stronger, weaker, or reversed.
For example, satisfaction is a mediator if the service experience increases satisfaction and satisfaction then increases recommendation intention. Income is a moderator if the effect of service experience on recommendation intention differs between the low-income and high-income groups.
For a mediator, you test the indirect effect, usually with bootstrap. For a moderator, you test an interaction term such as X × Z and examine whether the interaction coefficient is significant. You cannot use the result of one type of analysis to conclude the other. This distinction is discussed in (Baron and Kenny, 1986).
You also need to distinguish “full mediation” from “partial mediation.” These describe how much of the direct effect remains after M is added. They are not two different types of variables. If a study has one mediator, use the singular “mediator.” “Mediators” is appropriate when you are discussing several mediating variables or using the search term.
Common Errors
Looking only at the p-values for paths a and b. The two individual paths may be significant, but the final conclusion still needs to be based on the indirect effect and its bootstrap confidence interval.
Calling every variable between two arrows a mediator. Its position in the diagram is not enough. M needs a theoretical argument and a scale that fits the mechanism you want to test.
Concluding full mediation only because the direct effect is not significant. Report the indirect effect, direct effect, total effect, and confidence intervals. Do not hide a negative sign, a small coefficient, or a result that does not support the hypothesis.
Using the wrong tool. PROCESS in SPSS is suitable for many regression-based mediation models. SmartPLS is suitable when your model is PLS-SEM. AMOS and CB-SEM fit indices follow a different logic, so state the software and version clearly in your research methods section.
Choosing the sample or removing data after seeing the result. Do not delete responses simply because they make the mediator nonsignificant. If data must be removed, record the prespecified criterion, the number of observations removed, and the effect on sample size.
Frequently asked questions
What are mediators in SPSS?
Mediators are mediating variables used to test whether X affects Y through M. In SPSS, you can run separate regressions for each equation or use PROCESS to obtain the indirect effect and bootstrap confidence interval directly.
What is the difference between full and partial mediation?
Partial mediation is commonly used when the indirect effect is significant and the direct effect c' remains significant. Full mediation is commonly used when the indirect effect is significant but the direct effect is no longer significant. Report all effects instead of applying only a label.
How do you handle a mediator in SPSS?
Check the data coding and scale reliability first, then select the appropriate PROCESS model. When you run the analysis, set the number of bootstrap samples, identify X, M, and Y correctly, and read the Indirect effect table with BootLLCI and BootULCI.
Should you delete a mediator when the indirect effect is not significant?
You cannot conclude that immediately. Check the theoretical model, missing data, reverse-coded items, outliers, sample size, and scale selection. If the procedure is valid, a nonsignificant result still needs to be reported, because deleting a mediator only to support the hypothesis would distort the interpretation.
Can SmartPLS be used to test mediators?
Yes. In SmartPLS, examine Specific Indirect Effects, Path Coefficients, and the confidence intervals from Bootstrapping. Assess the measurement model first, then interpret the mediator in the structural model according to the PLS-SEM procedure.
Open your file again, draw X, M, and Y clearly, and record the values for Indirect effect, BootLLCI, BootULCI, c', and R-sq separately before writing the conclusion. If you need to run the analysis on your own .sav or .csv file and turn the output into the analysis section, DoThesis M4 analysis handles this step.