Population Forecasting with Nonstationary Multiregional Growth Matrices
Sweeney, Stuart H., Konty, Kevin J., Geographical Analysis
Though the mathematics of multiregional population projections were defined over twenty years ago, and the methodology has seen some adoption internationally, most practitioners in the United States still use rudimentary cohort component projections techniques. Both the stationarity assumption and the implicit five-year retrospective time scale imposed by the census migration data have probably contributed to the limited use of multiregional projections methods. This paper reviews previous attempts to overcome the stationarity assumption and proposes a decompositional approach using log linear models estimated via the ECM algorithm. The paper discusses the advantages of the decompositional approach and implements the model for intrastate migration in California.
Population forecasts are one of the primary inputs to a wide range of important planning functions carried out by states and localities in the United States. As has been argued eloquently by Isserman (1984) accurate population forecasts are fundamental to good planning. In long-range planning, large forecast errors translate into costs related to over- or underprovision of physical infrastructure. In rapidly growing states, such as California, there is added urgency for reliable short-term forecasts. Transportation, human resources, and environmental planning all rely on near-term forecasts of changes in the spatial population distribution.
In practice, the methods used to construct both long- and short-range population projections have remained rudimentary. Though multiregional population projection methodology was developed by Rogers in the 1960s, simple cohort-component methods are still the most widely taught and applied methodology in practice. This is true despite attempts to write gentle introductions to the methodology for a practitioner audience (Rogers 1985, Isserman 1984) and despite clear demonstrations of the bias introduced through the use of net migration in the cohort-component projection model (Rogers 1990).
Why is it that multiregional projection methods have not found wider use to date? There are several reasons. Like the large-scale models in Lee's requiem (1973), multiregional models are data hungry. Whereas simple cohort-component models can use either residually measured net migration or ignore migration altogether, multiregional models require a matrix of origin-destination flows. Demographically disaggregate flow data have been only available once per decade and the computational burden needed to access the data containing county-to-county flows, and to a lesser extent state-to-state flows, has been prohibitive for some.
The data landscape is becoming increasingly complex, presenting new challenges and new opportunities. In the United States, the decennial census long form, sampling approximately one-sixth of the population, has long been the primary source of internal migration data capable of supporting detailed socioeconomic and spatial disaggregation. (1) Given mounting political pressure related to privacy concerns and the high cost of the long form, the U.S. Census Bureau has proposed to supercede the long form with the American Community Survey (ACS) by the 2010 Census. (2) The ACS would have a monthly sample of 250,000 compared to approximately 17 million sampled by the 2000 Census long form. Though the monthly ACS will include a 1-year retrospective migration question, providing more timely data, the smaller sample size necessarily means some degradation of the socioeconomic and spatial detail traditionally available in the long form. The ACS/long form trade-offs reflect a general shift in the social science data col lection and dissemination.
Though there is an apparent loss in resolution and statistical power if the long form is indeed eclipsed by the ACS, it is likely that social science data resolution will advance on all fronts: spatial scales, temporal scales, and demographic characteristics. …