The study, conducted by Ekaterina Nikitina and Helena Mouriño, focused on a concrete classification problem: predicting hypertension from demographic and clinical variables. The researchers drew on the NHANES August 2021 to August 2023 cycle, restricting the analysis to adults aged 18 or older and excluding participants with missing hypertension status, age group, gender, or frequent alcohol use status. The resulting analytical dataset contained 5,431 participants and retained the naturally occurring missing values in three continuous variables: body mass index, fasting glucose, and total cholesterol. Hypertension was defined using averaged blood pressure measurements, with thresholds of at least 140 mmHg systolic or 90 mmHg diastolic, or through self-reported prior diagnosis or current use of antihypertensive medication, consistent with established clinical definitions.

The missingness itself was far from uniform. Fasting glucose was the dominant source of incompleteness, with 2,588 missing observations, or 47.65 percent of the sample. Total cholesterol had 488 missing values, about 8.99 percent, while body mass index was missing for only 47 participants, or 0.87 percent. Exploratory logistic regression models fitted to the missingness indicators revealed that incompleteness was not unrelated to observed characteristics. Participants with hypertension had higher estimated odds of missing BMI, with an odds ratio of 5.16, although the confidence interval was wide given the small number of missing BMI observations. Participants aged 60 to 69 had lower estimated odds of missing fasting glucose than those under 40. The authors were careful to note that these models identify associations with observed covariates but cannot establish the underlying missing data mechanism, which remains unobserved.

Five single-imputation methods were compared: mean imputation, linear regression imputation, K-nearest neighbour imputation with k equal to 5, random forest imputation using 500 trees, and CatBoost imputation based on gradient boosting with 1,000 iterations, tree depth 8, and a learning rate of 0.05. Complete case analysis, in which observations with missing values are simply discarded, was included as a reference strategy. The evaluation framework was deliberately rigorous. The dataset was partitioned into 70 percent training and 30 percent test subsets in a stratified manner, and the entire procedure was repeated 300 times with reproducible random seeds. Crucially, all imputation parameters and imputation models were estimated exclusively from the training data, and the fitted procedures were then applied to missing values in both the training and test datasets. The hypertension outcome was excluded from all imputation models to prevent outcome information from leaking into the imputation step.

After imputation, the continuous variables were converted into clinically defined binary predictors: obesity for BMI of at least 30 kg/m², hyperglycaemia for fasting glucose of at least 5.6 mmol/L, and high cholesterol for total cholesterol of at least 6.2 mmol/L. A logistic regression model was then fitted to predict hypertension, and a classification threshold was selected by maximising Youden’s J statistic on the training ROC curve. Predictive performance was evaluated using the area under the ROC curve, Cohen’s Kappa, sensitivity, and specificity. Because complete case analysis by definition evaluates only observations with complete predictor information, the researchers also ran a secondary common test analysis in which all six approaches were assessed on the same subset of test participants with complete data, standardising the test population without altering each method’s training procedure.

The predictive results were remarkably consistent. In the primary analysis, mean AUC values ranged from 0.781 for mean imputation to 0.784 for linear regression imputation, with KNN, random forest, and CatBoost producing nearly identical values around 0.783. Cohen’s Kappa ranged from 0.429 to 0.440, sensitivity from 0.762 to 0.775, and specificity from 0.671 to 0.692. Complete case analysis, evaluated on a substantially smaller effective sample, had a mean AUC of 0.774. When all approaches were evaluated on the same complete case test subset in the secondary analysis, the differences shrank further: mean AUC values clustered between 0.774 and 0.775, and the 95 percent empirical intervals for every pairwise difference in AUC, Kappa, sensitivity, and specificity included zero. In the primary analysis, the only pairwise comparison whose empirical interval excluded zero was the AUC difference between mean and linear regression imputation, and even that absolute difference was roughly 0.003.

Reconstruction accuracy told a different story. To assess imputation quality independently of predictive performance, the researchers implemented a masked validation framework on a complete case subset of 2,713 participants. Observed values in one continuous variable at a time were artificially masked at 20 percent and 50 percent levels, under two scenarios: completely random masking, and covariate-dependent masking in which the probability of masking depended on observed participant characteristics including age group, gender, frequent alcohol use, and the remaining continuous variables. The 50 percent level was chosen to approximate the observed missingness of fasting glucose. For each combination of variable, mechanism, proportion, and replication, the same observations were masked for all five methods, enabling paired comparison. Each scenario was repeated 300 times, and reconstruction accuracy was evaluated using root mean squared error and mean absolute error.

The masked validation revealed clearer separation between methods than the predictive analysis did. Random forest produced the lowest reconstruction errors for BMI across all four masking scenarios, with mean RMSE values between 6.648 and 6.805. Linear regression consistently achieved the lowest errors for fasting glucose, with mean RMSE values between 1.671 and 1.684, while random forest generally provided the next lowest. For total cholesterol, linear regression and random forest were closely matched, with random forest yielding the lowest RMSE under 20 percent MCAR masking and linear regression the lowest under 50 percent MCAR masking. Notably, neither KNN nor CatBoost demonstrated a consistent reconstruction advantage over these alternatives, despite their greater algorithmic flexibility. The relative ranking of the principal methods remained broadly stable across masking mechanisms and proportions, suggesting the results were not an artefact of a single masking design.

The disconnect between these two sets of findings is the study’s central message. Model-based methods, particularly random forest and linear regression, achieved lower reconstruction error than mean imputation, yet these improvements did not translate into substantial gains in downstream classification. The authors point to a plausible explanation rooted in the analytical framework itself: because the imputed continuous values were converted into clinically defined binary predictors before model fitting, differences between reconstructed values affect downstream prediction only when they are large enough to move an observation across a clinical threshold. This discretisation may reduce the sensitivity of the final model to moderate differences between imputation procedures. More generally, the relationship between reconstruction accuracy and downstream performance depends on the correlation structure of the data, the missingness process, and the analytical model used.

The practical implications are pragmatic rather than prescriptive. The findings suggest that imputation methods should not be selected solely on algorithmic complexity or reconstruction error, and that for predictive applications similar to the one studied here, simpler single-imputation approaches may perform comparably to more computationally intensive alternatives. Linear regression was particularly competitive, combining strong reconstruction performance with predictive performance equivalent to the machine learning methods. The authors emphasise that these conclusions are specific to the present dataset, masking scenarios, and downstream modelling framework, and should not be generalised without further validation. The study also focused on missing continuous predictors and did not evaluate multiple imputation, which propagates uncertainty associated with missing values and remains valuable when inferential rather than predictive objectives dominate. Still, the broader lesson stands: optimising how accurately missing values are reconstructed and optimising how well a model predicts are two different goals, and the choice of imputation strategy should be guided by the final analytical objective rather than by reconstruction error alone.

Subject of Research: Evaluation of missing data imputation methods and their disconnect between reconstruction accuracy and downstream predictive performance

Article Title: Reconstruction accuracy versus predictive performance: evaluating imputation methods under real-world missingness

Article References: Nikitina, E., & Mouriño, H. (2026). Reconstruction accuracy versus predictive performance: evaluating imputation methods under real-world missingness. International Journal of Data Science and Analytics, 22(1), 362. https://doi.org/10.1007/s41060-026-01295-3

Image Credits: AI Generated

DOI: 10.1007/s41060-026-01295-3

Keywords: missing data, imputation, machine learning, NHANES, hypertension, predictive modelling, epidemiology, data science, Reconstruction, accuracy, versus, predictive