When is listwise deletion acceptable in medical research?

Listwise deletion (complete case analysis) is acceptable when: (1) the data are MCAR — missingness is unrelated to any variable in the study; (2) fewer than 5% of values are missing; and (3) the analysis has sufficient statistical power even after the reduced sample. When any of these conditions are not met, listwise deletion produces biased estimates (in MAR and MNAR settings) and may substantially reduce statistical power. Multiple imputation is preferred in most medical research settings where the amount or pattern of missing data could plausibly introduce bias.

How many imputations should I use in multiple imputation?

The older rule of 5–20 imputations is now considered outdated. Modern guidance (White et al., 2011) recommends that the number of imputations should be at least equal to the percentage of incomplete cases. For a dataset with 20% missing cases, use at least 20 imputations. For datasets with up to 50% incomplete cases, use 50–100 imputations. More imputations increase reproducibility and reduce Monte Carlo error in the pooled estimates. Software such as R mice, SPSS Multiple Imputation, and Stata mi impute all easily accommodate 50–100 imputations.

What is MICE and how is it different from other imputation methods?

MICE (Multiple Imputation by Chained Equations), also called Fully Conditional Specification (FCS), is the most widely used algorithm for multiple imputation in medical research. Unlike joint modelling (which requires specifying a single multivariate distribution for all variables), MICE imputes each variable with missing data sequentially using a regression model conditioned on all other variables. For a continuous variable, it uses linear regression; for binary variables, logistic regression; for ordered categories, ordinal logistic regression; for counts, Poisson regression. This flexibility makes MICE suitable for mixed datasets with continuous, binary, ordinal, and count variables — exactly the type of data found in most clinical datasets.

Is mean imputation acceptable for medical data?

Mean imputation (replacing missing values with the variable's observed mean) is generally not acceptable in medical research for analytic purposes. It produces four systematic problems: it artificially reduces the variance of the imputed variable (biasing standard deviations and confidence intervals); it distorts correlation structures between variables (biasing regression coefficients); it does not account for uncertainty from the imputed values (producing standard errors that are too small and p-values that are too liberal); and it assumes MCAR, which is rarely defensible. Mean imputation may be used for descriptive tables only, clearly labelled, but never for the primary analysis.

What should I report about missing data in a clinical trial or thesis?

CONSORT and STROBE guidelines require reporting: (1) the number and percentage of missing observations for each key variable; (2) the pattern and potential reasons for missingness; (3) whether MCAR was tested (Little's test) and the result; (4) the missing data method used (listwise deletion, multiple imputation, FIML) and the justification; (5) for multiple imputation: the imputation model specification, number of imputations, software used; (6) a sensitivity analysis comparing results under different missing data assumptions. A flow diagram showing excluded participants and missing outcome data is required by CONSORT for clinical trials.

What variables should I include in the multiple imputation model?

The imputation model should include: (1) all variables that will be in the analysis model (outcome, exposures, confounders, covariates); (2) variables associated with the probability of being missing (auxiliary variables); (3) variables associated with the missing variable's values even after adjusting for the analysis model variables. Including the outcome variable in the imputation model is critical — omitting it biases imputed values toward the null and underestimates effect sizes. As a general rule, include more variables rather than fewer; a rich imputation model produces better imputed values than a sparse one, as long as the model remains estimable.

How to Handle Missing Data in Medical Research: Complete Statistical Guide

Q: What is the difference between MCAR, MAR, and MNAR?

MCAR (Missing Completely at Random) means the probability of a value being missing is unrelated to both the observed data and the missing value itself — for example, a blood sample dropped in the lab. MAR (Missing at Random) means missingness is related to observed variables in the dataset, but not to the unobserved missing value itself — for example, older patients being less likely to complete a questionnaire, where age is recorded. MNAR (Missing Not at Random) means the probability of missing is related to the missing value itself — for example, patients with worse outcomes dropping out because their condition deteriorated. MCAR is the safest assumption; MAR allows multiple imputation; MNAR requires sensitivity analysis and cannot be verified from the data alone.

Q: What is multiple imputation and how does it work?

Multiple imputation (MI) is a principled statistical method for handling missing data that creates m complete datasets by replacing each missing value with m plausible values drawn from the posterior predictive distribution. Each imputed dataset is then analysed separately using the intended analysis, and the results are combined using Rubin's Rules: the pooled estimate is the mean of the m estimates, and the pooled standard error accounts for both within-imputation variance (typical estimation error) and between-imputation variance (additional uncertainty due to missing data). This correctly propagates the uncertainty from imputation into the final estimates and standard errors, unlike single-value imputation methods.

Q: What is Little's MCAR test?

Little's MCAR test (Little, 1988) is a formal statistical test of whether data are missing completely at random. It compares the observed means of variables for different patterns of missingness. A non-significant result (p ≥ 0.05) supports the MCAR assumption. A significant result (p < 0.05) indicates that the data are NOT MCAR — they may be MAR or MNAR. Little's test is available in SPSS (Analyze → Missing Value Analysis → Little's MCAR test) and in R (mcar_test() from the naniar package). A non-significant Little's test does not prove MCAR definitively, but a significant result definitively rules it out and implies that multiple imputation or maximum likelihood estimation is needed.

Q: How much missing data is too much?

There is no universal threshold, but pragmatic guidance: less than 5% missing is rarely problematic with any method; 5–15% warrants multiple imputation; 15–25% requires multiple imputation with sensitivity analysis for MNAR; more than 25–40% makes valid inference increasingly difficult even with MI, and requires exploring the reasons for missingness and considering whether the variable or outcome can be used at all. When a primary outcome is missing in more than 20–25% of a trial's participants, the study's conclusions should be treated with substantial caution regardless of the imputation method used.

A multicentre hypertension trial enrolls 400 patients and randomises them to a new antihypertensive versus placebo. At the 12-week primary endpoint, 68 patients (17%) have missing blood pressure measurements — some withdrew due to side effects, others were lost to follow-up, and a few simply did not attend the final clinic visit. The trialist opens SPSS, sees the missing values highlighted in yellow, and hits "Analyze → Regression" without a second thought. SPSS silently drops all 68 patients from the analysis. The trial now has 332 participants, reduced power, and — more critically — a potentially biased treatment effect estimate if the patients who dropped out did so for reasons related to treatment response. Missing data is one of the most consequential and least-taught problems in clinical research. Understanding why data is missing, and choosing the correct method to handle it, is not a technical detail — it is the difference between valid and invalid inference.

Why Missing Data Is a Statistical and Clinical Problem

Missing data in medical research is nearly universal. A survey of 519 clinical trials published in major journals found that 95% reported some missing data, yet only 45% described the method used to handle it, and fewer than 20% reported the likely mechanism driving the missingness. This documentation gap is not merely cosmetic — the choice of missing data method can change the magnitude, direction, and statistical significance of a study's primary finding.

Missing data causes two distinct statistical harms:

Reduced statistical power: Every missing observation reduces the effective sample size. A study powered at n = 300 with 20% missing data has the effective power of an n = 240 study — which may be insufficient to detect the expected effect at 80% power.
Biased estimates: When missingness is related to the outcome or the exposure — the most common situation in clinical research — excluding missing cases produces estimates that systematically over- or underestimate the true effect. This bias cannot be corrected by a larger sample size.

The statistical community's response to these twin harms has evolved substantially since the 1970s. Modern best practice, now embedded in CONSORT (for trials) and STROBE (for observational studies) reporting guidelines, requires researchers to: (1) describe the pattern and probable mechanism of missingness; (2) choose a handling method appropriate to that mechanism; and (3) conduct a sensitivity analysis to test whether conclusions are robust to different missing data assumptions.

The Three Missing Data Mechanisms

Donald Rubin's 1976 taxonomy of missing data mechanisms remains the organising framework for all modern missing data methodology. The mechanism determines which handling methods are valid — making it the single most important question to address before choosing a statistical approach.

MCAR — Missing Completely at Random

The probability of a value being missing is unrelated to both the observed data and the unobserved missing value. Missingness is due to pure chance — a blood sample lost in the laboratory, a random data entry failure, a questionnaire page accidentally skipped.

Test: Little's MCAR test (p ≥ 0.05 supports MCAR)

Safest assumption — any method valid

MAR — Missing at Random

The probability of a value being missing depends on observed variables in the dataset, but NOT on the unobserved missing value itself. Older patients less likely to complete a questionnaire (age is observed); men less likely to report pain scores (sex is observed).

Test: Cannot be formally tested; supported by clinical reasoning about observed predictors of missingness

Supports multiple imputation

MNAR — Missing Not at Random

The probability of a value being missing depends on the unobserved value itself. Severely depressed patients not returning the depression questionnaire; patients with worst outcomes withdrawing from a trial. The most clinically common and statistically dangerous mechanism.

Test: Cannot be verified from observed data; requires sensitivity analysis and clinical judgement

Requires sensitivity analysis

The MNAR trap in clinical trials: Dropout in clinical trials is frequently MNAR. Patients who experience adverse effects (and thus have worse outcomes) drop out at higher rates. Patients whose disease improves dramatically may also drop out (feeling they no longer need treatment). Both mechanisms cause the missing outcome data to be systematically different from observed data — in opposite directions — and no imputation method can fully correct this without untestable assumptions.

Diagnosing the Missing Data Mechanism

Several approaches help characterise the likely mechanism in practice:

Little's MCAR test: A formal χ² test of the null hypothesis that data are MCAR. In SPSS: Analyze → Missing Value Analysis → EM (check "Little's MCAR test"). A non-significant result (p ≥ 0.05) supports MCAR; significant result (p < 0.05) rules it out.
Logistic regression for missingness: Create a binary indicator variable (1 = missing, 0 = observed) for each variable with missing data, then regress it on all other observed variables. Any significant predictor indicates the data are not MCAR — the significant predictors are the observed variables driving missingness (consistent with MAR).
Missing value pattern analysis: SPSS Missing Value Analysis provides a pattern matrix showing which variables are missing together, helping identify systematic patterns that clinical context can explain.
Clinical knowledge: The most underused diagnostic tool. Why do patients drop out of this type of study? If the answer is "because their condition was deteriorating" or "because they couldn't tolerate the treatment," the data are very likely MNAR.

Methods for Handling Missing Data

The landscape of missing data methods spans from dangerously naive single-value approaches to rigorous multiple-imputation frameworks. Understanding their assumptions and failure modes guides appropriate method selection.

Method	Mechanism required	Bias risk	Power	Recommended for
Complete Case Analysis (Listwise Deletion)	MCAR only	High (if MAR/MNAR)	Low (reduced n)	<5% missing, confirmed MCAR
Mean / Median Imputation	MCAR	High always	Moderate (n preserved)	Not recommended for analysis
Last Observation Carried Forward (LOCF)	Stable construct assumed	High (biased)	Moderate	Largely deprecated; regulators no longer accept as primary analysis
Single Imputation (regression)	MAR	Moderate	Moderate	Underestimates SE; not for primary analysis
Multiple Imputation (MI / MICE)	MAR (valid; some robustness under MNAR)	Low under MAR	Good (n preserved, uncertainty propagated)	Gold standard for 5–40% missing
Full Information Maximum Likelihood (FIML)	MAR	Low under MAR	Optimal	SEM, latent variable models; requires specialised software
Pattern Mixture Models / Selection Models	MNAR models	Low if model correct	Good	Sensitivity analysis for suspected MNAR; requires expert statistician

How Much Missing Data Is Too Much?

< 5%

Minimal impact

Any method acceptable

5–15%

Moderate concern

Multiple imputation recommended

15–25%

Substantial problem

MI essential + MNAR sensitivity analysis

25–40%

Severe problem

MI + pattern mixture models + consider data usability

> 40%

Critical — question validity

Specialist statistician required; conclusions severely limited

Multiple Imputation: The Current Gold Standard

Multiple imputation (MI) was formalised by Donald Rubin in 1987 and is now the recommended approach for handling missing data in medical research when the data are MCAR or MAR and the proportion of missing data exceeds 5%. It solves the two fundamental failures of single-value imputation: it preserves the variance structure of the data and correctly propagates the uncertainty from imputation into the final standard errors and confidence intervals.

Impute

Create m complete datasets by replacing each missing value with m plausible values drawn from the posterior predictive distribution. MICE imputes each variable using a separate regression model conditioned on all others.

Analyse

Run the intended analysis on each of the m imputed datasets separately, obtaining m sets of parameter estimates (β) and standard errors (SE). Each analysis uses the full n without any cases excluded.

Pool

Combine the m estimates using Rubin's Rules. The pooled estimate is the mean of m estimates; the pooled SE accounts for within-imputation variance plus between-imputation variance (reflecting uncertainty from missing data).

Rubin's Rules: Combining Results from Multiple Imputed Datasets

β̂ = (1/m) × ∑ β̂ᵢ

Pooled point estimate: the mean of estimates across m imputed datasets

V̂ᵗᵒᵗᵚᵩ = V̄ᵮ + (1 + 1/m) × B

Pooled variance: within-imputation variance V̄ᵮ plus scaled between-imputation variance B

m = number of imputed datasets (minimum = % incomplete cases; typically 20–100)

β̂ᵢ = estimate from the i-th imputed dataset

V̄ᵮ = mean of within-imputation variances (average of SE² across m analyses)

B = between-imputation variance = variance of the m estimates (captures imputation uncertainty)

MICE: Multiple Imputation by Chained Equations

MICE (also called Fully Conditional Specification) is the algorithm underlying most modern MI software. Rather than specifying a single joint distribution for all variables simultaneously (which requires a restrictive multivariate normal assumption), MICE imputes each incomplete variable sequentially using its own tailored regression model:

Continuous variables: Linear regression imputation
Binary variables: Logistic regression imputation
Ordered categorical: Ordinal logistic regression
Unordered categorical: Multinomial logistic regression
Count variables: Poisson regression

MICE cycles through these models iteratively (typically 10–20 cycles) until the imputed values stabilise. This flexibility makes MICE directly applicable to the mixed-type clinical datasets typical in medical research — combining continuous lab values, binary outcomes, ordinal symptoms scores, and categorical demographic variables in a single imputation procedure.

What to include in the imputation model: Include all variables in your analysis model (outcome, exposures, confounders), all variables associated with the probability of missingness, and all variables associated with the missing variable's values. Critically, include the outcome variable in the imputation model even when imputing covariates — omitting the outcome biases regression coefficients toward zero. Auxiliary variables (strongly associated with missingness or the missing values but not in the analysis model) improve imputation quality without affecting the analysis.

How Many Imputations?

The traditional guidance of m = 5–20 imputations is now considered inadequate. White et al. (2011) demonstrated that the number of imputations should equal at least the percentage of incomplete cases to achieve good reproducibility of results. With 30% of cases incomplete, use at least 30 imputations. Modern computing makes 50–100 imputations computationally trivial in all common software packages and is the current recommended practice for medical journal submission.

Clinical Examples

Hypertension RCT: Comparing Listwise Deletion vs Multiple Imputation

A 12-week parallel-group RCT compares a new antihypertensive (n = 200) vs placebo (n = 200). Primary outcome is change in systolic BP from baseline to week 12. Seventy-two patients (18%) are missing the week-12 BP measurement: 46 (23%) in the treatment group and 26 (13%) in the placebo group. The dropout pattern suggests patients with worse side effects (higher BP increase on treatment, or persistent BP elevations on placebo) are more likely to drop out — consistent with MNAR, though MAR models including baseline BP, age, and comorbidities are also plausible.

400

Randomised (n)

72 (18%)

Missing week-12 BP

23%

Missing in treatment arm

13%

Missing in placebo arm

Little's MCAR Test χ²(df=14) = 28.4, p = 0.013 → Data are NOT MCAR Logistic regression for missingness: baseline BP (OR=1.08, p=0.002) and treatment arm (OR=2.01, p=0.018) both predict missingness Conclusion: MAR assumption plausible (missingness explained by observed variables)

Analysis Method	n used	BP difference (mmHg)	95% CI	p-value
Listwise deletion	328	−8.4	−11.1 to −5.7	0.001
Mean imputation	400	−6.1	−8.2 to −4.0	0.001 (SE too small)
LOCF	400	−5.2	−7.4 to −3.0	0.001 (biased)
Multiple imputation (m=50)	400	−7.1	−10.0 to −4.2	0.001 (wider CI ✓)

MI produces the most honest estimate: the treatment effect is −7.1 mmHg with appropriately wider CI than listwise deletion (which excludes sicker patients, overestimating the effect).

Key learning: Listwise deletion overestimates the treatment effect (−8.4 mmHg) because it excludes patients who dropped out due to worse outcomes — exactly the patients who might have diluted the treatment effect. The 95% CI from listwise deletion is also narrower than the MI result because it doesn't account for the uncertainty introduced by missing data. Mean imputation artificially narrows the CI by borrowing information from the group mean without propagating imputation uncertainty. The MI result (−7.1, 95% CI −10.0 to −4.2) is the most defensible primary analysis under the MAR assumption; a MNAR sensitivity analysis using pattern mixture models would complete the CONSORT-required missing data reporting.

Hospital Registry: Missing Laboratory Values in an Observational Study

A retrospective cohort study uses an electronic health record database to examine the association between admission HbA1c and 30-day readmission in 1,850 hospitalised diabetic patients. HbA1c is missing in 412 patients (22.3%). The research team needs to determine whether to exclude these patients or impute before logistic regression analysis of readmission risk.

1,850

Total cohort (n)

412 (22.3%)

Missing HbA1c

p < 0.001

Little's MCAR test

MAR

Likely mechanism

Missing Data Pattern Analysis HbA1c more likely to be missing in: — Emergency admissions (OR=2.4, p<0.001) vs elective — Patients with shorter length of stay (OR=1.8 per day less, p=0.004) — Younger patients without documented diabetes follow-up (OR=1.6, p=0.01) Conclusion: missingness explained by observed variables → MAR plausible MICE Imputation Model (m = 25 imputations) Variables included: age, sex, admission type, LOS, comorbidity index, admission glucose, BMI, prior hospitalisations, 30-day readmission (outcome) MICE: HbA1c imputed with linear regression conditioned on all above Results Comparison Complete case (n=1,438): OR for HbA1c ≥9% vs <7%: 1.82 (95% CI 1.31–2.54) MI (n=1,850, m=25): OR for HbA1c ≥9% vs <7%: 1.68 (95% CI 1.24–2.28) Difference in OR: 0.14 (8% relative reduction in estimated effect after MI)

MI effect estimate is attenuated vs complete case (OR 1.68 vs 1.82) — consistent with sicker patients having missing HbA1c and overestimation in complete-case analysis

Why registry data is particularly vulnerable to missing data bias: Electronic health records are not designed for research — they are designed for clinical care. Lab values are ordered when clinicians deem them necessary, which is systematically related to patient severity and clinical pathway. Emergency admissions with rapid deterioration may have no HbA1c recorded simply because clinical priorities precluded it — making the missingness directly related to the clinical state (severity) that also predicts readmission. Including admission type and LOS in the imputation model allows MICE to impute HbA1c using information that captures part of this clinical severity signal, making the MAR assumption more defensible than naive complete-case analysis.

Questionnaire Study: Item Non-Response in a Depression Survey

A cross-sectional study administers the PHQ-9 (9-item depression screener) to 650 outpatients. Individual item non-response is 3–8% per item, but 94 participants (14.5%) have missing scores on at least one PHQ-9 item, making a total score impossible by simple sum. The researcher must decide whether to use mean substitution, prorate (scale up the available items), or impute before computing total scores.

650

Total sample (n)

94 (14.5%)

Incomplete PHQ-9

3–8%

Per-item missing rate

Item 9

Highest missing (8.1%) — suicidality item

Why Item 9 (Suicidality) Has the Highest Missing Rate PHQ-9 Item 9: "Thoughts that you would be better off dead..." Missing more often in: younger patients (OR=1.9), self-administered setting (OR=2.2) Clinical interpretation: selective non-response on sensitive item → likely MNAR Impact: excluding item 9 missers underestimates depression severity in study Three Approaches Compared Approach 1 — Prorating: PHQ-9 total = (sum of available items × 9 / items answered) Problem: assumes all missing items would score same as non-missing items Effect: underestimates scores when sensitive items selectively skipped Approach 2 — Mean imputation per item: replace with item mean Problem: underestimates variance; ignores within-person response patterns Approach 3 — MICE (recommended): impute each missing item conditioned on all other PHQ-9 items, plus age, sex, visit type, and other symptoms Accounts for: the higher-depression patients more likely to skip item 9 Result: PHQ-9 mean increases from 8.4 (complete case) to 9.1 (MICE) ≥10 prevalence: 31.2% (complete case) vs 34.8% (MICE)

MICE imputation reveals 34.8% clinical depression prevalence vs 31.2% by complete case — a 3.6 percentage-point difference driven by MNAR selective non-response on the suicidality item

The MNAR sensitivity analysis: Because item 9 non-response is likely MNAR (those most affected by suicidal ideation skip this question), the MICE estimate under MAR is itself potentially biased downward. A delta-adjustment sensitivity analysis imputes item 9 assuming the missing values would score 0.5 or 1.0 points higher than the MAR-imputed values and re-estimates the prevalence. If the clinical conclusion (significant association between depression scores and the study outcome) is robust across these sensitivity scenarios, confidence in the findings increases substantially. Always report MNAR sensitivity analyses for sensitive questionnaire items with selective non-response.

Thesis Writing Recommendations

Missing data handling deserves explicit treatment in the Methods section of any thesis chapter involving clinical data. Supervisors, examiners, and journal reviewers are increasingly familiar with missing data methodology — an undocumented or naive approach is a predictable viva question and peer review critique.

Methods Section

Document the following in sequence: (1) the number and percentage of missing observations for each key variable presented in a missing data summary table; (2) whether Little's MCAR test was conducted and the result; (3) a description of the likely mechanism (MCAR/MAR/MNAR) with clinical justification; (4) the chosen handling method and the justification; (5) for MI: the imputation model specification (all variables included), number of imputations, number of MICE cycles, software and version used; (6) a statement that a sensitivity analysis comparing MI results to complete-case results will be (or was) conducted.

Results Section

Model Reporting Paragraph — Missing Data (Multiple Imputation)

"Missing data analysis was conducted prior to primary analysis. HbA1c was missing for 412 participants (22.3%). Little's MCAR test was significant [χ²(df=8) = 31.4, p < 0.001], indicating data were not missing completely at random. Logistic regression of missingness on observed covariates identified emergency admission type (OR 2.4, 95% CI 1.8–3.1) and shorter length of stay (OR 1.8 per day, 95% CI 1.4–2.3) as significant predictors, consistent with a missing at random mechanism. Multiple imputation using MICE was conducted in IBM SPSS Statistics version 28, using 25 imputations and 10 iterations per imputation. The imputation model included all analysis variables and two auxiliary variables (prior hospitalisation count and admission glucose). Results were pooled using Rubin's Rules. A complete-case sensitivity analysis was pre-specified and is reported alongside the primary MI-based analysis."

Model Reporting Paragraph — CONSORT Flow (Clinical Trial)

"Of 400 randomised participants, 328 (82%) completed the week-12 primary outcome assessment; 72 (18%) had missing data (46 treatment, 26 placebo). Reasons for missing data included: withdrawal due to adverse events (n = 31, 43%), loss to follow-up (n = 24, 33%), and protocol deviation (n = 17, 24%). The rate of missing outcome data was higher in the treatment arm (23%) than the placebo arm (13%), a differential pattern consistent with treatment-related adverse events as a driver of dropout (MNAR possible). Primary analysis used multiple imputation (m = 50, MICE, including treatment allocation, baseline BP, age, sex, comorbidity index, and all adverse event indicators in the imputation model). Results from complete-case analysis are presented as a secondary analysis."

Common Mistakes Researchers Make

Mistake 1: Letting Software Handle Missing Data Silently

The most pervasive mistake. SPSS, R, and Stata all perform listwise deletion by default when missing values are present in the analysis variables. The researcher clicks "Run" without noticing that SPSS has quietly dropped 20% of the sample. The output shows no warning, and the reduced n often goes unreported. The entire analysis is conducted on a non-representative subsample.

Fix: Before running any analysis, check n. Does the "n" in your output match your full dataset? If not, identify which variables are causing the reduction and document it. Build the habit of running SPSS Missing Value Analysis (Analyze → Missing Value Analysis) on every dataset before beginning analysis.

Mistake 2: Using Mean Imputation for Analysis

Mean imputation is taught as the simplest fix for missing data, but it is statistically destructive for any analysis beyond descriptive tables. It reduces the variance of imputed variables (narrowing confidence intervals), distorts correlation structures (biasing regression coefficients), and produces standard errors that are too small (inflating Type I error). Despite being clearly inferior to multiple imputation for decades, mean imputation still appears regularly in published medical research.

Fix: Use mean imputation only for descriptive summary tables (clearly labelling imputed values), never for regression, ANOVA, or any inferential analysis. Switch to MICE for all analytic work — available natively in SPSS (Analyze → Multiple Imputation → Impute Missing Data Values), R (mice package), and Stata (mi impute chained).

Mistake 3: Assuming MAR Without Checking or Justifying

Multiple imputation is valid under MAR, and many researchers simply state "MAR was assumed" without any investigation or clinical justification. MAR cannot be confirmed from the observed data (only MCAR can be tested). Stating MAR without reasoning about why missingness is related to observed rather than unobserved variables is methodologically hollow and increasingly called out by reviewers.

Fix: (1) Conduct Little's MCAR test; (2) run a logistic regression of missingness indicator on observed covariates to identify observed predictors of missingness; (3) write a clinical justification: "Patients with longer LOS were more likely to have HbA1c recorded as part of a systematic review pathway — length of stay is observed, making MAR plausible." (4) State this reasoning in your Methods section and conduct a MNAR sensitivity analysis regardless.

Mistake 4: Excluding the Outcome from the Imputation Model

A surprisingly common error in MICE implementation: the researcher imputes missing covariate values but omits the outcome variable from the imputation model. This biases the imputed covariate values toward a distribution independent of the outcome, which attenuates regression coefficients toward the null. The resulting effect estimate is biased to a smaller, less clinically meaningful value.

Fix: Always include the analysis outcome in the imputation model when imputing covariates. In SPSS MI dialog, this means including the dependent variable in the imputation variable list. In R mice, the default predictorMatrix includes all variables by default — do not remove the outcome variable from the imputation specification.

Mistake 5: Using Too Few Imputations (m = 5)

The old rule of m = 5–10 imputations is now known to be inadequate for datasets with moderate-to-high missing data rates. With 20% missing data and m = 5 imputations, the Monte Carlo error in pooled estimates is substantial — repeating the same analysis with a different random seed can give noticeably different p-values. This undermines reproducibility.

Fix: Set m equal to at least the percentage of incomplete cases. With 20% missing cases, use m ≥ 20 imputations; with 40% missing, use m ≥ 40. Modern software handles 50–100 imputations in seconds to minutes. The additional computation cost is trivial; the reproducibility gain is substantial.

Mistake 6: Not Reporting a Sensitivity Analysis for MNAR

CONSORT and ICH E9(R1) guidelines for clinical trials require sensitivity analysis exploring the impact of MNAR on primary results. In practice, most published papers report only the MI analysis under MAR and do not explore what happens if the MAR assumption is violated. Reviewers at high-impact journals increasingly require a MNAR sensitivity analysis for any trial or study with significant dropout.

Fix: Conduct at minimum a tipping-point analysis ("how much more extreme would the missing values need to be before our conclusion changes?") or a delta-adjustment sensitivity analysis (add a clinically plausible shift parameter δ to the imputed values for the missing outcome group and re-estimate the primary analysis). Report these as sensitivity analyses in the Results section.

Scientific Reporting Guidance

The following regulatory and reporting guidelines govern missing data documentation in different medical research contexts:

CONSORT 2010 (updated 2025): Requires a flow diagram showing participants with missing outcome data at each time point; the missing data handling method; sensitivity analysis for missing data assumption. Item 12c specifically requires: "If done, how missing data were handled in the analysis."
STROBE statement (observational studies): Item 12e requires: "How missing data were addressed." Authors should describe patterns of missingness, the mechanism assumed, and the method used.
ICH E9(R1) — Estimands and Sensitivity Analysis: The 2019 ICH E9(R1) addendum for regulatory submissions requires pre-specification of the primary estimand (intent-to-treat vs per-protocol vs other), the missing data strategy associated with that estimand, and multiple sensitivity analyses. Multiple imputation under MAR is typically the primary analysis; pattern mixture models or tipping-point analysis are typical sensitivity analyses.
STROBE-MI (missing imputation extension): A 2023 extension to STROBE provides a 34-item reporting checklist specifically for studies using multiple imputation. Key items: complete-variable proportions, convergence diagnostics, imputation model specification, number of imputations, auxiliary variables used, software.

Practical Recommendations for Researchers

Prevent Missing Data Prospectively

The best missing data strategy is prevention. In prospective studies, build data completeness into the protocol: train staff on mandatory data capture, use electronic systems with mandatory fields for critical variables, plan reminder procedures for follow-up appointments, and include a data completeness audit in the trial monitoring plan. Every percentage point of missingness prevented at the design stage is better than any imputation method.

Run Missing Value Analysis Before Any Analysis

Make Missing Value Analysis the first step in every dataset. In SPSS: Analyze → Missing Value Analysis → request the EM table, patterns, and Little's MCAR test. In R: use naniar::vis_miss() for a visual pattern map and naniar::mcar_test() for Little's test. Document the output and include it in your thesis methods chapter as Table 1a (missing data summary).

Build Your Imputation Model Before Collecting Data

Pre-specify in your study protocol which variables will be included in the imputation model, the number of imputations planned, the MICE model for each variable type, and the sensitivity analyses to be conducted. This pre-registration prevents criticism that the imputation model was optimised post-hoc and aligns your analysis with ICH E9(R1) requirements for regulatory submissions.

Always Run a Complete-Case Sensitivity Analysis

Even when MI is the primary analysis, always report a complete-case (listwise deletion) sensitivity analysis alongside it. If the conclusions agree between MI and complete-case analyses, confidence in the primary finding increases. If they disagree substantially, investigate the direction and magnitude of the difference — it reveals how sensitive your conclusions are to the missing data assumption, which is itself an important result.

Check Convergence in MICE

MICE relies on iterative cycling to convergence. Check convergence by plotting the mean and SD of imputed values across iterations (SPSS MI dialog; R plot(mice_object)). Convergence is indicated by stable, non-trending traces that mix well. Non-convergence produces biased imputations and must be addressed by increasing iterations, removing collinear predictors from the imputation model, or using a different imputation method for the problematic variable.

Use R mice Package for Maximum Control

SPSS MI is accessible but limited in customisability. The R mice package (Van Buuren & Groothuis-Oudshoorn) provides the most flexible and well-documented MICE implementation: custom imputation methods per variable, predictor matrix editing, convergence plots, and easy integration with lm(), glm(), survival, and lavaan downstream analyses. The companion book "Flexible Imputation of Missing Data" (Van Buuren, free online) is the definitive reference.

Frequently Asked Questions

What is the difference between MCAR, MAR, and MNAR? +

MCAR (Missing Completely at Random): missingness is unrelated to any data — a randomly dropped sample. MAR (Missing at Random): missingness depends on observed variables (e.g., older patients less likely to complete a form, but age is recorded). MNAR (Missing Not at Random): missingness depends on the unobserved value itself — e.g., patients with the highest pain scores being least likely to return a pain questionnaire. MCAR is the safest assumption; only Little's test can formally support it. MAR allows valid multiple imputation. MNAR cannot be confirmed from observed data and requires sensitivity analysis with additional assumptions.

When is listwise deletion acceptable? +

Listwise deletion is acceptable only when: (1) data are MCAR (Little's test non-significant); (2) fewer than 5% of values are missing; and (3) the reduced sample still provides adequate statistical power. In most clinical research contexts, these conditions are not met — missingness is driven by patient characteristics or clinical pathways (MAR or MNAR), not random chance. When any of these conditions fail, listwise deletion produces biased estimates and should be replaced with multiple imputation.

What is multiple imputation and how does it work? +

Multiple imputation creates m complete datasets by replacing each missing value with m plausible values drawn from the posterior predictive distribution (using observed data as predictors). Each imputed dataset is analysed separately, then results are pooled using Rubin's Rules: the pooled estimate is the mean of m estimates, and the pooled SE accounts for both within-imputation variance and between-imputation variance (the additional uncertainty from imputation). This correctly propagates the uncertainty from missing data into the final CIs and p-values — unlike single imputation methods.

How many imputations should I use? +

The number of imputations m should be at least equal to the percentage of incomplete cases (White et al., 2011). With 20% missing cases, use m ≥ 20 imputations; with 40% missing, m ≥ 40. Modern software handles 50–100 imputations with minimal computation time. The old rule of m = 5 is inadequate for moderate-to-high missing data rates and is no longer recommended for journal submission.

What is MICE and when should I use it? +

MICE (Multiple Imputation by Chained Equations / Fully Conditional Specification) is the most widely used MI algorithm for mixed-type clinical datasets. It imputes each incomplete variable using a separate regression model tailored to that variable's type (linear for continuous, logistic for binary, ordinal logistic for ordered categories). Use MICE whenever you have multiple incomplete variables of different types — the typical situation in clinical data. It is available in SPSS, R (mice package), and Stata (mi impute chained).

What is Little's MCAR test? +

Little's MCAR test is a formal χ² test of whether the observed data are consistent with MCAR. A non-significant result (p ≥ 0.05) supports MCAR; a significant result (p < 0.05) rules it out. In SPSS: Analyze → Missing Value Analysis → EM → check "Little's MCAR test." In R: naniar::mcar_test(data). A non-significant Little's test does not prove MCAR definitively, but a significant result means you cannot justify listwise deletion and should use multiple imputation.

Is mean imputation acceptable? +

Mean imputation is not acceptable for inferential analysis. It reduces variance artificially (narrowing CIs, inflating test statistics), distorts correlation structures (biasing regression coefficients), and underestimates standard errors (producing liberal p-values). Mean imputation may be used for descriptive summary tables only, clearly labelled as "imputed with variable mean," but never as the basis for regression, ANOVA, or any other inferential test.

How much missing data is too much? +

There is no absolute universal threshold, but practical guidance: <5% is rarely a problem with any method; 5–15% warrants multiple imputation; 15–25% requires MI plus sensitivity analysis for MNAR; 25–40% makes inference increasingly uncertain even with MI; >40% makes the analysis's credibility questionable and requires expert statistical consultation, exploration of the reasons for missingness, and honest limitation acknowledgement regardless of the imputation approach used.

Should I include the outcome in the imputation model? +

Yes — always include the analysis outcome in the imputation model when imputing covariates. Omitting the outcome biases imputed covariate values toward a distribution independent of the outcome, attenuating regression coefficients toward the null (biasing results toward no effect). In SPSS MI, include the dependent variable in the variable list for imputation. In R mice, the default predictor matrix includes all variables — do not remove the outcome.

What does CONSORT require for missing data reporting? +

CONSORT requires: (1) a flow diagram showing the number of participants with missing primary outcome data at each time point; (2) explicit description of the missing data handling method in the Statistical Methods section; (3) a pre-specified sensitivity analysis testing the robustness of conclusions to the missing data assumption. ICH E9(R1) additionally requires the primary estimand to be pre-specified with its associated missing data strategy. For observational studies, STROBE item 12e requires: how missing data were addressed in the analysis.

Analyse Your Clinical Data Online

Use StatClinic to explore missing data patterns, run complete-case and imputed analyses side-by-side, and generate publication-ready output with 95% confidence intervals for your thesis or journal submission.

Open StatClinic →

How to Handle Missing Data in Medical Research:Complete Statistical Guide