Spatial fusion of DHS and MICS surveys with jittered and geomasked locations

arXiv:2405.04928 · stat.ME · Submitted 2024-05-08 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.

Tom: We're looking at the paper: "Spatial fusion of DHS and MICS surveys with jittered and geomasked locations".

Jane: Here's the gist of the paper. Anonymizing GPS locations can bias spatial model’s parameter estimates and attenuate spatial predictions when improperly accounted for, and is relevant in applications from public health to paleoseismology <ref:2405.04928#pg2>.

Tom: This work demonstrates that a newly introduced method for geostatistical modeling in the presence of anonymized point locations can be extended to account for more general kinds of positional uncertainty due to location anonymization, including both jittering (a form of random perturbations of GPS coordinates) and geomasking (reporting only the name of the area containing the true GPS coordinates) <ref:2405.04928#pg2>.

Jane: We further provide a numerical integration scheme that flexibly accounts for the positional uncertainty as well as spatial and covariate information <ref:2405.04928#pg2>.

Tom: Data Sources and Anonymization Types The study applies its method to women’s secondary education completion data in the 2018 Nigeria demographic and health survey (NDHS) containing jittered point locations, and the 2016 Nigeria multiple indicator cluster survey (NMICS) containing geomasked locations <ref:2405.04928#pg2>.

Paper summary: Tom: So, to recap, the authors are using this new scheme to model health and demographic indicators from two major surveys used in low and middle income countries: the two thousand eighteen Nigeria demographic and health survey, or DHS, which has jittered locations <ref:2405.04928#pg1,the 2018 Nigeria demographic and health survey>.

Jane: And they’re also using the two thousand sixteen Nigeria multiple indicator cluster survey, or MICS, which uses geomasked locations <ref:2405.04928#pg1,the 2016 Nigeria multiple indicator>.

Tom: The core thesis is that spatial modeling of these surveys gets better when you correctly account for this positional uncertainty instead of ignoring it.

Jane: They apply this to women’s secondary education completion data from both the DHS and MICS surveys in Nigeria.

Tom: The key result they show is that by accounting for the uncertainty, their model—specifically the MDM approach—estimated the urban effect to be much higher, with a posterior mean and ninety-five percent confidence interval width of one point zero three to zero point three two compared to other approaches.

Jane: But they also include a note that while this might look better on paper, their validation suggests that the reduced confidence interval widths for the MDM approach are actually more anticonservative than those of other methods, which could mean they aren't necessarily giving the most accurate estimates.

Tom: So they’re showing how to fuse these two different types of survey data—jittered and geomasked—into one model using a flexible integration scheme.

Lu: What I find interesting is that the method is flexible in terms of what kinds of positional anonymization it allows, and how it constructs those integration points and weights directly based on the distributional form of the uncertainty resulting from that anonymization.

Meng: From my side, I'm thinking about what this means for deployment. If we use this to model something like health access, we need to know what it costs to run this kind of modeling at scale.

Lalam: I see how this connects back to the literature because it builds on Altay et al., two thousand twenty-two by extending that method to handle uncertainty that can be much larger than what we usually see in DHS surveys <ref:2405.04928#pg1>.

Tom: Right, Lu's right about the flexibility, and Meng’s right about deployment concerns. Jane mentioned how they tested four different approaches—Md, MD, Mdm, and MDM—against each other using metrics like mean square error and continuous rank probability score.

Jane: And the validation showed that accounting for positional uncertainty tends to improve model areal predictions in terms of both CRPS and IS scores.

Tom: But the application findings showed a specific advantage for the MDM approach in estimating that urban effect, even if those confidence intervals were wider than expected.

Meng: If I'm shipping models to production, I need stability. So seeing an improvement in prediction metrics is good, but how stable is this whole framework when we introduce real-world noise?

Lu: That’s the next thing to think about. The paper suggests that the method is relatively quick compared to things like MCMC methods that might be needed for this kind of uncertainty handling.

Jane: So, what does this mean practically for researchers using these specific surveys, like DHS and MICS?

Lalam: It suggests that if you are modeling both DHS and MICS data together, or if the spatial random effects don't vary much over the areas where the observation positions might vary—like within geomasked areas—this method is a strong choice.

Tom: So, we’re looking at a way to handle location uncertainty from two different survey types in one model structure.

Conclusion: Tom: We've covered how this paper, "Spatial fusion of DHS and MICS surveys with jittered and geomasked locations," tackles the problem of positional anonymization biasing spatial models.

Jane: The authors demonstrate a new method that extends existing geostatistical modeling techniques to handle both jittering and geomasking uncertainties using a numerical integration scheme.

Tom: The main point we've hammered home is that this fusion allows researchers to model responses from these surveys jointly while accounting for their different forms of location anonymization in a principled way.

Jane: It’s important because it addresses the real-world problem where official statistics in low and middle income countries often have substantial bias, and third-party surveys like DHS and MICS are key data sources there.

Tom: The application results showed that the MDM approach yielded a higher urban effect estimate, even with its specific confidence interval characteristics.

Lu: From my view, the contribution lies in making the method flexible regarding the types of anonymization allowed and how it automatically constructs integration points based on those uncertainty distributions.

Meng: And from an engineering standpoint, it’s that this can be done relatively quickly compared to other complex methods like MCMC, which is important for practical use.

Jane: Ultimately, the paper advocates for choosing survey positional anonymization methods in tandem with the goals you have for the spatial resolution of your prediction when possible.

Tom: So, it's about giving people a way to handle this kind of data problem without losing predictive power across different data sources.

John Paigea, Geir-Arne Fuglstada, Andrea Rieblera

Department of Mathematical Sciences, NTNU

stat.ME

Submitted: 2024-05-08

Updated: 2026-11-09

Comments: main manuscript: 31 pages, 6 figures, 2 tables; supplemental materials: 10 pages, 4 figures, 7 tables

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 80/100

The gist: Anonymizing GPS locations can bias spatial model’s parameter estimates and attenuate spatial predictions when improperly accounted for…

Key concepts

Jittering
This occurs in DHS surveys where true cluster locations are randomly shifted by a small, uniform distance and direction. It simulates slight inaccuracies in recording the exact GPS coordinates of a location.
Geomasking
Used in MICS surveys, this method replaces precise GPS coordinates with only the name of the administrative area and whether it is urban or rural. It represents a more significant loss of positional detail than jittering.
Positional Uncertainty
This refers to the unknown error or variability in where a data point is actually located on a map. The paper treats this uncertainty flexibly, allowing the model to handle both small random shifts (jittering) and large area-based omissions (geomasking).
Numerical Integration Scheme
This is a mathematical technique used to combine the survey responses with the various forms of positional uncertainty. It allows researchers to calculate results by integrating over all possible true locations, given the observed uncertain locations.

Terminology

Summary

Anonymizing GPS locations can bias spatial model’s parameter estimates and attenuate spatial predictions when improperly accounted for, and is relevant in applications from public health to paleoseismology <ref:2405. p. 0

Where the paper says it

A joint model for DHS and MICS surveys: Spatial modeling with anonymized locations John Paigea,∗ , Geir-Arne Fuglstada , Andrea Rieblera aDepartment of Mathematical Sciences, NTNU, Trondheim, Norway Abstract Anonymizing the GPS locations of observations can bias a spatial model’s parameter estimates and attenuate spatial predictions when improperly accounted for, and is relevant in applications from public health to paleoseismology. In this work, we demonstrate that a newly introduced method for geostatistical modeling in the presence of anonymized point locations can be extended to account for more general kinds of positional uncertainty due to location anonymization, including both jittering (a form of random perturbations of GPS coordinates) and geomasking (reporting only the name of the area containing the true GPS coordinates).

04928#pg2>. This work demonstrates that a newly introduced method for geostatistical modeling in the presence of anonymized point locations can be extended to account for more general kinds of positional uncertainty due to location anonymization, including both jittering (a form of random perturbations of GPS coordinates) and geomasking (reporting only the name of the area containing the true GPS coordinates) <ref:2405. p. 0
Where the paper says it

A joint model for DHS and MICS surveys: Spatial modeling with anonymized locations John Paigea,∗ , Geir-Arne Fuglstada , Andrea Rieblera aDepartment of Mathematical Sciences, NTNU, Trondheim, Norway Abstract Anonymizing the GPS locations of observations can bias a spatial model’s parameter estimates and attenuate spatial predictions when improperly accounted for, and is relevant in applications from public health to paleoseismology. In this work, we demonstrate that a newly introduced method for geostatistical modeling in the presence of anonymized point locations can be extended to account for more general kinds of positional uncertainty due to location anonymization, including both jittering (a form of random perturbations of GPS coordinates) and geomasking (reporting only the name of the area containing the true GPS coordinates).

04928#pg2>. We further provide a numerical integration scheme that flexibly accounts for the positional uncertainty as well as spatial and covariate information <ref:2405. p. 0
Where the paper says it

In this work, we demonstrate that a newly introduced method for geostatistical modeling in the presence of anonymized point locations can be extended to account for more general kinds of positional uncertainty due to location anonymization, including both jittering (a form of random perturbations of GPS coordinates) and geomasking (reporting only the name of the area containing the true GPS coordinates). We further provide a numerical integration scheme that flexibly accounts for the positional uncertainty as well as spatial and covariate information.

04928#pg2>.

Data Sources and Anonymization Types

The study applies its method to women’s secondary education completion data in the 2018 Nigeria demographic and health survey (NDHS) containing jittered point locations, and the 2016 Nigeria multiple indicator cluster survey (NMICS) containing geomasked locations <ref:2405.04928#pg2>. For DHS surveys, cluster locations are jittered within a relatively short distance of the respective true locations by a uniform distributed random direction from 0 to 2π and by a uniform distributed radius from 0 to the maximum jittering distance <ref:2405. p. 8

Where the paper says it

The DHS cluster locations are jittered within a relatively short distance of the respective true locations. DHS cluster locations are jittered in a uniform distributed random direction from 0 to 2𝜋, 9

04928#pg2>. In contrast, MICS cluster locations are geomasked, meaning only the administrative area containing the cluster GPS coordinates along with the clusters urban/rural level is known <ref:2405. p. 5
Where the paper says it

NDHS cluster centroids in urban areas are jittered by up to 2 km, and 99% of rural locations are jittered by up to 5 km with 1% of them being jittered by up to 10 km. In MICS surveys, cluster locations are geomasked, which means that only the administrative area containing the cluster GPS coordinates along with the clusters urban/rural level is known (Khan and Hancioglu, 2019; National Bureau of Statistics (NBS) and United Nations Children’s Fund (UNICEF), 2017).

04928#pg2>.

Methodology for Positional Uncertainty

The authors extend a method proposed in Altay et al. (2022) to account for more general forms of positional anonymization, and where the resulting positional uncertainty is potentially much larger than in DHS surveys <ref:2405. p. 3

Where the paper says it

In this paper, we extend the method proposed in Altay et al. (2022) to account for more general forms of positional anonymization, and where the resulting positional uncertainty is potentially much larger than in DHS surveys.

04928#pg2>. They introduce a numerical integration scheme that extends the model to jointly model responses from the surveys while accounting for their different forms of location anonymization in a principled way <ref:2405. p. 2
Where the paper says it

, 2022b; Marquez and Wakefield, 2021), and finer and finer scale estimation of health and demographic indicators necessitates more and more data. DHS and MICS surveys use two different kinds of anonymization, and hence exhibit two different kinds of positional uncertainty, but a framework has not yet been developed to jointly model responses from the surveys while accounting for their different forms of location anonymization 3

04928#pg2>. The likelihood is extended by integrating over the true cluster locations, conditional on the observed locations and positional uncertainty distributions <ref:2405. p. 0
Where the paper says it

A joint model for DHS and MICS surveys: Spatial modeling with anonymized locations John Paigea,∗ , Geir-Arne Fuglstada , Andrea Rieblera aDepartment of Mathematical Sciences, NTNU, Trondheim, Norway Abstract Anonymizing the GPS locations of observations can bias a spatial model’s parameter estimates and attenuate spatial predictions when improperly accounted for, and is relevant in applications from public health to paleoseismology. In this work, we demonstrate that a newly introduced method for geostatistical modeling in the presence of anonymized point locations can be extended to account for more general kinds of positional uncertainty due to location anonymization, including both jittering (a form of random perturbations of GPS coordinates) and geomasking (reporting only the name of the area containing the true GPS coordinates).

04928#pg2>.

Model Validation and Performance

The validation process compares four approaches (Md, MD, Mdm, MDM) against each other to evaluate performance across different data combinations and uncertainty handling methods <ref:2405.04928#pg2>. The validation metrics considered include mean square error (MSE), continuous rank probability score (CRPS), interval score (IS), fuzzy empirical coverage, and 95% fuzzy CI width <ref:2405. p. 0

Where the paper says it

We apply the method to women’s secondary education completion data in the 2018 Nigeria demographic and health survey (NDHS) containing jittered point locations, and the 2016 Nigeria multiple indicator cluster survey (NMICS) containing geomasked locations. We show that accounting for the positional uncertainty in the surveys can improve predictions in terms of their continuous rank probability score.

04928#pg2>. Results show that accounting for positional uncertainty tends to improve model areal predictions in terms of both CRPS and IS <ref:2405. p. 0
Where the paper says it

We apply the method to women’s secondary education completion data in the 2018 Nigeria demographic and health survey (NDHS) containing jittered point locations, and the 2016 Nigeria multiple indicator cluster survey (NMICS) containing geomasked locations. We show that accounting for the positional uncertainty in the surveys can improve predictions in terms of their continuous rank probability score.

04928#pg2>.

Application Findings

The application to women’s secondary education in Nigeria using the 2018 NDHS and 2016 NMICS shows that the MDM approach estimated the urban effect to be much higher with a posterior mean (and 95% CI width) of 1.03 (0.32), compared to other approaches <ref:2405. p. 0

Where the paper says it

paige@ntnu. no (John Paige) Preprint submitted to Spatial Statistics May 9, 2024 arXiv:2405.

04928#pg2>. The validation from Section 4 suggests a more nuanced interpretation, as the reduced CI widths of the MDM approach are more anticonservative than those of other approaches rather than necessarily leading to more accurate estimates <ref:2405.04928#pg2>. The paper concludes by advocating for survey positional anonymization methods to be chosen in tandem with goals for the spatial resolution of prediction where possible <ref:2405.04928#pg2>.

Key Contributions

The proposed method is flexible in terms of the kinds of positional anonymization allowed, the likelihood families allowed, and the method by which integration points and weights are automatically constructed based directly on the distributional form of the positional uncertainty resulting from the anonymization <ref:2405.04928#pg2>. It can also be performed relatively quickly compared to alternative methods suitable for such contexts such as those involving MCMC <ref:2405.04928#pg2>. The method is recommended particularly in cases where the modeler wishes to model DHS and MICS surveys jointly, and in situations where the spatial random effects do not vary substantially over spatial scales on which the uncertain observation positions may vary, such as within geomasked areas <ref:2405.04928#pg2>. It is also possible that accounting for uncertainty from urbanicity classification may improve predictions, although that is a complex problem beyond the scope of this work <ref:2405. p. 6

Where the paper says it

Urbanicity classification is neither transformed nor normalized, since it is a binary variable. Urbanicity classifications are generated based on thresholding untransformed population density surface so as to ensure urban and rural populations are as close to the estimated numbers as possible, with any deviations removed through calibration as per Paige et al.

04928#pg2>. The results suggest that DHS surveys’ jittered cluster positions may lead to more stable and generally improved inference when compared to MICS surveys’ geomasked cluster positions due to the relatively small level of positional uncertainty resulting from jittering in DHS data <ref:2405. p. 9
Where the paper says it

Hence, the distribution for the MICS cluster positions given the true locations is much simpler, with 𝜋(𝑠 MICS 𝑐 | 𝑡 DHS 𝑐 ) = I{𝑠 MICS 𝑐 ∋ 𝑡 MICS 𝑐 }, where ‘∋’ denotes ‘contains’, and recalling that 𝑠 MICS 𝑐 is the geomasked area for cluster 𝑐 of the MICS survey. Likelihoods under positional uncertainty We extend the model proposed in Altay et al.

04928#pg2>.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Joint modeling of DHS and MICS data can be performed jointly to produce a joint spatial model accounting for anonymization of spatial positions in DHS and MICS surveys jointly. This allows systems to leverage both survey datasets simultaneously, which is crucial when dealing with sparse or costly data sources like MICS while benefiting from the larger sample sizes of DHS.

  2. Accounting for positional uncertainty can improve predictions, as the paper demonstrates that accounting for the positional uncertainty in the surveys can improve predictions in terms of their continuous rank probability score. This suggests that incorporating methods to handle jittering and geomasking will lead to more accurate spatial risk estimations.

  3. The proposed method is flexible because it can account for more general kinds of positional uncertainty due to location anonymization, including both jittering (a form of random perturbations of GPS coordinates) and geomasking. This flexibility allows AI systems to adapt the modeling framework to different anonymization techniques found in various surveys.

  4. The numerical integration scheme can be tailored based on the positional uncertainty, as it is flexible accounts for the positional uncertainty as well as spatial and covariate information. This means a single method can handle diverse data characteristics efficiently by adjusting how it samples or integrates over possible locations.

  5. The system can achieve computational efficiency compared to existing methods, noting that our proposed method is able to jointly model DHS and MICS data, and it can do so at a fraction of the time of existing approaches that can be applied in such contexts. This reduces the processing time required for complex spatial risk modeling.

  6. The system can provide more nuanced predictions by using metrics like Continuous Rank Probability Score (CRPS) or Interval Score (IS), which are strictly proper and proper scoring rules, leading to a better measure of predictive accuracy compared to simple mean square error.

Abstract

In low- and middle-income countries, survey data is a key data source for public health monitoring. Not accounting for the anonymization applied to the GPS locations in these surveys may lead to bias, and in turn sub-optimal public health decision-making and outcomes. We propose a spatial fusion model to account both for jittering (a form of random perturbation of GPS coordinates) and geomasking (reporting only the name of the area containing the true GPS coordinates). This allows us to combine data from surveys with different kinds of positional uncertainty, reduce bias, and give more precise estimates. We use this approach to estimate women's secondary education completion through a fusion model using both the 2018 Nigeria demographic and health survey (DHS), with jittered point locations, and the 2016 Nigeria multiple indicator cluster (MICS) survey, with geomasked locations. We show that accounting for the positional uncertainty in the surveys and accounting for both DHS and MICS surveys simultaneously can reduce bias while improving precision in parameter estimates and predictions via a simulation study and validation at cluster and areal levels. We recommend the proposed approach particularly in cases where the modeler wishes to model DHS and MICS surveys jointly, where continuously varying spatial covariates are used, and where local targeted spatial estimates are desired.

Sources

Related papers