In 1969, Saul Sternberg introduced the additive factors method (AFM), a logical framework for using reaction time data to infer the architecture of information processing. The key insight is simple but powerful: if two experimental factors affect different processing stages, their effects on RT should be additive (no interaction); if they affect the same stage, their effects should interact. This logic allows researchers to map out the stage structure of a task without directly observing the stages themselves.
The Logic of Additive Factors
The AFM rests on a serial-stage architecture in which total RT is the sum of durations of successive processing stages:
If factor A affects only stage i and factor B affects only stage j (i ≠ j):
E[RT_AB] − E[RT_A] − E[RT_B] + E[RT_baseline] = 0 (additivity)
If factors A and B both affect stage i:
Interaction term ≠ 0
The method is applied by factorially combining experimental manipulations and testing for statistical interactions in a standard ANOVA on mean RT. Additive effects are interpreted as evidence for separate stages; interactions as evidence for a common stage.
Classic Stage Model
Sternberg proposed a four-stage model for choice RT tasks: (1) stimulus encoding, affected by stimulus quality/degradation; (2) stimulus identification, affected by stimulus discriminability; (3) response selection, affected by S-R compatibility; (4) response execution, affected by response complexity. Early studies found remarkable additivity between factors assigned to different stages (e.g., stimulus quality and S-R compatibility), supporting the serial-stage architecture.
The AFM has been criticized on several grounds. McClelland (1979) showed that cascaded (overlapping) processing stages can also produce additive effects under certain conditions, undermining the uniqueness of the serial-stage interpretation. Roberts and Sternberg (1993) countered that while cascade models can produce additivity, they do so only under specific parameter settings, whereas serial models produce additivity as a structural property. The debate highlights that additivity is necessary but not sufficient evidence for serial stages.
Modern Applications
Despite its age, the AFM remains widely used. It has been applied to aging research (identifying which stages slow with age), clinical neuropsychology (localizing processing deficits in patient populations), and cognitive neuroscience (testing whether ERP components correspond to processing stages). Sternberg (1998, 2001) extended the logic to brain imaging, proposing that if two factors affect different stages, they should activate different brain regions — a prediction that has received some support from fMRI studies.
The AFM is best viewed as a discovery tool rather than a proof technique: it generates hypotheses about processing architecture that can then be tested with converging methods such as RT distribution analysis, electrophysiology, and computational modeling. When combined with the diffusion model or other process models, the stage structure inferred by the AFM can be given a more precise mechanistic interpretation.