Data Availability and Completeness: Strategies for Handling Missing Information

Incomplete data is a universal challenge across research, business analytics, healthcare, and policy. The phrase "the number of people who get is not specified" reflects a fundamental reality: datasets rarely capture everything we need. This guide provides a structured approach to assessing, addressing, and communicating data gaps.

Types of Missing Data (Rubin's Framework)

  • MCAR (Missing Completely at Random): Missingness unrelated to any observed or unobserved data. Example: Survey page lost in mail. Rare in practice.
  • MAR (Missing at Random): Missingness depends on observed data but not unobserved. Example: Higher-income respondents less likely to report income (income observed for others).
  • MNAR (Missing Not at Random): Missingness depends on the unobserved value itself. Example: People with severe depression less likely to complete depression survey. Most problematic.

Diagnosing Missingness Patterns

  • Missingness maps/heatmaps: Visualize patterns across variables and observations.
  • Little's MCAR test: Statistical test (limited power with large N).
  • Logistic regression of missingness indicators: Predict missingness from observed variables → evidence for MAR.
  • Sensitivity analysis: Test how conclusions change under different MNAR assumptions.

Strategies for Addressing Missing Data

1. Prevention (Best Strategy)

  • Study design: Minimize burden, pilot instruments, follow-up protocols.
  • Data collection: Real-time validation, mandatory fields, incentives.
  • Administrative linkage: Supplement surveys with registry/EHR data.

2. Complete Case Analysis (Listwise Deletion)

  • Valid only under MCAR (unbiased but inefficient).
  • Under MAR/MNAR: Biased estimates, reduced power.
  • Acceptable if <5% missing and MCAR plausible.

3. Single Imputation (Generally Discouraged for Analysis)

  • Mean/median/mode imputation: Underestimates variance, distorts distributions.
  • Regression imputation: Overestimates precision.
  • Last observation carried forward (LOCF): Biased for longitudinal data.
  • Use only for descriptive summaries or exploratory work.

4. Multiple Imputation (MI) — Gold Standard for MAR

  1. Impute: Create m (typically 20-100) complete datasets using chained equations (MICE) or joint modeling.
  2. Analyze: Run identical analysis on each dataset.
  3. Pool: Combine estimates using Rubin's rules (within + between imputation variance).

Requirements: Include all analysis variables + predictors of missingness + auxiliary variables. Check convergence, imputation diagnostics.

5. Maximum Likelihood / Full Information ML (FIML)

  • Directly estimates parameters using all observed data (no imputation step).
  • Standard in SEM (lavaan, Mplus), mixed models (lme4, nlme).
  • Assumes MAR; efficient, unbiased under correct model specification.

6. Inverse Probability Weighting (IPW)

  • Weight complete cases by inverse probability of being observed.
  • Requires correct specification of missingness model.
  • Useful for survey non-response, censoring.

7. Sensitivity Analysis for MNAR

  • Pattern-mixture models: Different parameters for missing vs. observed patterns.
  • Selection models: Joint model of outcome and missingness mechanism.
  • Delta-adjustment / tipping point: Shift imputed values systematically; find where conclusion changes.
  • Expert elicitation: Quantify plausible MNAR deviations.

Reporting Standards (Transparency)

  • Flow diagram: Screened → Eligible → Enrolled → Analyzed (CONSORT/STROBE).
  • Missingness by variable: Count, percentage, pattern.
  • Comparison: Complete vs. incomplete cases on observed variables.
  • Method: Explicit statement of approach, assumptions, software, diagnostics.
  • Sensitivity: Results under alternative assumptions.

Practical Decision Framework

  1. Quantify: How much missing? Which variables? Monotone or arbitrary pattern?
  2. Diagnose: MCAR test, missingness predictors, subject-matter plausibility of MAR vs. MNAR.
  3. Select method: MI (general purpose), FIML (SEM/longitudinal), IPW (survey), complete case (minimal MCAR).
  4. Implement carefully: Include auxiliary variables, check diagnostics, use sufficient m.
  5. Sensitivity analyze: At minimum, compare complete case vs. MI results.
  6. Report transparently: Enable reader to judge credibility.

Missing data is not a technical nuisance—it's a design and inference challenge. The goal is not to "fill in" numbers but to make valid inferences while honestly representing uncertainty.

Pediatrics - Related Articles