What Is a Mediator? How to Read Mediation in SPSS and SmartPLS
What Is a Mediator
You open a regression output or SmartPLS results and see terms such as indirect effect, direct effect, and total effect. A mediator is an intervening variable, commonly labelled M, placed between a cause variable X and an outcome variable Y. Mediation analysis tests whether X affects Y through M.
The basic model is: X → M → Y. For example, service quality X may increase satisfaction M, and satisfaction may then increase continuance intention Y. In this example, satisfaction is the mediator because it explains part of the mechanism linking service quality to continuance intention.
Mediation analysis usually examines three paths. Path a is the effect of X on M. Path b is the effect of M on Y when X is also included in the model. Path c is the total effect of X on Y before M is added. Path c' is the direct effect of X on Y after M has been added. The indirect effect is calculated as a × b, while the total effect is commonly expressed as c = c' + ab.
In practice, you can test a mediator with regression in SPSS, the PROCESS macro, AMOS, or a structural model in SmartPLS. SPSS is suitable when the model uses observed variables or scale scores. SmartPLS is suitable when the model contains latent variables and you want to assess the measurement model and structural model together.
Why a Mediator Matters in Quantitative Research
A mediator answers the question, “Through what mechanism does X affect Y?” This is different from simply testing whether X affects Y. A positive coefficient from X to Y shows a positive association, but it does not show which factor carries that effect.
For example, suppose you study the effect of perceived usefulness on intention to use an application. If trust is the mediator, the model may show that perceived usefulness increases trust, and trust then increases usage intention. This result takes the discussion further than the statement that perceived usefulness affects intention.
A model can produce three main mediation results. If the indirect effect is significant and the direct effect remains significant, M is usually described as a partial mediator. If the indirect effect is significant but the direct effect is no longer significant, the data support full mediation under the classic interpretation. If the indirect effect is not significant, you do not yet have statistical evidence that M transmits the effect from X to Y.
You need to distinguish a mediator from a control variable. A control variable is included to limit the influence of other factors, such as gender, age, or experience. A mediator is part of the theoretical mechanism that you deliberately specify in the model and hypotheses.
Mediation analysis is also different from outlier analysis. An outlier is an observation with an unusual value in the data, whereas a mediator is a variable in the research model. If your file contains extreme values, check outliers before interpreting the mediation paths.
What Counts as an Acceptable Mediator
A mediator does not have one threshold like Cronbach's Alpha or AVE. You need to assess the significance of the indirect effect, the bootstrap confidence interval, and the quality of M's scale. For mediation, the central question is whether the confidence interval for the indirect effect contains 0.
The table below summarises commonly used criteria. The values in the table are reference criteria from the sources listed, not default results for every study.
| What to check | How to assess it | Canonical source |
|---|---|---|
| Indirect effect | Significant when the bootstrap confidence interval does not contain 0 | (Preacher et al., 2008) |
| Bootstrap | You can use 5,000 bootstrap subsamples when running PLS-SEM | (Hair et al., 2022) |
| Outer loading of M's indicator | From 0.7 onward is a reference level in PLS-SEM | (Chin, 1998), (Hair et al., 2022) |
| Composite Reliability of M | From 0.7 onward | (Fornell and Larcker, 1981) |
| AVE of M | From 0.5 onward | (Fornell and Larcker, 1981) |
| HTMT between M and another construct | Below 0.85, or below 0.90 for closely related concepts | (Henseler et al., 2015) |
| VIF of the paths | Below 5 | (Hair et al., 2019) |
| f² of an effect | 0.02 small, 0.15 medium, 0.35 large | (Cohen, 1988) |
For example, suppose the indirect effect is 0.126 and the bootstrap 95% CI runs from 0.052 to 0.219. Because this interval does not contain 0, you can conclude that the indirect effect is significant under the bootstrap criterion. If the CI runs from -0.014 to 0.198, it contains 0, so there is not enough evidence to confirm a significant indirect effect.
You should not look only at the p-value for path a or path b and then declare mediation. It is possible for the individual paths to have low p-values while the product a × b is not significant, or for the reverse pattern to occur. Report the indirect effect and confidence interval first, then describe the direct effect to identify the form of mediation.
How to Read Mediation in Output
Reading Mediation in SPSS Output
If you use regression, you will usually run the first model with M as the dependent variable and X as the independent variable. The X → M path appears in the Coefficients table. Next, run a model of Y on X and M. The M → Y path is the coefficient for M, while the X → Y path in the model containing M is the direct effect.
If you use the PROCESS macro, the output commonly includes Model Summary, coefficients, Direct effect of X on Y, and Indirect effect(s) of X on Y. Confirm the names of X, M, and Y at the beginning of the output before reading the numbers. Many errors come from reversing the mediator and dependent variable when specifying the model.
Reading Mediation in SmartPLS 4
In SmartPLS 4, create three latent variables, connect X to M and M to Y, and usually add X to Y if you want to test the direct effect. Then run Calculate > Bootstrapping. In the results, open Specific Indirect Effects to see the indirect effect X → M → Y. Open Direct Effects to see X → Y, and Total Effects to see the total effect.
The table below is illustrative output, created to show how to read the output columns. It is not the result of an actual study.
| SmartPLS 4 output table | Path | Original Sample (O) | Sample Mean (M) | Standard Deviation (STDEV) | T Statistics | P Values |
|---|---|---|---|---|---|---|
| Specific Indirect Effects | X → M → Y | 0.126 | 0.128 | 0.041 | 3.073 | 0.002 |
| Direct Effects | X → Y | 0.184 | 0.181 | 0.058 | 3.172 | 0.002 |
| Total Effects | X → Y | 0.310 | 0.309 | 0.047 | 6.596 | 0.000 |
| Direct Effects | M → Y | 0.421 | 0.419 | 0.063 | 6.683 | 0.000 |
In the Specific Indirect Effects row, O is 0.126 and the p-value is 0.002. If the corresponding bootstrap confidence interval does not contain 0, the mediation path is statistically significant. The direct effect X → Y is also positive, with a p-value of 0.002, so this illustrative output fits the interpretation of partial mediation.
When reading the output, check the sign of the coefficient as well. A positive indirect effect means that the two paths produce an indirect effect in the same direction, or that both paths are negative. A negative indirect effect means that the effect through M runs in the opposite direction. Do not turn a negative result into a claim of a positive effect simply because the p-value meets the criterion.
You can also consult the explanation of moderator, because a moderator answers the question of when or under what conditions an effect changes. If your model contains both a mediator and a moderator, state clearly which variable represents the mediation mechanism and which variable changes the strength or direction of the effect.
What to Do When the Mediator Does Not Meet the Criteria
First, check the mediator's scale. If an outer loading is too low, check whether an indicator was entered with the wrong code, whether reverse coding has not been handled, or whether there are many missing values. With SPSS data, check Variable View, Data View, the frequency tables, and the coding of the variables before running the analysis again.
Next, examine the reliability and convergent validity of M. In SPSS, go to Analyze > Scale > Reliability Analysis to review Cronbach's Alpha, Corrected Item-Total Correlation, and the information shown when each item is deleted. Alpha from 0.7 onward is generally considered acceptable according to (Nunnally, 1978), while corrected item-total correlation from 0.3 onward is stated in (Nunnally and Bernstein, 1994). For an exploratory scale, Alpha from 0.6 may be considered according to (Hair et al., 2010), but you need to explain the context and should not delete an item simply to make the number look better.
How to write this in your thesis
If the measurement model meets the criteria but the indirect effect does not, return to the hypotheses and theoretical basis. X may genuinely have no effect on M, M may have no effect on Y after controlling for X, or the sample size may not be sufficient to detect a small effect. Do not change the mediator after seeing the results simply to obtain a desirable p-value.
One defensible option is to report the non-significant result and discuss causes that fit the data. You can also check data-entry errors, missing values, outliers, and multicollinearity. Deleting an observation requires a criterion set in advance, a record of what was removed, and a clear methodological reason.
If the model does not have a strong theoretical basis, discuss it with your supervisor before adding new paths. Software such as AMOS, JASP, and R can also run mediation, but changing software does not turn a non-significant indirect effect into a significant one. A dịch vụ chạy SPSS may save you time on the操作, but you still need to understand the model so you can explain X, M, Y, and the bootstrap procedure when the committee asks.
Distinguishing a Mediator from Easily Confused Variables
A mediator explains the mechanism between X and Y. A moderator changes the strength or direction of the relationship between X and Y. For example, satisfaction may be a mediator if service quality affects intention through satisfaction. Income may be a moderator if the effect of service quality on intention differs across income groups.
A moderating variable is commonly tested with an interaction term or an interaction path. A mediator is commonly tested with an indirect effect. The two concepts can appear in the same model, but the research questions, hypotheses, and output interpretation are different. See moderating to distinguish the terminology, or see what is a scale if you are unsure how the score for a scale was created.
A control variable is also easy to confuse with a mediator. Age, gender, and experience are often included to control for differences between groups. You should call a variable a mediator only when the theoretical model places it in the sequence X → M → Y and you have hypotheses for the relevant paths.
Common Errors
The first error is concluding that there is mediation simply because X affects Y. X → Y is a direct or total effect, and it does not prove that X affects Y through M. You need to inspect Specific Indirect Effects or the corresponding indirect effect table.
The second error is using CFI, TLI, and RMSEA from CB-SEM to evaluate a SmartPLS model. These are fit indices for CB-SEM according to (Hu and Bentler, 1999), so they should not be mixed into the assessment of PLS-SEM.
The third error is writing that a “mediator meets the criteria” as if mediation were one single statistic. A mediator is a variable's role in the model. The numbers you need to report include path coefficients, the indirect effect, the confidence interval, the p-value, and the measurement criteria for M itself.
The fourth error is deleting an indicator as soon as you see a low loading without checking its content. Removing an item can change the meaning of the construct. Save each version of the data, record the reason for deletion, and recheck reliability, AVE, and discriminant validity after every adjustment.
The final error is reporting a p-value of 0.000 as a probability equal to zero. In an SPSS or SmartPLS table, you can write p < 0.001 when the software displays 0.000, while preserving the coefficient's direction and magnitude.
Read also: outliers, moderating, moderator, what is a scale, what is std.
Frequently asked questions
What is a mediator in quantitative research?
A mediator is an intervening variable that explains how X affects Y. The basic model is X → M → Y, where the indirect effect is tested through the product of the X → M and M → Y paths.
How are a mediator and a moderator different?
A mediator explains the mechanism through which an effect is transmitted. A moderator shows that the effect of X on Y changes at different levels of another variable. You read a mediator through the indirect effect, while a moderator commonly involves an interaction effect or interaction path.
What indirect effect is acceptable?
There is no single numerical level that applies to every study. An indirect effect is considered significant when the bootstrap confidence interval does not contain 0, following the criterion in (Preacher and Hayes, 2008). You also need to examine the sign, size, p-value, and theoretical meaning of the effect.
Does a mediator need a significant direct effect?
A significant direct effect is not required for an indirect effect to exist. When the indirect effect is significant but the direct effect is not, you can describe the data as consistent with full mediation under the model's interpretation. Present the specific results clearly instead of assigning only a label to the model.
What should I do if the mediator has a problem?
Check variable coding, reverse-coded items, missing values, outliers, the scale, and multicollinearity before running the analysis again. If the measurement model meets the criteria but the indirect effect does not, report that result and explain it using the theory instead of deleting data or changing the model to search for a low p-value.
Should I run mediation in SPSS or SmartPLS?
SPSS is suitable for regression using composite or mean scores, especially when you use the PROCESS macro. SmartPLS is suitable when the mediator is a latent variable with multiple indicators and you want to run bootstrapping in a PLS-SEM model. Choose according to the model, data, and supervisor's requirements, rather than choosing the software that displays more attractive results.
Open your output again, identify X, M, and Y, and copy the Specific Indirect Effects, Direct Effects, and bootstrap confidence interval into the results table before writing the interpretation. If you need to run the analysis on your actual .sav or .csv file, M4 analysis by DoThesis supports this step.