Mediation Analysis

Are employees with more autonomy in remote work more productive? If that relationship arises because they can focus better on their work, how could we check it? Can we analyze what role work focus plays in the relationship between autonomy and productivity?

Mediation analysis is the method for analyzing the path through which the relationship between two variables runs via an intermediate variable.

Whereas correlation or simple regression shows the overall relationship between remote-work autonomy and productivity, mediation analysis splits that relationship into a direct effect and an indirect effect that runs through work focus. The indirect effect is usually estimated as the product of path coefficients, and its uncertainty can be assessed with a bootstrap confidence interval.

It is useful in research that seeks to explain theoretically the process behind a relationship, in fields such as organizations and HR, education, psychology and marketing. In this lesson we’ll use remote-work autonomy, work focus and employee productivity as an example to look at how the analysis works, direct, indirect and total effects, interpreting bootstrap results, how mediation differs from moderation, and the cautions about causal interpretation.

Mediation analysis checks what role a mediator M plays in the process by which an independent variable X relates to a dependent variable Y. Going beyond simply asking whether X and Y are related, it estimates the path through which that relationship comes about.

For example, if employees with more remote-work autonomy are more productive, you might wonder whether that is due to autonomy itself or because it lets them focus better on their work. In that case you set work focus as the mediator and analyze that path.

The correlation between autonomy and productivity alone can’t tell us what happens between the two variables. Mediation analysis splits the overall relationship into a direct path and an indirect path through the mediator, so the process proposed in the research hypothesis can be examined more concretely.

In organizational research in particular, how a policy changes employees’ behavior or work environment can matter more than whether the policy was introduced. But finding a statistical indirect effect doesn’t automatically prove an actual causal process.

In a simple mediation model based on linear regression, M is explained by X and the covariates, and Y is then explained by X, M and the covariates. The indirect effect is a × b, and under the same model assumptions and scales the total effect decomposes into c′ + ab.

Suppose remote-work autonomy, work focus and productivity scores were measured for 240 employees of a hypothetical company. All three variables are treated as continuous scale scores from 1 to 5, and job type and years of service are included as covariates.

The hypothesis is that higher autonomy leads to higher focus, and employees with higher focus are also more productive. Because this is a cross-sectional survey, the causal order is a hypothesis the researcher proposes, not a fact confirmed by the data alone.

The direct effect of 0.15 is the average difference in productivity associated with a 1-point difference in autonomy when work focus is included in the model. The indirect effect of 0.20 is the part of a 1-point difference in autonomy that is linked to productivity through the path of changes in focus.

The total effect of 0.35 is the sum of these two paths. But just because the indirect effect is larger than the direct effect, we can’t conclude that focus actually produced most of the overall relationship.

Because the indirect effect is the product of two regression coefficients, its sampling distribution may not be normal. So in practice, the bootstrap, which repeatedly resamples the data, is widely used to estimate a confidence interval for the indirect effect.

In the hypothetical example, the 95% confidence interval for the indirect effect does not include 0. So we can say that, in the specified model, a statistical indirect effect through focus was observed.

Even if the indirect effect is significant, if it is small its practical significance should be assessed separately. Conversely, there’s no need to conclude automatically that mediation is complete just because the direct effect isn’t significant.

A mediator is set up as an intermediate process that explains the X–Y relationship, and a moderator as a condition under which the X–Y relationship changes. Moderated mediation analysis, which combines the two roles, is also possible, but first make clear which role fits the research question.

A simple mediation model that uses only three observed scores can be estimated with ordinary regression. Structural equation modeling (SEM) can specify the direct and indirect paths simultaneously and can also handle latent variables measured by multiple items, along with measurement error.

Whichever method you choose, the temporal order of the mediator and the outcome, confounders and the validity of the research design must be reviewed separately.

The most common mistake is to say “we proved that focus is the cause of productivity gains from remote work” just because the indirect effect is significant. In cross-sectional data, autonomy, focus and productivity are measured at the same time, so it’s hard to establish which came first.

The more complex the model, the larger the sample needed and the more assumptions to review. It’s best to avoid adding paths without theoretical grounds.

A research report should include the sample size, measurement methods, covariates, estimation method, direct, indirect and total effects, confidence intervals and so on. Below is an example report for the hypothetical learning example.

This text was written to interpret the results only within the range the analysis supports, without asserting a causal relationship. In an actual paper it’s good to also state the number of bootstrap resamples, how the confidence interval was computed and the model assumptions.

The key isn’t arranging three variables; it’s explaining, through the research design and theory, why that variable can be seen as the intermediate process.

Mediation analysis splits the relationship between an independent variable and an outcome into a direct path and an indirect path through a mediator. In the remote-work example, we set up hypothesized relationships among autonomy, work focus and productivity and walked through checking the size of each path and the confidence interval of the indirect effect.

Even if the indirect effect is statistically significant, that alone doesn’t prove a causal mediation process. The temporal order in which variables were measured, confounders and the research design must be reviewed, and the presence of mediation shouldn’t be judged by whether the total effect is significant either. Interpret the results within the range the research design supports.