Unit: Digital Planet, The Fletcher School, Tufts University
Reports to: Head of Research
Hours: Part-time (20 hours) and full-time (35 hours), hourly available.
Rate: $30-$50, DOE
Duration: Immediate start through March 15, 2027, with possible extension
Location: Remote, working on US Eastern or Pacific hours. Must be based in the United States and authorized to work in the US without sponsorship.
About the work
Digital Planet is building a computational model and open-access interactive platform on AI readiness in low- and middle-income countries. The model classifies populations into segments according to what kind of AI-enabled service can realistically reach them, drawing on indicators of connectivity, affordability, device access, and use. It is fitted as a multi-country, multi-year Bayesian panel model covering 27 countries, with an indicator set that is still expanding. Estimates are produced at national and sub-national level and disaggregated by gender and location.
We are looking for an experienced modeler to work directly on the model itself, not only on the data feeding it. This is a temporary position of 20–35 hours per week, fully remote within the United States. The team itself works across US Eastern and Pacific time, so either coast works well, but we do ask for some flexibility to overlap with colleagues in both zones for check-ins and review sessions.
Responsibilities
• Work on specification, estimation, and testing of the probabilistic segmentation model, including diagnostics on calibration and cross-country consistency
• Investigate anomalies in model output and propose specification fixes, with the analysis to justify them
• Derive and test the weighting scheme that combines indicators into segment assignments, with weights estimated from data rather than assigned by judgment wherever possible
• Build and maintain reproducible pipelines to clean and harmonize indicator data from a range of international and national sources
• Reconcile indicators across differing vintages, definitions, and geographic levels; document assumptions and source hierarchies
• Produce diagnostic tables, charts, and sensitivity analyses for internal and external review
• Write clear code documentation and methodology notes suitable for public release
Required
• Bayesian analysis, specifically hierarchical and latent-variable specifications. Prior specification, partial pooling, posterior and convergence diagnostics, and judgment about how much structure a small number of units can support. With 27 countries, priors and shrinkage carry real weight, and experience with panel methods built for large-N asymptotics transfers only partly.
• Critical thinking. Able to interrogate a modeling choice or a data source rather than accept it, and to say plainly when something looks wrong.
• Strong communication. Can explain a method, a diagnostic result, or a caveat clearly, in writing and in conversation, to both technical and non-technical readers.
• Strong quantitative and coding skills. Comfortable reasoning about probability, uncertainty, and distributions, not only running estimation code.
• A track record of building comparable models. Not only familiarity with the methods, but prior work in which you specified and estimated a model of similar structure yourself, and derived or validated a weighting scheme. Measurement or latent-trait index models, composite indicators, and population segmentation or classification models are all close analogues.
• Comfort with unbalanced panel data. Indicator coverage differs by country and year, and the gaps are not random. You should have a considered view on when to impute, when to marginalize, and when to let the measurement model absorb the missingness.
Also important
• Python. Fluent for data and modeling work (pandas, NumPy, SciPy) and comfortable in a version-controlled repository. Experience with a probabilistic programming framework such as PyMC, NumPyro, or Stan is an advantage.
• Building for extension. The indicator set is still growing. Specifications and code should accommodate new series and revised definitions without a rebuild each time.
• Self-sufficiency. Can pick up an existing codebase, make progress with limited direction, and raise blockers early.
• Attention to detail. Rigorous about units, vintages, provenance, and documentation.
• Readiness to learn on the job. The framework, indicators, and country context are specific to this project. We expect a ramp-up period and will support it.
Nice to have
• Research ability. Can find and synthesize relevant literature to inform a methodological or measurement decision.
• A different angle. We are a small team that has been working inside this framework for a while. Someone who approaches problems differently from how we do is genuinely valuable to us.
• Experience with dynamic latent trait or cross-national measurement models of the kind used to build governance, democracy, or human rights indices
• Familiarity with household survey programs such as DHS, LSMS, or Global Findex, or with sub-national administrative data
• Graduate training or professional background in statistics, economics, data science, demography, or a related quantitative field
• Interest in digital access and AI readiness in low- and middle-income country contexts
To apply
Send a CV and a short description of one or two past modeling projects to christina.filipovic@tufts.edu. For each, tell us: what the model was meant to estimate, what you specified and estimated yourself, how you knew whether it was working, and what you would do differently now. We are more interested in how you reason about a model than in the length of your CV.
A link to a code sample, repository, paper, or technical appendix is welcome. Applications are reviewed on a rolling basis and we are looking to fill the position immediately.