QCast Episode 67: The Analysis of Direct and Indirect Pathways in Observational Studies

In this QCast episode, co-hosts Jullia and Tom examine the analysis of direct and indirect pathways in observational studies through the lens of mediation analysis. The discussion uses an ankylosing spondylitis example to show how fatigue, anxiety/depression, and work productivity loss can be framed as exposure, mediator, and outcome. For clinical development teams, this matters when a simple association does not explain how a relationship may be operating in the data.

The episode also deals with the main interpretation risks. A proposed mediator can look clinically plausible without being strongly supported by the model. Variable type, missing or poorly timed mediator data, uncertainty around indirect effects, and residual confounding can all affect how much weight should be placed on the findings. Getting the pathway question clear early helps avoid overstating what observational mediation analysis can show.

🎧 Listen to the Episode:

Key Takeaways

Direct and indirect pathways need clear statistical framing

In mediation analysis, a direct effect is the part of the relationship between exposure and outcome that does not operate through the mediator in the model. An indirect effect is the part that appears to pass through the mediator. This distinction is useful, but it should be kept separate from the more common basal ganglia use of ‘direct and indirect pathways’ in neuroanatomy.

The AS example shows why mediator choice matters

In the ankylosing spondylitis example, fatigue was the exposure, work productivity loss was the outcome, and anxiety/depression was the proposed mediator. The analysis examined whether the relationship between fatigue and work productivity loss was mainly direct or partly mediated through anxiety/depression. The findings suggested limited support for that mediated route in this dataset.

Pathway analysis depends on data quality and cautious interpretation

Mediation analysis is only as useful as the pathway question, model assumptions, and data supporting it. The mediator needs to be measured clearly and at suitable time points, and uncertainty around the indirect effect needs proper handling. In observational data, results should be interpreted as evidence consistent with a pathway, not proof of causality.

Full Transcript

Jullia
Welcome to QCast, the show where biometric expertise meets data-driven dialogue. I’m Jullia.

Tom
I’m Tom, and in each episode, we dive into the methodologies, case studies, regulatory shifts, and industry trends shaping modern drug development.

Jullia
Whether you’re in biotech, pharma or life sciences, we’re here to bring you practical insights straight from a leading biometrics CRO. Let’s get started.

Tom
Today we’re talking about the analysis of direct and indirect pathways in observational studies. When people hear ‘direct and indirect pathways’, what are we actually talking about here?

Jullia
Here, we mean pathways in mediation analysis. So, we’re not talking about the basal ganglia sense of direct and indirect pathways, where the phrase is used in neuroanatomy. We’re talking about how an observed relationship in data may operate. Essentially, does fatigue, for example, appear to relate to work productivity loss directly, or does part of that relationship run through something else, such as anxiety or depression?

Tom
So, it’s a way of asking what might sit between two things we can observe?

Jullia
Pretty much. See, observational analysis may show that X is associated with Y, but that doesn’t explain how the relationship arises. Mediation analysis asks whether a third variable, usually called M, could help explain that relationship between X and Y. In clinical research, that can be useful when the association alone is too blunt to answer the scientific question.

Tom
Can you give an example of that?

Jullia
So a good one to think about would be exercise and energy levels. Exercise might be associated with higher energy, but perhaps part of that association works through improved sleep. In that case, exercise is X, energy level is Y, and sleep is the mediator M. The analysis is asking whether the relationship between exercise and energy appears to operate, at least partly, through sleep.

Tom
And in the case we’re discussing today, the clinical example is ankylosing spondylitis, isn’t it?

Jullia
Yes, see, the example comes from an observational study in ankylosing spondylitis, or AS. AS is an inflammatory rheumatic disease, and work disability remains a concern for patients, even when inflammation control has improved. In this study, the question was whether fatigue was linked with work productivity loss directly, or whether part of that relationship was mediated by anxiety or depression.

Tom
So fatigue is X, work productivity loss is Y, and anxiety or depression is M?

Jullia
That’s right. Fatigue was the independent variable or exposure., work productivity loss was the outcome, and anxiety or depression was the proposed mediator. The hypothesis was that the effect of fatigue on work productivity loss might be mediated by anxiety or depression.

Tom
That sounds clinically plausible, but also quite easy to overinterpret. If a mediation model suggests a pathway, does that mean we’ve proved a causal mechanism?

Jullia
No, and actually that’s an important distinction to keep in mind. In an observational study, mediation analysis can help explore a possible mechanism, but it doesn’t prove causality on its own. There may be measured confounders, unmeasured confounders, or relationships that have been modelled imperfectly. So the analysis can support a hypothesis about a pathway, but the interpretation needs to stay cautious.

Tom
Let’s get into the direct and indirect effect part of it all. What is the direct effect in this type of analysis?

Jullia
Well, the direct effect is the effect of X on Y after accounting for the mediator M. In the AS example, it’s the relationship between fatigue and work productivity loss after accounting for anxiety or depression. It’s the part of the pathway that doesn’t operate through the mediator in the model.

Tom
And the indirect effect is the part that goes through anxiety or depression?

Jullia
Yes. The indirect effect describes how the effect of X on Y appears to flow through M. In a simple linear mediation model, that’s estimated as the product of two coefficients. One coefficient captures the X-to-M path, and the other captures the M-to-Y path, conditional on X.

Tom
So if fatigue is associated with anxiety or depression, and anxiety or depression is associated with work productivity loss after accounting for fatigue, that gives us the indirect path?

Jullia
That’s the idea. Put simply, it asks whether people with higher fatigue also tend to report anxiety or depression, and whether that’s linked with greater work productivity loss. In a straightforward linear setting, the product of those two paths gives the indirect effect.

Tom
Where does the total effect fit?

Jullia
The total effect is the overall relationship between X and Y before decomposing it into direct and indirect components. In a simple linear mediation model, the total effect can be thought of as the direct effect plus the indirect effect. But that tidy decomposition doesn’t always hold in the same form, especially with nonlinear models, exposure-mediator interactions, or non-additive scales.

Really, it’s easiest when X, M and Y are continuous and the models are on compatible scales. Like in the AS example, where work productivity loss was a continuous percentage and fatigue was measured on a visual analogue scale, but anxiety or depression was ordinal.

Here, anxiety or depression was assessed using the anxiety and depression dimension of the EQ-5D questionnaire, with categories of none, some, or extreme problems. Because the extreme category had a low event count, the mediator was dichotomised as none versus some or extreme. That means the usual continuous-variable approach no longer fits neatly.

Tom
This is where logistic regression might seem like the obvious route. But the material used an identity-link approach instead. Why?

Jullia
So logistic regression models log odds, not probabilities directly. For the indirect effect in this set-up, the analysis needed the difference in probability of anxiety or depression associated with a unit increase in fatigue. Using a binomial model with an identity link allows the coefficient to be interpreted on the probability scale, which makes the product-of-coefficients approach more straightforward.

Tom
But that comes with a limitation, doesn’t it?

Jullia
It does. See, an identity-link model does not constrain fitted probabilities to lie between zero and one. That means predicted probabilities need to be checked, and the indirect-effect estimate should be interpreted with caution. It is an approximate method, so it makes sense to use additional checks when interpreting the findings.

Tom
Now when we estimate an indirect effect, why isn’t the usual confidence interval approach always enough?

Jullia
So the indirect effect is often a product of two coefficients, and the sampling distribution of that product is not usually normal. It can be skewed, particularly in smaller samples or when one of the pathway coefficients is close to zero. Because of that, bootstrap methods are commonly used to estimate uncertainty around the indirect effect.

Tom
For listeners who don’t work with bootstrap methods every day, what is happening there?

Jullia
The model is refitted many times using repeated samples drawn from the dataset with replacement. Each sample gives an estimate of the indirect effect. Those estimates form a bootstrap distribution, which can then be used to construct a confidence interval. So the confidence interval comes from the data being resampled, rather than assuming the indirect effect follows a neat normal pattern.

Tom
The Sobel test also appears in this topic. How should people think about that?

Jullia
The Sobel test assesses whether the product of the two indirect-path coefficients differs from zero, but it relies on a normal approximation. In the AS analysis, an additional Z statistic method was used to corroborate significance, but it does not estimate pathway strength in the same way as the indirect effect estimate.

Tom
So pathway strength and statistical significance are related, but they aren’t the same thing?

Jullia
Exactly. You want to look at the estimated pathway strength, the confidence interval around that estimate, and the corroborating significance test. Together, they give a more balanced view of whether the hypothesised mediation is supported.

Tom
What did the AS example suggest once all of that was pulled together?

Jullia
So the strongest pathways from fatigue to work productivity loss appeared to be direct. The estimated indirect pathway through anxiety or depression was weak. The corroborating Z statistic did not provide clear evidence of mediation, so overall the findings offered limited support for the hypothesis that anxiety or depression mediated the relationship between fatigue and work productivity loss.

Tom
That’s interesting because the original clinical hypothesis still made sense. The data just didn’t strongly support that particular mediated route.

Jullia
Yes, and that’s a valuable result. Mediation analysis is not only useful when it confirms the suspected pathway. It can also show when a proposed mediator does not appear to explain much of the relationship. That can help teams avoid building too much interpretation around a pathway the data do not really support.

Tom
If we move from this example to clinical development more broadly, where might this kind of thinking show up operationally?

Jullia
It can show up whenever teams want to understand what might sit between an exposure and an outcome in observational data. That could be patient-reported outcomes, work productivity, symptom burden, treatment patterns, or other intermediate measures. The operational point is that the mediator needs to be captured clearly, at the right time points, and with enough data quality to support analysis.

Tom
So if the mediator is poorly measured, or missing at the wrong visits, the analysis is limited before the modelling even starts.

Jullia
Yes. If the visit schedule does not capture the mediator before the outcome, or if data entry is inconsistent, the pathway question becomes harder to answer. The same applies if queries are unresolved or if a key questionnaire has incomplete data. Mediation analysis depends on the statistical model, but it also depends on whether the study has collected the right information in a usable way.

Tom
Before we close, what would you want teams to keep in mind if they’re planning this type of analysis?

Jullia
They should be clear about the hypothesised pathway before fitting models. They should check whether the exposure, mediator, outcome, and covariates are measured in a way that supports the question. They should also decide how uncertainty will be assessed, especially for indirect effects, and be cautious about causal language if the analysis is observational.

With that, we’ve come to the end of today’s episode on the analysis of direct and indirect pathways in observational studies. If you found this discussion useful, don’t forget to subscribe to QCast so you never miss an episode and share it with a colleague. And if you’d like to learn more about how Quanticate supports data-driven solutions in clinical trials, head to our website or get in touch.

Tom
Thanks for tuning in, and we’ll see you in the next episode.

About QCast

QCast by Quanticate is the podcast for biotech, pharma, and life science leaders looking to deepen their understanding of biometrics and modern drug development. Join co-hosts Tom and Jullia as they explore methodologies, case studies, regulatory shifts, and industry trends shaping the future of clinical research. Where biometric expertise meets data-driven dialogue, QCast delivers practical insights and thought leadership to inform your next breakthrough.

Subscribe to QCast on Apple Podcasts or Spotify to never miss an episode.