Which of the following strategies used by modelers to enhance data accuracy has the GREATEST risk of bias and information loss?

Imputation by filling missing values with a central tendency statistic (mean, median, or mode) introduces significant bias. This method artificially reduces the variance of the dataset and can distort the covariance and correlation between variables, as it assumes the missing data follows the same distribution as the observed data. This often leads to underestimation of standard errors and biased model parameters. Furthermore, it causes information loss by discarding the potentially meaningful pattern of "missingness" itself, replacing it with a single, unrepresentative value. While other methods have drawbacks, this naive imputation carries the greatest combined risk of fundamentally skewing the data's statistical properties.

B: Deleting true duplicate entries is a standard data cleaning procedure that improves accuracy and reduces redundancy; it does not inherently introduce bias or lose unique information.

C: Separating a single field into multiple attribute columns is a data structuring (normalization) technique that increases information granularity and clarity, rather than causing loss or bias.

D: While binning (discretization) inherently causes information loss by reducing numerical precision, it is a deliberate data reduction technique. The resulting bias is often less severe than the distributional distortion caused by mean/median imputation.

1. Gelman

A.

& Hill

J. (2006). Data Analysis Using Regression and Multilevel/Hierarchical Models. Cambridge University Press. In Chapter 25

"Missing-Data Imputation

" the authors explain that single imputation methods like mean imputation "can lead to biased estimates of regression coefficients... and invariably leads to underestimation of standard errors" (p. 533).

2. Han

J.

Pei

J.

& Kamber

M. (2012). Data Mining: Concepts and Techniques (3rd ed.). Morgan Kaufmann. In Section 3.2.2

"Missing Values

" the text describes mean/median imputation as a popular but "simple" method

noting its disadvantage: "it may form a cluster with some outliers" and skew the data if the missing data is not missing at random (p. 90).

3. Stanford University

CS229: Machine Learning Course Notes. While not a single document

course materials on preprocessing frequently discuss the pitfalls of naive imputation. For instance

discussions on Expectation-Maximization for missing data highlight the limitations and biases of simpler methods like mean substitution. The general principle taught is that mean imputation ignores the relationship between variables and reduces data variance.