Abstract

Missing data remain an important challenge in agricultural surveys, particularly when questionnaires contain multiple interrelated modules covering household characteristics, land, crop and livestock production, labour, input use and farm management. Conventional imputation methods are often effective for individual variables but may have difficulty capturing complex nonlinear relationships and high-dimensional interactions across these domains. This study investigates the potential of deep learning to improve the imputation of missing values in agricultural survey microdata by exploiting auxiliary information from Liberia’s agricultural census.
The analysis uses information from the 2024 Liberia Agriculture Census together with data from the subsequent agricultural survey programme. Census variables describing agricultural households, land and parcels, crop and livestock activities, household characteristics and geographic location are used as auxiliary predictors for missing survey variables. A deep-learning imputation framework based on a denoising autoencoder is developed to accommodate mixed continuous and categorical data and to learn latent relationships among census and survey characteristics. Geographic and survey-design information is incorporated where appropriate to preserve important structural heterogeneity.
Model performance is evaluated through a controlled masking experiment in which observed survey values are removed under several missing-data scenarios and subsequently imputed. The deep-learning approach is benchmarked against conventional methods, including mean or median imputation, hot-deck procedures, multiple imputation by chained equations and tree-based machine-learning methods. Evaluation goes beyond record-level predictive accuracy. Continuous and categorical variables are assessed using appropriate error measures, while the ability of each approach to preserve distributions, relationships between variables and survey-weighted estimates of key agricultural indicators is also examined. Validation is conducted across geographic areas to assess the robustness and generalizability of the models.
The study provides an empirical assessment of whether high-dimensional census auxiliary information combined with deep learning can improve agricultural survey imputation without compromising the statistical properties required for official statistics. Particular attention is given to interpretability, reproducibility, imputation uncertainty and the preservation of design-based estimates. The findings are expected to inform statistical offices considering artificial intelligence and machine-learning methods for integrating census and survey data and improving the quality and efficiency of agricultural statistics.