We measure the scatteredness of development as the percentage of land that is not developed in the square kilometer surrounding the average house in each metropolitan area. The data we use are the best available for this purpose.

Despite its many virtues, the National Resource Inventory (NRI) is unsuitable for this measurement. Based on a sample of the United States, the NRI estimates the amount of land that is under different uses in the nation and in individual administrative areas. If all you want to do is estimate what percentage of land is developed in the United States or in individual states, then the NRI is perfectly suitable. It provides a slightly larger estimate of how much U.S. land was built up in 1992 than our data because it counts larger gaps between buildings (up to 500 feet) as developed. However, the NRI also gives a slightly lower estimate of the growth rate of developed land precisely because it cannot detect the infilling of these gaps. That does not make it better or worse, it just uses a different definition of what land is considered developed.

If, instead of state or national level numbers, you want to get an idea of how much land is developed at the level of individual counties, you start running into trouble with the NRI. The U.S. Department of Agriculture is the agency that produces the NRI data. Right in the first section of USDA’s user guide for the data, it says: “the NRI was NOT designed for analyzing issues at the county level” (emphasis in the original). In case this was not clear enough, the user guide goes on to illustrate this with an example of the confidence intervals that Professor Pendall mentions in his post: for the U.S. as a whole the margin of error for the amount of cultivated cropland is only 0.6%, but for Greyson County KY it jumps up to 46.1%. In our discussion of the NRI in “Causes of sprawl” we are just passing along the USDA’s report.

Our main purpose is not to measure how much land is developed, but instead to measure how scattered is development in individual metropolitan areas (and, in follow-up work, in individual neighborhoods). For this you need to be able to see where houses are located and then study land in the square kilometer around each house you see. Here the problem with the NRI is not its unreliability for small administrative areas, it is simply that you cannot reproduce our calculation of how scattered development is with counts of land in different uses at the level of administrative areas.

Turning from the NRI to our data, while there are many similarities between our 1970s and the 1992 data, we agree with Professor Pendall that there are also some subtle differences in the classification schemes that could have implications for the analysis. We have been very careful in dealing with this issue. The key here is to note that, while land is often redeveloped, it is almost never undeveloped. At the national level, according to the NRI, less than 0.8% of developed land was converted from urban to non-urban uses over the 15-year period 1982-1997. With virtually no undevelopment taking place, urban land in 1992 is the union of pre-1970s development and development taking place between the 1970s and 1992. Thus, we do not need to compare the 1970s and 1990s data directly. We only need to use the 1970s data to split 1992 development between pre-1970s and post-1970s development. To this end, we define `old development’ as land that was classified as urban in both the 1990s and 1970s. We define `new development’ as land that was classified as urban in the 1990s, but was not urban in the 1970s. We also use the 1970s data to account for any conversion between residential and commercial uses. Two additional observations reinforce our confidence in this procedure. First, the national growth rate of development that we record is in the same ballpark as what we see in the NRI for a similar 15-year period. Thus, on average, our classification of 1992 developed area into `new’ and `old’ is probably about right. Second, a comparison of maps produced from the data with images for the 1970s and 1990s shows remarkable accuracy. Since the two data sets, in spite of their slight differences, are intended to measure the same thing, we shouldn’t really be too surprised by this.

While we believe our data measures levels and changes in developed area quite accurately, the main purpose of our QJE article, however, is to study what causes differences in sprawl across U.S. metropolitan areas. For the purposes of our statistical analysis, the key is not whether there is some noise in the data but whether any such noise is systematically related to the unexplained component of sprawl. We have no reason to suspect such a correlation. Professor Pendall’s post mentions population growth through foreign immigration and federal lands. We looked at both population growth and proximity to public lands (and the results are mentioned in the text of our paper). We should also make it clear that we do not intend a “no problem” subtext. We do not make any judgment about whether there is too much or to little development, whether it is too scattered or too compact. We measure sprawl and explain its causes.

Diego Puga and Matthew Turner

my tacit comment was that the conclusions in Randall’s post weren’t quite correct.

Thank you for doing this work and making these concerns explicit – I should have brought those out with a little more detail.

BTW – John Carruthers is an old prof of mine. I’ll pass along the compliment. )

Regards,

Dan Staley

This article uses data sources that aren’t longitudinally compatible. I asked a colleague — Greg Taff, remote sensing and GIS maven at the UNC Chapel Hill Geography Department — to comment on the sources. Here’s what he had to say:

“The 2 datasets that the ‘Causes of sprawl’ article uses are not very compatible for the purpose of locating new exurban areas. The 1976 NLCD data requires 20% urban landcover to be classified as urban, and the 1992 NLCD data requires 30% urban landcover to be classified as urban. From experience in my lab at UNC, this 10% difference in definition can result in quite large classification differences. Thus, the 1992 NLCD classification procedure would classify far fewer areas as urban than would the 1976 NLCD classification procedure, if they were based on the same time point, since the 1992 procedure requires 10% higher urban cover to be classified as such. This may be why they are seeing less (particularly exurban) land development than you have seen with the National Resources Inventory [NRI] data.”

The NRI, by contrast, is a panel sample of the same data points over a series of years (1982, 1987, 1992, 1997 at the sub-national level; up to 2002 in national reporting). It avoids the problem of incompatibility and is, intrinsically, therefore a better source for measuring gross land-use change for multi-county areas than the combination of 1976 and 1992 data that the authors use. (By the way, the “1976″ data actually represent 12 years of data from 1971 to 1982, with 1976 being just the median year.)

I think the authors’ main concern is probably the “10 acre” rule, which classifies as “urban” any undeveloped land of 10 acres or less that is surrounded by urbanized land. In regions where development is very scattered (southern New England, for example, where wetlands and their regulation tend to force scattering), this classification process would tend to increase the total amount of land read as “urbanized,” but I don’t know how that would make a difference for the calculation of change over time. It would seem to be just as important at the beginning as at the end of the time period.

The authors also contend that the NRI is not appropriate for analysis at the sub-state level because it’s a sample. According to what I know, however, the broad land use categories are ground-truthed to a greater extent than the other data in the NRI, which is otherwise based mainly on aerial photography. (And anyway, why should we believe the data on Maricopa or San Diego County less than the data on Rhode Island? It’s a matter of the number of sample points.)

At the national level, the USDA NRCS does publish confidence intervals for the estimates of developed land. In 1982, 72.8 million acres were developed in the contiguous (+/- .4M @ 95% c.i.). By 2002, that figure jumped to 107.3 million acres (+/- .7M). Granted, that’s still “only” 5.5% of the contiguous U.S. land area, but it’s up from 3.8% in 1982. And a 47% increase in developed land is something we should pay attention to. (See http://www.nrcs.usda.gov/technical/land/nri02/landuse.pdf for the data.)

Also, evidence from the 2002 NRI suggests strongly that the pace of land development increased sharply between 1992 and 2002 compared with the previous decade. Since the paper stops in 1992, the “no problem” subtext is, I think, premature. 1976-1992 had two periods of substantial economic downturn, making it a less active period of land conversion than 1992-2002. The 1992-2002 period accounted for about 60% of the urbanized land for the two decades.

I was also a little surprised that the authors didn’t look at foreign immigration and growth in the foreign-born as a driver of population density, and they didn’t use the share of federal land either. These have been found to be significant and important influences on sprawl in other work.

John Carruthers’s work on sprawl, fiscal structure, and municipal fragmentation is still the best in the literature, in my opinion. No references to any of that in the paper, however.

Rolf

Marina Alberti at UW does landcover change modeling, some of it for NSF, but for Puget Sound region. Check out some of her work. Her older work is with the 30m LandSat pixel resolution, but she also has higher res. for other work as well. She has encountered the same issue wrt developed/grass pixel analysis & seems to have started to overcome it & has developed indicators of development and urban ecosystem degradation.

Best,

Dan Staley