The cost of floods in developing countries’ megacities: a hedonic price analysis of the Jakarta housing market, Indonesia

Abstract

Although many megacities in developing countries experience floods annually that affect a large number of people, relatively few empirical studies have evaluated the associated costs. This paper estimates such costs by conducting a hedonic price analysis—providing evidence regarding the impacts of floods on the housing market. A robust regression technique on a simple linear transformation model, and a maximum likelihood estimation technique on the spatial lag version of the simple linear transformation model, are utilised to estimate the correlation between the level of the 2007 floods and monthly housing rental prices in Jakarta, Indonesia. This paper sheds light on the fact that in developing countries’ megacities, the total cost of floods among households is significantly lower compared to the total amount of funding needed to permanently eliminate floods in these megacities. Hence, a constant exposure of the urban areas in developing countries to flood damage will most likely keep happening.

Appendices

Appendix A: Hedonic property value method

The hedonic price method provides an intuitive analytical tool for studying the effects of property attributes and spatially integrated amenities on housing prices. Lancaster (1966) pioneered the development of its theoretical foundations, derived from the theory of consumer demand. The central assumption is that consumer utilities are not based on the goods per se, but instead on the individual “characteristics” of goods—their composite attributes. Although Lancaster (1966) was the first to discuss hedonic utility, there was nothing about pricing models and the properties of market equilibrium. To fill this gap, Rosen (1974) studied the demand–supply interaction in which they bid (consumers) and offer (suppliers) the combination of attributes and prices of the goods that keep the market in equilibrium.

Additionally, Rosen’s (1974) studies form the basis for using the hedonic property price model to estimate the value of environmental amenities. The argument is that the attributes of residential properties—recognised as heterogeneous goods, such as structural, neighbourhood and environmental characteristics—are reflected in the price differentials that affect lessee preferences in a market clearing equilibrium condition (Rosen 1974). The advantage of using this method over other preference estimation techniques is that it makes use of actual market transactions to recover value estimates for non-market attributes (Bin et al. 2008a). These related to aesthetic sights and their closeness to recreational sites such as parks, and beaches, as well as the quality of the environment in terms of air, water and noise pollution.

According to this method, the hedonic price function is typically represented as:

$$P_{i} = f\left( {s, n,l, e} \right),$$

(6)

where $P_{i}$ is the price of property $i$ which is a function of structural characteristics (e.g., house size, number of rooms, quality of walls), $s$; neighbourhood characteristics (for example, ethnic composition, crime rate, flow of traffic), $n$; location characteristics (e.g., proximity to economic centres, distance to highways, accessibility to public transport), $l$; and environmental characteristics (such as air pollution and flooding), $e$. Therefore, characteristics that generate benefits for households, such as a larger number of rooms or home size, increase the property’s price, while characteristics that imply costs for households, such as a neighbourhood with a high crime rate, reduce the property’s price.

Given that the basis of the method is to find what portion of the price is determined by the hedonic variable, we obtain the environmental attribute (which is flooding) by calculating the partial derivative of the price with respect to the variable $e$, ∂$P_{i}$/∂$e$. It gives us the marginal implicit value for an additional unit of the environmental asset, and thus enables an estimate of its monetary value.

Appendix B: Mean comparison between IFLS and SUSENAS datasets

To support the representation of Jakarta’s population, we provide the means of certain variables from Indonesia–National Socio-Economic Survey (SUSENAS) 2007 that was accessed for this paper. The following table shows the means compared with those of IFLS 2007 (Table 4).

Table 4 Mean comparison between IFLS and SUSENAS datasets.

Appendix C: Distribution of independent variables

The below table represents the distribution of continuous variables when they are in levels. Those on the left, demonstrate the skewed towards zero on the axis, while those on the right demonstrate a spread across the axis (i.e., closer to normal distribution) (Fig. 2).

Fig. 2

Distribution of independent variables in levels and log transformed

Appendix D: Spatial average treatment effect model

Unlike previous studies, this study uses a continuous measure of flood water (in centimetres) in the empirical analysis. To find whether or not a house within the flooded area lowers the rental price at any certain water level, we construct a conventional binary measure of flooding at the neighbourhood level based on three different threshold water levels, e.g., 30, 50 and 60. A house is considered flooded when the water level in the area during the flood event in 2007 was 30 cm (or 50 cm or 60 cm) and above; otherwise, the house is considered not flooded. We then estimate the impact of flooding on location rental prices using the spatial lag model:

$$y = \beta_{0} + \rho \varvec{Wy} +\varvec{\beta}_{1} \varvec{x}_{1} +\varvec{\beta}_{2} \varvec{x}_{2} +\varvec{\beta}_{3} D_{\text{f}} + \varepsilon .$$

(7)

From Table 5, it can be seen that results are sensitive towards the definition of being flooded or not, i.e., flood water thresholds define whether the house is flooded or not. However, it should be considered that any marginal difference between the chosen flood levels (e.g., from 30 to 50 cm and 50 to 60 cm) could be irrelevant and trivial for a household, as they are still experiencing flooding with its associated damage to the home. Therefore, we prefer to use a continuous flood-level variable as our main variable of interest.

Table 5 Results of spatial average treatment effect.

Appendix E: Estimation results using spatial error model

The spatial error model takes the following form:

$$y = \beta_{0} +\varvec{\beta}_{1} \varvec{x}_{1} +\varvec{\beta}_{2} \varvec{x}_{2} +\varvec{\beta}_{3} D_{f} +\varvec{\beta}_{4} \left[ {D_{\text{f}} f} \right] + \varepsilon ;\quad \varepsilon = \lambda \varvec{W}\varepsilon + \mu ,$$

(8)

where $\mu$ is the independent and identically distributed (i.i.d.) error term and $\lambda$ is the spatial error parameter. The understanding of the spatial error model is close to the moving average model, whereby each observation error in the time series can also be affected by other observation errors. Ignoring the spatial error term, the OLS coefficients will be inefficient as it violates the assumption of independence among disturbance terms (Anselin 1988) (Table 6).

Table 6 Results of spatial error model

Appendix F: Owner-occupied and rental property comparison

When running separate regression for the dummy variable homeowner (1,0), we find that the study variable (LOGflood) is statistically significant at 5% for rental properties. This means that those renting are likely to be more vulnerable to flood impacts than those owning or buying. Owning a home can be considered as an indicator of income and economic resources that may support flood victims to cope with the effects of flooding (Table 7).

Table 7 Results of OLS for both owner-occupied and rental properties