TY - JOUR

T1 - Understanding the impact of covariates for trachoma prevalence prediction using geostatistical methods

AU - Sasanami, Misaki

AU - Almou, Ibrahim

AU - Diori, Adam Nouhou

AU - Bakhtiari, Ana

AU - Beidou, Nassirou

AU - Bisanzio, Donal

AU - Boyd, Sarah

AU - Burgert-Brucker, Clara R.

AU - Amza, Abdou

AU - Gass, Katherine

AU - Kadri, Boubacar

AU - Kebede, Fikreab

AU - Masika, Michael P.

AU - Olobio, Nicholas P.

AU - Seife, Fikre

AU - Souley, Abdoul Salam Youssoufou

AU - Tefera, Amsayaw

AU - Kello, Amir B.

AU - Solomon, Anthony W.

AU - Harding-Esch, Emma M.

AU - Giorgi, Emanuele

PY - 2025/6/1

Y1 - 2025/6/1

N2 - Background: Model-based geostatistics (MBG) is increasingly used for estimating the prevalence of neglected tropical diseases, including trachoma, in low- and middle-income countries. We sought to investigate the impact of spatially referenced covariates to improve spatial predictions for trachomatous inflammation—follicular (TF) prevalence generated by MBG. To this end, we assessed the ability of spatial covariates to explain the spatial variation of TF prevalence and to reduce uncertainty in the assessment of TF elimination for pre-defined evaluation units (EUs). Methods: We used data from Tropical Data-supported population-based trachoma prevalence surveys conducted in EUs in Ethiopia, Malawi, Niger, and Nigeria between 2016 and 2023. We then compared two models: a model that used only age, a variable required for the standardization of prevalence as used in the routine, standard prevalence estimation, and a model that included spatial covariates in addition to age. For each fitted model, we reported estimates of the parameters that quantify the strength of residual spatial correlation and 95% prediction intervals as the measure of uncertainty. Results: The strength of the association between covariates and TF prevalence varied within and across countries. For some EUs, spatially referenced covariates explained most of the spatial variation and thus allowed us to generate predictive inferences for TF prevalence with a substantially reduced uncertainty, compared with models without the spatial covariates. For example, the prediction interval for TF prevalence in the areas with the lowest TF prevalence in Nigeria narrowed substantially, from a width of 2.9 to 0.7. This reduction occurred as the inclusion of spatial covariates significantly decreased the variance of the spatial Gaussian process in the geostatistical model. In other cases, spatial covariates only led to minor gains, with slightly smaller prediction intervals for the EU-level TF prevalence or even a wider prediction interval. Conclusions: Although spatially referenced covariates could help reduce prediction uncertainty in some cases, the gain could be very minor, or uncertainty could even increase. When considering the routine, standardized use of MBG methods to support national trachoma programs worldwide, we recommend that spatial covariate use be avoided.

AB - Background: Model-based geostatistics (MBG) is increasingly used for estimating the prevalence of neglected tropical diseases, including trachoma, in low- and middle-income countries. We sought to investigate the impact of spatially referenced covariates to improve spatial predictions for trachomatous inflammation—follicular (TF) prevalence generated by MBG. To this end, we assessed the ability of spatial covariates to explain the spatial variation of TF prevalence and to reduce uncertainty in the assessment of TF elimination for pre-defined evaluation units (EUs). Methods: We used data from Tropical Data-supported population-based trachoma prevalence surveys conducted in EUs in Ethiopia, Malawi, Niger, and Nigeria between 2016 and 2023. We then compared two models: a model that used only age, a variable required for the standardization of prevalence as used in the routine, standard prevalence estimation, and a model that included spatial covariates in addition to age. For each fitted model, we reported estimates of the parameters that quantify the strength of residual spatial correlation and 95% prediction intervals as the measure of uncertainty. Results: The strength of the association between covariates and TF prevalence varied within and across countries. For some EUs, spatially referenced covariates explained most of the spatial variation and thus allowed us to generate predictive inferences for TF prevalence with a substantially reduced uncertainty, compared with models without the spatial covariates. For example, the prediction interval for TF prevalence in the areas with the lowest TF prevalence in Nigeria narrowed substantially, from a width of 2.9 to 0.7. This reduction occurred as the inclusion of spatial covariates significantly decreased the variance of the spatial Gaussian process in the geostatistical model. In other cases, spatial covariates only led to minor gains, with slightly smaller prediction intervals for the EU-level TF prevalence or even a wider prediction interval. Conclusions: Although spatially referenced covariates could help reduce prediction uncertainty in some cases, the gain could be very minor, or uncertainty could even increase. When considering the routine, standardized use of MBG methods to support national trachoma programs worldwide, we recommend that spatial covariate use be avoided.

KW - Evaluation unit

KW - Trachoma

KW - Covariates

KW - Disease mapping

KW - Geostatistics

KW - Neglected tropical diseases

U2 - 10.1186/s44263-025-00161-x

DO - 10.1186/s44263-025-00161-x

M3 - Journal article

VL - 3

JO - BMC Global and Public Health

JF - BMC Global and Public Health

SN - 2731-913X

IS - 1

M1 - 48

ER -