I'm not sure where else to ask this so I would like to pose it here and hope that it reaches an audience. As field experiments are typically designed to test the effect of treatments (tree species, tree spacing, fertilizer use, seeds, etc ) on responses (yield, biodiversity, profits, environmental variables), are causal methods a suitable approach to analyzing the data? Specifically, if a field experiment was designed to measure the effect of treatments on the environment (temperature, radiation, humidity) and yields, can I use mediation analysis to then determine a causal pathway between treatment, environment and yields when this was not the original goal of the experiment? My hypothesis is that the effect of treatment on yields is (at least partially) mediated through the environment, but I am struggling to fit treatments into the causal pathway. I haven't found much literature on the use of causal methods in ecological experimental research. Any thoughts or experience with this?
A mediation or indirect effect analysis itself cannot show causality in any way, because it is just a statistical tool. To claim causality you need a strong theoretical background. For example, you could reverse the mediation model or construct any other path combination and this might also show significant indirect effects. So, this alone is not enough to claim causality. And this is especially true for field studies. You write "experiments" but this would demand a random allocation and controlled conditions. I do not know, if this is true for your study.
But not all hope is lost for field studies, IF(!) you have strong theoretical knowledge about your variables and covariables. I would suggest to study the field of directed acyclic graphs (DAG), potential outcome analyses and inverse probability weighting etc. I think an execellent overview with R/Python/Stata codes is provided by Scott Cunningham's book Causal Inference: The Mixtape. It is also available as free online book. From there, you can dive deeper in the areas you might need for your study.
https://mixtape.scunning.com/
Yvonne,
beyond what Rainer said: the problems in even randomized experiments is that while your direct effect can be interpreted causally with high plausibility, incorporating a measured/observed mediator opens the possibility for all sorts of confounding referring to the link between the mediator and target variable. Hence, the direct effect (without the mediator) will be unbaised but the indirect effect may not be. If your treatment variable has no further direct effect on the target (i.e., it is a full mediation model), you could use this treatment as an instrument and test the error covariance between the mediator and targt (i.e, the Hausman test) but this requires that that full mediation is correct (i.e., the exclusion restriction criterion) and that the treatment has a strong effect on the mediator (i.e., the relevance criterion).
At any rate, I would wrap my head around possible confounders (common causes of the mediator and target) and use them as control variables.
And of course, all of this assumes that randomization really worked. In ecological settings, lots of things can happen even if the initial set up was clean.
HTH
Holger
Thanks @Rainer Duesing and @Holger Steinmetz ,
Since I could not control for all variables that could possibly affect my outcome, I agree that it would be hard to fit a mediation model. In my case, my experiment was randomized but I fear I do not have enough data to meet the criteria and I'm missing data on some relevant confounders. I will abandon this for now but read more on DAGs. Thanks for the book recommendation Rainer Duesing !
Hi Yvonne,
please keep in mind that you only would have to control for *common* causes of the mediator and outcome--not for all causes of the outcome.
In addition, of course you can fit a mediation model--not doing so would be throwing out the baby with the bathwater. Simply keep in mind that there may be not-controlled confounders (can you name some?) but keep the conclusion and interpretation of the model at a "modest" level. Science is never complete.
Best,
Holger
- Badges
- Science topic