GEOSTATS: Forecasting development from spatio-temporal data
- Mr. Spero,
I would not try to model future development at the census block level, particularly for the whole state of California. The errors in estimation will be so large as to make your projection completely unreliable and unusable. At the block level, the construction or demolition of one large apartment building could make any projection completely unrealistic. You would be much better off using larger aggregates (e.g., jurisdictions) since the base numbers are much more reliable and change more consistently from decade to decade.
Regarding the method, I'm a little puzzled at the data you have. The census long form has a question about the age of the building and, if my memory serves me right, gives the number of units existing that were built in certain decades (it's found on the STF3A sample file). But that's only a snapshot of what buildings remain, not the number that were originally built during those decades. That is, if a building was built in 1939 but replaced in 1959, the 2000 census would not document the original construction. If that is your data source, then you cannot use it for a projection since you don't have an accurate measure of the number of constructed units. On the other hand, if you actually did have the number of units constructed in any one decade, then it might be possible to make a projection. However, since building booms tend to go up and down according to the business cycles, I'm not sure it's a very reliable source. You might want to check with the Construction Industry Board in Burbank who follow construction trends in California.
The points you raise about spatial autocorrelation and density limits in a community are valid ones. Very few construction models have taken these into account (and would need to for a reliable estimate). But in addition to density limits and the adjacency of nearby units, there are a variety of local land use policies that significantly affect future construction levels (i.e., pro- and anti-growth policies). I've put together two different types of forecast that you might want to look at. In the first, I modeled housing unit change between 1980 and 1990 as a function of 1980 density, pro-growth policies, and growth control policies for virtually every jurisdiction in the state. Initially, I used a spatial lag regression model which takes into account spatial autocorrelation (using Luc Anselin's SpaceStat program), but found that the density variable captured most of the variance associated with the spatial autocorrelation (which turned out to be negative, incidentally, as high growth jurisdictions tended to be located adjacent to low growth jurisdictions). You can find this study in
Ned Levine, The effects of local growth management on regional housing production and population redistribution in California, Urban Studies. 1999. 36 12, 2047-2068.
Second, I've been constructing population and employment projections for the State of California between 1990 and 2020. The method starts with California Department of Finance projections for each county up to 2020 and then allocates the growth to each jurisdiction and the unincorporated areas of counties (within each county separately) based on the 1980-90 growth rate. Thus, the county totals are identical to that of the DOF, but the relative changes in jurisdictions are based on recent population trends. I then allocated the projected population of each jurisdiction to 3 mile square grids using kernel density estimation and my CrimeStat spatial statistics program (http://www.icpsr.umich.edu/NACJD/crimestat.html ).
Since the kernel density method takes into account spatial autocorrelation with a reasonable model of covariance between nearby jurisdictions, the result is a smooth gradient of population density change over the entire state. Kriging may have produced a slightly better estimate, but it didn't seem worth the effort since the errors in any projection are much greater than the differences between different models of interpolation. It would be a relatively easy task to convert the projected population into future housing units by dividing population by an estimate of average household size, though I didn't do that. You can see the projections on my web page
At 08:46 AM 9/11/2000 -0700, you wrote:
-- *To post a message to the list, send it to ai-geostats@.... *As a general service to list users, please remember to post a summary of any useful responses to your questions. *To unsubscribe, send email to majordomo@... with no subject and "unsubscribe ai-geostats" in the message body. DO NOT SEND Subscribe/Unsubscribe requests to the list!
I have a block level census data map (California, USA) with attributes that give the density (per sq mi) of structures for the years 1939, 49, 59, 69, 79, 89 and 90. I want to use the time-series to project future development, i.e., the number of new structures in each block, say, by 2020. What is the right way to model this? I hypothesize that very low and very high density areas are special cases, in that some places are not imminently developable due to geophysical barriers, etc, while very tightly packed areas (i.e., very small blocks) are essentially "built out." There is spatial autocorrelation to deal with, over an obviously anisotropic landscape, and the nature of urban development has shifted over time from urban infilling, to suburban expansion, to exurban growth in the foothills.
Any suggestions will be heartily appreciated.
Jim Spero, Fire Economics Analyst
Fire and Resource Assessment Program
California Department of Forestry and Fire Protection
1920 20th Street
Sacramento, CA 95814
Ned Levine, PhD
Ned Levine & Associates