¹ For instance, the 2013 disaster in Uttarakhand, North India resulted in a death toll estimated to be in the region of 6,000 people. According to the latest World Disasters Report, the number of people affected by flooding ranks number one among natural disasters and is more than double other events. Most of the affected people live in medium and low human development countries (IFRC, 2014).

² In particular, the IPCC states that ‘[C]limate-change-related risks from extreme events, such as heat waves, extreme precipitation, and coastal flooding, are already moderate (high confidence) and high with 1°C additional warming (medium confidence)’.

³ The La Plata River Basin is one of the most important river basins in the world, draining approximately one-fifth of the South American continent. With a maximum width of 220 km, the La Plata River is considered the widest in the world.

⁴ Floodplains are usually defined as those areas of land adjacent to waterways that would be inundated during a flood event which has a 1 per cent chance of occurring or being exceeded in every year.

⁵ These studies are predominantly for the United States where the Federal Emergency Management Agency (FEMA) requires mandatory insurance in flood-prone zones.

⁶ An exception is Harrison et al. (Reference Harrison, Smersh and Schwartz2001).

⁷ Hedonic models are part of what are usually referred to as ‘revealed-preference’ valuation methods since people's willingness to pay for changes in environmental quality are inferred from observed behavior in related markets – in this particular case, the real estate market.

⁸ In addition, spatial dependence may also stem from measurement problems in explanatory variables, omitted variables and other forms of model misspecifications.

⁹ Setting λ=0 yields the Spatial Lag Model P=ρ W ₁ P+Xβ+Fγ+ε. Setting ρ=0 yields the model P=Xβ+Fγ+ε, with ε=λ W ₂ε+μ, which is sometimes referred to as the spatial error model. Setting λ=0 and ρ=0 causes the model to reduce to a linear regression model.

¹⁰ Measurement error may occur when the spatial unit of observation does not coincide with the spatial extent of the economic behavior, and then systematically relates to location.

¹¹ These matrices are often row-standardized to constrain the sum of elements of each row to be equal to 1, so that the above model conceptually means that the price of each property is affected by a form of weighted average prices of nearby properties.

¹² This type of weighting matrix is the most commonly used in spatial hedonic models applied to urban real estate markets. Alternatively, an α=1 implies that the weight decreases linearly with the distance between two observations. Other weighting schemes assign equal weights to every unit within a certain radius (often referred to as short-distance or radial-distance weights) and zero otherwise, but this might be problematic given the discontinuity implied by the definition. Another common alternative is to use the n-nearest neighbors, often treating each neighbor equally. Below we comment on the impact on our results of using different weighting schemes.

¹³ Following LeSage and Pace (Reference LeSage and Pace2009), the total effect can be written as ${\delta P \over \delta X_{1k}} \ldots {\delta P \over \delta X_{nk}} = \lpar I_{n}-\rho W \rpar ^{-1} \lsqb \beta_{k}I_{n} \rsqb $ , where β _k is the k-th element in the vector β. This total effect can be decomposed into a direct and an indirect effect. On the one hand, the direct effect captures the impact of the change in one explanatory variable on the price. This effect is different for each observation. The authors suggest reporting this effect as the average of the element of the main diagonal of the matrix $\lpar I_{n}-\rho W\rpar ^{-1}\lsqb \beta_{k}I_{n}\rsqb $ . On the other hand, the indirect effect (or spatial spillover) is the average of the sum of each row, without considering the elements of the main diagonal of $\lpar I_{n}-\rho W\rpar ^{-1}\lsqb \beta_{k}I_{n}\rsqb $ . The significance of these effects can be computed using Monte Carlo simulations.

¹⁴ The city was awarded two gold medals in the categories ‘city of the future’ and ‘better performance built’ at the 1889 Exposition Universelle in Paris.

¹⁵ The LPMA includes the towns of Tolosa, Ringuelet, Gonnet, City Bell, Villa Elisa, Melchor Romero, Abasto, Gorina, Jose Hernandez, Angel Etcheverry, Arturo Segui, Los Hornos, Lisandro Olmos, Villa Elvira and Altos de San Lorenzo, all of which have community centers that operate as local government offices.

¹⁶ Those houses and plots located in gated communities were not considered in this work because other factors are involved in determining their price, such as security services, sports and recreational facilities and even the location of the lot within the estate.

¹⁷ Calculations were performed with the Stata command TRAVELTIME, which finds travel distances and time between points by alternative means of transport using Google Maps (see Ozimek and Miles, Reference Ozimek and Miles2011).

¹⁸ Essentially, suburb dummies capture differences in crime rates and in the quality of the transport infrastructure. To a lesser extent they also capture differences between school districts. However, school choice is not restricted to housing location, as usually happens in developed countries, especially in the United States.

¹⁹ Housing values for those properties bought with a mortgage are usually registered with the real transaction value because there is a financial institution involved. However, such transactions are rare in Argentinean housing markets. According to the Chilean Association of Banks, while housing loans in 2012 represented almost 18 per cent of gross domestic product in Chile, and 6.8 per cent in Brazil, they only reached 1.5 per cent in Argentina.

²⁰ Of course, (unobserved) under-registration would not be an issue for the empirical exercise if the practice was uniformly applied throughout the market, e.g., if the transaction price of every house was under-reported, for instance by 10 per cent. However, there is no evidence that such under-reporting is uniformly done across the market.

²¹ Argentinean housing markets are not characterized by significant price negotiations between buyers and sellers. In general, real estate properties are sold for their asking price, and if the house is not sold shortly after being listed, sellers usually prefer not to decrease the price but to keep the house on the market for a longer period. Even in more sophisticated housing markets like in the United States, where buyers and sellers do engage in more negotiations, a significant proportion of houses are actually sold at their listed price. For instance, Case and Shiller (Reference Case and Shiller2003) report that 48.4 per cent of houses in four major US cities were sold for their asking price.

²² These orders of magnitude might seem large but in fact are quite common in the literature.

²³ There are houses in the sample which have rather large lot sizes. In the regressions we estimated the models with and without these properties and the results remained unchanged.

²⁴ To get a sense of the dramatic flooding of April 2013, one needs to know that the amount of rain that fell in less than five hours was the equivalent of approximately 20 per cent of the yearly rainfall (and about three times the normal level for April).

²⁵ In reality flooding risks vary within the floodplain, usually decreasing with distance from the stream or river. However, most of the empirical studies described in section 2 model flooding risks as a dichotomous variable, i.e., whether the property is located in the floodplain or not. Notable exceptions in the literature are Bartosova et al. (Reference Bartosova, Clark, Novotny and Taylor2000) and Rambaldi et al. (Reference Rambaldi, Fletcher, Collins and McAllister2013), whose works use a continuous flood risk measure.

²⁶ Some of these properties were actually flooded during the 2013 event. That event was, however, truly unique in the city's history.

²⁷ The estimation of SARAR models was made using the Econometrics Toolbox developed by James P. LeSage for MATLAB. The latest version of the toolbox functions can be found at http://www.spatial-econometrics.com. These functions implement ML, Bayesian and IV/GMM estimation of a host of spatial econometric models.

²⁸ In a simple hedonic OLS model, parameter interpretation is straightforward since the parameters can be interpreted as the partial derivatives of the dependent variable with respect to the explanatory variable, and indirect effects are set to zero by definition.

²⁹ Appendix table A1 reports marginal effects for spatial models based on alternative weighting schemes, i.e., the W matrices in equations (1) and (2). The upper part of the table shows results using a linear decay parameter (α=1) and the lower section of the table shows the results associated with a weighting scheme that assigns equal weight to each unit within a radius of 1,000 m, and zero otherwise. In general, using a weighting scheme other than the inverse of the square distance does not change the main results of the paper: houses located in flood-prone zones have a lower market value than similar houses in riskless areas. Marginal effects are higher, particularly the indirect marginal effects. Total marginal effects for houses are 19–71 per cent larger, depending on the specification and the weighting scheme. Perhaps the most striking difference is that under a radial-distance weighting scheme (third row in the lower panel of table A1), coefficients on moderate risk variables are statistically significant, and are approximately half of those associated with high hazard variables.

³⁰ Other relevant attributes usually included in this type of analysis are the quality of the air and the access to transportation infrastructure, especially subway or train stations. Regarding air quality, La Plata seldom exceeds any ambient pollution threshold because its flat topography helps to efficiently disperse pollutants. Even though some geographic variation in pollution levels might exist, there are no data to capture it. However, it is very unlikely that any spatial variation in pollution is correlated with flooding risks. With respect to access to transportation infrastructure, there is no subway in the city, and the main train station is located very close to the CBD; therefore any effect on housing values should already be captured by that variable.

³¹ In an unreported specification we tried the travel time from the property to the CBD instead of the distance, and results remained unchanged.

³² As before, using alternative weighting schemes does not qualitatively change the results (see footnote 29 for the case of houses). Marginal effects for undeveloped parcels of land reported in appendix table A1 are similar to those reported in panel B of table 4.

³³ The reason for separating houses and vacant plots in the empirical implementation is more a practical one. We do not have the same information for both types of real estate. For instance, we do not have the FOT ratio for houses, or whether they have access to essential services like piped water or wastewater services. Pooling both types of properties in a single regression would have forced us to discard relevant information.

³⁴ Note that interacting a dummy variable, indicating whether it was a house or not, with the flood risk variables in the full model would have sufficed. However, in this particular case the dummy variable would not only be capturing the type of real estate but also the time trend since data on houses and vacant plots come from different years, and probably other unobserved characteristics.

Notes: Marginal effects for panels A and B are based on corresponding model specifications in Tables 2 and 3, respectively. The upper section of the table uses a weighting scheme such as w _ij =d _ij ^−α, where α=1, and the lower section of the table uses a radial distance with a bandwidth of 1 km.