Missing Data Methods in Educational Testing

1,871 words · 8 min read · 4 references cited

Missing data is the rule, not the exception, in educational testing. Examinees skip items they do not know, run out of time on long tests, encounter technical glitches that drop responses, or have items they were never administered under multistage or adaptive designs. Whatever the cause, the resulting response matrix has holes, and what an analyst does about those holes affects every downstream ability estimate. The choice of method matters most where it is hardest to make: in sparse datasets, where the proportion missing is high or the response matrix is structurally incomplete by design.

Xiao and Bulut (2020) ran one of the more thorough Monte Carlo evaluations of how four common missing-data methods perform under realistic IRT calibration conditions. Their principal finding — full-information maximum likelihood (FIML) outperforms imputation-based alternatives across most realistic conditions — confirms what the methodological literature has been saying for two decades, but their secondary findings sharpen the practical guidance about when each method’s failure modes start to bite.

Why missingness is a substantive problem, not a bookkeeping one

Rubin’s (1976) taxonomy is the framework everything else builds on. Data are missing completely at random (MCAR) when the probability of missingness depends on neither observed nor unobserved values; missing at random (MAR) when it depends on observed values but not on the missing values themselves; and missing not at random (MNAR, sometimes NMAR) when it depends on the missing values themselves even after conditioning on what is observed. The taxonomy matters because the validity of every missing-data method depends on which mechanism is operating, and the mechanism is not directly testable from the data — it has to be reasoned about from substantive knowledge of how the missingness arose.

MCAR: missingness depends on neither observed norunobserved valuesMAR: missingness depends on observed values onlyMNAR: missingness depends on the missing valuesthemselves
Figure 1. Rubin's (1976) taxonomy classifies missingness by what the probability of a missing value depends on — the mechanism that determines which methods stay unbiased.

For ability estimation in IRT, the worst case is when low-ability examinees disproportionately omit hard items because they do not know the answers. The mechanism is MNAR: the probability of a missing response depends on the latent ability of the respondent, which is exactly the quantity being estimated. Naive treatments — scoring missing as wrong, or dropping respondents with any missing data — both bias the ability estimates downward for the affected examinees. Listwise deletion is rarely viable in real testing because it discards too many respondents; the question is what to do with the data that remain.

The four methods Xiao and Bulut compared

Full-information maximum likelihood (FIML) uses every available response and integrates over the unobserved ones inside the IRT likelihood. It does not impute missing values; it respects which values are observed and conditions the likelihood on the observed pattern. FIML is consistent under MCAR and MAR by construction, and its standard errors are correct to the asymptotic order. The cost is computational: with high missingness, the integral becomes high-dimensional and the optimizer has to work harder to converge.

Zero replacement treats every missing response as a wrong answer. It is the simplest method possible — a one-line patch to the data matrix before calibration runs as if no data were missing. Theoretically it should bias estimates downward whenever low-ability respondents are over-represented in the missing set, which is almost always. In practice, the bias is real but bounded; under specific patterns it produces ability estimates that are not far from the truth, and the result is computationally trivial.

MICE-CART and MICE-RFI are multivariate-imputation-by-chained-equations methods using classification and regression trees (CART) or random forest imputation (RFI) as the per-variable conditional model. MICE iteratively imputes each variable with missing values from a regression on the other variables, cycling through the variables until imputations stabilize. CART and RFI use machine-learning regressors as the conditional models, which is more flexible than parametric MICE alternatives but introduces the usual machine-learning vulnerabilities (overfitting on small samples, sensitivity to tuning). The mice R package by van Buuren and Groothuis-Oudshoorn (2011) is the canonical implementation and supports both backends.

What the simulation showed

Xiao and Bulut crossed missing-data mechanism (MCAR, MAR, MNAR), missing proportion (5%, 15%, 30%, 40%), test length (20, 40, 60 items), and sample size (500, 1,000, 3,000) — a fully crossed factorial design that covers most realistic operational scenarios. For each cell they generated 2PL response data, induced missingness according to the mechanism, applied each of the four methods, and recorded the root-mean-square error (RMSE) of the recovered ability parameters against ground truth.

The principal result: FIML produced the lowest RMSE across most conditions, regardless of mechanism. Even under MNAR — where FIML’s consistency is no longer guaranteed — it still outperformed the imputation alternatives in absolute RMSE terms, presumably because the bias from MNAR was smaller than the noise from imputation in the conditions tested.

Zero replacement was the worst method on average, with RMSE consistently higher than the alternatives, but it had a counterintuitive property: at high missingness proportions (40%), it became competitive. The reason is that imputation methods need enough observed data to fit a credible imputation model; when 40% of responses are missing, the imputation model is being fit on a thin substrate and produces noisy fills, while zero replacement at least delivers a deterministic answer. The crossover is not a recommendation to use zero replacement at high missingness — FIML still wins — but it explains why the simple method is hard to displace in some operational settings.

MICE-CART and MICE-RFI performed similarly to each other, with CART slightly ahead in most conditions but the differences small relative to the gap between either ML-based method and FIML. Both improved as test length increased (more observed items per respondent gives a richer imputation model) and as missingness decreased. Under MAR they were close to FIML; under MNAR they fell behind, as expected.

What this means for practice

The practical implication is straightforward: use FIML when it is available. Modern IRT software — Stan, PyMC, mirt in R, flexMIRT, IRTPRO — all support FIML natively for the standard 1PL/2PL/3PL models. The integration cost is modest in software that is already optimized for the missing-data pattern, and the asymptotic guarantees under MAR are real.

The exceptions are scenarios where FIML is structurally unavailable: ability estimation downstream of an unrelated software pipeline that does not expose the FIML option, or models with complex non-IRT components where the FIML integral would be intractable. In those cases, MICE with a flexible conditional model is the next-best option, with the proviso that imputation quality degrades when missingness is high or the test is short. Zero replacement should only be used as a transparent baseline against which other methods are compared, not as a recommended production method.

For sparse-by-design data — multistage testing, computerized adaptive testing where examinees see only a subset of items — the missingness is typically MAR by construction (the routing rules depend on observed responses, not unobserved abilities), and FIML is the standard treatment. The Xiao-Bulut findings extend cleanly to this case: FIML is consistent and efficient, and there is no reason to introduce imputation as an extra layer.

The MNAR caveat

The honest qualifier is that all four methods, including FIML, are biased under MNAR. Rubin’s (1976) original distinction between ignorable and non-ignorable missingness is still binding: when missingness depends on unobserved quantities even conditional on observed ones, no missing-data method can recover unbiased estimates without additional modeling assumptions about the missingness process itself. Pattern-mixture models, selection models, and shared-parameter models can each address MNAR but require the analyst to specify a non-identifiable component of the model — usually via sensitivity analysis across plausible specifications (Enders, 2010).

Low abilityOmit hard itemsMissing (MNAR)Estimate biaseddown
Figure 2. The worst case for ability estimation is MNAR: low-ability examinees omit the hard items they cannot answer, and naive treatments then bias their ability estimates downward.

For high-stakes testing where MNAR is plausible — adaptive tests where ability-driven omissions are common, or accommodations testing where systematic non-response is a feature of the population — sensitivity analysis is the responsible reporting standard. FIML or MICE results are presented as the primary finding, with secondary analyses showing how the conclusions move under alternative MNAR specifications. This is more work than running a single method and reporting the answer, but the cost of mis-reporting an ability estimate that depends on a wrong missingness assumption is paid by the examinees, not by the analyst.

Where this connects to broader psychometric methodology

Missing-data handling is one of several places where the IRT calibration workflow has methodological choices with substantive consequences. The choice of estimator (Bayesian hierarchical with ADVI vs MMLE vs JMLE), the assumed prior structure for item parameters, the treatment of differential item functioning, and the handling of item distributions for reliability estimation all interact with the missingness method. The central lesson from Xiao and Bulut (2020) — FIML is robust enough to be the default, with explicit alternatives for unusual conditions — generalizes: modern IRT estimation rewards making methodological choices explicitly, defending them in writing, and reporting sensitivity to alternatives that a sophisticated reader might prefer.

Frequently Asked Questions

Is FIML the same as multiple imputation?

No. FIML uses every observed value directly in the likelihood without filling in the missing ones. Multiple imputation generates several complete datasets by imputing missing values, fits the model to each, and pools the results. Both are valid under MAR; FIML is more efficient when applicable because it avoids the imputation step.

When is zero replacement defensible?

As a transparent baseline against which other methods are compared, or in operational settings where the test specification scores omitted items as wrong by definition (some criterion-referenced certification programs do this). Outside those cases, it biases ability estimates downward and the bias is hard to quantify without simulation.

Does the choice of imputation backend matter?

Less than the choice of FIML vs imputation in the first place. Xiao and Bulut (2020) found that MICE-CART and MICE-RFI produced similar RMSE; CART had a small edge. The bigger lever is using FIML when it is available, and reserving imputation for scenarios where it is not.

What if the missingness is MNAR?

No standard method is unbiased under MNAR without additional assumptions about the missingness process. The defensible workflow is sensitivity analysis: report the primary FIML or MICE result, then show how the conclusions move under explicit MNAR specifications. This is the methodological recommendation in Enders (2010) and the consensus practice in modern missing-data analysis.

How much missingness is too much?

There is no universal threshold, but Xiao and Bulut’s results suggest that imputation-based methods degrade noticeably above 30% missing, while FIML remains competitive up through 40%. Beyond 40% the parameter estimates become noisy regardless of method, and the question shifts from “which method to use” to “is the design adequate to support inference at all”.

References

  • Enders, C. K. (2010). Applied missing data analysis. Guilford.
  • Rubin, D. B. (1976). Inference and missing data. Biometrika, 63(3), 581–592. https://doi.org/10.1093/biomet/63.3.581
  • van Buuren, S., & Groothuis-Oudshoorn, K. (2011). mice: Multivariate imputation by chained equations in R. Journal of Statistical Software, 45(3), 1–67. https://doi.org/10.18637/jss.v045.i03
  • Xiao, J., & Bulut, O. (2020). Evaluating the performances of missing data handling methods in ability estimation from sparse data. Educational and Psychological Measurement, 80(5), 932–954. https://doi.org/10.1177/0013164420911136

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 →

Why is missingness a substantive problem, not just bookkeeping?

Rubin's (1976) taxonomy is the framework everything else builds on. Data are missing completely at random (MCAR) when the probability of missingness depends on neither observed nor unobserved values; missing at random (MAR) when it depends on observed values but not on the missing values themselves; and missing not at random (MNAR, sometimes NMAR) when it depends on the missing values themselves even after conditioning on what is observed. The taxonomy matters because the validity of every missing-data method depends on which mechanism is operating, and the mechanism is not directly testable from the data — it has to be reasoned about from substantive knowledge of how the missingness arose.

Why do the four methods Xiao and Bulut compared matter?

Full-information maximum likelihood (FIML) uses every available response and integrates over the unobserved ones inside the IRT likelihood. It does not impute missing values; it just respects which values are observed and conditions the likelihood on the observed pattern. FIML is consistent under MCAR and MAR by construction, and its standard errors are correct to the asymptotic order. The cost is computational: with very high missingness, the integral becomes high-dimensional and the optimizer has to work harder to converge.