Flood Protection and Land Value Creation – Not all Resilience Investments Are Created Equal

This paper investigates the land value creation potential from flood mitigation investments in a theoretical and applied setting, using the urban area of Buenos Aires as a case study. It contributes to the literature on the wider economic benefits of government interventions and the dividends of resilience investments. Using a simple urban economics framework that represents land and housing markets, it finds that not all flood mitigation interventions display the same potential for land value creation: where land is more valuable (city centers for example), the benefits of resilience are higher. The paper also provides ranges for land value creation potential from the flood mitigation works in Buenos Aires under various model specifications. Although the estimates vary largely depending on model parameters and specifications, in many cases the land value creation would be sufficient to justify the investments. This result is robust even in the closed city configuration with conservative flood damage estimates, providing that the parameters remain reasonably close to the values obtained from the calibration. Finally, acknowledging that fully calibrating and running an urban simulation model is data greedy and time intensive – even a simple model as proposed here – this research also proposes reduced form expressions that can provide approximations for land value creation from flood mitigation investments and can be used in operational contexts.


Introduction
Cost-benefit analyses of projects aiming at reducing flood risks traditionally consider benefits in the form of avoided loss of life and capital destruction, while the costs consist of the investment and maintenance spending of the project. There is a strong suspicion however that the benefits of urban resilience projects at large and in particular those focusing on urban floods are underestimated. The triple dividend framework (Tanner et al. 2015), for example, makes the case that besides the reduced costs from floods (1st dividend), Disaster Risk Management (DRM) interventions can unlock economic potential through the reduction in background risk (2nd dividend) and allow for possible development co-benefits (3rd dividend), such as shelters doubling as community spaces. Whether we consider the triple dividend framework or other closely linked terminologies such as "wider economic benefits" or "secondary benefits" of resilience, the notion that benefits from disaster risk management interventions go beyond avoided losses is gaining traction with both economists and practitioners alike. It opens the possibility of addressing (some of) the shortcomings of traditional Cost-Benefit Analyses while also potentially providing a better assessment of the value of various DRM interventions.
The second dividend in particular relies on the reduction of the background risk which can encourage more investments and less risk averse behaviors, both of which are central to economic development. In particular, in urban settings subject to floods, the risk reduction enabled by the DRM project can promote more investments in the flood prone area in the form of more or better housing and business development. For example, in an area flooded annually but located close to many jobs, some households may decide that the benefits of accessible jobs outweigh the costs of the floods. They would then choose to settle there but will be unlikely to invest much in their dwellings and well-being given annual capital destructions. If the flood risk disappears (or is drastically reduced) because of better flood protection, many more households are likely to choose to settle in the area and will invest more in their dwellings as the damages would be lower.
The reduction in risks is captured through land value appreciations when land markets are functional. Estimates suggest that measures such as canal improvements, storm and wastewater upgrades aiming at reducing flood impacts can lead to land value appreciations in the range of 11% -18% (OECD 2011). This opens up the possibility of land value capture from the authorities in order either to pay for the costs of the resilience project or to finance other public priorities. This benefit should be captured in Cost-Benefit analyses.
But local land value creation from resilience investments is not the same thing as land value creation in the urban area. Indeed, reduction in flood risks, will, at the scale of the urban area change the relative attractiveness of land plots, so that while the newly protected areas will see an increase in attractiveness (translated into higher land values) others will see a decrease. To some extent, by protecting some areas from natural disasters and changing the relative attractiveness of locations within the urban area, resilience investments will trigger a transfer in land values across space. 1 Whether these transfers lead to an aggregate increase in land values in the urban area, and if so of what magnitude, is a more complex question. But the aggregate net effect is precisely the question of interest for urban planners 1 3 and city officials hoping to leverage land value creation for fiscal purposes or for recovering part of the resilience investment costs.
Much has been written on the topic of land value creation and capture linked to urban transport investments both to quantify the effects of increased accessibility on land and property values Stiglitz 1979, 1981;Stokenberga 2014;Viguié and Hallegatte 2014;Gupta et al. 2020) and to discuss practical ways in which public authorities can recuperate part of these appreciations through fiscal means (Suzuki et al. 2015;Medda 2012;Peterson 2008;Germán and Bernstein 2018). In comparison, despite an increased interest in the "wider economic benefits" of government interventions (World Bank 2018), and with a few notable exceptions (Grafakos et al. 2019;Smolka 2013), relatively little is known about the potential for disaster risk management investments to create land value appreciation and ultimately to create fiscal space for local governments. This research contributes to filling this gap by focusing on the potential for land value creation from flood mitigation works in urban areas.
We use a simple applied urban economics model (NEDUM-2D) to investigate land value creation from flood mitigation investments (Viguié and Hallegatte 2012;Viguié et al. 2014). This methodology has several advantages. First, because we rely on a simulation model, we can conduct assessments ex-ante, without having to wait until the completion of the investments and the adjustments of land markets to measure and provide estimates of potential land value creation. This is important as one of the objectives of this research is to provide tools that can be used to perform enhanced Cost-Benefit Analyses which are typically required at the project design phase (for example for World Bank project appraisals). Second, such a simulation model captures some of the general equilibrium effects of interventions in the urban area and can document not only the local land value appreciations but also the transfer of land values across locations. As such it provides a more complete picture than hedonic evaluations. We do not argue that this methodology is a substitute to data collection and empirical studies to measure land value creation from resilience investments -in fact, it would be important to confront our ex-ante estimates with on-the-ground measurements in the future -rather that it is a useful complement.
Our main results show (1) that the location of floods in the urban area matter for the potential for land value creation from flood risk mitigation investments, (2) land value creation depends on the amount of in-migration triggered by increased resilience and improved living standards; and the model provides a range using two extreme modeling assumptions: the closed or open city settings, (3) that while land value creation estimates vary widely depending on modeling choices, in most cases, they are either superior to, or represent a significant share (33% or more) of, the initial flood mitigation investments. This is true even in the closed city configuration with conservative flood damage estimates, providing that the parameters remain reasonably close to the values obtained from our calibration. This demonstrates the importance of accounting for land value creation in Cost-Benefit Analyses.
The paper is structured as follows. To illustrate the dynamics at play, "Resilience investments and land value creation in a simple model" section introduces a perfectly circular schematic urban economics model and investigates the impacts of floods and flood mitigation on aggregate land values and household utilities in open and closed cities under different land ownership assumptions and by varying the locations of flood-prone areas. "Application to Buenos Aires" section briefly introduces the NEDUM-2D model and documents its application on the urban area of Buenos Aires. It also details the flooding information used in this study and how it is integrated into the simulation model. "Results" section presents the results in terms of avoided damages and potential land value creation from flood mitigation investments in Buenos Aires. "Approximating land value creation potential" section presents a reduced form expression that can be used to approximate the potential land value creation estimates from the fully calibrated NEDUM-2D model. "Robustness of the analysis and driving forces" section relies on Decision Making under Deep Uncertainty techniques to test the robustness of our central results to changes in modeling parameters and data interpretation. Finally, "Conclusions and discussion" section discusses this research effort, provides avenues for future work and concludes.

Modeling Principles
In this section, we present the principal features and equations of the standard urban economic modeling framework that accounts for floods. We do so briefly, because the basic urban economic model has been extensively described in many papers and books (e.g., (Brueckner 2011;Fujita 1989)) and the introduction of floods in the framework leads to only marginal changes. We use a model inspired by von Thünen (1826), adapted by Alonso (1964), Mills (1967), and Muth (1969) and comprehensively described in Fujita (1989). Assuming that a city is defined by a number of jobs located in a single location, the CBD, this static model is based on two trades-offs, made by two categories of economic actors, and together they characterize a static equilibrium in the urban system: • Households choose their housing location in the agglomeration by arbitrating between larger and cheaper dwellings further from the city center, and increased commuting costs to the city center where all jobs are assumed to be located. • Landowners choose where and how much to invest (i.e., what buildings to construct), as a function of expected rents at each location.

Households
Each household is composed of one representative worker living at a distance r from the city business district (CBD) where all jobs are located. Each worker has to commute once a day to the CBD, and this commuting entails transportation costs T(r). All households are supposed to earn the same income Y and -at the equilibrium -they have the same level of satisfaction described by a utility function U that depends on both the level of consumption of a composite good z and of the size of their dwelling q. The level of rents per square meter (or, equivalently, the annualized real estate price per square meter) is R(r), at each location in the city. Each household maximizes his or her utility function under the budget constraint described in (1) by choosing where to settle in the urban area r, where r is the distance to the CBD, how much housing space q to consume and what to spend on other goods z. The term L represents the annual amount accruing from land rents per household which is recycled in the urban economy in the form of increased incomes.

Developers and Landowners
Developers purchase land from landowners at price P(r), and in doing so incur an annual opportunity cost of capital iP, i, being the real interest rate. Developers first decide whether to allocate land to agricultural purposes or to housing. In the first case their revenue is R a L, where R a is the agricultural rent and L the area of land they own. In the second case, they choose what amount of capital K(r) to invest at each distance from the CBD to produce a housing surface H(r). In this framework, they are thus also the owners of the buildings. The amount of floor space built and the annual land rents depend on an exogenous specific construction technology which displays constant returns to scale but diminishing returns to capital: H(r) = F(K(r), L(r)). These decreasing returns to capital will cap the building heights and as consequence the total amount of residential floor space. Developers will only invest large amounts of capital to build tall buildings when the anticipated rents R(r) they expect will offset the extra building costs. The edge of the city is given by an exogenous and constant agricultural rent R a below which it is no longer profitable to build. At the edge of the city (r f ) we therefore have R(r f ) = R a .
Under the classic assumption of zero-profit condition for developers, which supposes a competitive construction market, the price of land is given by eq. (3): Developers maximize their profit function (2), which consists in identifying the optimal amount of capital to invest at each location (4): The term ρ represents the joined effect of real-estate depreciation and annual taxes paid by landowners on real-estate capital. i is the real interest rate so that iK represents the annual opportunity cost of capital.
The parameter f j expresses the damage to structures caused by floods at location j in the city. It is comprised between 0, when location j is not hit by floods, and 1. The cost of floods is expressed as the capital destroyed when floods hit, or, alternatively, as the capital needed to repair the structure to its pre-flood condition. In this model, the only impact of floods is through capital destruction. All other consequences (e.g., casualties and fatalities) are disregarded, assuming they can be avoided with relevant tools, e.g., early warning and evacuation. These damages from floods depend on the location in the urban area and are not necessarily identical everywhere. They are in turn the result of two factors, the damage caused by floods in location j when these happen, δ j , and the return period of the flood, τ j , with the return period of flood typically ranging from frequent annual events to much more exceptional events occurring every 100 years or more, such that: The term L found in (1) can schematically take on one of two values depending on the type of model used: city models with Absentee landowners or city models with public ownership of land. In the city models with absentee landowners, landowners are supposed to live outside the urban area which means that land rents are not recycled in the local economy in the form of increased incomes. In this case L = 0 . At the opposite end of the spectra, if all land plots are supposed to be owned by a local government (or by all households in equal shares) then L = ∫

Functional Forms and Further Assumptions
We use Cobb-Douglas functional forms with constant returns to define households' utility and the housing service construction function, a choice that is widely shared in urban economics 2 : We also make the following additional simplifying assumptions for "Locations of floods impact household utility and aggregate land values" and "How are the results affected in open city settings or with absentee landowners?" sections but lift them in subsequent sections: • Transport costs increase linearly with distance from the city center, T(r) = p * r, where p is the generalized cost of transportation including the cost of time. • The agricultural rent R a is supposed to be null. This simplification is not necessary but it makes most of the calculations much easier without altering results.
At each distance r from the city center, and assuming a perfectly circular city, we can define the available land as the following increasing function of r: L(r) = lr where l=2π. Although the cities presented here are unrealistic because they assume no land and housing regulations and constraints and rely on linear transport costs, they use realistic parameter values that are calibrated so as to reproduce the main features of the urban area of Buenos (6) U(z, q) = z q where , > 0 and + = 1

Locations of Floods Impact Household Utility and Aggregate Land Values
The appropriate functional form for the housing construction function has been widely debated in the literature with some scholars arguing that a CES with a substitution elasticity between capital and land inputs below one is a better fit because capital to land ratios increase at a slower rate than land prices (Larson and Yezer 2015). However, measuring land values and invested capital appropriately is uneasy and creates empirical challenges which are likely to bias estimates of the substitution elasticity downward. Works that have paid closer attention to this measurement problem tend to find much higher substitution elasticities and ones that are very close to 1, meaning that Cobb-Douglas functions, although not perfect, are an appropriate representation for housing production functions (Thorsnes 1997;Epple et al. 2010 Aires (see "Robustness of the analysis and driving forces" section and the online supplement). The eight cities represented in Fig. 1 differ by the location of the floods: spread across the city (a, b), localized in one quadrant but proportional to the distance to the city 1 3 center (c, d), localized in the periphery (e, f) or localized in central areas (g, h). They also differ by the severity of the floods, with the left column showing in dark gray unmitigated floods destroying 20% of the structures located in flood prone areas (a, c, e, g), and the right column showing mitigated floods in light gray that destroy 10% of the structures located in flood prone areas. It should be noted that the area affected by floods in all localized flood scenarios (c to h subplots) is identical at 25% of the urbanized area, so that differences in household utility and aggregate land values can only be explained by the location of floods and their severity; not the total flood-prone area.
A number of conclusions can be drawn from this figure. The results in terms of household utility and aggregate land values are shown for one specific setting, namely the Closed City with Public ownership of land (CCP). In this setting, the population is exogenous and unaffected by damage caused by floods, i.e. increased or reduced floods are assumed not to affect migration to and from other parts of the country. In parallel, it is assumed that revenues accruing from land rents are entirely recycled into the local economy in the form of a complementary income that each household benefits from in equal shares.
First, household utility is unambiguously reduced when floods are present compared to a "no-flood" situation. But the magnitude of the decrease depends on the severity of the flood and the location of the flood-prone areas. The more severe the flood, the lower the household utility for a given distribution of flood-prone areas. This can be explained by higher construction costs linked to having to maintain or replace structures, which leads to higher rents. The loss in household utility is highest when floods affect the whole city. The loss of utility is also very large when floods are located in the central parts of the urban area. For localized floods that affect only peripheral areas or one quadrant of the city, household utility losses are lower.
Second, aggregate land values are unaffected by floods when these are distributed homogeneously throughout the urban area (a, b) or proportionally to the distance to the city center (c, d). This result is consistent with Avner and Hallegatte (2019), that show, with Cobb-Douglas functions, that higher structure depreciation rates caused by the costs of floods (or risk-based insurance) lead to a decrease in built housing surface that is exactly compensated by higher rents, leaving landowners' profits unaffected. In other terms, flood losses are fully transferred to tenants.
Thirdly, aggregate land values are impacted in different ways when floods are localized in a specific part of the city. When flood-prone areas are peripheral (e, f), aggregate land value are shown to increase slightly (+0.102% -+0.099%). When flood-prone areas are central (g, h), aggregate land values decrease significantly (−12.2% --7.56%) compared to a "no-flood" situation. The reason behind this is that in a closed city framework, as is the case here, the population of the urban area is fixed and makes locational decisions by trading-off localized housing rents and transport costs under the constraint of housing supply that depend on landowners' profit maximization. Where floods occur, the supply of housing space is more limited because the costs of constructions are higher. So that when floods are in the periphery, households cannot enjoy as large dwellings as they would do in the absence of floods and consequently some of them choose to move closer to the city center, making this area, which already displays higher land values, more attractive, triggering an absolute increase in aggregate land values in the city. There is a transfer of land values from the periphery to the center. When floods are central, the opposite happens. Landowners and developers will build less in the most valuable central areas, and more in the periphery, therefore triggering a redeployment of the population and a transfer of land values toward less valuable lands. The net effect is a decrease in aggregate land values. This result is important because it shows that floods have complex effects on land values. On one side, floods have the aggregate effect of reducing housing supply which leads to higher rents and tends to inflate land values. On the other hand, lower localized supply of housing leads to a new locational trade-off for households which will choose to settle elsewhere in the urban area and potentially in lower value places, which would tend to depreciate land values. The net effect depends on the location of the floods.
Lastly, and linked to the previous conclusion, mitigating the frequency or the damage caused by urban floods, for example through public interventions aiming to achieve resilience will have opposite impacts on aggregate land values depending on the location of the floods. When floods are in the periphery, decreasing their severity, they will also decrease aggregate land values. This can be seen from aggregate land values going from +0.102% to +0.099% compared to a "no-flood" situation in figures e) and f). Arguably the effect remains small. Conversely, mitigating central floods, will make the most valuable parts of the city's land more attractive again, thereby leading to an increase in aggregate land values: from −12.23% to −7.56% compared to a "no-flood" situation in subplots g) and h) of Fig. 1. This result is key as it frames the possibility of creating land values from investing in resilience: not all risk reduction investments create value in aggregate (although, locally, it is true). Those that protect the most valuable land in urban areas will create land value in a closed city setting.

How Are the Results Affected in Open City Settings or with Absentee Landowners?
The results in "Locations of floods impact household utility and aggregate land values" section were derived in a specific setting of classic urban economics -namely a Closed City with Public ownership of land (CCP). But how are these results affected in the other three main urban economics settings -Closed City with Absentee landowners (CCA), Open City with Public ownership of land (OCP) and Open City with Absentee landowners (OCA).? The absentee landlords' assumption means that land rents are not recycled into the urban economy and do not accrue in the form of increased incomes to local residents, L = 0 in eq. (1). The Open City's assumptions means that the urban national system is interdependent, such that any pressure upwards or downwards on household utility will result in in-or out-migration respectively such that all households in any city of the system display the same utility level at any given time: u = u ; whereas the population N, becomes endogenous.
When flood prone areas are distributed proportionally to the distance from the city center (configurations a) to d) in Fig. 1), it is possible to show analytically, using Cobb-Douglas functions and building on Avner and Hallegatte (2019), how household utility (u), aggregate land values (ALV) and city populations (N) are affected by floods compared to a no-flood situation in all four polar city configurations (CCA, CCP, OCA and OCP).
C f and γ in Table 1 are parameters that intervene in calculations with , with ρ f being identical over all flood-prone areas (there is only one flood type) and = 1 a . In closed cities (CCA and CCP), two results, already reported in Avner and Hallegatte (2019), are noteworthy. First, household utility is decreased by a factor C f 1 <1, compared to a 'no-flood' situation, translating the fact that reduced housing supply will increase unitary housing rents and lower housing consumption, all else equal. Second, Aggregate 1 3 land values (ALV) remain at their 'no-flood' level as the decrease in housing supply and increase in housing construction costs are exactly compensated by higher housing rents. This result, documented in "Locations of floods impact household utility and aggregate land values" section, is true in the CCA situation also.
In open cities (OCA and OCP), household utility is exogenously given such that it is equal in flooded or non-flooded settings. However, city population will adjust through inor out-migration in order for household utility to remain constant. Two results here are also noteworthy. First, both population and aggregate land values are decreased by a factor C f < 1, with homogenously distributed flood-prone areas, compared to a no-flood situation in both OCA and OCP settings. It should be noted that C f > C f 1 , so that the decrease in population and aggregate land values is higher in open cities than the decrease in utility in closed cities. Second, aggregate land values per household remain at their 'no-flood' level, When floods are central or peripheral, analytic calculations are more difficult to derive. We use simulations informed by the same set of parameters as in "Locations of floods impact household utility and aggregate land values" section and described in "Robustness of the analysis and driving forces" section and the online supplement to provide estimates of how floods affect household utility, aggregate land values and populations compared to a no flood city, when floods are not necessarily distributed proportionally to the distance to the city center (Table 2).
For closed cities, impacts from floods for household utility losses and aggregate land values compared to a 'no-flood' baseline are identical between CCA and CCP models when floods are homogenously distributed in the urban area (cities (a) to (d)). 3 And results for floods which are distributed in the city center or in the periphery are very similar in CCA and CCP models for utilities and aggregate land values (cities (e) to (h)). For open cities, the same conclusions hold between OCA and OCP models: identical impacts on land values and populations for cities (a) to (d) and very similar results for these same variables when floods are central or peripheral (cities (e) to (f)). Table 1 Household utility (u), Aggregate land values (ALV), and city population (N) when floods affect a city compared to a no-flood situation in the four classic urban configurations (open or closed cities with absentee landowners or public ownership of land) when flood-prone areas are proportional to the distance to the city center, corresponding to cases a), b), c) and d) in Fig. 1 Closed Cities Open Cities Absentee Landowners CCA OCA This suggests that the absentee landowner or public ownership of land assumption has limited impacts on our results. One small exception here consists of the results for aggregate land values in OCA and OCP models when floods are peripheral (cities (e) and (f)). Whereas the impacts of floods on aggregate land values are nearly non-existent in both cases, the sign of the change is different: negative, meaning aggregate land value destruction from floods in an OCA setting, positive, meaning aggregate land value creation in an OCP setting.
The difference in impacts from floods are much more significant between open or closed city models. First, because utility is fixed in open city models, populations decrease with floods, leading to much less severe competition for land and smaller aggregate land values. This is true even when floods are homogenously distributed (cities (a) to (d)) where no impacts are reported in closed cities. For these cities ((a) to (d)), the loss of utility in closed cities is measured by C f 1 = C f a , whereas the loss of population and aggregate land values is measured by C f in open cities. With both a < 1 and β < 1, aβ is small (0.3 × 0.4 = 0.12 with our set of parameters), implying that C f 1 is much closer to 1 than C f . This means that the impacts of floods on aggregate land values is much higher in open city settings when population can adjust through in-or out-migration than in closed city settings when population is fixed.

NEDUM-2D Model and Application to the Urban Area of Buenos Aires 4
We use the NEDUM-2D model (Viguié and Hallegatte 2012;Viguié et al. 2014) to investigate the implications of flood mitigation investments on land values in the urban area of Table 2 Household utility, aggregate land values and population variation (in %) compared to a no-flood situation depending on the location of flood prone areas and the severity of floods (damage inflicted to structures when a flood occurs) in the four polar settings of classic urban economics Buenos Aires. NEDUM-2D (Non-Equilibrium Dynamic Urban Model) is an extension of the standard monocentric 5 urban economic model with one income group 6 such as defined by Fujita (1989) but informed with real world data such as transportation times. We have already introduced its main principles, features, equations as well as assumptions in "A simple urban economics model accounting for floods" section so we will focus here briefly on how it departs from the one-dimensional traditional model, what data we provide it with for real-world city applications and how it performs in reproducing observations, in this case on Buenos Aires.
The version of NEDUM-2D we use in this analysis differs from the standard model developed in Fujita (1989) for two main reasons. First, the theoretical model described by Fujita (1989) represents spatial differences solely as a function of the distance to the city center. As the name suggests NEDUM-2D is two-dimensional meaning that it is an urban model that represents urbanization on a map rather than on a single axis. NEDUM-2D uses a grid with cells of variable size (classically 1km 2 or 500mx500m). It can therefore account for spatial differences in land use and accessibility at a much finer scale. The model can represent differences between two cells situated at the same distance from the city center such as the amount of land that can be built upon, land use features and topological constraints such as parks or rivers, or measured transportation times and costs.
Second, the classic urban economy model only represents one means of transport. In NEDUM-2D there are three main transport options: private cars, public transport and walking. For each location in the urban area, citizens choose between walking, public transport and private vehicles, or a combination of these as a means for commuting. The competition between these modes is organized on the basis of their generalized costs (i.e. the total cost including both the cost of time and the monetary costs incurred during the trip to the city center). It is assumed in this study that modal switch does not affect congestion levels and therefore leaves commuting times unchanged. In order to reflect the heterogeneity of citizens' preferences in terms of transport modes we employ a discrete choice model (De Palma and Thisse 1987).
With a limited amount of data describing the size of the population, households' average income, the transport systems summarized by generalized transport costs to reach the CBD for each mode at the grid cell level, land use including areas that cannot be built, housing regulations such as FARs, construction costs, households and developers' behavior, the NEDUM-2D model can reproduce the structure of Buenos Aires in a convincing way. 7 For this exercise we rely on the data collection and processing as well as the calibration efforts undertaken for a previous study on the Buenos Aires area (Avner et al. 2017). The interested reader will find the main calibration and validation details in the online supplement and a complete documentation, including data descriptions and modeling choices, is available in the main text and appendices of Avner et al. (2017).
The previous study is however enriched by a number of features such as the use of a higher resolution grid (500mx500m), the multiple city configurations studied (Open or Closed cities, public ownership of land or absentee landowners), and the introduction of flood damages (see "Flood data and damages before and after resilience investments" section). Figure 2 provides one element of the model validation; more validation information is presented in the online supplement. It shows that the simulated and the actual urban area in 2012 coincide well within the larger Greater Buenos Aires region (which is called GBA+, in black in the figure). The model captures the general size of urbanized area but also its shape and specific urbanization directions along the transportation network. There are however some discrepancies even within the GBA+ boundaries. In particular it can be seen that the map misses some areas toward the North and the North West of the GBA+ region. This can mainly be explained either by the existence of local secondary employment centers that attract settlements or the presence of local amenities.

Flood Data and Damages before and after Resilience Investments
The flood hazard maps used in this analysis were produced by the Government of Buenos Aires in 2018 as part of their Plan Director de Ordinamiento Hidráulico (PDOH) or watershed management plan. The flood hazard maps, corresponding to the Cildáñez, Maldonado and Vega basins, cover an area which is slightly larger than the Ciudad Autonoma de Buenos Aires (CABA) or city of Buenos Aires proper and, therefore not the whole area of study which corresponds to the Area Metropolitana de Buenos Aires (AMBA). It should be noted here that our study focuses on floods that occur mostly in CABA, a vastly smaller area than our area of study. There are two sets of flood maps: before and after works planned in the PDOH, aiming to reduce flood instances. The works planned to mitigate flooding consist mostly of stormwater drainage and retention capacity investments in the three water basins. In both cases a flood hazard map (in polygon form) is produced that provides the extent of a flood for a given return period and for a given flood depth. Return periods range from frequent events occurring every two years to rare events occurring once every 100 years. The list of return periods considered is the following: 2, 5, 10, 20, 50 and 100 years. Flood depths range from 15 cm to 160 cm. The full list of flood depths considered is: 15, 25, 40, 90 and 160 cm. In total, with each flood map being produced after flood protection works and in their absence, there are 59 flood maps. 8 Typically, the area covered by a given flooddepth for a specific flood return period is smaller after flood mitigation actions have been implemented.
Each flood map is intersected (overlayed) with a grid that has homogenous 500 m × 500 m sized grid cells so that the flooded share of land, s j, τ, d, w , of each grid cell (j), is known for each of the 59 flood maps, depending on the return period of the flood (τ), the depth of the flood (d) and whether flood protection works have taken place or not (w). Given that for a given grid cell, a given return period and a given public works scenario, the share of land flooded by 160 cm, is also by construction flooded by 15 cm of water, there is some obvious double counting of flooded areas that needs to be addressed. In a geometric sense, areas that are only flooded by 15 cm of water (not more) constitute the outer concentric ring of areas which are also flooded by 25 cm of water, by 40 cm and so on 9 (see Fig. 3 for a visual representation of flood maps for a centennial event before flood mitigation investments). We address this double counting by defining s _ ring j, τ, d, w as follows: The total damage of floods at the grid cell level expressed in terms of percentage of the (re)construction costs can be expressed through the flood depreciation factor ρ f , which accounts for the share of land in each pixel j which is flood prone as well as the frequency (8 We would expect 60 distinct flood maps for each flood depth, return period and in the absence or after flood protection works: 5x6x2 = 60. However, after flood protection works, there are no locations that get flooded every two years with a water depth of 160 cm, therefore we have 59 flood maps rather than 60. 9 This pattern has some very limited exceptions: out of the sum of 10,699 instances where a grid cell is impacted by a specific flood event (before resilience investments), only in 97 of these do we observe unusual behaviors where a bigger flood-depth is not contained within the area of a smaller flood-depth for any given return periods. We also find that out of the 961 grid cells that get flooded in part or in whole by at least one flood event (before flood mitigation), in only 31 of these does the flood data show bigger flood depth areas that are larger than smaller flood depths. These localized discrepancies can possibly be explained by some small errors in the hydrological modeling or the transcription of these results into shapefiles. We deal with these outliers by assuming that a concentric ring for a smaller flood-depth cannot have a negative value: s j, τ, d, w ≥ s j, τ, d + 1, w .
of the flood τ, and the damage δ d inflicted to buildings which are flooded, as a function of flood depth 10 : The missing piece is interpreting damages in percentage of reconstruction costs as a function of flood depths δ d . For this we rely on flood depth-damage curves reported by two studies, 1) Hallegatte et al. (2013) based on Huang (2005) and 2) Englhardt et al. (2019), linearly interpolating the damages they report as a function of flood depth to match the flood-depth data we have. There is considerable uncertainty in assessing how floods and water depth affect structures (Jongman et al. 2012) and the two studies we use reflect this, with a large spread in values for a given flood depth (see Fig. 4). In both cases we retain for our study the estimates that concern the sturdiest type of building material 11 (category 'masonry' in Hallegatte et al. (2013) )) and yet for a 1 m50 water depth, structural damage is estimated at around 8% and 38% respectively. We choose to rely on the flood damage estimates derived from Hallegatte et al. (2013) for our central scenario, as this is less likely to overestimate the benefits from flood protection investments, but we also show aggregate results derived by using the estimates from Englhardt et al. (2019). With these assumptions we have defined a depreciation rate linked to flood occurrences at the pixel level, depending on flood protection investments, ρ fj, w . We can now turn to the results of the model in terms of land value creation triggered by resilience investments.

Potential Land Value Creation in Closed and Open Cities
As discussed in "Locations of floods impact household utility and aggregate land values" section, the assumption about land ownership (Absentee landowners or Public ownership of land) in the urban area has limited impacts on the results of our simulations focusing on the potential for land value creation linked to resilience investments. On the contrary the assumption of open or closed city models has important implications. For this reason we choose to present our results for the CCP and OCP settings only, in our central scenario using lower bounds flood damage estimates, as there is some evidence that housing ownership rates are high in Buenos Aires (Fay 2005). Figure 5 shows the distribution of land value changes triggered by the flood protection investments in both CCP and OCP settings in our central exercise where we use flood depth-damage estimates from Hallegatte et al. (2013). While in both cases, the mean land value change is positive, 0.29$/m 2 for CCP and 1.74$/m 2 for OCP, signaling aggregate land value creation, the pattern of change is vastly different across both cities.
In the closed city model, more than 95% of grid cells (13,704 out of 14,465) see their land value decrease as a response to the resilience investments. Only 3.3% of grid cells (473) see an increase and the remaining 238 are unaffected. There is a vast transfer of land  Englhardt et al. 2019) values toward areas that benefit from increased resilience to floods from the rest of the urban area. While the average increase in land value per m 2 is large (39.94$/m 2 ), the average decrease is mild (−1.07$/m 2 ). The top histogram of Fig. 5 in particular, demonstrates the importance of accounting for land and housing market equilibrium effects. The positive variation in land values amounts to US$ 2.19 billion according to our central estimations in a CCP setting. The corresponding negative variation in land values is close to US$ -1.81 billion. The actual land value creation is the difference between these two amounts: US$ 379 million. Ignoring the land value transfer/destruction linked to land and housing market adjustments would vastly overestimate the aggregate land value appreciation in the urban area (US$ 2.19 billion vs US$ 379 million). These results strongly support the use of equilibrium models rather than hedonic pricing strategies when the objective is to understand aggregate land value changes from investments rather than purely local ones.

Open City with Public ownership of land (CCP)
In the open city model, conversely, only 20 out of 14,465 grid cells (0.14%) experience a decrease in land value, while grid cells that show an increase in land values represent 98% of the urban area. The average increase in land value is much smaller than in the closed city case (+1.78$/m2) but spread across a much wider area, while the average decrease in land values is much higher (−7.25$/m2) for a limited number of grid cells. Figure 6 shows the spatial distribution of land value variations triggered by the flood mitigation investments in both the closed and open city settings. In the closed city setting, a very small share of the urban area will see increases in land values which can reach up to 60% of their pre-investment levels, essentially the zones that see a direct reduction in flood risks from resilience investments. The rest of the urban area will experience a mild decrease in land values (<2%). In the open city setting, most of the urban area will experience a moderate increase in land values with a few pixels experiencing a more substantial hike. Table 3 shows the impact of flood protection investments on aggregate land values, household utility and population size in both the CCP and the OCP setting and for two flood depth-damage curves; the lower bound one retained for our central scenario that relies on Hallegatte et al. (2013) and the alternative higher bound one that uses the findings from Englhardt et al. (2019).
In total, as shown in Table 3, aggregate land values increase in both the CCP and OCP settings. For our central scenario using lower bound damage estimates, aggregate land values increase by 0.10% and 0.79% respectively. With our central baseline calibration parameters, 12 this translates into net land value creations of US $0.38 and $2.94 billion respectively. Using higher bound damage estimates, the increases in land values are much higher, +0.48% and + 4.57% in the CCP and OCP settings respectively, amounting to US $1.93 and $16.95 billion net land value creation from resilience investments. The large range of 12 A large sensitivity analysis, using techniques adopted for Decision Making under Uncertainty, is performed on most calibrated parameters of the model to explore the robustness of our results as well as which parameters drive our results. This analysis is documented in section 6. It is shown that parameter b, which intervenes in the construction cost function and the real interest rate i have strong impacts on our results and that land value creation ranges from approximately US$ 6 thousand to US$ 4.5 billion depending on their values (excluding 3 of 3000 runs that return negative values). +4.01% these figures shows the importance of the choice of modeling closed or open cities but also of the choice of the flood vulnerability curves. The closed city assumption is more realistic in the short term, but some degree of in-migration due to reduced flood risks, consistent with the open city assumption, could happen in the longer term. In comparison, the flood mitigation works are estimated at US $338 million. 13 This means that the lower bound land value appreciation estimation alone would nearly be able to cover the cost of the resilience investments.
The flood protection investments not only increase aggregate land values, opening up the possibility for land value capture, but are unsurprisingly also welfare improving. In the CCP setting, household utility increases by 0.08% and 0.48% for lower and higher bound flood damage estimates, whereas in the OCP setting the investments trigger an in-migration of approximately 28,000 and 165,000 households (or a 0.68% and 4.01% population increase).
Additional results presented in the online supplement document that whereas relaxing housing regulations in parallel to investing in flood mitigation increases households' utility, it does not translate into higher aggregate land values compared to a scenario in which resilience investments are implemented alone in a closed city setting. This (maybe) surprising result stems from the fact that there are two opposite forces at play: on one hand flood protection in attractive areas leads to higher land values because of lower expected repair costs to buildings and more incentives to build intensively; on the other hand, relaxing housing regulations will reduce housing scarcity and contribute to lowering housing rents in the urban area and thus land values.

Comparison of Land Value Creation and Avoided Damages
The NEDUM-2D model can also shed light on the avoided damages from flood protection investments, that is the amount of capital that does not need to be invested to repair structural damages, i.e. restore structure to their pre-flood integrity. To do so we compute the following formula for aggregate avoided damages AAD.
In eq. (9), f j,a and f j,b are the flood depreciation rates after and before the resilience investments respectively for grid cell j, while K j, b is the invested capital in structures in grid cell j before flood mitigation interventions.
The annualized avoided damages from floods in both the closed and open cities with public ownership of land (CCP and OCP) are US$ 47.4 million and US$ 102 million respectively using the Hallegatte et al. (2013) and the Englhardt et al. (2019) flood-depth damage functions. Avoided damages are identical in CCP and OCP settings because we rely on capital investments pre-resilience investments only, K j, b , which is by construction identical in the baseline in CCP and OCP settings as we start from CCP's household utility in the baseline scenarios. In comparison, annual land value increases in the CCP settings are US$ 7.6 million and US$ 38.6 million with Hallegatte et al. and Englhardt et al. flood damage functions respectively. Corresponding numbers in the open city setting (OCP) are US$ 58.7 million and US$ 339 million. 14 In CCP settings net land value creation potentials therefore represent approximately 16% and 20% of avoided damages whereas in OCP settings, that account for in-migration as a result of increased household welfare because of improved resilience to floods, land value creation represents between 124% and 172% of avoided damages, depending again on the flood depth damage curve function selected. As highlighted earlier, the CCP results are more realistic in the short term, but in the long-term some degree of in-migration can be expected as a result of flood mitigation works, consistent with the OCP setting. In all cases, numbers that we derive from land value appreciation with respect to avoided damages are significant but they are especially high in open city settings. If instead of focusing on aggregate land value creation at the scale of the urban area, we look at land value increases only in the areas of the city that benefit directly from resilience investments, where flood damages decrease, we find that local annualized land value increases are always superior to avoided damages.
We conclude this sub-section with a note of caution: whereas it is interesting and useful to compare land value creation potential and avoided damages from flood mitigation works, researchers and cost-benefit practitioners alike should refrain from adding these two benefits as the risk of double-counting is real, with land value increases capturing to a large extent or in full the reduced damages from floods. Exploring what is the correct approach to include these two benefit categories will be the topic of future work.

Approximating Land Value Creation Potential
While the NEDUM-2D model used above is in many aspects a simple urban model, relying only on core urban economics, calibrating, and using it in other urban areas to estimate land value appreciations from resilience might be out of reach of local planning agencies because of lack of data, staff time or the right mix of technical skills. For this reason, we develop in this section a reduced form approximation of localized and aggregate land value changes that relies on a more limited set of data and parameters and can be computed outside of the fully calibrated model. From eq. (3), we have the annual price of land iP: With housing density per unit of land, h = Ak b , and a + b = 1, landowners' program maximization leads to: which then gives: The numbers displayed for land value creation in sections 4.1 and 4.2 differ because section 4.1 reports land value creation potential for the lifetime of the structure investments whereas section 4.2 reports annual land value creation in order to be comparable with annual avoided damages. Total land value creation is equivalent to annual land value creation divided by the real interest rate, they both entertain a very direct relationship (see equation (3)).

and:
We can also express the unitary housing rent at each distance from the city center as a function of the unitary housing rent in the most central location of the urban area, R 0 .
Combining eqs. (13) and (14), we can express the unitary land value, P(r), as: Let us denote all variables that differ between before and after resilience investments as NR, for "non-resilient", and R, for "resilient" respectively. As transport costs and most parameters are not altered by flood mitigation investments, we can then compare land values in each grid cell P R and P NR as follows: The second term in eq. (16) can easily be interpreted as the fraction of capital costs imposed by floods without flood mitigation works compared to a situation in which flood resilience investments have taken place. It will be equal to one when grid cells are either unaffected by floods or where resilience works do not affect flood depths and ensuing damages, or superior to one when flood mitigation works translate into lower damages and flood depreciation rates. The first term in eq. (16) translates the reduction in central housing rents, and equivalently the gains in households' utility, that stem from increased housing supply in valuable locations. This first term will be equal to or lower than one where flood resilience investments have tangible impacts on the housing stock and the relative attractiveness of specific locations in the urban area. Combining the first and the second term, the net impact on localized land values is not straightforward.
Unfortunately, whereas the second term is easy to estimate from the use of flood maps combined with flood depth -damage curve functions, the first term is impossible to compute without relying on a fully fledged urban economics model as we have described and used throughout this study. It however can be proxied for, through several simplifications. We report below how the changes in central housing rents because of resilience investments can be approximated and provide calculation details in the online supplement.
In eq. (17), f 1 , f 2 , …, represent fractions of the total urban population living in grid cells 1, 2,…, so that f i = n i N , where n i is the population residing in grid cell i, and N is the total population in the urban area. And x 1 , x 2 , … represent the difference in flood depreciation parameters divided by the initial complete unitary capital costs, so that If we combine (16) and (17) and aggregate over the whole urban area, we can compute the ratio of aggregate land values with and without flood mitigation works, ALV R and ALV NR as: With the formula in eq. (18) it is possible to proxy for the impacts of flood mitigation investments without developing and calibrating a fully-fledged urban economics model and with only a limited set of data. The data needed to inform eq. (18) are the construction function parameter b, the real interest rate i, the depreciation rate of structures ρ (commonly assumed to be 0.01), the flood depreciation rates, i.e. the percentage of damages that floods inflict on structures with and without resilience investments which can be retrieved as in this paper through the use of before and after flood maps combined with flood depth -damage curve functions, and finally the distribution of population in urban areas at the grid cell level, n i , which can be acquired either with population censuses or through global population data sets such as WorldPop or Landscan 15 for example.
Using this reduced form approximation and using the model's own prediction of the spatial distribution of households in the urban area n i , we estimate annual and aggregate land value creation from resilience investments to be respectively US$ 8.5 million and US$ 425 million. These numbers are reasonably close to our fully fledged model results of US$ 7.6 million and US$ 379 million but they overestimate them by 12% in both cases. There is always a tradeoff between the benefits and the drawbacks of model complexity. Given the simplicity of producing estimates for the reduced form expression in terms of data and modeling requirements, we consider this to be a promising result and approach that could be used in locations where data is scarce so as to produce quick estimates of land value creation potential from resilience investments.

Robustness of the Analysis and Driving Forces
Systematic sensitivity analyses of model results are useful for two reasons. First, whereas most parameter values are the result of a careful calibration procedure, there nonetheless remains irreducible uncertainty, and it is useful to explore the consequences of deviations from the central estimations. Second, aggregate land value creation from resilience investments materializes across a long-time horizon, when housing and land markets have adjusted, new construction has occurred, and households' locational choices have evolved. 16 When we look into the future, even assuming that the static parameters informing the model are perfectly calibrated, there are uncertainties pertaining to their stability over time as various agents' behaviors can vary, for example as a response to exogenous changes in construction techniques.
To evaluate the robustness of the results, we recalculate aggregate land value creation not only using the calibrated model parameters, but also across various combinations of parameterizations. In particular, we recalculate land value creations across 3000 scenarios in the Closed City with Public ownership of land setting (CCP) and using the conservative flood-depth damage curves provided in Hallegatte et al. (2013). We restrict the sensitivity analysis to one city setting and one flood-depth damage curve because we have already explored the impact of different choices in previous sections and we wish to isolate the contribution of model parameters to land value creation estimates. We choose the CCP setting and the lower bound flood-depth damage curve as these yield more conservative estimates of land value creation from flood mitigation estimates. The choice of the closed city setting is also more likely to be valid in the short to medium-term as potential in-migration cannot happen instantaneously. Each scenario comprises different model parameterization sampled 17 from the uncertainty range as presented in Table 4. 18 It should be noted here that while we explore the impacts of a wide range of parameter values on the model's results, these results should not be interpreted as carrying the same weight. This is because many parameter values (or the combination of parameter values) that we test are highly unlikely and depart both significantly from our calibration values and in most cases from the values found in the literature. The robustness analysis is therefore geared more toward understanding how the model behaves in response to changes in parameter than it is toward understanding how optimistic/pessimistic our central results are.

How Does Land Value Creation Change across all Sensitivity Scenarios?
We first observe the distribution of aggregate land value change across the 3000 different parameterizations. As shown in Fig. 7, aggregate land value change ranges from US$ -6 million to more than US$ 600 million, whereas the annualized land value creation ranges between US$ -0.5 million to more than US$ 15 million. The results found from using the calibrated parameters (US$ 379.04 million for aggregate land value creation and US$ 7.58 million for the annualized one) lie on the right side of the distribution (the 94th percentile of the distribution). Put differently, the outcomes from the calibrated parameters could be categorized as optimistic, as there are only 180 (of 3000) scenarios for which outcomes are at least as high as these calibrated baseline outcomes. However, when looked across scenarios, the outcomes are quite promising. The average aggregate 16 Local land value increases, as opposed to aggregate net land value change, can occur in the very short term as soon as the flood mitigation works have been announced even, consistent with the assumption of land prices corresponding to their highest and best use and developers/landowners' perfect foresight. We are however interested in aggregate land value creation after land and housing market mediation, which can only manifest itself over time. 17 We use a Latin Hypercube Sampling (LHS) function and make use of the option to minimize the sum of between-column squared correlations. 18 Table 4 also presents the values and the signification of the main parameters of the NEDUM-2D model applied to Buenos Aires in our baseline calibration. It also explains how these were obtained (through calibration processes, informed by data, borrowing from the literature or through assumptions), and how these parameter values relate to those found in the literature.

Seizing Opportunities: Which Scenarios Lead to High Land Value Creation?
Here we identify conditions leading to favorable outcomes in terms of land value creation. We define favorable outcomes using two thresholds: scenarios where land value creation is at least one third of the investment cost (Fig. 8) or even more than the investment cost itself (Fig. 7, in the online supplement). We find that the lower threshold is surpassed in almost 18%, and the upper threshold in almost 7% of the entire scenarios. Figure 8 and online supplement Fig. 7 show the scenario identification results for the two thresholds. There are two pathways leading to these favorable outcomes. The similarity between both pathways is that they require the real interest rate to be fairly low (<3.7%) and the β parameter (share of household income net of transport costs spent on housing) to be medium to high. In the first pathway, this is complemented with low to medium b parameter (i.e., more costly to build tall structures) and medium to high A parameter. The second pathway is characterized by medium value of b parameter (between 0.58-0.77, so not too low and not too high) and low to high value of the A parameter. This implies that when it is relatively costly to build tall structures, the annual capital cost of land is relatively cheap and households spend a non-marginal fraction of their budget on housing, land value creation from resilience works is likely to represent a significant share of the initial investment costs. In other words, land is highly valuable when building housing space elsewhere is expensive, so that flood protection works will result in large land value creation. In the online supplementary material we document the mirroring finding that land value creation from flood mitigation works is low when it is cheap to build tall structures and households only spend a small fraction of their budgets on housing.
It is important to note that small deviations from our centrally calibrated parameters, b, A, and β, are unlikely to change whether the land value creation estimates from NEDUM-2D allow to recover at least a third of the initial investment costs. Our central land value creation results are therefore robust to changes in these parameter values. The one parameter in our simplified modeling framework that is hard to ascertain and which leads to very different outcomes in terms of land value creation is the real interest rate, and it has historically been very unstable in Argentina over the last few decades.

What Are the Driving Forces of Land Value Creation and Total Avoided Damages?
Here we try to identify, among the four uncertain parameters, which one has the highest influence on the outcome variables. We use the extra trees algorithm for this purpose (Jaxa-Rozen and Kwakkel 2018), and the results are shown in Fig. 9. Overall, we find that the b parameter, which represents the elasticity of housing production with respect to capital, is the most significant driving force across the outcome variables. This translates the fact Fig. 8 Identification of scenarios leading to favorable outcomes: aggregate land value creation of at least 33% of investment cost that land is highly valuable when it is costly to build more elsewhere, that is when land and capital are very imperfect substitutes, i.e. b is low. In these cases, mitigating floods will result in large land value appreciations, all else equal. For annualized land value creation, this is followed by the real interest rate. On the contrary, the most significant driver for aggregate land value creation is the real interest rate, followed by the b parameter. It is important to highlight that the interest rate intervenes in multiple parts of the model, such as in estimating the price of land and thus determining the optimal amount of capital to invest at each location.

Conclusions and Discussion
Using a simple urban economics framework, in both a simplified form and calibrated for the urban area of Buenos Aires, this paper shows that the potential for land value creation from resilience investments is real and significant, in particular when flood-prone areas to be protected are valuable because of their proximity to employment centers. Under central, and often conservative, modeling assumptions, land value appreciation alone would justify the upfront investments costs of flood mitigation works. We also show that, in our central setting, net land value creation from resilience is one order of magnitude lower than the value generated from avoided damages to residential structures. 19 However, because appreciated land values as a result of public interventions can appear as more tangible than avoided structural destruction, it might be a more practical and politically acceptable Fig. 9 Driving forces of the different outcome variables. The values indicate the total gain in impurity from using the indicated input parameter as the splitting feature in the tree algorithm. Higher values imply higher importance of the input feature in explaining the variance in the outcome variable 19 In most scenarios of the sensitivity analysis, this is not the case as most scenarios show land value creation potential to be higher than avoided damages (see online supplement). channel through which to implement taxes aiming at recovering part or all of the protection investments costs, and thus more appealing to local decision makers. Another important result from our simulations is that ignoring the land value transfer/destruction linked to land and housing market adjustments would vastly overestimate the aggregate land value appreciation in the urban area (US$ 2.19 billion vs US$ 379 million in our central simulations in a CCP setting). These results tend to support the use of equilibrium models rather than hedonic pricing strategies when the objective is to understand aggregate land value changes from investments rather than purely local ones. If the objective is to estimate local gains only, for example to understand the possibilities of taxing winners, then both methods are useful.
This effort provides an additional practical contribution. While the model presented and used in this paper is, in many aspects, simple, it still requires detailed data sets and implementation time that are likely out of reach for teams in development agencies or cities aiming to produce a quick, yet robust, estimation of land value creation potential from their resilience-focused projects. For this reason, we developed a reduced form version of the model that limits the data needs and can be applied as a first approximation to understand the orders of magnitude of land value appreciation from resilience projects. This reduced form expression provides estimates that were of the same magnitude when tested against the fully calibrated model. This paper systematically explores the impacts of modeling choices on land value appreciation from resilience investments. It does so, by investigating the impacts of the location of floods on the results of the analysis; by varying the conceptual framework, testing the polar assumptions of open and closed cities and of absentee landlords or public ownership of land; and by using two contrasting flood-depth damage curve functions.
In addition, a detailed sensitivity analysis in the tradition of Robust Decision Making studies is performed to evaluate the impact of parameter values on results. In combination, these analyses highlight which are the most critical parameters and modeling choices when estimating land value creation potential from resilience investments. In terms of the parameters, particular attention should be paid to the elasticity of housing production with respect to capital and the real interest rate influencing people's decision in valuing land prices. The choice of the flood depth -damage curve function also has important implications for the model's results. Finally, the question of whether reduced flood impacts permitted by resilience investments translate into in-migration or not and if so to what extent, captured in our simulations through the election of the closed or open city modeling assumption, is critical to the estimation of land value creation potential: one order of magnitude separates these estimations. Both cases are extreme, albeit useful frameworks, and the appropriate setting probably lies somewhere between the two. The question that remains unanswered, is: is it realistic for local flood mitigation works to trigger in-migration and, if so, to what extent? Several points deserve discussion at this stage. First, we placed ourselves in a static setting which allows us to look at the long-term impacts of resilience on land value creation potential and to untangle the impacts of these investments from other socio-economic evolutions including demographics. These changes will however only materialize over time as the urban form adjusts to its new risk-profile and building decisions are implemented in reaction. A dynamic version of this model could be used as a next step to investigate the trajectory of land value appreciation. This could be useful for decision makers who are curious about when the initial investments costs could be recovered under various land value taxation schemes.
Second, and more fundamentally, we have assumed a perfect land market, where land prices reflect general attractiveness and flood risk levels, and where construction decisions are only guided by construction costs, including flood depreciation costs, and anticipated housing rents. There are at least two elements that could prevent real-world land markets from behaving as in our applied model: myopic land markets and transaction costs.
With myopic land markets, land values do not fully reflect flood risks as we have posited in our modeling framework, so that construction decisions depart from our simulations. This can happen either because risk information is not widely available or because this information is overlooked by sellers and buyers. In this case it is likely that protection from floods would also only partially be reflected in land values and construction decisions, reducing the land value creation potential from resilience investments. To what extent land markets account for flood risks is still a relatively open question with most studies finding that flood risks do reduce land and property prices to various degrees (Bin et al. 2008;Zhang 2016;Zhang and Leonard 2019;Dubé et al. 2021;Ortega and Taṣpınar 2018). But there at least several papers that either find mixed and weak evidence for property flood risk discount (Beltrán et al. 2018;Hino and Burke 2020) or find evidence of land/housing markets adjusting in the aftermath of natural disasters but "forgetting" about the risk several years down the road (Atreya et al. 2013;Bin and Landry 2013). Ideally our model would be able to account for such myopic land markets. This will be the subject of future work.
A second major bottleneck to the fluid functioning of land markets in developing country cities lies in land right issues and transaction costs involved in converting small scale to larger scale structures (Henderson et al. 2017;Henderson et al. 2016). Land rights issues can arise from multiple claims on specific land plots and the coexistence of private property systems. Such costs can act as a strong disincentive to investing in upgrading structures when and where land becomes more valuable. In our context where resilience investments lead to increased attractiveness of certain locations because of the reduction in background risks, there exists the possibility that such an improvement does not actually translate into increased investments in structures, and land values appreciations remain hypothetical. While we do not have evidence of such land right issues in the central area of Buenos Aires where the flood protection investments are taking place, this topic is nonetheless central for replicating this study elsewhere. Here too, future work will aim and introducing land market frictions in the model.