What do reaction times mean in implicit association research?

Imagine two concepts repeatedly appearing together on a screen. One pairing seems straightforward. With another you hesitate briefly before responding. That difference may appear in how quickly you press the correct key. Perhaps one combination is easier for you to process. Or perhaps the keys were easier to remember, the words were more familiar, or you were already tired by the second block.

That is both the attraction and the difficulty of reaction-time research. The researcher observes a measurable outcome - the interval between stimulus and response - and wants to draw an inference about a process that cannot be observed directly. The strength of that inference depends on the study design, not on how many decimal places are recorded.

Neurofactor's existing article How do you measure associations? compares free association, self-report and indirect measures. This article examines a narrower technical question: what can the difference in time actually tell us?

A response latency is an interval, usually expressed in milliseconds (ms). Its starting point might be the display of a word or image. Its endpoint is the correct or incorrect keypress registered by the software. A latency of 700 ms means approximately seven tenths of a second passed between those two recorded events.

Several processes may contribute to that interval: perceiving the stimulus, recognising its category, processing relevant information, making a choice and carrying out the motor response. The task does not automatically measure these components separately. It records the overall task response. The precision of the timestamp also depends on the presentation hardware, browser or software, input device and technical checks.

A millisecond is therefore a unit of measurement, not a psychological interpretation. 'This participant responded in 700 ms' is more directly supported than 'This participant trusts the brand'. The second statement requires theoretical assumptions and additional evidence.

A conventional Implicit Association Test (IAT) requires participants to categorise items using response keys. In one combined condition, selected concepts share a key; in another, the pairings are reversed. Greenwald, McGhee and Schwartz introduced this procedure in 1998 [1].

Consider two fictional service providers, A and B, and attributes such as 'reliable' and 'uncertain'. In the task, participants classify words or images as quickly and accurately as possible. In one phase, A and 'reliable' share a response key. In the other, B and 'reliable' share a key. The research question is not whether someone presses a key quickly once. It concerns a systematic difference in performance between two predefined combinations.

Element	What is recorded?	What it does not automatically establish
Stimulus	Which word or image and category are displayed	Whether the participant likes it
Response	Which key is selected	The precise thought behind the choice
Reaction time	How long the response takes	A pure measure of association strength
Error	Whether the response breaks the task rule	That the person lies or suppresses something consciously
Condition difference	How performance varies between pairings	That one combination causes subsequent behaviour

The comparison is useful only when the concepts, stimuli, instructions and task structure adequately represent the research question. Different task variants can involve different psychological processes. 'We measured reaction times' is not a complete description of a method.

One participant may respond quickly across all tasks because they are familiar with keyboards. Another may be more careful and slower. Even for one person, latency varies with attention, tiredness and key position. An absolute mean for one condition combines all of these influences.

For that reason, a comparison within the same person across controlled conditions is often more informative than one isolated response time. Even then, you must establish whether the difference is reasonably consistent, what error patterns occur and whether the categories represent the intended construct.

Worked illustration: 680 versus 790 milliseconds

Imagine an entirely illustrative task that has not actually been conducted. Mean correct response time for pairing A is 680 ms and for pairing B it is 790 ms. The raw difference is 110 ms. In this fictional example, responses in condition B are, on average, slower.

What does this not show? It does not establish that A produces 16 per cent 'more trust', that preference is 110 points stronger or that people will ultimately buy A. Calculating a defensible IAT score requires decisions about which trials to include, error handling, latency variation and scoring direction. These numbers demonstrate only the logic of a condition difference.

Research using the conventional IAT commonly reports a D-score. Greenwald, Nosek and Banaji compared scoring methods in 2003 and introduced an improved procedure that relates differences between combined conditions to variation in response times and incorporates errors [2].

At a conceptual level:

Standardised condition difference ≈ difference between mean response times / relevant response-time variability.

This is a simplified explanation, not a full scoring recipe. The published D procedure includes rules for practice blocks, separate block pairs, extreme latencies and error treatment. Other indirect task variants may require different methods. A Neurofactor-specific RIAT or association score must not be described as a conventional IAT D-score unless the implemented task and validation justify that label.

The sign of a score depends on how the conditions are coded. A positive D-score does not universally mean 'positive', and zero does not mean that a person has no associations. Always state the pairings, the scoring direction and what zero represents in the specified task.

A common mistake is to assume that the fastest participant has the strongest association. Someone may answer quickly because they check their responses less carefully. Someone else may spend more time avoiding mistakes. Researchers call this a speed-accuracy trade-off.

Klauer and colleagues analysed IAT results with a diffusion model and found that condition differences were not attributable to a single process. Differences in evidence accumulation, response caution and non-decision components could all contribute to the observed latency [4].

Error data are therefore essential. A condition with a lower mean response time but many more errors is not automatically 'stronger'. Two groups with identical means may also differ in accuracy, outliers or participant-level variation.

A sound study protocol specifies in advance how incorrect responses are handled, when observations may be excluded and how performance is summarised across conditions. Those rules should follow the chosen method, not the result the organisation hopes to find.

A performance difference may involve the target association. It may also reflect other factors in part or in full:

1. Familiarity with the materials: a familiar word or logo may be recognised more quickly than an unfamiliar one.
2. Category and stimulus quality: ambiguous words or unevenly difficult images make one condition harder.
3. Motor and technical variation: keys, input devices and display timing affect response latencies.
4. Practice and order: participants learn the rules as the procedure continues. IAT studies show that additional practice can reduce order-related method effects [3].
5. Fatigue and attention: later blocks can be slower even when an association has not changed.
6. Strategy and response caution: people may trade speed against avoiding mistakes [4].
7. Language and culture: word length, meaning and translation may differ across target groups.
8. Context and instructions: preceding questions, examples and expectations may influence task performance.

These limitations do not make reaction-time research worthless. They make rigorous design, appropriate counterbalancing, documented equipment and checks of alternative explanations indispensable.

A carefully established difference may support	The difference does not establish on its own
Particular category combinations are processed relatively more fluently in this task	One unique mental cause of the difference
Task performance differs across controlled conditions	Exactly what someone thinks or feels
Groups differ on a measure with adequate reliability	Every individual in the group has the same pattern
Performance changes across specified conditions	That the manipulation causes a purchase or job application
An indirect measure adds a distinct layer of information	That self-report and real-world behaviour are unnecessary

Here, 'implicit' means that the construct of interest is inferred indirectly from task performance. It does not automatically mean unconscious, uncontrollable or more truthful than direct reporting. See Implicit and explicit attitudes: what is the difference?.

A research report must answer two different questions. Reliability concerns the consistency of a measure under specified conditions. Validity concerns whether the task and score support the construct that is being interpreted.

A reaction-time measure may reliably differ between conditions while still partly reflecting an unforeseen stimulus imbalance. Conversely, a meaningful effect may be difficult to reproduce when there are too few valid trials, excessive noise or a small sample. Both measurement quality and evidence for interpretation must be visible.

The IAT literature reports relationships between indirect measures and behaviour, but these depend on the topic, criterion and design. A 2009 meta-analysis by Greenwald and colleagues found generally limited-to-moderate average predictive associations across different outcome types [5]. Kurdi and colleagues reported heterogeneous relationships between IAT scores and intergroup behaviour in 2019 [6]. Those studies address particular domains; they do not provide a generic conversion formula for brands, job applicants or customers.

Any claim that 'this association affects purchasing behaviour' should ideally be supported by an appropriate behavioural criterion and a design that tests that relationship. A causal claim requires more than a correlation between two measures.

A useful research brief should make each of the following steps explicit.

1. Define the construct and comparison

Which relationship is actually being investigated? For example, do two service providers differ in how readily the same audience pairs them with 'security' or 'risk'? State the expected direction and plausible alternative explanations. Use more precise terms than 'the subconscious'.

2. Select and evaluate the stimuli

Choose words and images that adequately represent the intended concepts. Check familiarity, ambiguity, language, visual properties and difficulty. Comparing two brands is not a clean test when one is entirely unknown and this is ignored.

3. Predefine the task protocol

Document response keys, blocks, instructions, practice rounds, order variants, equipment and scoring rules. Base the choices on relevant literature and the particular application. Do not invent a threshold after seeing the data to make the result sound persuasive.

4. Collect response times and errors

Preserve sufficient trial-level data to check the computation. Record which stimulus and condition produced each response and identify implausible latencies or technical problems. Minimise identifiable data and follow appropriate privacy and research procedures.

5. Evaluate measurement quality and uncertainty

Investigate missing or excluded trials, error patterns, variability, possible order effects and whether the measure is suitable for group comparisons. Report relevant uncertainty around group results rather than presenting only a single mean.

6. Draw only the conclusion warranted by the design

Explain what actually differed, what interpretation is plausible and what alternatives remain. Where appropriate, compare the result with explicit answers, observed behaviour, interviews or other measures. Combining methods provides different perspectives, not automatic certainty.

Imagine an organisation asking whether candidates connect a role more readily with 'freedom' or 'uncertainty'. A researcher could design an indirect classification task pairing suitable descriptions of the role with carefully chosen attributes. The question is not: who is lying about their motivation? It is: what relative performance patterns arise for these specific pairings, and how can they responsibly be interpreted?

The research team would first check that the stimuli are understandable and comparably familiar, that order is suitably balanced and that candidates do not rely on different response strategies. Interviews or direct questions could explore what 'freedom' actually means to participants. Behavioural data could then separately test whether the task outcome is associated with applying for the role.

This is an illustrative research design, not a completed Neurofactor project or an observed outcome. Without a study, we cannot claim that any association dominates. Even with a reliable performance difference, a change in application rates would remain unproven.

No. A reaction-time task measures performance in a behavioural classification procedure. EEG records electrical activity at the scalp, whereas eye tracking records properties of gaze behaviour. Appropriate studies can examine these signals together, but they are different measurements. A longer latency is not, in itself, EEG evidence of stress or eye-tracking evidence of attention.

The research question comes first. An association task may suit an investigation of relative pairing patterns. A study of visual attention, actions in an environment or experience over time may need other instruments. Explore Neurofactor's research methods to see how the methods differ.

• 'Faster means more positive.' A relative evaluation is justified only to the extent supported by the task and coding.
• '700 ms means high association strength.' Without a comparison, protocol and reference frame, this is meaningless.
• 'One D-score predicts one purchase.' A relative task score is not an individual purchase prediction.
• 'Few errors prove that the effect is real.' Stimuli, task order, technical influences and measurement quality still matter.
• 'Implicit always means unconscious.' The measurement label does not justify that claim.
• 'We measured the brain.' A keyboard task records behaviour, not neural activity directly.
• 'The same score works for every RIAT variant.' Scoring rules belong to a specified, supported procedure.
• 'A difference proves a cause.' Effects on behaviour require their own appropriate study design.

When evaluating a reaction-time study, request:

1. The exact research question, categories and examples of stimuli.
2. Number of participants, inclusion criteria and relevant audience characteristics.
3. The full protocol, including practice and order.
4. Timing implementation and devices used.
5. Trial counts, errors and exclusions.
6. Score definition, coding direction and methodological justification.
7. Dispersion, uncertainty and reliability.
8. Plausible alternative explanations and limitations.
9. Any relevant relationship with explicit or actual behavioural data.
10. A conclusion no stronger than the data permit.

If the report does not show how the number emerged from the task, the number has not been explained well enough to guide a high-stakes decision.

Reaction times are informative when treated as observable performance within a carefully designed task. What matters is not one exceptionally fast response but a defensible pattern of differences between conditions - including errors, variability, order and possible alternative explanations.

A sound association study therefore demonstrates more than what it measured. It also makes clear why the step from milliseconds to meaning is justified. That is where scientifically grounded insight begins.

Further reading: How do you measure associations? · Implicit and explicit attitudes · Neurofactor research methods