A major shift in U.S. land development avoids significant losses in forest and agricultural land

Analysis is primarily conducted with the National Resources Inventory (NRI), a plot-level longitudinal land-use database compiled by the U.S. Department of Agriculture's (USDA) Natural Resources Conservation Service covering the 1982–2015 period and comprising a sample of over 800 000 points (USDA 2018). A key feature of the NRI database is that it tracks changes between all major land-use categories. Our main focus is on trends in urban and built-up (or 'developed') land, including conversion from the four major undeveloped NRI land classes: forest, cropland, pasture, and range. Additional detail on the land-use classes contained in the NRI data are provided in the supplemental materials (SM) appendix A (available online at stacks.iop.org/ERL/17/024007/mmedia). The NRI data have been used in a number of previous empirical studies of land-use change (e.g. Lewis and Plantinga 2007, Lubowski et al 2008, Lawler et al 2014, Bigelow and Kuethe 2020). We aggregate the plot-level NRI data to the county level (the finest geographic resolution possible) to generate our main findings concerning the spatial and temporal pattern of land development across the U.S. Data on other variables that enter the analysis come from well-known publicly available sources, such as the U.S. Census Bureau and Energy Information Administration, and are described in greater detail in SM appendix A. Our county-level dataset includes the 3024 counties in the conterminous U.S. accounted for in all datasets used in the analysis.

We begin by presenting a set of descriptive trends that depict how the rate of land development has changed over the 1982–2015 study period. In doing so, we draw comparisons with concurrent trends in large-scale socioeconomic drivers (population, income, and gasoline prices). To study the spatial distribution of the land development trend, we decompose the National trend into strata based on the 1980 quartiles of population density, household income, and commuting cost, and provide further breakdowns by U.S. state, region, starting land use, and urban classification as measured by the USDA Economic Research Service's (ERS) 2013 Urban Influence Codes (USDA, ERS 2013). We also consider the changes in the spatial distribution of the ratio of population to developed area, a metric we term the 'developed area population density', between 1982–2000 and 2000–2015. Calculation of the developed area population density is described in detail in SM appendix A.

The second section of results uses the plot-level NRI land-use transition probabilities to simulate what the 2015 landscape would have looked like if the pre-2000 land development rate had continued through the first 15 years of the 21st century. To compute the cumulative avoided land development, we adopt an approach used in prior land-use simulation studies (Lewis and Plantinga 2007) that accounts for the fact that any avoided land development must alter the composition of undeveloped uses in a way that the total landscape size remains fixed. This process generates transition probabilities P_c,t,j,k defining the probability that an acre of land in county c in use j converts to use k in time t, which we use to describe the likelihood that an acre of land in a county either stays in its starting use or converts to an alternative use over a 15-year (1982–1997) time step. SM table B1 presents these conversion probabilities relative to urban conversions in both the pre- and post-2000 periods. We omit the 1997–2000 period in calculating P_c,t,j,k so that our 15 year conversion probability is applied to a consistent time step. A simulated U.S. landscape in 2015 is generated by fixing the transition probabilities P_c,t,j,k at these initial levels and using them to project how the landscape would have evolved from 2000 to 2015. We then compare our simulated 2015 landscape with the observed 2015 landscape to illustrate how changes in the probability of land development have affected the amounts of land remaining in forest, crop, pasture, and range uses, which account for 95% of all new land developed over our study period (SM figure B1).

The final section of results uses an econometric model to decompose county-level land development trends into constituent components stemming from population growth, income, and commuting costs, which are the major drivers of land development cited in the urban economics literature (e.g. Brueckner 2000, Nechyba and Walsh 2004). The model we estimate is a linear two-way fixed effect regression model of the following form:

$$\begin{align} {\widetilde {{{De}}v}_{c,t}}& = {{{\beta }}_0} + {{{\beta }}_1}P{G_{c,t - j}} + {{{\beta }}_2}{I_{c,t - j}}\nonumber\\ & \quad + {{{\beta }}_3}P_{c\left( s \right),t - j}^{\,G} + {{{\beta }}_4}P_{c\left( s \right),t - j}^{\,G} \times T{T_c} \nonumber\\ & \quad + \gamma 1\left\{ {Stat{e_c} = s} \right\}t + {{{\tau }}_t} + {{{\alpha }}_c} + {{{\varepsilon }}_{c,t}}.\end{align} \tag{ 1 }$$

In equation (1), the dependent variable, ${\widetilde {{\text{De}}v}_{c,t}}$, is measured as the inverse hyperbolic sine (IHS) of the change in developed acres in county $c$ over the year leading up to year $t$ (i.e. the net change in total developed area between $t$ and $t - 1$). Using the IHS transformation allows us to include county-year observations with zero land developed and yields a similar interpretation to a model with a logged dependent variable (Bellemare and Wichman 2020). Explanatory variables enter the model in lagged form, where $j$ denotes the number of years prior to year $t$ at which the variable is measured. Included in the model are population growth over the previous year in county $c$ ($P{G_{c,t - j}}$), median household income in county $c$ (${I_{c,t - j}}$), the price of gasoline the state, $s$, where county $c$ is located ($P_{s\left( c \right),t - j}^{\,G}$), and an interaction term between the price of gasoline and average commuting travel time ($P_{c\left( s \right),t - j}^{\,G} \times T{T_c}$). For the interaction term, $T{T_c}$ is fixed in each county at its baseline 1980 level to avoid confounding changes in unobserved factors affecting commuting time with changes in development (as in, e.g. Molloy and Shan 2013). The purpose of the interaction is to allow for the possibility that gasoline prices have a bigger effect on development in areas with higher average commuting costs.

An important element of our econometric modeling strategy is how we use the panel structure of the data to include county fixed effects (${\alpha _c}$), year fixed effects (${\tau _t}$), and state-specific linear time trends ($\gamma 1\left\{ {Stat{e_c} = s} \right\}t$), where $1\left\{ {Stat{e_c} = s} \right\}$ is an indicator variable for the state, $s$, in which county $c$ is located. County fixed effects absorb features of each county that affect land development and do not change over the time period of our analysis (e.g. climate and geographic amenities that draw migrants such as coastlines, mountains, and other outdoor amenities). Year fixed effects absorb land development drivers that are spatially-invariant but time-varying (e.g. interest rates and macroeconomic shocks, such as the Great Recession). State time trends absorb land development drivers that are specific to each state and alter the trajectory of development over the time period of our analysis (e.g. state government policies that affect land-use, changing regional attractiveness for migration). To account for spatial correlation in the model error term, ${{{\varepsilon }}_{c,t}}$, standard errors are clustered by state.

To depict results from the econometric estimation, we first compute the average change in each explanatory variable over two periods: (a) the last two decades of the 20th century (1980–2000) and (b) the first 15 years of the 21st century (2001–2015). For income and gasoline price, which enter the model in level form, we use the average annual growth in each variable for each of the two periods. For population growth, which already enters the model in change form, we compute the average value in each of the two periods (as opposed to the change in population growth). We then use the estimated model parameters to estimate the percentage effect on land development stemming from the observed average annual change in each explanatory factor over the early (1983–2000) and later (2001–2015) portions of the study period. To highlight the importance of the observed population, income, and commuting cost drivers, relative to other state-specific influences, we combine the county and year fixed effects, along with the state-specific trends, to produce an annualized state-level marginal effect of unobserved land development drivers.

Growth in the developed land area of the United States increased throughout the 1980s and into the 1990s before peaking and undergoing a persistent multidecade decline (figure 1(A)). As of 2015, the current rate of land conversion (0.47 million acres per year), is less than one-quarter of the peak development rate that occurred over 1992–1997 (2.04 million acres per year). The downward trajectory of land development after 1997 predates the Great Recession of 2007–2009 and contrasts with the contemporaneous trend in new housing starts, suggesting that the declining rate of land development is marked by an increase in the density of new housing built rather than a slowdown in construction (SM figure B2).

Zoom In Zoom Out Reset image size

In the last two decades of the 20th century (1982–2000), U.S. population growth lagged behind growth in the developed land base (figure 1(B)), which is consistent with recently documented global trends (Seto et al 2010, Angel et al 2011). In relative terms, the population elasticity of land development, which we measure as the ratio of the annualized percentage change in developed land and the annualized percentage change in population, shows that land development has become increasingly population-inelastic, with the elasticity measure declining from 2.59 in 1982 to 0.7 in 2015 (figure 1(C)). This implies that the stock of developed land has become increasingly dense over time, which is also corroborated by the cumulative and annual change ratios of population and developed land shown in SM figure B3.

The observed trend in land development is in general agreement with trends in two of its major drivers: income and commuting costs. The last two decades of the 20th century were marked by a rise in real median household income, but household income growth was relatively stagnant between 2000 and 2015 (figure 1(D)). Gasoline prices, a primary component of our measure of commuting cost, declined rapidly in the early 1980s and remained low until the early 2000s, when they increased sharply, and have since generally remained higher, though more volatile, than in the later 20th century. Taken together with the trend in land development, the trends in income and gasoline prices accord with basic economic intuition, suggesting that (1) consumption of land increases with income and (2) commuting costs impose a constraint on the geographic extent of growing urban areas.

Lastly, we find that the shift towards denser development patterns has occurred broadly across urban and rural areas that collectively contain a large majority of the US population. Specifically, 83% of the 2015 U.S. population is found in areas that got denser (as measured by the ratio of total population to total developed land area) over 2000–2015 compared to 1982–2000 (SM table B1). Overall, 90% of counties with any developed land area during our study period (SM table B2), and all but one state (Nevada; SM figure B4), have developed areas that became more densely populated over 2000–2015. We further decompose the spatial distribution of development using a county-level federal urban influence classification system. At least 84% of counties assigned to each of the 12 urban influence classes are associated with developed areas that got more densely populated over 2000–2015 relative to 1982–2000 (SM table B2).

In general, we find a remarkably consistent decline in land development over 2000–2015 across various stratifications of the U.S. land base. The development rate peaked in the mid-late 1990s across the quartiles of 1980 county-level population density distribution (figure 2(A)). The densest counties experienced the most dramatic reduction in land development, falling from a peak of 1.29 million acres per year over 1992–1997 to a 2015 level of just 0.23 million acres per year (an 82% reduction). As of 2015, over 80% of land in the highest density quartile and over 95% of land in the bottom three quartiles remains undeveloped (SM figure B5), suggesting that the decline in new land development is not entirely driven by a lack of remaining physical land onto which existing developed areas may expand. A similar pattern of land development has also taken place across the distribution of median household income in 1980 (figure 2(B)). Intuitively, total land development is positively correlated with median household income. Across all quartiles, current levels of annual development represent 73%–79% decreases from their peaks.

Zoom In Zoom Out Reset image size

Counties characterized by above-median average 1980 commuting times generally saw the largest gains in developed area and, subsequently, the largest declines in the developed area growth rate after the turn of the 21st century (figure 2(C)). The pattern of declining land development in counties with the highest commuting times is consistent with prior research documenting a widespread slowdown in the decline of commuting costs around the year 2000 due to increases in congestion and the growing trend of urban renewal (Cuberes et al 2021), as well as increases in gas prices (figure 1(D)).

Figure 2(D) shows the land development rate for land being converted from each of the four major pre-development uses (forest, crop, pasture, and range). In absolute acreage terms, the current forest land development rate of 0.19 million acres per year has fallen the most off its peak of 0.87 million acres per year, amounting to a 78% rate reduction. This implies that ecosystem services from forest land, in particular, have been most affected by recent changes in the rate of land development. Qualitatively similar patterns emerge for crop, pasture, and, to a lesser extent, range. Lastly, the trend in annual land development persists across broad geographic regions of the U.S. (figure 2(E)). Most development in the conterminous U.S. takes place in the Southeast and Northeast/Midwest regions, with the two accounting for roughly 80% of total development in each year in the study period. This regional development pattern accords with the rankings by pre-developed use, as a majority of counties in the Northeast/Midwest and Southeast are associated with forest as the dominant pre-developed use (SM figure B6).

Figure 3 decomposes the avoided land development by state, starting land use, and RPA region. Had the land-use change trajectories from the end of the 20th century continued across 2000–2015, there would have been an additional 7 million acres of land that would have been developed. States with large amounts of avoided land development are generally located east of the Mississippi River or on the Pacific coast. The 36.6% total reduction in newly developed land implies avoided losses across all major undeveloped land-use categories. In general, there is more avoided deforestation than avoided losses in agricultural lands. Avoided deforestation amounted to 3.56 million acres, while avoided cropland loss is 2.06 million acres, most of which is concentrated in the Northeast/Midwest and Southeast regions. The 1.16 million acres of avoided pasture loss is spread more evenly throughout the conterminous U.S. Compared to the other three uses, avoided losses in rangeland are minimal. The spatial distribution of avoided development is attributable to the existing land base. Forty-one percent of avoided deforestation occurs in the Southeast region corresponding to the southeast's 39% share of all U.S. forestland. Similarly, 42% of avoided crop loss occurs in the Northeast/Midwest where approximately 54% of all U.S. cropland is located.

Zoom In Zoom Out Reset image size

Population growth has a (weakly) positive effect on land development (table 1). However, population growth was not markedly different in the early and later portions of the study period (0.89 and 0.82 1000s of persons per year, respectively). We estimate that the average annual population change over 1983–2000 increased annual development by approximately 0.63%, while the same effect in the later portion of the study period was 0.62%. Median household income increased by an average of 0.67 per year ($1000s of $USD) over 1983–2000, and our estimates indicate that this increased land development by approximately 3.29% per year. Income stagnation over 2001–2015 led to a minimal decline in land development of 0.09% per year.

Zoom In Zoom Out Reset image size

Of the three drivers, changes in commuting costs play the largest relative role in driving the observed patterns of land development. Computed at the average county-level commuting time of 19 min, the average annual gasoline price decrease of $0.05 during the last two decades of the 20th century boosted annual land development by 6.06%, while the increase in gasoline prices in the second half of the study period ($0.03 annually) decreased land development by 2.84% per year. The interaction term between gasoline price and 1980 commuting time in the econometric model shows that the relationship between commuting costs and land development depends on the average amount of time commuters spend traveling to their place of work. In the early part of the study period, the average change in gasoline price increased development by 6.11%–7.12% for counties with average commuting times longer than 15 min, while counties with shorter commuting times were unaffected by gasoline price changes (figure 4). Similarly, in the second half of the study period, when gasoline prices were increasing, the estimated decrease in land development was largest for counties characterized by longer commuting times. Counties with average commuting times over 12 min saw a decrease in development ranging from −2.63% to −3.31%, with shorter-commute counties not seeing any impact. Compared to the average level of development over 1983–2000, the commuting cost impact estimates imply that the average annual increase in commuting costs over 2001–2015 avoided a cumulative total of 4.19 million acres of new land development (p = 0.012), or roughly 59% of the total avoided land development estimated with the simulation model.

A notable drawback of our estimation strategy is that we are unable to account for changes in land-use regulations, as well as inherently unobservable time-varying factors such as household locational preferences. To determine the magnitude and importance of these factors not explicitly accounted for as measurable independent variables in our model, the final column of table 1 shows the weighted average annual effect of unobservable factors (with weights corresponding to the number of counties in each state), which we estimate by separately combining the year fixed effects and state-specific trends for the 1983–2000 and 2000–2015 periods. For the early period, our results indicate that if population, income, and commuting costs had remained fixed over this period, annual land development would have been 9.07% lower, on average. After the year 2000, however, the unobservable state trends have a positive average annual effect, but it is not statistically significant. In SM tables C2 and C3, we present annual unobservable effects for each state, which show that the unobservable-induced decline in land development in the early period was most prominent in in several states with fairly stringent land-use regulations (e.g. New Jersey, Massachusetts, and Oregon), which is consistent with the expected effects of these policies. Moreover, the avoided increase in development in these states after 2000 was smallest, but the effects are not significant.

SM table C1 presents the raw model coefficients and standard errors used to generate the percentage effects. SM table C4 shows results from omitting ${{{\alpha }}_c}$, ${{{\tau }}_t}$, and $\gamma 1\left\{ {Stat{e_c} = s} \right\}t$ from (1) and provides justification for their inclusion in the final specification. The percentage effects displayed in table 1 are based on a two-year lag of each explanatory factor. SM table C5 shows results from alternative lags of one and three years. See SM appendix C for additional details on the development of the econometric model and supplemental estimation results.