To understand the relationship among (1) psychological mechanisms, (2) variables caused by the independent variable (IV), and (3) whether a variable belongs to the theoretical construct of the IV, it is essential to begin with a fundamental premise of experimental research. In experiments, researchers do not merely manipulate a physical stimulus; rather, they manipulate a theoretical construct. For example, when a study manipulates an anthropomorphic interface, the goal is not to change a specific visual element, but to operationalize and manipulate the construct of anthropomorphism itself. The experimental stimuli are therefore implementations of this theoretical construct. Recognizing this distinction is crucial for understanding the difference between mediators and confounds.

When an experiment produces results, researchers often observe variables that change across IV conditions and also influence the dependent variable (DV). Determining the role of such variables can be done through two sequential logical steps.

The first step is to examine the causal origin of the variable: whether the variable is actually produced by the manipulation of the IV. If a variable is not a consequence of the IV, but simply differs systematically between experimental conditions and also affects the DV, then it constitutes a confound. For example, in a study manipulating an anthropomorphic interface, the anthropomorphic condition might inadvertently produce an interface that is brighter (brightness). If brightness influences users’ evaluations, then brightness satisfies two conditions: it varies with the IV condition and it affects the DV. However, it is not the intended outcome of the manipulation of anthropomorphism; it is merely an unintended difference in stimulus design. In this case, brightness is a stimulus confound, because it provides an alternative explanation for the observed IV–DV relationship.

The second step applies once a variable is confirmed to be caused by the IV. At this stage, the key question becomes theoretical: does this variable represent the psychological mechanism through which the IV is expected to influence the DV? If the answer is yes, the variable is a mediator [1], because it explains how the IV exerts its effect through a psychological process. For example, in research on anthropomorphic interfaces, theory often suggests that anthropomorphism leads users to perceive technology as a kind of social entity, thereby increasing social presence. If increased social presence subsequently enhances willingness to continue using the product, then social presence is a mediator: it is both caused by the IV and theoretically expected to be the psychological mechanism through which anthropomorphism operates.

Importantly, however, not every psychological response triggered by the IV is a mediator. A variable may indeed be caused by the IV and still fail to qualify as a mediator if it is not part of the theoretically predicted mechanism. In such cases, the variable functions as a psychological confound. For instance, users interacting with an anthropomorphic interface may perceive the interface as more friendly (perceived friendliness). This perception may indeed be triggered by the IV and may also influence the DV. Yet if the theoretical mechanism of anthropomorphism focuses on social presence, rather than friendliness, then perceived friendliness is not the mediator the theory seeks to test. Instead, it becomes a psychological confound because it offers an alternative explanation for the effect of the IV on the DV.

The distinction between stimulus confounds and psychological confounds therefore lies only in the level at which they occur. Stimulus confounds are differences in the physical properties of the stimuli, such as brightness, voice pitch, or visual complexity. Psychological confounds are differences in users’ perceptions or reactions, such as friendliness, perceived innovativeness, or perceived complexity. Despite this difference in level, they are logically equivalent because both satisfy the defining conditions of a confound: they vary with the IV, influence the DV, and do not belong to the theoretical construct being manipulated.

(Now let’s consider the example of perceived complexity, which illustrates how these distinctions apply in practice. Perceived complexity lies at the boundary between stimulus and psychological levels. If “complexity” refers to objective properties of the interface (such as the number of elements or hierarchical depth), it is a stimulus feature. By contrast, perceived complexity refers to the user’s subjective experience and is therefore a psychological variable. However, whether it functions as a mediator or a confound still depends on theory. If a study’s theoretical model proposes that interface complexity leads to perceived complexity, which increases cognitive load and affects evaluation, then perceived complexity is a mediator. If, however, the study is primarily concerned with another construct (such as modularity) and perceived complexity merely arises as an unintended psychological response to the design (which has confused me the most), then perceived complexity functions as a psychological confound. This example again illustrates the central principle: the critical distinction between mediator and confound does not depend on whether the variable is psychological, but on whether it belongs to the theoretical mechanism through which the IV is expected to influence the DV.)

From this reasoning we can derive a unified decision logic.

When a variable changes with the IV and influences the DV, we first ask whether it is caused by the IV. If it is not, it is a confound, typically at the stimulus level. If it is caused by the IV, we then ask whether it belongs to the theoretically expected psychological mechanism of the IV. If it does, it is a mediator; if it does not, it remains a confound, specifically a psychological confound. In other words, mediators are part of the theoretical mechanism, whereas confounds are changes outside that mechanism that nevertheless influence the DV. The distinction therefore depends not merely on causal relationships, but on how the theoretical construct and its psychological processes are defined.

Last, a further important clarification concerns the difference between a variable’s role in the causal structure and its role in the statistical model. These two levels are often confused.

At the level of causal structure, a variable’s role is determined by theory and causal relationships. For instance, in an anthropomorphic interface study, brightness may be a stimulus confound, social presence a mediator, and perceived friendliness a psychological confound. These roles reflect how the variables relate to the theoretical construct and the causal pathways in the experiment.

At the level of the statistical model, however, researchers may include variables in regression or structural models for analytical purposes. When a variable is statistically controlled, it functions as a covariate, regardless of its causal role. For example, perceived friendliness might be included in a regression model to control for its influence; in that case it is a covariate in the statistical model but remains a confound in the causal structure. Similarly, a mediator such as social presence might also be included as a control variable in certain analyses. In statistical terms it becomes a covariate, but in the causal logic it remains a mediator. Thus, covariate refers to a statistical treatment, whereas mediator and confound refer to roles within the causal structure.

Understanding this distinction helps prevent a common misunderstanding: the fact that a variable is controlled in a statistical model does not determine its causal role. Instead, causal roles must first be identified at the level of theory and causal reasoning; statistical models then determine how those variables are handled analytically.

[1] A mediator is not defined by whether a variable is psychological or behavioral, but by whether it represents the theoretical mechanism through which the independent variable (IV) influences the dependent variable (DV). In causal terms, a mediator lies on the pathway IV → mediator → DV, explaining how the effect occurs. Importantly, mediators can take several forms depending on the type of mechanism involved. The most common type in behavioral and social science research is the psychological mediator, which captures internal mental states such as perceptions, beliefs, or motivations (e.g., anthropomorphic interface → social presence → willingness to continue using). Other forms also exist. Cognitive mediators involve mental evaluations or understanding (e.g., product transparency → perceived understanding → trust). Affective mediators involve emotional responses (e.g., humorous AI → amusement → engagement). Behavioral mediators involve observable actions that transmit the effect (e.g., dynamic interface → longer dwell time → curiosity). Perceptual mediators involve subjective interpretations of stimulus properties (e.g., modular product design → perceived repairability → willingness to continue using). Regardless of type, a variable qualifies as a mediator only when it is caused by the IV and theoretically specified as part of the mechanism linking the IV to the DV. If a variable changes with the IV but does not belong to this theoretical pathway, it should instead be treated as a confound.