Data Quality and Performance Classification in Multisectoral Nutrition Monitoring: A Composite Framework and Empirical Application
Related Experiment Videos
Gesessew Kibr1,2,3, Bamilaku Alamirew4, Zerayehu Sime5
1Department of Project Planning and Management, Yom Institute of Economic Development, Postgraduate College, Addis Ababa, Ethiopia.
Background:
The consequences of using unfit data are substantial: misallocated resources, false signals of progress, delayed corrective action, and wasted investments.
Objectives:
This study developed and validated a composite traffic-light framework for jointly classifying data quality and performance in multisectoral indicator systems, and assessed the quality of 98 indicators from the Seqota Declaration scorecard in Addis Ababa, Ethiopia.
Methods:
A retrospective audit examined 98 indicators across 7 sectors, 4 subcities, and 4 y (2014-2017 Ethiopian Calender). The framework assessed: 1) completeness (≥75%), 2) temporal variation (coefficient of variation >0%), 3) subcity disaggregation (duplication detection with ±0.05% point tolerance), and 4) performance classification (GREEN ≥85%, YELLOW 55%-84%, RED <55%). Statistical validation employed nonparametric tests (Kruskal-Wallis, Wilcoxon, Jonckheere-Terpstra) on 11 priority indicators. A Monte Carlo simulation (1000 iterations) compared parametric and nonparametric tests across sample sizes (n = 4, 10, 20, 30, 50, 100) and distributions (normal, skewed, heavy-tailed). Sensitivity analyses tested alternative thresholds. Quality assurance followed the World Health Organization Data Quality Review Framework.
Results:
Of 98 indicators, 13.3% were GREEN, 46.9% YELLOW, 8.2% RED, and 31.6% EXCLUDED. Critically, 40.8% of indicators were duplicated across ≥3 subcities in ≥2 y. Women's Affairs (83.3%), Governance (80.0%), and Social Protection (66.7%) had the highest duplication rates. The disaggregation dimension reclassified 9 indicators from GREEN to YELLOW. Sensitivity analysis confirmed robustness. Statistical validation showed significant differences across categories (Kruskal-Wallis: χ2 = 29.381, P < 0.001; Jonckheere-Terpstra: z = 5.29, P < 0.001). Monte Carlo simulation confirmed that for n = 4, nonparametric tests maintain appropriate type I error rates (∼4.9%), whereas parametric tests show inflated type I error (6.2%-7.1%) with nonnormal data.
Conclusions:
This is a data-quality-dominant dataset, and systematic duplication across subcities challenges the assumption of genuine disaggregation. The composite framework provides a replicable, transparent approach to classifying indicators before performance analysis. For small-sample longitudinal data (n = 4), nonparametric tests are appropriate.
