How much do demographic models actually predict versus explore?
Core argument: Model choice accounts for 98% of output variance in 2050 global population projections, driving divergent forecasts ranging from 4–25 billion.
We explore demographic predictions by propagating all plausible choices that can be made during the analysis through the modelling process. This approach involves navigating the so-called ’garden of forking paths’—mimicking in silico what would happen if multiple investigators were to examine the same problem. For this, we now abandon [the Chinese government mathematician] Song Jian’s ‘historic’ model and turn to models currently in use: the Cohort-Component and UN WPP models, the Lee-Carter model and the Lotka-Volterra model. Note that in standard use, these tools are used in isolation, see e.g. the FAO’s How to Feed the World in 2050, resting on a single UN WPP population trajectory shielding the reader from the compounding effect of their uncertainty. Unsurprisingly, the exercise capturing the modelling of the modelling process for global population projections to 2050 and 2075 reveals distinct characteristics regarding the sensitivity and projection outcomes of the different models. Once abandoned the straitjacket of Song’s approach, uncertainty is free to manifest itself. The overall uncertainty distributions of the projected populations show a wide range of possible outcomes, reflecting the inherent uncertainties in demographic projections.

