Differential Item Functioning and Response Process

1,682 words · 7 min read · 3 references cited

A test item that scores differently for two groups of equally able examinees is called a differential item functioning (DIF) item, and identifying these items is now a routine part of large-scale assessment quality control. The hard part has never been the detection — statistical tests for DIF have been mature for thirty years — but the interpretation: knowing why an item flagged as DIF behaves the way it does. Expert content reviewers and statistical DIF flags often disagree, leaving test developers with a list of suspicious items and no clear story about what is driving the difference. A 2024 study by Li, Shin, Kuang, and Huggins-Manley shows that response process data — the digital traces of how examinees interact with computerized items — can fill in part of this missing layer.

What DIF is, and why detection is not enough

DIF occurs when examinees from two groups (commonly defined by gender, native language, or demographic background) who have the same level of the trait being measured nonetheless have different probabilities of getting an item right. The matching on ability is the key feature: a raw difference in pass rates is not DIF, because the groups may genuinely differ in ability. A conditional difference, after equating ability, is what flags the item.

Statistical flag: scoreoutcomesContent expert: itemtextWhy the item shows DIF
Figure 1. Standard DIF methods flag a suspect item but cannot say why: the statistical flag sees only score outcomes, the content expert only the item text.

Standard DIF methods — Mantel-Haenszel, logistic regression, IRT-based approaches — are good at producing this flag. What they cannot do is explain it. A flagged item might function differently because of:

  • genuinely construct-irrelevant content that one group encounters more readily;
  • differences in test-taking strategy, pacing, or familiarity with the response format;
  • statistical noise plus multiple testing across hundreds of items;
  • group differences on a secondary, correlated trait that the item happens to tap.

Expert reviewers — content specialists asked to read flagged items and judge whether the difference is meaningful — often produce low agreement with statistical flags and with one another. The result is a chronic gap between “this item is statistically suspicious” and “this is what the suspicion means.”

What response process data adds

Computerized assessments such as the Programme for International Assessment of Adult Competencies (PIAAC) generate detailed log files: time stamps for every action, sequences of clicks and keystrokes, time on each item, time per response option, and patterns of revisiting earlier items. Goldhammer, Hahnel, and Kroehne’s 2020 methodological treatment of PIAAC log files maps out how these traces can be transformed into structured features for analysis — moving from raw event streams to interpretable variables like total response time, dwell time on specific options, and the frequency of particular action sequences.

Unlike a simple right/wrong score, these features describe the process by which an examinee arrived at an answer. If two groups solve an item correctly through different routes, or if one group spends systematically more time on a specific tool or distractor, those patterns carry information that a final-score-only DIF analysis cannot see.

What Li and colleagues found

The 2024 Li et al. study used the PIAAC 2012 computer-based numeracy assessment to investigate gender DIF. Their analytic approach combined:

  • random forest models — to capture potentially non-linear and interactive effects of process features on DIF;
  • logistic regression with ridge regularization — to handle high-dimensional feature sets and produce more stable coefficient estimates;
  • variation in the assumed proportion of DIF items — to test how the methodology behaves under different prevalence assumptions, since real test forms vary.

Their core empirical claim is that the combination of timing features and action-sequence features is informative for distinguishing how the groups respond to flagged items. Neither timing nor sequence alone carried as much information as the combination. The most useful features were not pre-specified: the modeling approach surfaced them empirically from the log file data.

The interpretive payoff is that, for a given DIF-flagged item, the analyst can now describe not only that it functions differently but also which aspects of the response process differ between groups — for example, whether one group takes systematically longer, dwells on different distractors, or follows a different sequence of within-item actions. This is the missing bridge between statistical flag and substantive explanation.

Why low statistical-versus-expert-review agreement makes more sense in this light

The chronic disagreement between statistical DIF flags and expert content review can be partially explained by what each method has access to. The statistical flag knows the score outcomes; the content expert knows the item text. Neither has direct access to how examinees solved the problem. If the source of differential functioning lies in the process — e.g., one group skipping over a built-in calculator tool that helps with the item — neither the score nor the item text alone will reveal it.

Process data closes this gap by making the unobserved middle layer observable. The implication, consistent with Li et al.’s framing, is that some historic disagreements are not failures of either method but reflect a genuine information gap that process data can fill.

The ecological context: DIF is not only about items

Woitschach, Zumbo, and Fernández-Alonso (2019) make a complementary point at a different level of analysis. Treating DIF as purely an item-level property, they argue, ignores the multilevel structure of educational assessment data: students are nested in classrooms, schools, regions, and educational systems, and group-level context can drive differential functioning that no item-only analysis will resolve. Their multilevel modeling approach treats DIF as partly an emergent property of the testing context, not a pure item flaw.

Statistical DIF detection: flag items with differingconditional pass ratesExpert review: judge item content forconstruct-irrelevant featuresResponse process analysis: examine how groups engagewith itemsMultilevel ecological analysis: model howgroup-level context shapes patterns
Figure 2. A complete account of why an item shows DIF increasingly draws on four layers, from statistical detection to multilevel ecological analysis.

Combined with the process-data approach, the picture is layered:

  • Statistical DIF detection identifies items where conditional pass rates differ across groups.
  • Expert review evaluates item content for plausible construct-irrelevant features.
  • Response process analysis (Li et al.) examines how groups engage with items.
  • Multilevel ecological analysis (Woitschach et al.) examines how group-level context shapes the patterns observed.

A complete account of why a particular item shows DIF will increasingly draw on more than one of these layers.

What this means for assessment practice

For test developers and large-scale assessment programs, several practical implications follow:

  • Process data should be retained and analyzed, not discarded. Many computerized assessments generate detailed log files that are then archived without systematic analysis. The Li et al. results suggest meaningful DIF interpretation information is sitting in those files.
  • Combined feature sets matter. Single-feature analyses (timing alone, or sequence alone) appear to leave information on the table.
  • The technique does not eliminate the need for expert review. Process features describe what differs; substantive judgment is still required to interpret whether the difference reflects construct-relevant variation or construct-irrelevant noise.
  • The method is not yet operational. Random-forest-and-ridge-regression analysis of log file features is research-grade rather than routine. Operationalization for production assessment programs would require methodological standardization that does not yet exist.

Limits of the current evidence

The Li et al. study is, at the time of writing, a single empirical investigation on one assessment domain (PIAAC numeracy, 2012 cycle) and one DIF dimension (gender). Several open questions remain:

  • Generalization across domains. Whether timing-and-sequence features carry similar information for literacy, problem-solving, or domain-specific tests is not yet established.
  • Generalization across DIF dimensions. Gender DIF may have different process correlates than language-based or age-based DIF.
  • Stability of findings. The features identified as most informative by random forest are not guaranteed to replicate. Different cycles or samples could highlight different features.
  • Causal interpretation. Process features describe correlations between behavior and DIF flags. They do not, on their own, establish that the process difference causes the score difference, only that it accompanies it.

Frequently Asked Questions

What is differential item functioning in plain language?

A test item that two equally able people, from different demographic groups, are not equally likely to get right. The “equally able” qualifier is what distinguishes DIF from a raw group difference in scores.

Why is interpreting DIF harder than detecting it?

Statistical detection produces a flag, but the flag does not indicate why the item is functioning differently. Reasons can range from substantive content bias to test-taking-strategy differences to statistical noise from multiple testing. Distinguishing these requires information that pure score data does not contain.

What is response process data?

The detailed digital trace of how an examinee interacts with an item: time stamps, click sequences, dwell times on options, use of in-item tools, and revisits. Computer-based assessments like PIAAC routinely produce this data even when it is not part of the scoring.

Does this mean DIF analysis should now always include process data?

Not yet operationally. The methods are research-grade and require analytical infrastructure that most assessment programs do not have in production. The direction of travel is toward incorporation, but the field is not there.

Could process-data DIF analysis introduce new biases?

Yes, in principle. Process features themselves can be confounded with examinee characteristics like familiarity with the response interface or testing motivation. Treating process data as an unbiased window onto cognition would be a mistake; it is one more layer of evidence, not the final answer.

Is gender DIF on numeracy items always meaningful?

No. Some flagged items reflect minor measurement noise; others reflect substantive content issues; others reflect process differences that may or may not be construct-relevant. The whole point of layered interpretation is that “flagged for DIF” is the start of an analysis, not the conclusion.

References

  • Li, Z., Shin, J., Kuang, H., & Huggins-Manley, A. C. (2024). Exploring the Evidence to Interpret Differential Item Functioning via Response Process Data. Educational and Psychological Measurement, 85(4), 783–813. https://doi.org/10.1177/00131644241298975
  • Goldhammer, F., Hahnel, C., & Kroehne, U. (2020). Analysing Log File Data from PIAAC. In D. B. Maehler & B. Rammstedt (Eds.), Large-Scale Cognitive Assessment (pp. 239–269). Springer. https://doi.org/10.1007/978-3-030-47515-4_10
  • Woitschach, P., Zumbo, B. D., & Fernández-Alonso, R. (2019). An ecological view of measurement: focus on multilevel model explanation of differential item functioning. Psicothema, 31(2), 194–203. https://doi.org/10.7334/psicothema2018.303

Related questions

What is item response theory?

Every time you take a standardized test — an IQ assessment, a college entrance exam, a professional certification — the questions have been calibrated using sophisticated statistical models that most test-takers never learn about. Item Response Theory (IRT) is the mathematical framework behind virtually all modern psychological and educational testing, and understanding its basics illuminates why tests work the way they do. Read more →

What is attenuation-corrected reliability?

Most psychometrics textbooks teach the classical "correction for attenuation" — Spearman's century-old technique for estimating what the correlation between two psychological constructs would be if the tests measuring them were perfectly reliable. The technique is simple: divide the observed correlation by the square root of the product of the two reliabilities. The technique is also limited: it adjusts the relationship between two scales, but assumes the reliability values plugged into the denominator are themselves accurate. A 2022 paper by Jari Metsämuuronen in Applied Psychological Measurement argues that this assumption is broken in practice. Reliability estimates produced by Cronbach's alpha and similar formulas are themselves attenuated by the same mechanical errors that attenuate correlations — and in some datasets, alpha may be deflated by 0.40–0.60 units of reliability. Metsämuuronen's contribution is a class of deflation-corrected reliability estimators that apply the classical attenuation logic inside the reliability formula rather than only to correlations between scales. Read more →

How are JCCES General Knowledge items structured?

The General Knowledge (GK) subtest of the Jouve Cerebrals Crystallized Educational Scale (JCCES) measures factual breadth — the accumulated stock of information about the world that crystallized intelligence theory treats as a core component of acquired cognitive ability. A subtest of this kind has to satisfy two structural requirements: items should span a meaningful range of difficulty, and they should order along a single underlying continuum of factual breadth rather than tapping multiple unrelated dimensions. This study examined the item structure of the JCCES GK subtest using multidimensional scaling (MDS) on response data from 588 respondents, and recovered the empirical signature of a well-ordered unidimensional construct: a horseshoe-shaped scaling pattern that is the canonical evidence of a Guttman simplex — items ordered cleanly along a single difficulty continuum. Read more →

How well does coefficient alpha perform with non-normal data?

Cronbach's coefficient alpha is the most-reported reliability statistic in psychology and educational measurement. It is also one of the most-misunderstood. The classical formula assumes that test items measure a single construct with equal factor loadings (tau-equivalence), uncorrelated errors, and continuously distributed scores. Real psychological measurement rarely meets all three assumptions: most scales use Likert responses (discrete), have items with unequal contributions to the construct (congeneric), and produce score distributions that depart from normality. The natural question is how badly alpha breaks under these violations and which alternatives perform better. A 2023 simulation study by Xiao and Hau in Educational and Psychological Measurement provides a systematic answer, with implications for the routine reliability reporting that fills psychometric methods sections. Read more →

What is DIF, and why isn't detecting it enough?

DIF occurs when examinees from two groups (commonly defined by gender, native language, or demographic background) who have the same level of the trait being measured nonetheless have different probabilities of getting an item right. The matching on ability is the key feature: a raw difference in pass rates is not DIF, because the groups may genuinely differ in ability. A conditional difference, after equating ability, is what flags the item.

What does response process data add?

Computerized assessments such as the Programme for International Assessment of Adult Competencies (PIAAC) generate detailed log files: time stamps for every action, sequences of clicks and keystrokes, time on each item, time per response option, and patterns of revisiting earlier items. Goldhammer, Hahnel, and Kroehne's 2020 methodological treatment of PIAAC log files maps out how these traces can be transformed into structured features for analysis — moving from raw event streams to interpretable variables like total response time, dwell time on specific options, and the frequency of particular action sequences.