Division of Epidemiology and Biostatistics, School of Public Health, University of Witwatersrand, Johannesburg, South Africa.
METRe Group, Department of International Health, Liverpool School of Tropical Medicine, Liverpool, United Kingdom.
PLoS One. 2021 Aug 6;16(8):e0253375. doi: 10.1371/journal.pone.0253375. eCollection 2021.
Model-based small area estimation methods can help generate parameter estimates at the district level, where planned population survey sample sizes are not large enough to support direct estimates of HIV prevalence with adequate precision. We computed district-level HIV prevalence estimates and their 95% confidence intervals for districts in Uganda.
Our analysis used direct survey and model-based estimation methods, including Fay-Herriot (area-level) and Battese-Harter-Fuller (unit-level) small area models. We used regression analysis to assess for consistency in estimating HIV prevalence. We use a ratio analysis of the mean square error and the coefficient of variation of the estimates to evaluate precision. The models were applied to Uganda Population-Based HIV Impact Assessment 2016/2017 data with auxiliary information from the 2016 Lot Quality Assurance Sampling survey and antenatal care data from district health information system datasets for unit-level and area-level models, respectively.
Estimates from the model-based and the direct survey methods were similar. However, direct survey estimates were unstable compared with the model-based estimates. Area-level model estimates were more stable than unit-level model estimates. The correlation between unit-level and direct survey estimates was (β1 = 0.66, r2 = 0.862), and correlation between area-level model and direct survey estimates was (β1 = 0.44, r2 = 0.698). The error associated with the estimates decreased by 37.5% and 33.1% for the unit-level and area-level models, respectively, compared to the direct survey estimates.
Although the unit-level model estimates were less precise than the area-level model estimates, they were highly correlated with the direct survey estimates and had less standard error associated with estimates than the area-level model. Unit-level models provide more accurate and reliable data to support local decision-making when unit-level auxiliary information is available.
基于模型的小区域估计方法可以帮助生成地区一级的参数估计，在这些地区，计划的人口调查样本量不足以支持直接估计 HIV 流行率，且无法达到足够的精度。我们计算了乌干达各地区的 HIV 流行率估计值及其 95%置信区间。
我们的分析使用了直接调查和基于模型的估计方法，包括 Fay-Herriot（地区层面）和 Battese-Harter-Fuller（单位层面）小区域模型。我们使用回归分析来评估估计 HIV 流行率的一致性。我们使用估计的均方误差和变异系数的比率分析来评估精度。这些模型应用于 2016/2017 年乌干达基于人口的 HIV 影响评估数据，以及来自 2016 年批量质量保证抽样调查和地区卫生信息系统数据集的产前护理数据，分别用于单位层面和地区层面的模型。
基于模型和直接调查方法的估计值相似。然而，与基于模型的估计值相比，直接调查的估计值不稳定。地区层面模型的估计值比单位层面模型的估计值更稳定。单位层面和直接调查估计值之间的相关性为（β1=0.66，r2=0.862），而地区层面模型和直接调查估计值之间的相关性为（β1=0.44，r2=0.698）。与直接调查估计值相比，单位层面和地区层面模型的估计值误差分别降低了 37.5%和 33.1%。
尽管单位层面模型的估计值不如地区层面模型的估计值精确，但它们与直接调查估计值高度相关，且与估计值相关的标准误差比地区层面模型小。当有单位层面的辅助信息时，单位层面模型可以提供更准确和可靠的数据，以支持地方决策。
