Experimental linguistics uses planned comparisons to investigate language: how people judge sentences, interpret meanings, distinguish sounds, recognise words and produce utterances. Its gold standard is an experiment whose manipulation, measurement and analysis support a clearly bounded conclusion. A striking difference between two averages becomes useful evidence only when we understand what else could have produced it.
To design a good language experiment, start with a question that competing explanations answer differently. Construct suitable language materials, decide how participants encounter them, choose an observable response and specify how uncertainty will be evaluated. This guide works through those decisions using original teaching examples. All numerical results below are invented for calculation and interpretation practice; they are not findings from newly recruited participants.
The Gold Standard collection connects dependable methods across subjects. Psycholinguistics explains the mental processes involved in language use. Here the focus is the architecture of an experiment, including applications to grammar, sound and meaning. The aim is to help you design a defensible small investigation, read a research paper more critically, or recognise why a persuasive demonstration still needs a better comparison.
Choose your experimental linguistics route
Chapters 1–3: Define the question and build materials
Chapters 4–6: Design the comparison and measurement
Chapters 7–9: Understand variation and uncertainty
Chapters 10–12: Plan, repair and test transfer
Chapter index
Define the question and build materials · Chapters 1–3
1. Turn an interesting observation into a testable contrast
Design the comparison and measurement · Chapters 4–6
4. Use a factorial design to test a relationship
5. Counterbalance versions without confusing it with randomisation
Understand variation and uncertainty · Chapters 7–9
7. Treat participants and items as sources of variation
8. Calculate uncertainty without overstating the answer
9. Plan the sample around the uncertainty you need to reduce
Plan, repair and test transfer · Chapters 10–12
10. Make exclusions and exploratory decisions visible
11. Diagnose the experiment before diagnosing the language user
Answer questions and complete a study · Chapters 13–14
CHAPTER 1 OF 14 · Define the question and build materials
1. Turn an interesting observation into a testable contrast
Ask what would distinguish the explanations
Suppose someone says that a sentence containing a pronoun feels harder to understand than a sentence naming the person. That observation admits several explanations. The pronoun might have more than one plausible referent. The surrounding story might be confusing. One sentence might contain unfamiliar vocabulary. The observer might simply prefer explicit writing. An experiment should separate a useful subset of these possibilities.
Our running example asks whether introducing a second plausible person changes the processing cost of a pronoun relative to a repeated name. The question is narrower than whether pronouns are difficult. It predicts a relationship between the available referents and the expression used to identify one of them. That relationship can be tested through a comparison with four conditions.
Before collecting anything, write the anticipated result in ordinary language: the difference between pronoun and name should change when a competing referent is introduced. Then state a rival account. If additional story complexity slows every continuation equally, both pronoun and name conditions may become slower by similar amounts. These predictions give the comparison a purpose.
Keep causal claims attached to the manipulation
An experiment deliberately varies something. An observational study might compare people with different language histories without assigning those histories. Both can be valuable, but their conclusions have different foundations. Randomly assigning a sentence version can help estimate its effect under the task conditions; it does not randomly assign someone’s education, multilingual experience or reading habits.
Imagine that bilingual volunteers respond faster than monolingual volunteers. Calling bilingualism the cause would require addressing how the groups were formed and what else differs between them. Adding a randomised sentence manipulation does not remove every limitation of the group comparison. Write down which variables were assigned and which were measured before deciding what causal language is justified.
For a first investigation, choose a question small enough that the comparison can remain clear. A well-supported statement about one construction and one participant population is more informative than a universal claim built from uncontrolled examples. Scientific Method places this reasoning within the wider practice of testing explanations against evidence.
CHAPTER 2 OF 14 · Define the question and build materials
2. Define the construct, the task and the recorded value
A theory is not a column in a spreadsheet
Ambiguity, grammatical knowledge and lexical access are theoretical ideas. A keypress, a rating or a recorded utterance is an observation. The operational definition explains how a task connects the two. Without that bridge, researchers can measure something very precisely while remaining unsure what it means.
For the pronoun example, possible outcomes include acceptability ratings, reading time in a specified region, and answers to comprehension questions. These are different measurements. A low naturalness rating need not mean that a reader failed to understand the intended message. A long reading time might accompany successful resolution. A correct final answer may conceal a difficult intermediate step.
Choose a primary outcome that answers the question. If the hypothesis concerns where processing effort arises during a sentence, a single end-of-task satisfaction score is a weak match. If the question concerns the interpretation ultimately selected, a timing difference without an interpretation measure may leave the central issue unresolved.
Write the trial record before writing the conclusion
For a hypothetical reading experiment, a useful record contains an anonymous participant code, item code, list assignment, condition, displayed region, response time, comprehension response and technical-event flags. Keep the actual stimulus version available through the item and condition codes. A column labelled only “score” is insufficient because another analyst cannot reconstruct how it was obtained.
Specify units. A time of 0.65 seconds equals 650 milliseconds, but software exports do not always share a unit. Specify the timing origin too: stimulus request, actual display onset and sound onset are different events. Decide whether the response ends the trial or merely advances to another region.
For acceptability work, Schütze and Sprouse’s discussion of judgment data explains why a reported judgment is evidence requiring interpretation, rather than a direct inspection of grammar. Make instructions explicit: “How natural does this sentence sound in this context?” asks something different from “Would you use this sentence in a formal letter?”
Try explaining the outcome to a colleague without using the construct’s name. If you cannot describe what the participant actually does, the measurement plan needs another pass.
CHAPTER 3 OF 14 · Define the question and build materials
3. Build stimulus sets that survive close inspection
Create versions of the same item
Consider a story introducing Nadia, a curator. A second version introduces Nadia and another woman, a guard. The continuation begins either “Later, she examined the display” or “Later, Nadia examined the display.” These four versions illustrate the planned referent-by-expression comparison. They are starting materials, not a validated stimulus set.
The two-person context must actually make the alternative referent plausible. Mentioning the guard in a remote aside might not do that. Conversely, making the guard the obvious subject of the previous sentence could strongly favour her, creating a different comparison. Inspect discourse prominence, recency, grammatical position and the plausibility of each person’s action.
Match what can reasonably be matched, and document what cannot. The two contexts differ in their content because that is part of the manipulation. The name and pronoun differ in length because they are different expression types. A factorial design can help distinguish some patterns, but it does not magically cancel every confound. For example, length-related effects might themselves depend on context.
Do not let one memorable sentence represent a language
Develop multiple item families using different people, activities and settings while preserving the intended relationship. Keep a material sheet explaining the target manipulation and likely alternatives for each family. If every competing person is a guard and every target person is a curator, occupational roles may become entangled with the experimental conditions.
A small pilot can reveal that a supposedly neutral verb strongly favours one role, that an unfamiliar name is difficult to read, or that participants interpret an instruction differently from the researcher. Revise those materials before the main study. Record the revision and avoid treating pilot results selected for looking promising as independent confirmation.
Use fillers for a defined reason, such as reducing repetition of the critical construction or providing a wider range of sentence types. They should not accidentally teach a shortcut. If every sentence containing “she” requires the same response, the task may reward noticing that surface cue instead of doing the intended language work. Corpus Linguistics offers complementary methods for inspecting naturally occurring examples and their contexts; corpus frequency can inform material selection without replacing the experimental comparison.
A useful repair exercise is to give someone the stimulus sheet without condition labels. Ask what differences they notice. Their observations can expose unintended contrasts that the labels made you overlook.
CHAPTER 4 OF 14 · Design the comparison and measurement
4. Use a factorial design to test a relationship
Work through the four cells
A two-by-two factorial design crosses two factors, each with two levels. In our constructed example, the factors are referent context, with one or two plausible people, and expression, with a pronoun or repeated name. Crossing them means that both expressions appear in both contexts.
Suppose the invented mean times for a predefined continuation region are 620 milliseconds for one-person/pronoun, 640 for one-person/name, 740 for two-person/pronoun and 660 for two-person/name. Define the expression contrast as pronoun minus name. In the one-person context it is 620 − 640 = −20 milliseconds. In the two-person context it is 740 − 660 = 80 milliseconds.
The difference between those differences is 80 − (−20) = 100 milliseconds. Descriptively, the pronoun’s relative cost increased by 100 milliseconds when the competitor was introduced. This interaction is closer to the stated hypothesis than the observation that the two-person/pronoun cell is simply the slowest.
Distinguish a numerical pattern from evidence about its reliability
These four means contain no information about sampling uncertainty. They cannot tell us whether the interaction is estimated precisely, whether it depends on two unusual items, or whether the participants varied substantially. The correct next step is analysis of the underlying observations using a model suited to the design and response distribution.
An interaction also depends on the scale. A difference of differences in milliseconds is not identical to a difference of differences in log time, where comparisons correspond to ratios on the original scale. State the scale before interpreting the relationship. Choose transformations and models for defensible measurement and distributional reasons, not because one version produces a preferred result.
Finally, do not conclude that two conditions differ because one separate test was statistically significant and another was not. The claim that effects differ requires a direct comparison of those effects. The worked arithmetic shows what that comparison is trying to estimate; a suitable uncertainty analysis establishes how much the observed pattern supports it.
CHAPTER 5 OF 14 · Design the comparison and measurement
5. Counterbalance versions without confusing it with randomisation
Give each person a coherent experience
If a participant reads all four versions of the Nadia story, later responses may reflect memory of earlier versions. A common solution is to show each person one version of each item family while distributing versions across people. This preserves comparisons across the study without asking every reader to encounter every rewritten story.
Consider an original planning example with 48 item families, four conditions and four lists. Divide the item families into four groups of 12. On List 1, assign groups 1, 2, 3 and 4 to conditions A, B, C and D respectively. On List 2, rotate those assignments to B, C, D and A; continue with C, D, A, B and then D, A, B, C.
Every list now contains 12 items per condition. Every item family appears in each condition across the four lists. With 60 participants allocated evenly, each list has 15 participants, and each item-condition version receives 15 observations before exclusions. Each condition contributes 60 × 12 = 720 critical-trial observations, giving 2,880 overall.
Balance versions, then consider trial order
This cyclic arrangement balances item versions. It does not, by itself, balance every preceding-trial effect or every serial position. Randomising or suitably constraining trial order is a separate decision. For example, you might prevent several critical trials of the same type from appearing consecutively, provided the rule is defined in advance and implemented consistently.
Random assignment to lists helps avoid systematic participant differences between versions. Balance can still be disrupted by incomplete sessions or unequal exclusions. Inspect the achieved allocation and report it. Do not silently move remaining participants between conditions after seeing their responses.
The 2,880 observations are not 2,880 independent people. Each participant contributes many responses, and each item family is encountered by many participants. This arithmetic establishes the data structure and workload, not an automatic guarantee of statistical power. A good allocation sheet should let another person trace both a participant’s experience and an item’s appearances across the full experiment.
CHAPTER 6 OF 14 · Design the comparison and measurement
6. Choose a method for the claim it can support
Match the response to the linguistic question
An acceptability task asks for a judgment about a presented expression under specified instructions. A lexical-decision task asks whether a stimulus is a word in the relevant language. A reading task records aspects of how a sentence is traversed. None of these measurements is interchangeable with general language ability.
The specialist guide to auditory lexical decision explains why spoken duration, nonword construction and the response reference point matter. Suppose two word sets differ in both frequency and recording duration. Measuring latency from sound onset can combine differences in available acoustic information with the lexical contrast. Simply subtracting duration afterward is not automatically a complete repair.
For the running sentence experiment, self-paced reading offers region-by-region timing. Decide where the critical region begins and ends and whether a following region is also theoretically relevant. A cost may emerge after the expression itself. Searching every region and reporting only the strongest difference changes the evidential status of the finding.
More equipment does not remove the need for a comparison
The visual-world paradigm connects spoken language with looks towards displayed referents. The pictures, their positions and their names become part of the experiment. A visually conspicuous target could attract attention before the critical spoken information arrives, so the display needs its own controls.
For an original design exercise, imagine testing whether “cup” briefly competes with “cushion” during spoken recognition. A picture set cannot be chosen only because those words share initial sounds. Participants must recognise the pictures and use the expected names. A picture some call “mug” changes the intended relationship. Norming the labels is therefore part of building the comparison.
Combining methods can strengthen an explanation when each measurement tests a useful prediction. It can also produce more opportunities to select a pleasing result. State which outcomes are primary, which are supporting and which are exploratory before interpreting their agreement or disagreement.
CHAPTER 7 OF 14 · Understand variation and uncertainty
7. Treat participants and items as sources of variation
Repetition creates structure in the data
Some readers are generally faster than others. Some stories are generally harder than others. Readers may also differ in how strongly the manipulation affects them, and some item families may produce a larger contrast than others. These are distinct sources of variation, not merely inconvenient noise to remove.
In a mixed-effects analysis, participant and item intercepts can represent baseline differences, while suitable random slopes can represent variation in condition effects. The exact specification depends on which factors vary within each grouping unit and what the design can support. A random effect is not a declaration that a variable is unimportant.
Barr and colleagues’ 2013 study demonstrated risks from neglecting relevant random-effect structure in confirmatory analyses. Matuschek and colleagues’ later simulations examined the balance between error control and power. These papers motivate careful modelling; they do not make one formula appropriate for every dataset. Model convergence, singular fits and the intended generalisation require attention.
Diagnose a claim that rests on a few materials
Imagine an experiment with 80 readers but only two item families. Both stories use unusual occupations. Even if their average condition difference is clear, the evidence for ordinary reference resolution across many everyday situations remains limited. Recruiting another 80 readers measures responses to those same stories more precisely; it does not automatically broaden the range of language materials.
Inspect variation by participant and by item as well as the overall contrast. A plot showing item-level differences may reveal that one family behaves unlike the others. Investigate whether it contains a coding mistake or a meaningful linguistic exception. Do not remove it merely because it weakens the result.
The practical question is what population the claim concerns: these participants, comparable readers, these stories, or a wider construction class. State that target early. It determines which variation the design must sample and which uncertainty the analysis needs to represent.
CHAPTER 8 OF 14 · Understand variation and uncertainty
8. Calculate uncertainty without overstating the answer
A deliberately simplified paired calculation
Consider a separate teaching example using a fixed set of materials. Twenty-five independent participants each provide an average time in condition A and condition B. Define each participant’s difference as B minus A. Suppose the mean of those 25 differences is 36 milliseconds and their sample standard deviation is 60 milliseconds.
The standard error of the mean paired difference is 60 divided by the square root of 25, which equals 12 milliseconds. Using a two-sided 95% t interval with 24 degrees of freedom gives an approximate critical value of 2.064. The margin is therefore 2.064 × 12 = 24.768 milliseconds. The interval is 36 ± 24.768, or approximately 11.2 to 60.8 milliseconds.
This calculation follows the paired-difference framework in NIST’s analysis of paired observations. Its application assumes an appropriate sampling model, independent participants and a distribution of participant differences for which the t procedure is reasonable. It deliberately treats the material set as fixed; it does not solve the crossed participant-and-item generalisation problem from the previous chapter.
Say what the interval means
Under the model, the procedure has 95% long-run coverage for the population mean difference across repeated samples. The particular interval is an uncertainty summary, not a claim that 95% of individual readers have differences inside it. Nor does it establish a 95% probability that the proposed linguistic explanation is correct.
The estimated difference is positive in this illustration, but its practical importance depends on the task. A 36-millisecond shift in one laboratory region does not directly translate into a meaningful change in everyday reading performance. The theoretical prediction, measurement reliability and alternative explanations remain relevant.
If an interval includes zero, do not automatically declare equivalence. A wide interval can include both negligible and substantial effects. A claim of practical equivalence needs a justified negligible-effect range and an analysis appropriate to that claim. The repair is to ask how precise the estimate is and which meaningful possibilities remain compatible with the observations.
CHAPTER 9 OF 14 · Understand variation and uncertainty
9. Plan the sample around the uncertainty you need to reduce
There is no universal participant count
The number of participants needed depends on the effect of interest, variation, design, measurement reliability and planned analysis. The number and diversity of items matter too. Westfall, Kenny and Judd’s work on participants and stimuli shows why power calculations that ignore stimulus variation can mislead in crossed designs.
For a real project, identify an effect size worth detecting or an estimation precision worth achieving. Use relevant prior evidence cautiously; a small pilot’s observed effect can be very unstable. A simulation-based plan can vary plausible participant differences, item differences, residual variation and missingness, then evaluate the proposed analysis under those assumptions. Report the assumptions rather than presenting the resulting sample size as a universal rule.
Return to the 48-item, 60-participant example. The allocation works arithmetically, but the numbers were chosen to illustrate balance. They are not a recommendation that 60 readers are sufficient for reference-resolution research. Before recruitment, test whether the design can answer the scientific question at a defensible level of precision.
Count time as well as people
Suppose a session has 48 critical trials and 48 fillers. If the planning assumption is 12 seconds per trial on average, the trial sequence takes 96 × 12 = 1,152 seconds, or 19.2 minutes. Add six minutes for consent, instructions and practice, plus two minutes of breaks: the estimated session is 27.2 minutes. Pilot the actual duration because participants may differ substantially.
A longer session can increase fatigue, dropout and repeated exposure to the target construction. Adding items is therefore a design decision with costs as well as benefits. More observations are useful when they improve the relevant information, not merely when they enlarge a spreadsheet.
Decide the recruitment and stopping rule before examining the target result. If replacement participants are allowed after predefined technical exclusions, specify how replacements are allocated. Repeatedly checking for a preferred threshold and stopping when it appears requires a statistical design that explicitly accounts for that procedure.
CHAPTER 10 OF 14 · Plan, repair and test transfer
10. Make exclusions and exploratory decisions visible
Write a plan that could be followed by someone else
Preregistration records the planned questions, methods and analysis before the relevant results are known. The Center for Open Science’s guidance distinguishes planned tests from ideas developed after inspecting data. Both can contribute to knowledge, but readers should be able to tell which is which.
For our hypothetical experiment, the plan would identify the expression-by-context interaction, its primary measurement region, participant eligibility, item lists, stopping rule, technical-failure handling and statistical model. It would also state how nonconvergence or missing responses will be handled. “We will analyse reading time” leaves too many important choices unspecified.
Consider an original exclusion scenario. The program sometimes loses browser focus, and the next keypress is recorded after a long interruption. A rule based on a logged focus-loss event has a clear procedural rationale. Deleting the slowest trials only in the condition that undermines the hypothesis does not. A numerical cutoff also needs justification for this task; there is no single reaction-time boundary appropriate to every language experiment.
Preserve an account of what changed
Suppose 60 people begin, three sessions contain no usable critical trials because of a technical failure, and two participants do not meet a documented eligibility criterion. Report the flow from 60 to 55, along with the rules and their timing. Keep trial-level exclusions separate from participant-level exclusions so the final denominators can be reconstructed.
If a programming error is discovered after collection, explain it and its consequences. A justified deviation is better than an inaccurate claim that the original plan was followed. Where feasible, show whether reasonable alternative handling changes the conclusion. Sensitivity analyses should clarify dependence on decisions rather than provide a menu from which to select the most attractive result.
An unexpected pattern, such as a larger effect late in the session, can become a valuable new hypothesis. Label it exploratory, describe how it was noticed and test it in new data. The discovery is not weakened by an honest account of its origin.
CHAPTER 11 OF 14 · Plan, repair and test transfer
11. Diagnose the experiment before diagnosing the language user
Follow a failure to its most immediate explanation
If an expected difference is absent, inspect whether the manipulation was delivered correctly. Did both conditions accidentally display the same text? Did a list assignment reverse a label? Were recordings preloaded? Did the saved unit match the analysis unit? Check these possibilities before proposing a new theory about comprehension.
If accuracy is near perfect, the task may have a ceiling that hides relevant variation. If performance is near guessing, unclear instructions or unfamiliar materials may have overwhelmed the target contrast. A pilot should test the participant’s understanding of the task as well as the software’s ability to collect a response.
For browser-based timing, jsPsych’s official timing guidance and Anwyl-Irvine and colleagues’ timing study document why hardware, software and measurement requirements matter. Check the current setup rather than assuming that a benchmark for older browsers guarantees identical performance today. A recorded millisecond value is a measurement with a delivery process behind it.
Work through a repair case
Imagine that one condition’s audio files are larger and are downloaded only when each trial begins, while the other condition’s files are already cached. The observed delay could reflect a delivery difference. The repair begins by separating file readiness, actual playback onset and response timing in the logs, then correcting the implementation and validating it before new collection.
Do not simply subtract an assumed download delay from every trial. Without a valid measurement of the relevant event, that correction may create a new error. Preserve the original records and describe which conclusions the affected data can still support.
At the human level, slow responses are not diagnoses. They may reflect deliberate checking, uncertainty about instructions, a keyboard issue or a language variety unfamiliar to the researcher. An experimental task developed to compare conditions is not automatically a validated individual assessment. The strongest repair often improves the task or narrows the claim rather than attaching a deficit label to the participant.
CHAPTER 12 OF 14 · Plan, repair and test transfer
12. Transfer the design from reading to listening
The changed-context challenge
You now want to test the Nadia comparison with spoken stories instead of written ones. Keep the theoretical question: does a competing referent alter the relative processing cost of a pronoun? Before reading the answer, identify three features of the written experiment that cannot be transferred unchanged.
One answer is stimulus timing. Written text can appear as a region on a screen, whereas speech unfolds. A pronoun and a name may reach their informative points at different times. Measuring every response from the start of the sentence may combine reference resolution with differences in acoustic duration. The relevant timing reference must be justified for the listening task.
A second answer is prosody. Stress, intonation and pauses can influence the interpretation of a pronoun or repeated name. Recordings should not unintentionally make one referent more prominent only in one condition. Natural recordings and carefully controlled edits each have advantages and limitations; document the production method and inspect the results.
A third answer is access to the context. A reader may be able to reread, depending on the presentation method. A listener may hear the passage only once. Allowing replay changes the task again. Decide whether the question concerns initial interpretation or comprehension with opportunities to review.
A transfer plan preserves the relationship, not the apparatus
Pilot the spoken materials for intelligibility and intended discourse structure. Define the response measure afresh. If using pictures, check their naming and visual properties; if using a question after the story, acknowledge that it measures a later outcome. Do not describe a new task as an exact replication simply because it reuses the same words.
If the listening result differs from the reading result, several explanations remain possible: modality, prosody, task demands, participant sampling or measurement precision. A follow-up should distinguish some of these instead of treating disagreement as immediate proof that one study was wrong.
This is the educational transfer test developed in Education: learning transfers into life: can you carry the underlying comparison into a changed situation while noticing which assumptions no longer hold? A copied procedure is not enough; the learner must reconstruct why the comparison was informative.
CHAPTER 13 OF 14 · Answer questions and complete a study
13. Questions readers ask about language experiments
Does an experiment have to use a laboratory?
No. A controlled comparison can be conducted in several settings, including appropriately designed online tasks. The setting should suit the measurement and participant population. A quiet laboratory does not rescue confounded materials, while an online experiment needs suitable control over delivery, eligibility and technical variation. Describe the actual conditions under which evidence was collected.
Are acceptability ratings just opinions?
They are systematic reports of a response to language under specified instructions. That makes them data, with measurement properties and interpretive limits. Researchers can investigate their consistency and how they change across conditions. They should not treat a participant’s rating as a direct reading of an abstract grammar or as a verdict on the worth of a language variety.
Can I average a seven-point rating scale?
An average can be a descriptive summary, but the numbers represent ordered response categories whose spacing may not behave as equal intervals. The analysis should consider the scale and distribution, and ordinal models may be appropriate. Report the response options and inspect category use. A precise-looking decimal does not resolve the measurement assumptions.
Does a large dataset eliminate confounding?
No. If every difficult sentence also has unfamiliar vocabulary, many observations can estimate that combined contrast precisely while leaving its cause unresolved. Better sampling and better control answer different problems. First identify what changed together; then redesign the comparison or limit the interpretation to the combined change actually tested.
What is the difference between reproducibility and replication here?
Use the terms explicitly. Here, reproducibility means obtaining the reported analysis from the same data and procedures; replication means collecting new evidence to examine the finding again. Reanalysis can expose coding problems. New participants and suitably chosen new materials can test generalisation. Neither operation alone answers every question about an effect.
Can an AI model replace human participants?
Its outputs can answer a question about that model under a documented prompt and configuration. They are not interchangeable with human judgments. If the research question concerns people, replacing people changes the object of study. Model-assisted material generation also requires human inspection: fluent sentences can still contain unintended cues, odd meanings or near-duplicate items.
What should a beginner learn first?
Learn to state the contrast, identify alternative explanations and trace each recorded value back to a trial. Then learn the analysis required by the design. A small, transparent demonstration with synthetic data is a useful first step. It allows practice with allocation, calculations and reporting without making unsupported claims about real populations.
CHAPTER 14 OF 14 · Answer questions and complete a study
14. Build a complete investigation someone else can inspect
Assemble the evidence before writing the headline
Begin with a one-page question and prediction. Add a material sheet with item families and versions, a list-allocation sheet, a trial specification and an analysis plan. Test the full path using synthetic records: stimulus selection, response storage, exclusions, calculation and final display. A successful software launch is only one part of this rehearsal.
For the running example, the final report should let a reader recover the four-condition comparison, understand why the continuation region was selected, inspect the stimulus variation and see how participant and item uncertainty entered the analysis. Include the achieved sample, missingness, departures from the plan and results that do not support the preferred explanation.
Write a bounded conclusion. A defensible form might be: under the stated task and sampling conditions, the estimated expression contrast changed with referent context, with the reported uncertainty. The next sentence should identify unresolved alternatives. Avoid turning a laboratory contrast into a universal rule that pronouns should be removed from everyday writing.
Respect the people whose language makes the study possible
The Linguistic Society of America’s ethics statement emphasises voluntary participation, informed consent, confidentiality and responsibilities towards communities. Establish the applicable review process before recruitment. Explain recording, data access and withdrawal arrangements clearly. Sharing materials and analysis code can support scrutiny without making identifiable speech recordings public.
For a learning project, a useful first deliverable is a simulated experiment and a critique of its limits. Moving to human participation requires suitable supervision and procedures. Pay attention to accessibility and to whether instructions or materials unfairly exclude the very population the question concerns.
Follow a route that fits your next question
For theory about comprehension and production, continue to Psycholinguistics. For a concrete measurement method, revisit the guides to self-paced reading, auditory lexical decision and the visual-world paradigm. Read the linked primary studies and official methodological resources alongside these explanations; the numerical exercises here remain constructed teaching cases.
The How X Works hub connects these methods to other systems whose mechanisms need testing. The practical achievement is the same: another person can understand what was varied, what was observed, why the comparison matters and where the conclusion stops. That is what makes experimental linguistics a dependable way to learn about language.