Skip to content
Back to search results

Self-Paced Reading Task

Introduced by
Marcel Just, Patricia Carpenter, and Jacqueline Woolley
Year
1982
Type
Online measure of sentence-processing difficulty
Family
Psycholinguistics of sentence comprehension
Font size

What does it do?

The self-paced reading task — Just, Carpenter, and Woolley's moving-window paradigm — measures sentence comprehension as it unfolds.

Participants read a sentence one word (or phrase) at a time, pressing a button to reveal each segment and hide the previous one, and the software logs how long each segment stays on screen. Those per-region reading times index processing difficulty: when the parser stumbles — at a garden-path disambiguation, an unexpected word, or a long-distance dependency — reading slows, and the slowdown localizes the difficulty to a specific region of the sentence. It is one of the simplest and most widely used online measures in psycholinguistics.

The intuition behind the method

When you hit a hard word, you slow down — your reading time betrays your processing. Comprehension is invisible from the outside, but if you force readers to pull each word into view themselves, the time they linger before asking for the next word becomes a stopwatch on the mind. Lead a reader down a garden path ('the horse raced past the barn fell') and the moment the sentence forces a reanalysis, the button press lags. Self-paced reading turns that lag into data: a clean, cheap proxy for where and how much the parser struggles, region by region, without any special equipment.

History

Marcel Just, Patricia Carpenter, and Jacqueline Woolley introduced the moving-window self-paced reading paradigm in their 1982 Journal of Experimental Psychology: General paper, presenting it alongside eye-tracking as a way to track reading comprehension processes in time. The method spread rapidly through psycholinguistics because it required no specialized hardware, becoming a standard tool for studying garden-path sentences, syntactic ambiguity, and dependency processing. As statistical practice matured, the field moved from by-participant and by-item ANOVAs to linear mixed-effects models for reading times and, following arguments such as Jaeger's, to logit mixed models for the categorical comprehension data — and web platforms later made large-sample online self-paced reading routine.

When to use it

Use self-paced reading when you want a moment-by-moment, low-cost measure of where comprehension difficulty arises in a sentence — for garden-path and ambiguity resolution, agreement and dependency processing, prediction effects, and second-language sentence processing. It is ideal when you need an online measure but lack an eye-tracker, or when you want to run online at scale. It is less suitable when you need natural reading (the word-by-word button press is unnatural and disrupts normal eye movements), when regressions and rereading matter (self-paced reading cannot look back), or when very fine temporal resolution within a word is required — situations where eye-tracking is preferable.

Limitations and considerations

  • The word-by-word button press is unnatural and slows reading, distorting the normal eye-movement record.
  • Non-cumulative windows prevent rereading, so regressions and recovery — central to comprehension — are invisible.
  • Effects often appear one or more regions after the critical word (spillover), complicating localization.
  • Temporal resolution is coarser than eye-tracking and cannot separate early (lexical) from late (integrative) processing within a region.

Source conditions do not replace independent scientific review or validation by an analysis engine.

Strengths

  • Cheap and equipment-light — needs only a computer and a button, and runs easily online at scale.
  • Provides an online, region-localized index of processing difficulty as the sentence unfolds.
  • Simple to design and analyze, with a transparent dependent measure (per-region reading time).
  • Sensitive to a wide range of effects — garden paths, agreement attraction, surprisal, integration cost — replicating eye-tracking findings in many cases.

Applications

  • Studying garden-path sentences and the cost of syntactic reanalysis at the disambiguating region.
  • Measuring the processing of long-distance dependencies, filler-gap constructions, and agreement attraction.
  • Testing predictability and surprisal effects from language models against human reading times.
  • Investigating second-language and bilingual sentence processing relative to native baselines.
  • Comparing reading difficulty across populations (e.g., aging, aphasia, dyslexia) on controlled materials.

How it works

How it works

1

Each sentence is divided into presentation regions — single words in the strictest version, or short phrases. Critical and control conditions are built as minimal pairs differing only at a defined critical region (and often a following spillover region), with matched length and frequency elsewhere. A comprehension question follows many trials to ensure participants read for meaning rather than racing through, and fillers mask the manipulation.

2

Words appear in a moving window: at the start, dashes mask every character; each button press reveals the next region and re-masks the previous one (non-cumulative) or leaves it visible (cumulative). The reading time for a region is the interval between the press that reveals it and the press that reveals the next, \(\text{RT}_{\text{region}} = t_{\text{reveal next}} - t_{\text{reveal current}}\), so the participant's own pace yields a per-region latency.

Formula
\text{RT}_{\text{region}} = t_{\text{reveal next}} - t_{\text{reveal current}}
3

Reading times are aggregated by region and condition, and the key dependent measure is the reading-time difference between conditions at the critical and spillover regions, \(\Delta\text{RT} = \overline{\text{RT}}_{\text{critical}} - \overline{\text{RT}}_{\text{control}}\). A positive difference — longer reading in the harder condition — marks processing difficulty localized to that region; raw times are typically log-transformed or residualized against word length and position before comparison.

Formula
\Delta\text{RT} = \overline{\text{RT}}_{\text{critical}} - \overline{\text{RT}}_{\text{control}}
4

The region times are submitted to inferential analysis. Modern practice fits linear mixed-effects models with participants and items as crossed random effects, predicting (log) reading time from condition and controlling for length and position; comprehension-question accuracy, a categorical outcome, is analyzed with logistic mixed models. As Jaeger argues, such logit mixed models — rather than ANOVA on proportions — are the appropriate tool for the categorical accuracy data that accompany the continuous reading times.

Common pitfalls

  • Failing to define and analyze spillover regions, so a real effect that surfaces after the critical word is missed.
  • Comparing raw reading times without controlling for region length and serial position, conflating those with the manipulation.
  • Dropping or ignoring comprehension-question accuracy, leaving fast 'reading' that was really skimming undetected.
  • Analyzing accuracy with ANOVA on proportions rather than a logit mixed model, distorting inference on the categorical data.

Frequently asked questions

How does self-paced reading compare with eye-tracking during reading?

Both are online measures of reading, but eye-tracking records natural eye movements — including skipping, regressions, and rereading — at high temporal resolution and distinguishes early from late measures, whereas self-paced reading imposes a serial word-by-word button press that prevents looking back and slows reading unnaturally. Self-paced reading wins on cost, simplicity, and online scalability and often replicates eye-tracking effects, but eye-tracking gives a richer, more natural, and finer-grained picture. Many researchers use self-paced reading for breadth and eye-tracking for depth and naturalness.

What is spillover, and why does it matter?

Spillover is the tendency for a processing-difficulty effect to appear not on the critical word itself but on the following region or two, because comprehension of a word continues after the reader has moved on. In self-paced reading this is common, so analysts pre-register and analyze one or more post-critical (spillover) regions in addition to the critical region. Ignoring spillover can make a real effect look absent or mislocate it; analyzing it correctly attributes the slowdown to the right manipulation.

Should the window be cumulative or non-cumulative?

In the non-cumulative moving window — the most common form — each new word appears and the previous word is re-masked, so only one region is visible at a time; this isolates per-region times but forbids rereading. In a cumulative window, revealed words remain on screen. Non-cumulative presentation gives cleaner region-by-region reading times and is the standard for processing studies, while cumulative presentation is occasionally used when keeping prior context visible is desired, at the cost of contaminating later-region times with rereading.

Related methods

Sources

  1. Just, M. A., Carpenter, P. A., & Woolley, J. D. (1982). Paradigms and processes in reading comprehension. Journal of Experimental Psychology: General, 111(2), 228–238.

    Open DOI reference
    Copied.
  2. Jaeger, T. F. (2008). Categorical data analysis: Away from ANOVAs (transformation or not) and towards logit mixed models. Journal of Memory and Language, 59(4), 434–446.

    Open DOI reference
    Copied.