Mental Processes

Abstract

A mental process is an internal operation that transforms information between a stimulus and a response; because it cannot be observed directly, it must be inferred from behaviour, most powerfully from the time a response takes. This article sets out the logic of that inference: Donders' subtractive method, which isolates a process by comparing tasks that differ by one operation, and Sternberg's additive-factors method, which reads the pattern of reaction-time effects to decide whether two manipulations act on one processing stage or on separate stages. It then examines two organising distinctions the chronometric programme uncovered, between automatic and controlled processing and between serial and parallel architectures, and the limits each inference faces. Three demonstrations let the reader run a subtraction, test additive factors, and trace the power law of practice.

Keywords: mental chronometry, reaction time, information processing

Between the light that strikes the eye and the word that leaves the mouth, something happens that no instrument records directly. The stimulus is measurable and the response is measurable, but the operations in between — encoding the input, comparing it with memory, choosing an action — are hidden inside the organism. Cognitive psychology takes those hidden operations as its proper subject and treats the observable margins, above all the duration of a response, as the trace from which their number, order, and nature can be reconstructed (Neisser, 1967). This article is about how that reconstruction is done: the reasoning that turns a pattern of reaction times into a claim about the mind's internal architecture, the classic methods that formalise it, and the recurring caution that the same behaviour can often be produced by more than one arrangement of processes.

What a Mental Process Is

A mental process is an internal operation that receives, transforms, stores, or acts on information. The term covers the whole span from registering a stimulus to selecting a response — encoding, comparison, retrieval, decision, and the preparation of action — and it names the operations that intervene between what goes in and what comes out. For much of the early twentieth century such operations were held to lie outside science: behaviourism restricted psychology to observable stimulus-response relations on the ground that anything in between could not be measured. The reframing that founded modern cognitive psychology rejected that restriction, casting the mind as an information-processing system in which input is successively transformed through a series of stages, each doing a definable job, on the way to a response (Neisser, 1967).

Treating cognition as information processing brings two commitments that organise everything below. The first is that processing is staged: a response is the end of a chain of operations, not a single indivisible act, so the central empirical task is to discover how many stages there are, what each computes, and in what order they run. The second is that the system has a limited capacity: only so much information can be held or operated on at once, a limit George Miller fixed at about seven items for immediate memory and used to argue that the bottleneck is a general feature of the processor rather than of any one task (Miller, 1956). Capacity limits matter here because they are what make the automatic-controlled distinction, later in this article, a distinction with teeth: an operation that needs capacity competes with others, and one that has escaped the limit does not. Table 1 previews the four analytic tools the rest of the article develops.

Types of Mental Process

The methods above cut a single act of cognition across time, into stages that run one after another. A different and older question cuts it by kind: what families of mental operation are there? The National Library of Medicine's Medical Subject Headings answers that second question with a formal classification, placing Mental Processes at tree position F02.463 and hanging the recognised subtypes beneath it. These categories are orthogonal to the chronometric decomposition that is this article's main subject: a subtype such as perception or learning is itself made of encoding, comparison, and decision stages, so the two cuts describe the same activity at different grains rather than competing. Table 2 lists the direct children of the descriptor; several are treated at length in their own articles.

Two cautions keep this taxonomy in its place. It is a classification for indexing, built to organise the literature, not a theory claiming these are the mind's natural joints or that they are mutually exclusive — mentalization and theory of mind plainly overlap, and perception shades into cognition. And a subtype is a family of function, not a processing stage: naming perception as a type of mental process says nothing about the encoding, comparison, and decision operations the rest of this article uses reaction time to expose within it.

Inferring the Unobservable

If a mental process takes time, then time carries information about the process. This is the founding premise of mental chronometry: because each operation in the chain occupies a measurable interval, the total reaction time is the sum of the durations of the stages that produced it, and systematic changes in that total, under controlled changes to the task, expose the stages themselves (Posner, 1978). The response is thus read not as a simple index of speed but as an accumulated record, and the analyst's job is to partition it. Figure 1 shows the picture that underlies every method in this article: a stimulus enters, passes through a sequence of internal stages each consuming some time, and issues in a response whose latency is their sum.

The strategy is more general than reaction time alone, because some processes leave a graded trace in time that reveals their internal form. When Roger Shepard and Jacqueline Metzler asked observers to judge whether two rotated shapes were the same object, the time to decide rose linearly with the angular difference between them, as though the observer were rotating a mental image through the intervening angle at a roughly constant rate before matching (Shepard & Metzler, 1971). The orderly relation between an external variable and response time is evidence not merely that a process took time but that it had a specific, analog character — it traversed the intermediate orientations rather than leaping to the answer. Chronometry, in short, offers two kinds of leverage: the total latency can be decomposed into stages, and its lawful variation with a stimulus parameter can constrain what a single stage is doing.

All of this leverage rests on one precondition that is easy to overlook: a reaction time is interpretable only when accuracy is held roughly constant. A responder can always trade speed for accuracy, answering faster at the cost of more errors or more slowly to be surer, and this speed-accuracy tradeoff means a difference in mean reaction time can reflect a shift in how cautiously the response is being made rather than any change in the underlying processing (Wickelgren, 1977). Every subtraction and every additive-factors comparison in this article therefore assumes that error rates are matched across the conditions being compared; where they are not, the latency difference is confounded with a difference in response caution, and the claim about stages is unsafe. Holding accuracy constant, or modelling speed and accuracy jointly, is the price of reading time as a record of process. The modern form of that joint modelling is the family of sequential-sampling models, of which Roger Ratcliff's diffusion model is the most influential: it casts a decision as the gradual accumulation of noisy evidence toward one of two response boundaries, and by fitting the whole distribution of both correct and error latencies it separates the rate at which evidence accumulates from the response caution that sets the boundaries — turning the tradeoff into distinct estimated parameters rather than a confound to be controlled away (Ratcliff, 1978). In the decades since, the diffusion model has become a standard analytic tool across cognitive and neural studies of decision making, fitted routinely to the joint distribution of choices and latencies to recover the components of a decision from behaviour (Ratcliff & McKoon, 2008). The two classic methods that follow instead exploit the simpler decomposition of total latency, each under this precondition.

The Subtractive Method

The oldest method is subtraction, introduced by the Dutch physiologist Franciscus Donders in 1868. Donders distinguished three tasks. In a simple reaction (his a-reaction) one stimulus is met with one response, and the participant responds as fast as possible: this measures detection and motor execution with no choice involved. In a choice reaction (his b-reaction) each of several stimuli demands its own response, so the participant must both tell the stimuli apart and select the matching action. In a go/no-go reaction (his c-reaction) there are several possible stimuli but only one requires a response, which is withheld for the others, so the participant must discriminate the stimuli but need not select among responses. Donders' insight was that these tasks form a graded series, each adding one operation to the last, so the duration of that operation can be recovered by subtraction (Donders, 1969). The response-selection stage that subtraction isolates was later found to obey a quantitative law of its own: choice reaction time rises with the logarithm of the number of equally likely alternatives, lengthening by a roughly constant increment each time that number doubles (Hick, 1952; Hyman, 1953). The Hick–Hyman law ties selection time to the information the stimulus resolves, measured in bits, giving the stage a slope as well as a duration.

The arithmetic is direct. The go/no-go task adds stimulus discrimination to the simple task, so the discrimination time is the go/no-go reaction minus the simple reaction. The choice task adds response selection to the go/no-go task, so the response-selection time is the choice reaction minus the go/no-go reaction. In this way a difference between two averaged latencies becomes an estimate of the duration of one hidden mental operation — the first time in the history of psychology that an unobservable process was given a number. The method rests on a strong premise that Donders himself stated and that later work named the assumption of pure insertion: that a stage can be added to a task without changing the duration of any other stage. Where the premise fails — where inserting a discrimination also slows encoding, say — the subtraction misattributes the extra time, and this vulnerability is exactly what the additive-factors method was designed to sidestep. The first demonstration lets the reader set the three reaction times and read off the two inserted processes.

Additive Factors and Processing Stages

Saul Sternberg revived and repaired Donders' program a century later. His experimental engine was the memory-scanning task: a person holds a short set of items — the memory set — in mind, then sees a probe and decides as fast as possible whether it was in the set. Sternberg found that the time to answer rose as a straight line with the size of the set, adding a near-constant increment of roughly 38 milliseconds for each extra item, and that the slope was the same whether the probe was present or absent. He read the linearity as the signature of a serial comparison process that comes to a target and the equal present-absent slopes as evidence that the scan is exhaustive — every item is checked even after a match is found, because stopping to check would cost more than it saves (Sternberg, 1966).

The deeper contribution was methodological. Rather than inserting whole stages and subtracting, Sternberg proposed manipulating two factors at once and asking how their effects on reaction time combine (Sternberg, 1969). The logic is that if two factors influence different stages, their effects are independent and simply add: each shifts the total by a fixed amount regardless of the other. If two factors influence the same stage, they interact, and the effect of one depends on the level of the other. Additivity therefore becomes a test for separate stages, and an interaction a test for a shared one — an inference that needs no assumption of pure insertion, because nothing is being removed. In the scanning task, degrading the quality of the probe slows responding by a constant amount at every set size, adding to the set-size effect rather than multiplying it: probe quality acts on an early encoding stage and set size on a later comparison stage, so the two are separate. Concretely, with an intercept near 400 milliseconds and a slope of 38 milliseconds per item, a set of four gives 400 + 38 x 4 = 552 milliseconds; degrading the probe adds a constant 50 milliseconds to encoding, giving 602 milliseconds, and the set-size slope is unchanged. That parallel shift is the additivity signature, and the second demonstration lets the reader produce it.

The inference from additivity carries an assumption of its own, weaker than pure insertion but still substantive: that processing proceeds in discrete, non-overlapping stages, each finishing before it hands its completed product to the next. James McClelland challenged exactly this, proposing a cascade model in which stages run continuously and overlap in time, every stage passing partial, still-accumulating output forward rather than waiting to complete (McClelland, 1979). Continuous processing of this kind can also produce additive reaction-time effects, so additivity localises where a factor acts without proving that the stages it distinguishes are truly discrete. The method trades the strong premise of pure insertion for the weaker premise of separable stages — a real gain in security, not an escape from assumption.

Automatic and Controlled Processing

Decomposing a task into stages says little about how much of the processor each stage consumes, and that question produced the most consequential distinction in the field. Walter Schneider and Richard Shiffrin trained people to search for target items among distractors under two regimes. Under consistent mapping, a stimulus that was ever a target was never a distractor, so the same items always called for the same response; under varied mapping, targets and distractors were swapped between trials, so an item might be a target now and a distractor later. With varied mapping, search stayed slow and its time rose with the number of items and the size of the memory set, the marks of a serial, capacity-limited process. With consistent mapping and enough practice, search became fast and nearly flat across set size, as though the targets now announced themselves in parallel without drawing on central capacity (Schneider & Shiffrin, 1977). The pair named the two regimes controlled and automatic processing and argued, in the companion paper, that consistent mapping is what lets practice convert one into the other (Shiffrin & Schneider, 1977).

Michael Posner and Charles Snyder had drawn a parallel line by a different route, proposing that a process counts as automatic to the degree that it occurs without intention, without giving rise to conscious awareness, and without interfering with other ongoing activity (Posner & Snyder, 1975). The three criteria matter because they can come apart, and a large later analysis argued that they routinely do: automaticity is not a single switch but a bundle of features — speed, unintentionality, efficiency, unavailability to consciousness — that dissociate, so a process can be automatic in one sense and controlled in another (Moors & De Houwer, 2006). What made the practice effect more than a curiosity was its regularity. Across an enormous range of skills, the time to perform a task falls with practice as a power function — rapid early gains that flatten into ever smaller improvements — a relation stable enough to be called the power law of practice (Anderson, 1982). The name is not settled, however: Andrew Heathcote and colleagues showed that the power form is largely an artefact of averaging, because individual learning curves are typically better fit by an exponential, and averaging over people who each improve exponentially at their own rate manufactures the power function seen in the group (Heathcote, Brown, & Mewhort, 2000). The speed-up itself is not in doubt; its exact functional form, and what that form implies about the underlying mechanism, are. Gordon Logan gave it a mechanism with his instance theory: each encounter with a problem lays down a separate memory of its solution, and performance improves because the system shifts from computing the answer by a slow general algorithm to retrieving a stored answer directly, with retrieval growing faster and more reliable as instances accumulate (Logan, 1988). Automatization, on this account, is not the greasing of a fixed procedure but a change in the kind of process that produces the response. The third demonstration traces the resulting curve.

Serial versus Parallel Processing

Underlying both classic methods is a picture of stages that run one after another, and the automatic-controlled work adds the possibility that some processing runs in parallel instead. Whether a given task is carried out serially, one operation finishing before the next begins, or in parallel, several proceeding at once, is a question about mental architecture — and it is far harder to answer than it looks. James Townsend proved the difficulty precisely: for a wide class of models a serial process and a parallel process can produce identical mean reaction times, so that the two architectures, in his phrase, look like Tweedledum and Tweedledee and cannot be told apart from averages alone (Townsend, 1990). A flat search function need not mean parallel processing, and a rising one need not mean serial processing; a limited-capacity parallel model, dividing a fixed resource among more items, produces the same rising curve a serial scan does.

Distinguishing the architectures therefore requires more than mean latency. It requires attention to capacity — whether adding items slows the processing of each, as a shared resource predicts, or leaves it untouched, as independent parallel channels predict — and to the full shape of the reaction-time distribution under factorial manipulations, where serial and parallel models make different predictions that their means conceal. Sternberg's own conclusion that memory scanning is serial and exhaustive was later contested on exactly these grounds, with parallel models shown to fit the linear set-size functions equally well. The lesson is not that the question is unanswerable but that it is underdetermined by the coarsest data: the same behaviour is consistent with more than one internal arrangement, and separating them is a problem of experimental design and formal modelling rather than of reading a slope. This caution generalises to every inference in this article, and the Discussion returns to it.

Worked Example

Donders' subtraction can be worked by hand, and the first demonstration reproduces the calculation. Suppose an experiment measures three mean reaction times. The simple reaction, one stimulus and one response, comes out at 220 milliseconds. The go/no-go reaction, in which the participant responds to one stimulus and withholds the response to another, comes out at 290 milliseconds. The choice reaction, in which each of two stimuli demands its own distinct response, comes out at 350 milliseconds.

The subtractive logic assigns each increment to one inserted operation. The go/no-go task differs from the simple task by the addition of stimulus discrimination — the participant must now tell two stimuli apart — so the duration of discrimination is 290 - 220 = 70 milliseconds. The choice task differs from the go/no-go task by the addition of response selection — the participant must now choose which of two responses to make — so the duration of response selection is 350 - 290 = 60 milliseconds. From three averaged latencies the method has extracted numerical estimates of two operations that are nowhere directly visible: a 70-millisecond discrimination stage and a 60-millisecond selection stage. The estimates are only as good as the assumption of pure insertion behind them; if requiring a choice of response also lengthens discrimination, the 60-millisecond figure absorbs that spillover and overstates selection. This is precisely the confound the additive-factors method removes, which is why the two methods are complementary rather than rival: subtraction gives a duration under a strong assumption, additive factors gives a stage assignment under a weaker one.

Discussion

The study of mental processes is the study of how to make disciplined claims about operations no one can watch. Its central move is to treat behaviour as an accumulated record and to design comparisons that partition that record into interpretable pieces — Donders by subtracting matched tasks, Sternberg by crossing factors and reading their interactions, Schneider and Shiffrin by varying consistency of mapping to expose the boundary between capacity-limited and capacity-free processing (Donders, 1969; Sternberg, 1969; Schneider & Shiffrin, 1977). Each method buys its conclusion with an assumption, and the history of the field is largely the history of trading stronger assumptions for weaker ones: pure insertion gives way to additivity, a single automaticity switch gives way to a bundle of dissociable features, a confidently serial scan gives way to an acknowledgement that parallel models fit the same curve (Moors & De Houwer, 2006; Townsend, 1990).

Two themes persist. The first is that architecture is underdetermined by mean performance: because serial and parallel processes, and algorithmic and retrieval-based ones, can mimic each other in the average, progress depends on richer data — full distributions, capacity manipulations, the graded traces that analog processes such as mental rotation leave in time (Shepard & Metzler, 1971; Logan, 1988). The second is that the processor is defined as much by its limits as by its operations: capacity is the scarce resource that makes controlled processing costly and automatic processing valuable, and much of cognition can be read as the system's contrivances for economising it (Miller, 1956). What the chronometric tradition established, and what still underwrites the measurement of cognition today, is that the hidden machinery of thought is not beyond empirical reach: with the right comparison, a reaction time is a window onto a process, provided the analyst remembers how many processes could have cast the same shadow.

Common Misconceptions

A reaction time measures the speed of a single mental process.

A reaction time is the sum of the durations of many stages — encoding, comparison, decision, response selection, and motor execution — not the reading of one. Recovering the duration of any single operation requires a decomposition method such as subtraction or additive factors, which is the entire point of the chronometric program (Donders, 1969; Sternberg, 1969).

Automatic and controlled processing are two discrete, mutually exclusive categories.

Automaticity is better understood as a bundle of separable features — speed, unintentionality, efficiency, and unavailability to consciousness — that do not always travel together. A process can be automatic in one sense and controlled in another, so the label marks a region on several graded dimensions rather than one side of a switch (Moors & De Houwer, 2006).

A flat reaction-time function proves parallel processing and a rising one proves serial processing.

Serial and parallel models can produce identical mean reaction times, and a limited-capacity parallel process yields the same rising set-size function as a serial scan. Mean latency alone cannot decide the architecture; capacity analysis and the full response-time distribution are needed (Townsend, 1990).

Glossary

Additive-factors method.

Sternberg's technique of crossing two experimental factors and inferring processing structure from their combined effect on reaction time: additive effects indicate separate stages, an interaction a shared stage.

Automatic processing.

Fast, parallel processing that makes little or no demand on central capacity and runs with little intention or awareness; developed through consistent practice, as in consistent-mapping search.

Capacity.

The limited pool of processing resource available at one time; controlled processes draw on it and compete for it, whereas automatic processes have largely escaped it.

Cascade model.

McClelland's alternative to discrete stages, in which processing runs continuously and stages overlap in time, each passing partial output forward before it finishes; it can mimic the additive effects the additive-factors method reads as evidence of separate stages.

Choice reaction time.

The latency to respond when each of several stimuli requires its own distinct response, so the task involves both discrimination and response selection; Donders' b-reaction.

Controlled processing.

Slow, serial, capacity-limited processing that is deployed intentionally and flexibly; required under varied mapping and whenever stimulus-response relations are novel or inconsistent.

Diffusion model.

Ratcliff's sequential-sampling account of two-choice decisions, in which noisy evidence accumulates to one of two boundaries; fitting the full latency distribution separates the rate of evidence accumulation from the response caution that governs the speed-accuracy tradeoff.

Hick–Hyman law.

The regularity that choice reaction time increases with the logarithm of the number of equally likely alternatives — with the information the stimulus transmits, in bits — so response-selection time grows by a constant increment each time the alternatives double.

Instance theory.

Logan's account of automatization on which each encounter stores a separate memory trace, and practice improves performance by shifting it from algorithmic computation to direct retrieval of stored instances.

Mental chronometry.

The use of reaction time to infer the timing, number, and organisation of the mental processes that intervene between stimulus and response.

Mental process.

An unobservable internal operation that receives, transforms, stores, or acts on information on the path from stimulus to response.

Power law of practice.

The regularity that the time to perform a task decreases with practice as a power function, with large early gains that diminish steadily as skill develops.

Pure insertion.

The assumption behind the subtractive method that a processing stage can be added to a task without altering the duration of the other stages; its failure biases subtractive estimates.

Reaction time.

The interval between the onset of a stimulus and the completion of a response, treated in chronometry as the summed duration of the intervening processing stages.

Response selection.

The stage at which a discriminated stimulus is mapped to the action it calls for; isolated by Donders as the difference between choice and go/no-go reactions.

Serial exhaustive search.

A comparison process that checks memory-set items one at a time and continues through all of them even after a match is found; Sternberg's proposed account of memory scanning.

Simple reaction time.

The latency to make a single prepared response to the onset of a single stimulus, involving detection and motor execution but no choice; Donders' a-reaction.

Speed-accuracy tradeoff.

The trade a responder can always make between responding faster and responding more accurately; because it can move reaction time with no change in processing, chronometric inference requires accuracy to be matched across the conditions compared.

Stage.

A distinct processing operation that occupies its own interval within the stimulus-to-response chain and can, in principle, be isolated by chronometric methods.

Subtractive method.

Donders' technique of estimating the duration of a mental operation as the difference in mean reaction time between a task that includes it and a matched task that does not.

Key Researchers

Franciscus C. Donders (1818-1889). Physiologist at Utrecht University; he founded mental chronometry, devising the subtractive method that estimates the duration of an individual mental operation from the difference between matched reaction-time tasks. Wikipedia - Wikidata

Gordon D. Logan. Centennial Professor of Psychology at Vanderbilt University; he proposed the instance theory of automatization, recasting the shift from controlled to automatic performance as a move from algorithmic computation to single-step memory retrieval. Faculty Page - Wikipedia - ORCID

Ulric Neisser (1928-2012). Psychologist at Cornell University; his 1967 synthesis named and defined the field, framing cognition as the processes by which sensory input is transformed, reduced, elaborated, stored, recovered, and used. Wikipedia - Wikidata

Michael I. Posner. Professor Emeritus of Psychology at the University of Oregon; he advanced chronometric methods for studying mind and, with Snyder, drew the influential distinction between automatic and consciously controlled processing. Faculty Page - Google Scholar - Wikipedia

Walter Schneider. Professor of Psychology at the University of Pittsburgh; with Shiffrin he established the distinction between controlled and automatic processing, showing that consistent stimulus-response mapping over practice yields fast, capacity-free search. Faculty Page - Google Scholar

Richard M. Shiffrin. Distinguished Professor of Psychological and Brain Sciences at Indiana University Bloomington; with Schneider he developed the theory of controlled versus automatic processing, and earlier co-authored the Atkinson-Shiffrin memory model. Faculty Page - Google Scholar - Wikipedia - Wikidata - ORCID

Saul Sternberg. Professor Emeritus of Psychology at the University of Pennsylvania; he originated the additive-factors method and the high-speed memory-scanning task, using reaction-time patterns to infer discrete stages of human information processing. Faculty Page - Wikipedia - ORCID

Frequently Asked Questions

What is a mental process in cognitive psychology?
It is an internal operation that receives, transforms, stores, or acts on information between a stimulus and a response, such as encoding, comparison, retrieval, decision, or response selection. Because it cannot be observed directly, it is studied by inference from behaviour (Neisser, 1967).

Why can mental processes not be observed directly?
Only the stimulus entering the system and the response leaving it are measurable; the operations in between are hidden inside the organism. Cognitive psychology reconstructs them from patterns in behaviour, above all from how long a response takes (Posner, 1978).

What is Donders' subtractive method?
It estimates the duration of a mental operation as the difference in mean reaction time between a task that includes the operation and a matched task that does not. Donders used simple, go/no-go, and choice reactions to isolate discrimination and response selection (Donders, 1969).

What is the additive-factors method?
Proposed by Sternberg, it crosses two experimental factors and inspects how their effects on reaction time combine: if the effects add, the factors act on separate processing stages; if they interact, the factors act on the same stage (Sternberg, 1969).

What is the difference between automatic and controlled processing?
Controlled processing is slow, serial, and capacity-limited, and is used when stimulus-response relations are novel or inconsistent; automatic processing is fast, parallel, and capacity-free, and develops through consistent practice (Schneider & Shiffrin, 1977; Shiffrin & Schneider, 1977).

What is the power law of practice?
It is the finding that the time to perform a task falls with practice as a power function, with rapid early improvement that flattens over time. Logan explained it as a shift from computing answers by algorithm to retrieving stored instances from memory (Anderson, 1982; Logan, 1988).

Can reaction time tell whether processing is serial or parallel?
Not from mean latency alone, because serial and parallel models can produce the same average reaction times, and a limited-capacity parallel process mimics a serial scan. Separating them needs capacity analysis and the full reaction-time distribution (Townsend, 1990).

How is mental chronometry used today?
The chronometric logic of decomposing reaction time into stages remains the standard way to measure the components of cognition, and reaction time is now combined with neuroimaging and electrophysiology to locate those stages in the brain in time as well as in function (Posner, 2005).

References