Ciencia de Datos
Imputación múltiple libre de distribución basada en descomposición en valores singulares robusta: Una comparación empírica
Este estudio desarrolla, valida y compara un enfoque de imputación múltiple (IM) basado en variantes robustas de la Descomposición en Valores Singulares (SVD), específicamente rSVD84, rSVD01 y un híbrido rSVD84 01, para abordar datos faltantes bajo diferentes mecanismos y niveles de contaminación. El objetivo es lograr imputaciones confiables bajo escenarios MCAR, MAR y MNAR, manteniendo la robustez frente a valores atípicos. Los modelos propuestos se comparan con Bootstrap Bayesiano Aproximado (ABB) e Imputación Multivariada por Ecuaciones Encadenadas (MICE) mediante simulaciones controladas y conjuntos de datos del mundo real. Los resultados de la simulación indican que, bajo contaminación moderada, los modelos basados en SVD tienen un rendimiento competitivo: superan consistentemente a ABB y se mantienen libres de distribución, aunque generalmente no alcanzan la eficiencia de MICE. Bajo valores atípicos extremos, los modelos SVD demuestran una clara ventaja, manteniendo una precisión estable a medida que aumentan los valores faltantes, siendo rSVD84 el que exhibe la mayor robustez. El escenario MNAR sigue siendo el más desafiante para todos los métodos, produciendo una menor precisión predictiva y una mayor variabilidad entre imputaciones. En conjuntos de datos reales multivariados, los métodos basados en SVD producen resultados mixtos y a menudo tienen un rendimiento inferior en relación con MICE y ABB, probablemente porque la aleatorización de estilo bootstrap inherente a estos métodos es más adecuada para las estructuras de genotipo por entorno (GxE) que para los datos multivariados genéricos. En muestras muy pequeñas, estos modelos no pueden explotar completamente su mecanismo de aleatorización, lo que lleva a imputaciones inestables. En general, los métodos robustos de SVD preservan la precisión predictiva en presencia de valores atípicos extremos, pero a costa de una mayor varianza entre imputaciones, lo que indica una mayor incertidumbre de imputación. Si bien la imputación robusta basada en SVD puede aproximarse al rendimiento de las líneas base establecidas, rara vez logra el error de predicción más bajo, lo que sugiere que las fortalezas observadas en las simulaciones no siempre se generalizan a diversas aplicaciones del mundo real.
Full text 101 pages with text
Read the full thesis Record in the repository
Contents
- General Objectivep. 14
- Specific Objectivesp. 14
- Multiple Imputationp. 15
- Rubin’s Rulesp. 15
- Classical MI Methods (MICE, Model-Based, Nonparametric)p. 16
- Evolution in recent literaturep. 17
- Missing Data Mechanismsp. 18
- Missing Data Patternsp. 18
- Diagnosing Missing Data Patternsp. 19
- Simulating Missing Data Mechanismsp. 21
- Outliersp. 21
- Definitions and Types of Outliersp. 21
- Outlier Detection Methodsp. 22
- Generating Outlier Simulationsp. 23
- Singular Value Decomposition (SVD)p. 23
- SVD Theoretical Foundations and Applicationsp. 23
- SVD as imputation methodp. 24
- Sensitivity of SVD to Outliers and Missing Datap. 24
- Robust Alternatives to SVD (rSVD84 and rSVD01)p. 25
- rSVD84p. 25
- rSVD01p. 26
- Simulation studyp. 28
- Real Data Studyp. 29
- Evaluation methodp. 30
- Proposed MI methodp. 30
- Results for simulated data with 5 standard deviation environmentp. 31
- Results for simulated data 100 standard deviation environmentp. 40
- Results for real data with 5 standard deviation environmentp. 47
- Results for real data with 100 standard deviation environmentp. 55
- Boxplot of Pe under MAR at p = 10%p. 31
- Boxplot of Pe under MAR at p = 20%p. 32
- Boxplot of Pe under MAR at p = 30%p. 32
- Boxplot of Pe under MCAR at p = 10%p. 33
- Boxplot of Pe under MCAR at p = 20%p. 33
- Boxplot of Pe under MCAR at p = 30%p. 33
- Boxplot of Pe under MNAR at p = 10%p. 34
- Boxplot of Pe under MNAR at p = 20%p. 34
- Boxplot of Pe under MNAR at p = 30%p. 35
- Boxplot of V under MAR at p = 10%p. 35
- Boxplot of V under MAR at p = 20%p. 36
- Boxplot of V under MAR at p = 30%p. 36
- Boxplot of V under MCAR at p = 10%p. 37
- Boxplot of V under MCAR at p = 20%p. 37
- Boxplot of V under MCAR at p = 30%p. 37
- Boxplot of V under MNAR at p = 10%p. 38
- Boxplot of V under MNAR at p = 20%p. 38
- Boxplot of V under MNAR at p = 30%p. 39
- Boxplot of Pe under MAR at p = 10%p. 40
- Boxplot of Pe under MAR at p = 20%p. 40
- Boxplot of Pe under MAR at p = 30%p. 41
- Boxplot of Pe under MCAR at p = 10%p. 41
- Boxplot of Pe under MCAR at p = 20%p. 42
- Boxplot of Pe under MCAR at p = 30%p. 42
- Boxplot of Pe under MNAR at p = 10%p. 43
- Boxplot of Pe under MNAR at p = 20%p. 43
- Boxplot of Pe under MNAR at p = 30%p. 43
- Boxplot of V under MAR at p = 10%p. 44
- Boxplot of V under MAR at p = 20%p. 44
- Boxplot of V under MAR at p = 30%p. 45
- Boxplot of V under MCAR at p = 10%p. 45
- Boxplot of V under MCAR at p = 20%p. 46
- Boxplot of V under MCAR at p = 30%p. 46
- Boxplot of V under MAR at p = 10%p. 46
- Boxplot of V under MAR at p = 20%p. 47
- Boxplot of V under MAR at p = 30%p. 47
- Multivariate datasets chosen to perform the real-data studyp. 29
- Summary of G×E datasetsp. 29
- Pe by percentage and method — Mechanism MAR (REP 1–5)p. 49
- Pe by percentage and method — Mechanism MCAR (REP 1–5)p. 50
- Pe by percentage and method — Mechanism MNAR (REP 1–5)p. 50
- Pe by percentage and method — Mechanism MAR (REP 6–10)p. 51
- Pe by percentage and method — Mechanism MCAR (REP 6–10)p. 51
- Pe by percentage and method — Mechanism MNAR (REP 6–10)p. 52
- Pe by percentage and method — Mechanism MAR (REP 11–15)p. 52
- Pe by percentage and method — Mechanism MCAR (REP 11–15)p. 53
- Pe by percentage and method — Mechanism MNAR (REP 11–15)p. 53
- Pe by percentage and method — Mechanism MAR (REP 16–17)p. 53
- Pe by percentage and method — Mechanism MAR (REP 16–17)p. 54
- Pe by percentage and method — Mechanism MCAR (REP 16–17)p. 54
- Pe by percentage and method — Mechanism MNAR (REP 16–17)p. 54
- Pe by percentage and method — Mechanism MAR (REP 1–5)p. 55
- Pe by percentage and method — Mechanism MCAR (REP 1–5)p. 56
- Pe by percentage and method — Mechanism MNAR (REP 1–5)p. 56
- Pe by percentage and method — Mechanism MAR (REP 6–10)p. 57
- Pe by percentage and method — Mechanism MCAR (REP 6–10)p. 57
- Pe by percentage and method — Mechanism MNAR (REP 6–10)p. 58
- Pe by percentage and method — Mechanism MAR (REP 11–15)p. 58
- Pe by percentage and method — Mechanism MCAR (REP 11–15)p. 59
- Pe by percentage and method — Mechanism MNAR (REP 11–15)p. 59
- Pe by percentage and method — Mechanism MAR (REP 16–17)p. 59
- Pe by percentage and method — Mechanism MCAR (REP 16–17)p. 60
- Prediction error (Pe) under MAR (grouped by missingness percentage)p. 68
- Prediction error (Pe) under MCAR (grouped by missingness percentage)p. 69
- Prediction error (Pe) under MNAR (grouped by missingness percentage)p. 70
- centage)p. 71
- centage)p. 71
- centage)p. 71
- Prediction error (Pe) under MAR (grouped by missingness percentage)p. 74
- Prediction error (Pe) under MCAR (grouped by missingness percentage)p. 75
- Prediction error (Pe) under MNAR (grouped by missingness percentage)p. 76
- centage)p. 71
- centage)p. 71
- centage)p. 71
- V by percentage and method — Mechanism MAR (REP 1–5)p. 80
- V by percentage and method — Mechanism MCAR (REP 1–5)p. 81
- V by percentage and method — Mechanism MNAR (REP 1–5)p. 82
- V by percentage and method — Mechanism MAR (REP 6–10)p. 83
- V by percentage and method — Mechanism MCAR (REP 6–10)p. 84
- V by percentage and method — Mechanism MNAR (REP 6–10)p. 85
- V by percentage and method — Mechanism MAR (REP 11–15)p. 86
- V by percentage and method — Mechanism MCAR (REP 11–15)p. 87
- V by percentage and method — Mechanism MNAR (REP 11–15)p. 88
- V by percentage and method — Mechanism MAR (REP 16–17)p. 89
- V by percentage and method — Mechanism MCAR (REP 16–17)p. 89
- V by percentage and method — Mechanism MNAR (REP 16–17)p. 90
- V by percentage and method — Mechanism MAR (REP 1–5)p. 91
- V by percentage and method — Mechanism MCAR (REP 1–5)p. 92
- V by percentage and method — Mechanism MNAR (REP 1–5)p. 93
- V by percentage and method — Mechanism MAR (REP 6–10)p. 94
- V by percentage and method — Mechanism MCAR (REP 6–10)p. 95
- V by percentage and method — Mechanism MNAR (REP 6–10)p. 96
- V by percentage and method — Mechanism MAR (REP 11–15)p. 97
- V by percentage and method — Mechanism MCAR (REP 11–15)p. 98
- V by percentage and method — Mechanism MNAR (REP 11–15)p. 99
- V by percentage and method — Mechanism MAR (REP 16–17)p. 100
- V by percentage and method — Mechanism MCAR (REP 16–17)p. 100
- V by percentage and method — Mechanism MNAR (REP 16–17)p. 101