Alternative digital, administrative, or sensor-based sources can fill gaps in traditional surveys but often introduce sampling bias that distorts projections. The Intergovernmental Panel on Climate Change and the United Nations Global Pulse have documented the growing use of such alternative data for policy and research, highlighting both opportunities and risks. Understanding who is captured and who is omitted is central to trustworthy inference.
Causes and mechanisms of sampling bias
Bias arises when data generation is linked to variables of interest. Mobile phone metadata, social media activity, or transaction records typically overrepresent urban, younger, and wealthier populations in many regions. Judea Pearl UCLA has emphasised that unobserved selection mechanisms create confounding between the selection process and outcomes, while Donald Rubin Harvard University developed frameworks showing how nonrandom assignment undermines direct comparisons. Cultural and territorial factors magnify these effects: language preferences, platform censorship, and differential infrastructure access create systematic undercoverage of marginalized rural communities and indigenous territories, producing projections that can misinform resource allocation.
Adjusting projections for sampling bias
Practical adjustments start with assessing representativeness against reliable benchmarks such as censuses or administrative registries. Andrew Gelman Columbia University advocates multilevel regression with poststratification as a robust approach to reweight nonrepresentative samples to known population margins. Propensity score weighting and calibration methods developed from Rubinian causal frameworks mitigate selection effects by modeling the probability of inclusion conditional on observed covariates. Bayesian hierarchical models can propagate uncertainty from weighting into final projections, and sensitivity analyses guided by Judea Pearl UCLA style causal graphs reveal the impact of plausible unobserved selection. When benchmarks are scarce, combining multiple alternative sources and applying small area estimation can reduce variance while explicitly modelling bias.
Consequences and ethical dimensions
Failure to adjust leads to systematically biased forecasts, policy missteps, and inequitable service delivery, particularly for populations already disadvantaged. Transparent documentation of data provenance, methodological choices, and remaining uncertainties strengthens credibility. Projections should therefore treat alternative data as complements to, not replacements for, representative sampling, and should articulate the limits of inference when coverage gaps remain.