In many public health and dental studies, specific treatments or interventions are designed to lead to changes in such things as a patient's knowledge, attitude, behavior, or biological and social factors, which in turn improve dental health itself.
These elements are technically known as intermediate variables or mediators.
A change in these mediators is thought to lead to improved dental outcomes.
This process of change is called mediation.
Understanding whether a treatment works as well as how a treatment works can help researchers make advances in future research and tailor specific treatments for application in specific populations.
In order to understand the mediation process, researchers seek to measure whether the results of a treatment are due to the effect of mediators identified in the initial study design or a result of unmeasured mediators.
For example, does providing oral health education to parents improve their children’s oral health? If so, is this effect due to increased tooth brushing, more frequent dental visits, decreased sugar consumption, or some other unknown factor?
The statistical models that researchers commonly use to analyze the direct or indirect effect of a given treatment on outcomes around or through the mediators they have targeted for change often break down, because they are based on different assumptions that fail to account for actual conditions in the real world.
One of these assumptions is known as sequential ignorability.
In sequential ignorability, researchers assume that the randomization process also randomizes the mediators in relation to a given treatment or intervention.
Unfortunately, this assumption may be incorrect, because researchers cannot actually randomize mediators after randomizing patients for treatment.
This is why more recent causal models replace the sequential ignorability assumption with other assumptions (e.g., no interaction between intervention and mediator).
Most standard and causal models for mediation analysis focus on continuous or binary mediators and outcomes.
However, the outcome variable in dental studies such as the number of decayed, missing, or filled tooth surfaces or the number of dental visits, often consists of whole number counts with many zeros.
These numbers can represent anything from decayed, missing, or filled (dmf/DMF) teeth or surface indices to the number of dental visits.
The goal of the Mediation Analysis study is to develop statistical models that can account for the complexities of the real world as represented by the count and zero-inflated count data common in dental studies.
Researchers will use an instrumental variables approach to allow for both measured and unmeasured confounding.
This has the potential to help researchers interpret results in new ways and improve on standard statistical models by mapping data sets more accurately.