Design of a Data-driven Intervention Dashboard for SDG Localization

1 Introduction

Sustainable Development Goals (SDGs) are designed by taking national frameworks into account but the success of SDGs strongly depends on their implementation at local and regional levels [Oosterhof, 2018]. The 17 SDGs with 169 targets and 231 unique indicators 1 may not be feasible or applicable uniformly worldwide. Every community has its unique challenges regarding socio-economic issues, cultural practices, environmental concerns, and government structures. The local stakeholders have a better understanding of the issues of their territory. Thus, the process of SDG localization promotes an understanding of the local context for the implementation of the SDGs at sub-national levels by encouraging national agencies to work in cooperation with the local and regional authorities. The provision of training, expertise, and resources to strengthen the local communities and institutions is part of the SDG localization process. Thus, local and regional governments need effective collaboration within the national framework such that transformational changes can be implemented considering local priorities and concerns.

1 https://unstats.un.org/sdgs/indicators/indicators-list/

To address the SDG localization problem in the Indian state of Karnataka, we use Open Government Data (OGD) for our intervention dashboard. In our present study, we focus on maternal mortality/deaths (MD) indicator from SDG 3: Good Health and Well-Being. Our intervention dashboard is built by combining health-related data from five different OGD sources - Karnataka At a Glance 2 (KAG), National Family Health Survey 3 (NFHS-4), Karnataka Health Profile 4 , Men and Women in Karnataka 5 and eJanMa - Births, Deaths and Stillbirths Registration 6 . The intervention dashboard demonstrates predictive and prescriptive modeling to understand the impact of policy interventions [Bassin et al., 2021; Bassin, 2022] on maternal health and also provides recommendations on budgetary allocations for various related schemes. Section 2 explains the methodology for performing prescriptive analysis. Section 3 presents a case study on reducing MD by executing the best course of action - 'what to do'.

Related Literature. Data visualization tools have been extensively developed for monitoring and reporting the progress of SDG indicators. [Garwood et al., 2020] cater a dynamic collaborative dashboard for business schools and universities for SDG information sharing. [Bissio and Watch, 2016] report on diverse and independent dimensions such as gender equity, basics capabilities, etc. Dashboards on equity monitoring in vaccination coverage [Arsenault et al., 2017], urban water security [Zainuddin et al., 2023], health coverage [Azcuna, 2016] and, climate change [Canepa et al., 2022] are primarily engaged as reporting tools for SDGs. Sustainable Development Report [Sachs et al., 2023], SDGUrban Index and Dashboard 7 , India SDG Dashboard 8 and, NIUA SDG Dashboard 9 also act as monitoring and reporting tools to evaluate the status and achievement of various SDGs. We present an intervention dashboard 10 to study the differential impact on different regions of the state based on interventions at multiple levels. Also, a prescriptive modeling approach is followed to prescribe the magnitude of interventions required to lower MD annually. Consequently, the impact and stability of recommendations are also calculated. The dashboard goes beyond reporting and monitoring SDG indicators. It assists policymakers with listing factors deserving the most attention, regions with the highest proportion of beneficiaries, and thus a suitable distribution of the funds. Hence, our work provides actionable insights to policymakers by going beyond predictions, monitoring, and reporting tools.

2 https://kgis.ksrsac.in/kag/

3 https://rchiips.org/nfhs/NFHS-4reports/Karnataka.pdf

4 https://nhm.karnataka.gov.in/info-2/Demography+and+ Evaluation+Cell/Publications/en

5 https://des.karnataka.gov.in/info-2/Publication+Training+and+

Co-ordination+(PTC)/Reports/en

6 https://ejanma.karnataka.gov.in/Sitemap.aspx

2 Prescriptive Modeling with Stability Mapping

We divide the methodology into prescriptive modeling with a 'what-to-do' analysis and stability mapping as a subsequent scenario.

2.1 Prescriptive Modeling

To optimize SDG localization strategies, this study takes a prescriptive modeling approach. For a given geographical unit say a, multiple factors (F a 1 , F a 2 , ..., F a k ) associated with the outcome variable, Oa are evaluated based on the multivariate linear regression model as shown in equation 1:

For SDG localization, it is crucial to identify factors that influence MD and regulate them to mitigate the problem. To approach a given target for MD, equation 2 determines the new value for a factor Fj for district a, (denoted as new F aj ), based on the given state target, Otar, and the current value of the indicator, Ocurr.

where, oldF aj is the original value of a factor. αj denotes the sensitivity of the factor Fj where αj ∈ [0, 1]. αj is external to the model and can be calibrated by the policymaker to adjust the impact of a factor on the outcome variable.

2.2 StabilityMapping

Our previous work [Rachuri et al., 2023] on SDG 2 indicators, demonstrated how a single intervention impacts multiple crops in different Agro-climatic Zones. We proposed a novel approach of Stress Modeling to understand the behavior of a system before and after an intervention. We use a modified approach here to infer the impact and stability of interventions.

7 https://sdgindiaindex.niti.gov.in/urban/

8 http://www.sdgindia2030.mospi.gov.in/dashboard/

9 https://niua.in/niua-sdg-dashboard

10 https://wsl.iiitb.ac.in/intervention-dashboard-maternal-deaths/

The stress of a district is the disparity between its impact and the impact of its neighboring districts. Interventions even with a high impact on a given region may not be sustained if the disparity in outcomes is too high. Stable interventions are those that result in minimal stress.

For stability modeling, we prepare an undirected graph G = (V,E), where the nodes V represent the set of administrative regions under consideration. Currently, each node is a district of the state of Karnataka. Edges, (u, v) represent adjacency between two districts. Let the set of neighbors of a given node v be represented as Γ(v). In a generic sense, if m indicators were to be impacted with an intervention, each node v ∈ V is associated with a vector c(v) of m dimensions. Given this, the stability of a node v is given as,

where, L(., .) is a suitable norm function to calculate the distance between two vectors. Here, we have just one indicator, [MD] associated with each v.

3 Case Study: InterventionModeling for Maternal Deaths

For MD, we analyzed 115 attributes across all the five datasets mentioned in section 1. Of these, based on consultations with domain experts, we narrowed down to four significantly correlated factors, namely: Births Delivered by a Caesarean Section (CS), Births Delivered in a Health Facility (HF), Female Leprosy Patients (LP), and Female Tuberculosis Deaths (TB). Hence, equation 1 takes the form:

Figure 1 illustrates these associations.

As explained in the demo video 11 , we divide our intervention dashboard into four parts to illustrate an efficient and lucid approach to policymakers for the given SDG indicator, Maternal Deaths.

3.1 Differential Impact Analysis

We analyze how the predicted change in a target variable for a district compares with the neighboring districts' change. Thus, by changing factor Fj by x%, the predicted change in the target variable Oa is calculated as,

11 https://youtu.be/0Slg39gf1og

3.2 Prescriptive Analysis

To reduce maternal deaths from the current state average to a chosen target value, the policymaker may choose to set sensitivities for each parameter based on its influence on the indicator. In figure 2, αHF = 1.0, αCS = 0.8, αLP = 0.0 and αTB = 0.4. The resultant bar chart shows the new prescribed values vs. the old values for the factors. We see the biggest drop in female tuberculosis deaths. Hence, TB deaths should decline drastically to lower the overall MD in the state. We also notice that institutional deliveries need to be carried out throughout the state uniformly 12 to increase survival prospects for mother and baby. Thus, with a combination of prescribed interventions, the state average for maternal deaths can be lowered to 10 from 19.42.

3.3 Stability vs. Impact Analysis

Corresponding to the new prescribed values, districts are distributed into high/low impact and stability zones. Figure 3 shows 5 districts that fall into the most desirable zone with high impact and high stability for the new set of interventions. Dharwad (DWD) district shows high impact but relatively low stability. This means DWD may not be sustained in this state for the long term. Thus, policymakers need to engage more with the districts with low stability to carry out sustained improvements toward maternal health and well-being.

3.4 Budget Allocation

The prescribed values from the dashboard aid in efficient budget allocation for the various schemes that will promote awareness toward maternity benefit programs. Hence, the amount (A) allocated for a scheme will be,

12 https://hfw.delhi.gov.in/fw/janani-shishu-suraksha-karyakram

where k denotes the total number of factors and βj denotes the coefficient of a factor. From figure 2, schemes related to HF will get the largest share of the budget followed by CS and then TB.

Schemes and Beneficiaries

For n districts, the amount allocated to a district a for a scheme corresponding to factor Fj is calculated as,

where F a j is the value of the factor Fj for district a.

The total number of beneficiaries, B, for a district under the scheme corresponding to factor Fj is calculated as,

where uFj is the unit cost of the scheme.

4 Conclusions, Limitations and FutureWork

Policymakers need data-driven recommendations to reach expected targets. The intervention dashboard illustrates data stories for strategic decision-making by prescribing the magnitude of interventions for mitigating maternal mortality and thus distributing funds appropriately in relevant schemes. This aids in optimizing region-specific SDG localization strategies where some regions may have disparate responses towards the same set of interventions.

As a public-facing demonstration, we have limited ourselves to using datasets available only at the district level. Previously, multiple dashboards in collaboration with the Planning Department of the Government of Karnataka were developed using proprietary datasets at a more granular level.

Hereafter, we aim to enhance the dashboard with indicators from other domains and pursue a deeper understanding of lateral impact and internal stress among the indicators based on some common interventions.

Ethical Statement

There are no ethical issues. All the datasets used for our study are available in the public domain and free to download from government websites.

Acknowledgments

The authors thank the Planning and Statistics Department, Government of Karnataka, India, and the Machine Intelligence and Robotics (MINRO) Center funded by the Government of Karnataka, India for supporting this work. The authors also like to thank Dr. Mukund Raj from the United Nations Development Programme (UNDP) for his expertise and insightful discussions, and Manish Reddy Koppula for his assistance with modeling and analysis.