A long-term consistent socioeconomic dataset of Chinese cities generated by Bayesian spatiotemporal modeling with multi-source Earth observations

Tang, Zhangying; Tang, Xianteng; Liao, Lingfeng; Yan, Guoqiang; Wang, Zhenyan; Wu, Yuju; Xie, Mingyu; Zhang, Yumeng; Wang, Chengwu; Wang, Zhoufeng; Zeng, Yangting; Song, Chao; Pan, Jay

doi:10.5194/essd-2026-306

Preprints

https://doi.org/10.5194/essd-2026-306

Preprints

08 May 2026

| 08 May 2026

Status: this discussion paper is a preprint. It has been under review for the journal Earth System Science Data (ESSD). The manuscript was not accepted for further review after discussion.

A long-term consistent socioeconomic dataset of Chinese cities generated by Bayesian spatiotemporal modeling with multi-source Earth observations

Zhangying Tang, Xianteng Tang, Lingfeng Liao, Guoqiang Yan, Zhenyan Wang, Yuju Wu, Mingyu Xie, Yumeng Zhang, Chengwu Wang, Zhoufeng Wang, Yangting Zeng, Chao Song, and Jay Pan

Abstract. Within the Healthy Cities and Sustainable Development Goals (SDGs) agendas, socioeconomic data are fundamental for tracking regional development. China, however, lacks a complete, long-term subnational socioeconomic dataset due to severe spatiotemporal missingness in official statistical yearbooks. We compiled 35 official socioeconomic indicators for 366 Chinese cities from 2000 to 2021, incorporated remote-sensing-derived covariates as auxiliary information, and applied a Bayesian spatiotemporal interacting varying intercepts (BSTIVI) model to capture the target variables’ spatial, temporal, and coupled spatiotemporal dependence. Model performance was evaluated using global Bayesian criteria and cross-validation, while local error distributions and temporal trends were visualized to examine imputation outcomes. Based on the completed dataset, we further derived a composite development index using entropy weighting and assessed spatial inequality with the Gini coefficient, coefficient of variation and hotspot analysis. The results show that BSTIVI achieved markedly better fit than traditional multiple linear regression (MLR). In cross-validation, 32 of 35 indicators achieved R² >= 0.95, RMSE and MAE remained low. The resulting data product showed strong imputation performance in both spatial and temporal dimensions. Analyses of the completed dataset revealed marked spatial inequality and clustering in urban socioeconomic development across China during 2000–2021. We ultimately produced the first long-term city-level socioeconomic dataset for China, comprising 35 indicators and one composite index, with Bayesian credible intervals for imputed values. This study provides both a new city-level data resource for China and a transferable framework for imputing missing subnational socioeconomic data worldwide, thereby supporting Earth system research and SDG implementation.

Received: 22 Apr 2026 – Discussion started: 08 May 2026

Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.

Download & links

Zhangying Tang, Xianteng Tang, Lingfeng Liao, Guoqiang Yan, Zhenyan Wang, Yuju Wu, Mingyu Xie, Yumeng Zhang, Chengwu Wang, Zhoufeng Wang, Yangting Zeng, Chao Song, and Jay Pan

Status: closed

RC1:
'Comment on essd-2026-306', Anonymous Referee #1, 19 May 2026
Overall evaluation:

This manuscript addresses an important issue: the lack of long-term, consistent, city-level socioeconomic data for China. The authors attempt to compile 35 socioeconomic indicators for 366 Chinese cities from 2000 to 2021, use auxiliary Earth observation variables, and apply a Bayesian spatiotemporal model to impute missing values. The topic is relevant to urban studies, regional development assessment, and SDG monitoring, and the proposed dataset could be valuable if its quality and methodological reliability are fully demonstrated. However, the manuscript currently has several major weaknesses regarding the theoretical basis of the indicator system, the transparency of original data collection, the explanation of missing-data mechanisms, the selection of auxiliary variables, the benchmark model comparison, and the consistency of the final spatial coverage. These issues directly affect the credibility of the dataset and the reliability of the derived composite development index.
The selection of the 35 socioeconomic indicators and the construction of the composite development index are not sufficiently justified. The manuscript does not clearly explain why these specific indicators were selected, what theoretical or policy framework they are based on, or how they collectively represent urban socioeconomic development. If the index is constructed simply by combining a set of economy-related variables, without considering the conceptual dimensions of development, the relationships among indicators, multicollinearity, information redundancy, and the actual socioeconomic meaning of each variable, the resulting composite index may have limited interpretability and practical value. The entropy weighting method only provides a statistical weighting scheme; it does not justify the indicator system itself. The authors should clarify whether the indicator selection is based on SDG indicators, healthy city frameworks, regional development theory, previous literature, or official evaluation systems. They should also test the correlation structure among indicators and assess whether the composite index is sensitive to indicator selection or weighting methods.

The temporal coverage and missing-data problem require much more explanation. The study only uses data from 2000 onwards, but statistical yearbooks may contain longer historical records. The authors should explain why the dataset starts in 2000 and whether earlier years were unavailable, inconsistent, or excluded for methodological reasons. More importantly, the manuscript states that many indicators have substantial missing values. However, some of these variables appear to be basic routine statistical indicators that local statistical departments would normally collect every year. The authors should therefore clarify the reasons for missingness. Are the missing data due to the absence of official statistics, changes in administrative boundaries, changes in statistical definitions, non-reporting by local governments, incomplete publication in yearbooks, or incomplete data collection by the authors? This distinction is critical. If the missingness mainly results from incomplete access to yearbook materials rather than true absence of official data, the nature and reliability of the imputation task would be quite different. The authors should provide a detailed missing-data summary by indicator, city, and year, and discuss the missing-data mechanism more rigorously.

The manuscript does not provide enough information on how the city-level statistical yearbooks were obtained. It only states that the data were collected from local statistical yearbooks, but does not specify the exact data acquisition channels. Were the yearbooks obtained from official statistical bureau websites, printed yearbooks, library archives, commercial databases, local government portals, or other sources? This information is essential for assessing data quality and completeness. If the authors relied mainly on publicly available electronic yearbooks or online data, missingness may partly reflect incomplete digitization, especially for earlier years and smaller cities. Therefore, the authors should provide a clear description of the data collection workflow, source types, quality control procedures, and criteria for determining whether a value is truly missing.

The selection of auxiliary variables used for imputation lacks sufficient theoretical justification. The manuscript uses meteorological and remote-sensing-related variables as auxiliary covariates, but it does not clearly explain why these variables are appropriate predictors for the 35 socioeconomic indicators. In particular, among the ten auxiliary variables, a large proportion are meteorological variables, together with NDVI and a few other Earth observation variables. The authors should explain the theoretical assumptions linking these variables to annual changes in socioeconomic indicators. For some indicators, such as fiscal revenue, education, healthcare, employment, or industrial output, the direct relationship with climate variables may be weak or indirect. If an auxiliary variable has no plausible causal or predictive relationship with a socioeconomic indicator, its inclusion may introduce noise or spurious correlations. Moreover, the same set of ten auxiliary variables appears to be used for imputing all 35 indicators, which may not be appropriate because different socioeconomic variables have different drivers. The authors should conduct indicator-specific variable selection, report the statistical contribution or posterior effects of the auxiliary variables, and provide sensitivity tests to show whether these covariates actually improve imputation performance.

The terminology used to describe the auxiliary variables is inaccurate. The manuscript and tables appear to refer to all auxiliary variables as “remote sensing factors” or similar terms. This is not correct, because most of the selected variables are meteorological or climatic variables, while only a small number are truly remote-sensing-derived variables such as NDVI. This is a basic but important terminology issue. The authors should revise the wording throughout the manuscript, tables, and figure captions. A more accurate expression would be “auxiliary environmental covariates,” “Earth observation and meteorological variables,” or “remote-sensing and climate-derived covariates,” depending on the actual data sources.

The model comparison is insufficient to support the claim that the Bayesian spatiotemporal model is superior. The manuscript mainly compares the proposed BSTIVI model with multiple linear regression. However, multiple linear regression is a relatively weak baseline for complex spatiotemporal imputation tasks. Demonstrating better performance than MLR does not necessarily prove that the Bayesian model is the best or even a clearly superior option. There are many alternative methods that could be considered, including random forests, gradient boosting, Gaussian processes, spatiotemporal kriging, matrix completion, multiple imputation methods, and other machine-learning or hierarchical spatiotemporal models. The authors should either provide a stronger justification for using MLR as the only benchmark or add comparisons with several more competitive baseline methods. Without such comparisons, the conclusion about the superior performance of the proposed model should be stated more cautiously.

The spatial coverage and completeness of the final dataset are not clearly demonstrated. The manuscript claims to produce a dataset for 366 Chinese cities, but some figures appear to show spatial gaps, with certain prefecture-level cities lacking data. This seems inconsistent with the stated full coverage of 366 cities. The authors should clarify exactly which cities are included in the dataset and provide a complete city list, preferably with administrative codes and boundary information. They should also distinguish between observed values and imputed values for each city and year. In addition, the manuscript should explain why the final composite development index appears to be available for all cities even though some underlying indicators or spatial units seem to have missing data in the figures. Clear maps or tables showing the spatial distribution of observed data, missing data, and imputed data would greatly improve transparency and reproducibility.
Citation: https://doi.org/10.5194/essd-2026-306-RC1
RC2:
'Comment on essd-2026-306', Anonymous Referee #2, 29 Jun 2026
The authors present a valuable long-term socioeconomic dataset for Chinese cities. The integration of remote sensing data with a Bayesian spatiotemporal interacting varying intercepts (BSTIVI) model is robust and novel. I have some suggestions regarding the methodological details, uncertainty propagation, and data usability.
In the data pre-processing section, the authors state that they "carefully examined the time series of each socioeconomic indicator for each city…" to identify and remove outliers. This description is too subjective and vague for a data descriptor paper. I suggest explicitly stating the exact criteria or algorithms used for outlier detection and removal in the manuscript to ensure reproducibility.

In Section 2.1.2, the manuscript states that remote sensing data were aggregated to the city scale. However, the exact aggregation method is not specified. Please clarify whether this was a simple arithmetic mean, median, maximum, or a weighted calculation. Different aggregation methods can introduce substantial variations in the model results.

Regarding the spatiotemporal interaction structure (Wit) in the BSTIVI model, does the model assume this structure to be stationary? Considering that China has experienced dramatic economic transformations, rapid urbanization, and shifting regional policies over the study period, the underlying spatiotemporal dependency structures of socioeconomic indicators are likely highly non-stationary. A brief discussion on this assumption and its potential impact on the long-term imputation stability would strengthen the manuscript.

One of the advantages of the Bayesian framework presented here is the generation of credible intervals, providing an interpretable measure of uncertainty for the imputed values. However, when constructing the composite index using the entropy weight method, it is unclear if this uncertainty was considered.

While the authors note that ignoring spatiotemporal non-stationarity may reduce prediction accuracy for extreme values, it would be helpful to discuss the direction of this bias. Specifically, does the model tend to overestimate or underestimate values for highly developed coastal megacities versus underdeveloped western cities?

Minor Comments:
(1) Missing Variable Definition. Equation (2), the term Rst appears but is not explicitly defined in the text. Please provide the definition for clarity.
(2) Data Usability and Format. I have reviewed the dataset provided by the authors. While the CSV files are useful, providing the data in a spatial format (e.g., Shapefiles) is highly recommended to enhance intuitive visualization and broader usability.
Citation: https://doi.org/10.5194/essd-2026-306-RC2

Status: closed

RC1:
'Comment on essd-2026-306', Anonymous Referee #1, 19 May 2026
Overall evaluation:

This manuscript addresses an important issue: the lack of long-term, consistent, city-level socioeconomic data for China. The authors attempt to compile 35 socioeconomic indicators for 366 Chinese cities from 2000 to 2021, use auxiliary Earth observation variables, and apply a Bayesian spatiotemporal model to impute missing values. The topic is relevant to urban studies, regional development assessment, and SDG monitoring, and the proposed dataset could be valuable if its quality and methodological reliability are fully demonstrated. However, the manuscript currently has several major weaknesses regarding the theoretical basis of the indicator system, the transparency of original data collection, the explanation of missing-data mechanisms, the selection of auxiliary variables, the benchmark model comparison, and the consistency of the final spatial coverage. These issues directly affect the credibility of the dataset and the reliability of the derived composite development index.
The selection of the 35 socioeconomic indicators and the construction of the composite development index are not sufficiently justified. The manuscript does not clearly explain why these specific indicators were selected, what theoretical or policy framework they are based on, or how they collectively represent urban socioeconomic development. If the index is constructed simply by combining a set of economy-related variables, without considering the conceptual dimensions of development, the relationships among indicators, multicollinearity, information redundancy, and the actual socioeconomic meaning of each variable, the resulting composite index may have limited interpretability and practical value. The entropy weighting method only provides a statistical weighting scheme; it does not justify the indicator system itself. The authors should clarify whether the indicator selection is based on SDG indicators, healthy city frameworks, regional development theory, previous literature, or official evaluation systems. They should also test the correlation structure among indicators and assess whether the composite index is sensitive to indicator selection or weighting methods.

The temporal coverage and missing-data problem require much more explanation. The study only uses data from 2000 onwards, but statistical yearbooks may contain longer historical records. The authors should explain why the dataset starts in 2000 and whether earlier years were unavailable, inconsistent, or excluded for methodological reasons. More importantly, the manuscript states that many indicators have substantial missing values. However, some of these variables appear to be basic routine statistical indicators that local statistical departments would normally collect every year. The authors should therefore clarify the reasons for missingness. Are the missing data due to the absence of official statistics, changes in administrative boundaries, changes in statistical definitions, non-reporting by local governments, incomplete publication in yearbooks, or incomplete data collection by the authors? This distinction is critical. If the missingness mainly results from incomplete access to yearbook materials rather than true absence of official data, the nature and reliability of the imputation task would be quite different. The authors should provide a detailed missing-data summary by indicator, city, and year, and discuss the missing-data mechanism more rigorously.

The manuscript does not provide enough information on how the city-level statistical yearbooks were obtained. It only states that the data were collected from local statistical yearbooks, but does not specify the exact data acquisition channels. Were the yearbooks obtained from official statistical bureau websites, printed yearbooks, library archives, commercial databases, local government portals, or other sources? This information is essential for assessing data quality and completeness. If the authors relied mainly on publicly available electronic yearbooks or online data, missingness may partly reflect incomplete digitization, especially for earlier years and smaller cities. Therefore, the authors should provide a clear description of the data collection workflow, source types, quality control procedures, and criteria for determining whether a value is truly missing.

The selection of auxiliary variables used for imputation lacks sufficient theoretical justification. The manuscript uses meteorological and remote-sensing-related variables as auxiliary covariates, but it does not clearly explain why these variables are appropriate predictors for the 35 socioeconomic indicators. In particular, among the ten auxiliary variables, a large proportion are meteorological variables, together with NDVI and a few other Earth observation variables. The authors should explain the theoretical assumptions linking these variables to annual changes in socioeconomic indicators. For some indicators, such as fiscal revenue, education, healthcare, employment, or industrial output, the direct relationship with climate variables may be weak or indirect. If an auxiliary variable has no plausible causal or predictive relationship with a socioeconomic indicator, its inclusion may introduce noise or spurious correlations. Moreover, the same set of ten auxiliary variables appears to be used for imputing all 35 indicators, which may not be appropriate because different socioeconomic variables have different drivers. The authors should conduct indicator-specific variable selection, report the statistical contribution or posterior effects of the auxiliary variables, and provide sensitivity tests to show whether these covariates actually improve imputation performance.

The terminology used to describe the auxiliary variables is inaccurate. The manuscript and tables appear to refer to all auxiliary variables as “remote sensing factors” or similar terms. This is not correct, because most of the selected variables are meteorological or climatic variables, while only a small number are truly remote-sensing-derived variables such as NDVI. This is a basic but important terminology issue. The authors should revise the wording throughout the manuscript, tables, and figure captions. A more accurate expression would be “auxiliary environmental covariates,” “Earth observation and meteorological variables,” or “remote-sensing and climate-derived covariates,” depending on the actual data sources.

The model comparison is insufficient to support the claim that the Bayesian spatiotemporal model is superior. The manuscript mainly compares the proposed BSTIVI model with multiple linear regression. However, multiple linear regression is a relatively weak baseline for complex spatiotemporal imputation tasks. Demonstrating better performance than MLR does not necessarily prove that the Bayesian model is the best or even a clearly superior option. There are many alternative methods that could be considered, including random forests, gradient boosting, Gaussian processes, spatiotemporal kriging, matrix completion, multiple imputation methods, and other machine-learning or hierarchical spatiotemporal models. The authors should either provide a stronger justification for using MLR as the only benchmark or add comparisons with several more competitive baseline methods. Without such comparisons, the conclusion about the superior performance of the proposed model should be stated more cautiously.

The spatial coverage and completeness of the final dataset are not clearly demonstrated. The manuscript claims to produce a dataset for 366 Chinese cities, but some figures appear to show spatial gaps, with certain prefecture-level cities lacking data. This seems inconsistent with the stated full coverage of 366 cities. The authors should clarify exactly which cities are included in the dataset and provide a complete city list, preferably with administrative codes and boundary information. They should also distinguish between observed values and imputed values for each city and year. In addition, the manuscript should explain why the final composite development index appears to be available for all cities even though some underlying indicators or spatial units seem to have missing data in the figures. Clear maps or tables showing the spatial distribution of observed data, missing data, and imputed data would greatly improve transparency and reproducibility.
Citation: https://doi.org/10.5194/essd-2026-306-RC1
RC2:
'Comment on essd-2026-306', Anonymous Referee #2, 29 Jun 2026
The authors present a valuable long-term socioeconomic dataset for Chinese cities. The integration of remote sensing data with a Bayesian spatiotemporal interacting varying intercepts (BSTIVI) model is robust and novel. I have some suggestions regarding the methodological details, uncertainty propagation, and data usability.
In the data pre-processing section, the authors state that they "carefully examined the time series of each socioeconomic indicator for each city…" to identify and remove outliers. This description is too subjective and vague for a data descriptor paper. I suggest explicitly stating the exact criteria or algorithms used for outlier detection and removal in the manuscript to ensure reproducibility.

In Section 2.1.2, the manuscript states that remote sensing data were aggregated to the city scale. However, the exact aggregation method is not specified. Please clarify whether this was a simple arithmetic mean, median, maximum, or a weighted calculation. Different aggregation methods can introduce substantial variations in the model results.

Regarding the spatiotemporal interaction structure (Wit) in the BSTIVI model, does the model assume this structure to be stationary? Considering that China has experienced dramatic economic transformations, rapid urbanization, and shifting regional policies over the study period, the underlying spatiotemporal dependency structures of socioeconomic indicators are likely highly non-stationary. A brief discussion on this assumption and its potential impact on the long-term imputation stability would strengthen the manuscript.

One of the advantages of the Bayesian framework presented here is the generation of credible intervals, providing an interpretable measure of uncertainty for the imputed values. However, when constructing the composite index using the entropy weight method, it is unclear if this uncertainty was considered.

While the authors note that ignoring spatiotemporal non-stationarity may reduce prediction accuracy for extreme values, it would be helpful to discuss the direction of this bias. Specifically, does the model tend to overestimate or underestimate values for highly developed coastal megacities versus underdeveloped western cities?

Minor Comments:
(1) Missing Variable Definition. Equation (2), the term Rst appears but is not explicitly defined in the text. Please provide the definition for clarity.
(2) Data Usability and Format. I have reviewed the dataset provided by the authors. While the CSV files are useful, providing the data in a spatial format (e.g., Shapefiles) is highly recommended to enhance intuitive visualization and broader usability.
Citation: https://doi.org/10.5194/essd-2026-306-RC2

Zhangying Tang, Xianteng Tang, Lingfeng Liao, Guoqiang Yan, Zhenyan Wang, Yuju Wu, Mingyu Xie, Yumeng Zhang, Chengwu Wang, Zhoufeng Wang, Yangting Zeng, Chao Song, and Jay Pan

Data sets

City-Level Socioeconomic Indicators and Composite Development Index for China, 2000-2021 Zhangying Tang, Xianteng Tang, Lingfeng Liao, Guoqiang Yan, Zhenyan Wang, Yuju Wu, Mingyu Xie, Yumeng Zhang, Chengwu Wang, Zhoufeng Wang, Yangting Zeng, Chao Song, and Jay Pan https://doi.org/10.5281/zenodo.18217116

Zhangying Tang, Xianteng Tang, Lingfeng Liao, Guoqiang Yan, Zhenyan Wang, Yuju Wu, Mingyu Xie, Yumeng Zhang, Chengwu Wang, Zhoufeng Wang, Yangting Zeng, Chao Song, and Jay Pan

Viewed

Total article views: 255 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
209	30	16	255	13	13

HTML: 209
PDF: 30
XML: 16
Total: 255
BibTeX: 13
EndNote: 13

Views and downloads (calculated since 08 May 2026)

Month	HTML	PDF	XML	Total
May 2026	166	24	12	202
Jun 2026	32	2	3	37
Jul 2026	11	4	1	16

Cumulative views and downloads (calculated since 08 May 2026)

Month	HTML	PDF	XML	Total
May 2026	166	24	12	202
Jun 2026	32	2	3	37
Jul 2026	11	4	1	16

Viewed (geographical distribution)

Total article views: 246 (including HTML, PDF, and XML) Thereof 246 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 16 Jul 2026

Short summary

China’s city statistics often have major gaps, making it hard to track long-term socioeconomic change. We combined official records with satellite observations to create a public dataset for 366 cities from 2000 to 2021, covering 35 socioeconomic indicators and uncertainty estimates. The dataset reveals strong regional inequality and provides a new resource for research on urbanization, sustainability, and city development.


Total:	0
HTML:	0
PDF:	0
XML:	0