Nearest neighbor imputation algorithms: a critical evaluation

Abstract

BackgroundNearest neighbor (NN) imputation algorithms are efficient methods to fill in missing data where each missing value on some records is replaced by a value obtained from related cases in the whole set of records. Besides the capability to substitute the missing data with plausible values that are as close as possible to the true value, imputation algorithms should preserve the original data structure and avoid to distort the distribution of the imputed variable. Despite the efficiency of NN algorithms little is known about the effect of these methods on data structure.MethodsSimulation on synthetic datasets with different patterns and degrees of missingness were conducted to evaluate the performance of NN with one single neighbor (1NN) and with k neighbors without (kNN) or with weighting (wkNN) in the context of different learning frameworks: plain set, reduced set after ReliefF filtering, bagging, random choice of attributes, bagging combined with random choice of attributes (Random-Forest-like method).ResultsWhatever the framework, kNN usually outperformed 1NN in terms of precision of imputation and reduced errors in inferential statistics, 1NN was however the only method capable of preserving the data structure and data were distorted even when small values of k neighbors were considered; distortion was more severe for resampling schemas.ConclusionsThe use of three neighbors in conjunction with ReliefF seems to provide the best trade-off between imputation error and preservation of the data structure. The very same conclusions can be drawn when imputation experiments were conducted on the single proton emission computed tomography (SPECTF) heart dataset after introduction of missing data completely at random.

Changes in subtypes of overt thyrotoxicosis and hypothyroidism following iodine fortification

“…From their medical history, we knew they did not suffer from the entities e‐m. In accordance with our previous study, we used nearest neighbour hot deck‐imputation to classify this small group of patients (6.9%) into the subgroups a‐d (GD/MNTG/mixed‐type/STA: n=10/9/8/1). Nearest neighbour hot deck‐imputation did not alter any of the results of the present study.…”

Predictors of objectively measured physical activity in 12‐month‐old infants: A study of linked birth cohort data with electronic health records

“…In this selected cohort, some variables presented with missing values (see Figure S2 for distribution of missing values across the studied variables). The k ‐nearest neighbour ( k NN) method was used to impute the missing values with 5 neighbours (ie, k = 5) . The k NN method is an efficient imputation technique by which each missing value on some data fields is replaced with a value calculated from related neighbouring cases across the whole records of a dataset.…”

“…The k NN method is an efficient imputation technique by which each missing value on some data fields is replaced with a value calculated from related neighbouring cases across the whole records of a dataset. It has demonstrated good capacity for preserving the original data structure . For each continuous univariable variable, a two‐sample t test was used in between two subgroups at 5% significance level, and a Pearson chi‐squared test was used in categorical variables with 5% significance level.…”

Subgrouping Factors Influencing Migraine Intensity in Women: A Semi‐automatic Methodology Based on Machine Learning and Information Geometry

“…This algorithm is an efficient method to fill in missing data. Each missing value on a record is replaced by a value from related cases in the whole set of records that depends on the type of variable used: categorical missing values are replaced by the mode and quantitative ones are replaced by the mean . The number of neighbors was fixed to 10 before conducting experiments.…”