Downscaling livestock census data using multivariate predictive models: Sensitivity to modifiable areal unit problem

• Published in: PloS one Vol. 15; no. 1; p. e0221070
• Main Authors: Da Re, Daniele, Gilbert, Marius, Chaiban, Celia, Bourguignon, Pierre, Thanapongtharm, Weerapong, Robinson, Timothy P, Vanwambeke, Sophie O
• Format: Journal Article
• Language: English
• Published: United States Public Library of Science 27.01.2020; Public Library of Science (PLoS)
• Subjects: Analysis; Animals; Aquatic birds; Bias; Biology and Life Sciences; Census; Censuses; Chickens; Clustering; Computer and Information Sciences; Correlation analysis; Datasets; Ducks; Earth Sciences; Epidemiology; Geospatial data; Goodness of fit; Humans; Livestock; Models, Statistical; Multivariate analysis; Physical Sciences; Poultry; Prediction models; Remote sensing; Research and Analysis Methods; Researchers; Sampling; Sampling methods; Sensitivity analysis; Spatial Analysis; Spatial data; Statistical analysis; Thailand; Variables; Thailand; Belgium; Bangkok Thailand
• ISSN: 1932-6203; 1932-6203
• DOI: 10.1371/journal.pone.0221070

The analysis of census data aggregated by administrative units introduces a statistical bias known as the modifiable areal unit problem (MAUP). Previous researches have mostly assessed the effect of MAUP on upscaling models. The present study contributes to clarify the effects of MAUP on the downscaling methodologies, highlighting how a priori choices of scales and shapes could influence the results. We aggregated chicken and duck fine-resolution census in Thailand, using three administrative census levels in regular and irregular shapes. We then disaggregated the data within the Gridded Livestock of the World analytical framework, sampling predictors in two different ways. A sensitivity analysis on Pearson's r correlation statistics and RMSE was carried out to understand how size and shapes of the response variables affect the goodness-of-fit and downscaling performances. We showed that scale, rather than shapes and sampling methods, affected downscaling precision, suggesting that training the model using the finest administrative level available is preferable. Moreover, datasets showing non-homogeneous distribution but instead spatial clustering seemed less affected by MAUP, yielding higher Pearson's r values and lower RMSE compared to a more spatially homogenous dataset. Implementing aggregation sensitivity analysis in spatial studies could help to interpret complex results and disseminate robust products.