Flood Risk Information, Actual Floods and Property Values: A Quasi-Experimental Analysis

1 Introduction

The global frequency and severity of natural disasters have shown an increase over the past several decades. As a result, common natural disasters such as floods, bushfires, cyclones and earthquakes have had an increasingly severe impact on both properties and lives (Gentle et al., 2001). Among them, floods are one of the most common. Australia has also experienced highly variable extremes in weather conditions and a rise in the frequency of natural disasters. Of these, the Brisbane floods of 2011 were one of the worst natural disasters in Australia's recent history in terms of economic damage. In the wake of such major disasters, relevant authorities and independent researchers have reported on the causes and damage inflicted, but have generally considered only direct tangible costs (see Queensland Government, 2011). In particular, although the losses to private property have been estimated, there is little consideration given to the change in property values due to flooding. This paper, therefore, focuses on the impact of such natural disasters on property market behaviour using the 2011 Brisbane floods as a case study.

The property market is an important indicator of an economy's health involving a wide range of stakeholders, which include real estate agents, banks, insurance companies, policymakers and the public. Moreover, the maintenance of property values is a key element in the sustainability of local communities (Lamond, 2008). Property price changes affect the expected lifetime wealth of home owners and the collateral values of homes (Miller et al., 2011). This suggests that it would be in the interest of governments to invest in appropriate infrastructure and policies to minimise the adverse impacts of natural hazards such as floods on the property market.

Numerous studies have looked at the behaviour of the property market, particularly in relation to natural disasters. For example, the impact of flooding and flood plains on property markets has been investigated by Samarasinghe and Sharp (2010), Bin and Landry (2013), Petrolia et al. (2013), Rambaldi et al. (2013) and Eves and Wilkinson (2014), a number of which have investigated the impact of floods on property values. Importantly, a meta-analysis of flood risks in the USA has shown that there is a difference between ex-ante and ex-post analysis (Daniel et al., 2009). However, none of the research has investigated the impact of floods on property values after a flood event (such as the 2011 floods) in the Brisbane City Council (BCC) region.

Although the hedonic pricing (HP) theory has been extensively used for investigating the impact of floods on property prices, most studies have ignored inherent deficiencies in traditional HP theory. In this paper, quasi-experimental and spatial econometric techniques are employed to avoid methodological issues and to distinguish specific objectives. For example, adoption of a quasi-experimental hedonic method is useful to take into account endogeneity and omitted variable biases (see Parmeter & Pope, 2013).

This study was motivated by the 2011 Brisbane floods. Prior to this event, the BCC mapped flood risk areas and made them available to the public in October 2009 – a decision which may have influenced property market behaviour. However, existing studies are limited by considering only either the actual flood incidence or availability of flood risk information. This study will be the first to analyse the effect on the property market of flood risk map information being made public and to compare this with the effect on the property market of the actual flood incidence. This is tested using the difference-in-differences (DID) method, which takes account of different time variables and treatment effects. Prior to this, we examine the impact of flood-related variables (among other factors) on property values. Furthermore, we examine the temporal variation of flood impacts on property values. This is because several studies show that any negative effect on property values following a natural event disappears with time (see Lamond, 2008; Bin & Landry, 2013).

2 Literature Review

The demand for property depends on certain characteristics, primarily structural, environmental and neighbourhood, with property prices being a function of those characteristics. A review of the role of such structural characteristics in determining property prices is found in Sirmans et al.'s (2005) study. Several studies have shown their importance in influencing property values (see Acharya & Bennett, 2001). Thus neighbourhood amenities such as parks, green spaces, water bodies, reservoirs, and open spaces have been found to be important determinants of property prices. Previous studies have shown that residential properties closer to open spaces and recreational facilities have higher values than similar properties farther away (see Acharya & Bennett, 2001). In addition to physical factors such as water bodies and parks, pollution and environmental quality are also important considerations for people when deciding on residential location (see Gayer, 2000; Kim et al., 2003).

Other researchers have studied the effect of natural disasters and climate change on property markets (see Schlenker et al., 2005; Michael, 2007; Butsic et al., 2011). Butsic et al. (2011), for example, applied the HP model to simulate the impact of climate change on real estate prices in Canada and showed a negative impact on housing prices. Numerous studies have focused on investigating the impact of flooding and flood risks on property values, indicating that flood risk tends to discount property values (see Bin & Landry, 2013). However, the findings vary due to the use of different methodologies and the nature of impacts (Lamond, 2008), and discounted property values vary according to the country and location.

Previous studies of flood risk on properties cover a range of issues, including climate change predictions, analysing past flood events, existence of flood risk maps and following actual flood events. Such studies have evaluated the impact of flood risks on the property market rather than focusing on an actual flood event (see Samarasinghe & Sharp, 2010; Rambaldi et al., 2013). Other studies have been conducted immediately following a major flood incident (Bin & Landry, 2013; Eves & Wilkinson, 2014). A recent qualitative analysis of price behaviour of the Brisbane property market has suggested that floods create negative effects on the average listing price (Eves & Wilkinson, 2014). Another important study (Page et al., 2014) investigated the risk-seeking behaviour of flood-affected households following the 2011 Brisbane floods, and concluded that households in flood-affected areas were 50 per cent more likely to take risks than their neighbours who were unaffected. The data for this study were collected from 220 individual home owners two months after the floods.

Hedonic pricing price analysis is a widely used method which indirectly estimates flood risks, a methodology which is continuously being developed to address econometric questions. Importantly, many HP studies have indicated that, where cross-sectional data are used, the value of a property in one location may also be affected by the property values in its neighbouring area – a phenomenon described as spatial autocorrelation (see Anselin, 1988, 2010).

Some studies have adopted before-and-after analyses to examine the effect of floods on property values (see Bin & Polasky, 2004). More precisely, these analyses have compared one group over two time periods. Before-and-after analysis does not control for changes of market behaviour between the two time periods, although it is an effective tool for measuring other unobserved variables. Bin and Landry (2013) examined the effects of hurricanes Fran and Floyd in North Carolina, USA, within a DID framework. However, this study did not consider the impact of floods on demographic submarkets.

Previous studies have confirmed that following a flood event, there is a significant negative effect on the value of properties at risk. But as many studies have suggested, flood risk tends to disappear over time (see McCluskey & Rausser, 2003; Bin & Landry, 2013). Thus, particularly when flooding is infrequent, the actual flood impact is most pronounced during the event and immediately after. However, many factors such as suburb characteristics also need to be taken into consideration. The following section presents a methodology designed to isolate the impact of floods and the disclosure of flood risk information on property values – a linkage which has not been tested before in the literature. This section then continues with the analysis of the temporal distribution.

3 Methods and Data Collection

As mentioned in the introduction, this study examines the impact of flood risk variables on property values, analyses the effect on the property market of release of flood risk information to the public and compares it with the effect on the property market of the actual flood incidence. It also examines the temporal variation of flood impacts on property values. Below we discuss the methodological approaches and data used in the study.

3.1 Hedonic Property Price Approach

In HP models, the price of a commodity can be attributed to a vector of characteristics which may include property characteristics, environmental and socio-economic factors. Given that the market is in equilibrium, the marginal price of a housing attribute is equal to the marginal willingness to pay for a change in the attribute. The hedonic theory provides a methodology for identifying the structure of prices of each property attribute.

Price theory predicts that the willingness to pay to avoid the disutility of flood risk needs to be reflected in a property price discount. When two houses are identical in all aspects except the flood risk, the house without the flood risk can be expected to sell at a higher price. That difference is the cost of the flood or flood risk on the property market. The model takes the form

(1)

where p is the market price of the property, α is the constant term, β is the vector of coefficients that reflects the influence of each characteristic, x is the vector of characteristics of the property (e.g. structural, neighbourhood and flood variables), and ε is the error term.

The next challenge is to analyse and compare the impact of two incidences (the release of flood maps and actual flood event) simultaneously within a DID framework. In this analysis, data on property transactions were collected first before flood risk information was made available (from between 2006 and September 2009; i.e., before October 2009) until the flood incidence (October 2009 to January 2011) and secondly post-floods (January 2011 to March 2013).

3.2 Quasi-Experimental Approach

The release of flood prone area maps for the BCC suburbs in October 2009 and the actual floods in 2011 provide an interesting experiment to assess the impact of flood risks on property markets compared to only examining the effect of either one of these events. For this purpose, the DID framework was employed. According to Woodridge (2007), this method can be applied to both repeated cross-sectional data and panel data. Our study has demonstrated the applicability of this approach for both random as well as natural experiments (see Bin & Landry, 2013; Gawande et al., 2013; Parmeter & Pope, 2013).

The DID methodology rests on the premise of using two groups and two time periods, where the second time period of one group is treated (the treated group in this study being the flood-affected properties) and the other is not (the control group in this study is non-flood-affected properties). With respect to the before-and-after (flood) treatment, for the treated groups (i.e., flood-affected properties), the average price difference is the treatment effect, which can be expressed as

(2)

In order to remove biases from the comparison over time (time trend), a control group was included. The comparison with the control group also removed differences between the treated and control groups. The outcome difference for the control group can be expressed as

(3)

The real treatment effect was obtained by subtracting the difference in average outcome in the control group (for the two time periods) from the difference in average outcome in the treated group (before and after the incident):

(4)

3.3 Study Area and Data

As a case study, BCC suburbs were selected for empirical investigation. According to Bureau of Transport Economics (2011) estimates, Queensland experiences the second most frequent natural disasters of any Australian state. Brisbane, the capital city of Queensland, is prone to floods as a result of overflow from the Brisbane river and cyclones. In the years following the 1974 floods, the peak water level recorded in Brisbane was in January 2011, at 4.46 metres. Before this flood, the BCC mapped flood risk areas, making them available to the public in October 2009. Although property buyers may not have any experience of the 1974 floods, the release of flood risk maps in October 2009 may reduce information asymmetry, which is assumed will be reflected in the property market. The release of flood maps and the 2011 floods were used as a basis for our case study. A study area, including heterogeneous suburban income groups within the BCC region, was selected using a stratified random sampling procedure. Property transaction data for selected suburbs within the BCC region were collected for the period 2006–2013, and the supplementary secondary data, published by different sources, were collected using Geographical Information System (GIS) techniques. The data collection procedure is summarised in Figure 1.

For this study all the relevant property data were available from RP Data Information Service (www.rpdata.com).1 This database provided a comprehensive overview of Australian property sales history and attributes, transactions and other information, including transaction data for multiple and single dwellings, and commercial and vacant properties. Researchers have often used market transaction price as a dependent variable in HP analysis. However, there was no single recognised comprehensive database with which to estimate the HP function. In this case further information available from different sources in different formats (i.e., maps) was merged using ArcGIS. Selected variables and descriptive statistics are shown in the Appendix.

The property market consists of a number of structural variables, surrounding amenities, environmental characteristics and social factors. These attributes include number of bedrooms, number of bathrooms, garage spaces, lot size, type of construction material (wooden or brick), carpeting or not, and existence of a pool. Neighbourhood characteristics of a house include median weekly income, travelling distance to the nearest school, travelling distance to the nearest shopping centre, and distance to amenities (i.e. distance to park). All these characteristics have an influence on property values.

The focus of this study is, however, on flood-related variables and their impact on property values, analysing the impact on property values of the release of flood risk information to the public, comparing it with the effect on property values of the actual flood incidence, and examining the temporal variation of flood impacts.

3.4 Empirical Estimation

We begin by estimating the impact on property values of flood related variables, namely distance to river and elevation. Flood risk is defined in two discreet ways in the literature: as a dummy variable for flood-prone areas and as the direct distance to the water body (Samarasinghe & Sharp, 2010). For the first set of regression models in this paper, flood incidence was represented by a dummy variable (FLOOD), denoted by 1 for properties in flood-affected (flood zone) areas and 0 for non-flood-affected areas. It was hypothesised that both property buyers and sellers were aware of flood-affected zones and of the discounted prices for properties in flood-affected areas when compared to properties in non-flood affected areas. Thus, the FLOOD variable can be represented by

(5)

where x_i represents a vector of structural, environmental and neighbourhood characteristics and year variables.

In addition to the FLOOD variable which shows flood-affected areas, we specified another variable to capture the flood risk of properties. The flood risk is specified as the distance to the river (DIS_RIVER):

(6)

Previous research has shown that proximity to water bodies can have a positive and a significant impact on property values (e.g., Gopalakrishnan et al., 2011), but they also have a risk of being affected by floods. Hence, properties close to a river had amenities as well as disamenities. In addition, properties located close to the river that are also high in elevation are likely to command higher market prices than properties that have lower elevation and are farther away. This is expressed as

(7)

where the product of distance (DIS_RIVER) and elevation (ELEV) forms the flood risk variable (DIS_ELEV). It was hypothesised that the properties with a higher DIS_ELEV value had a lower risk than properties with a lower value of DIS_ELEV. The results of this model are shown in Table 1. Note that in addition to reporting ordinary least squares (OLS) estimates, we also account for spatial interactions based on Moran's I statistics and the Lagrange multiplier (LM) test.2 The OLS results (columns 2–4) are compared with the maximum likelihood (ML) estimation results (columns 6–9).

Table 1. Hedonic Property Price Analysis Considering Flood-Affected Areas and Distance to River and Elevation Combined (Flood Risk)

	OLS estimation				ML estimation
	Group 1		Group 2		Group 1		Group 2
	Model 1	Model 2	Model 1	Model 2	Model 1	Model 2	Model 1	Model 2
CONSTANT	11.701328***	11.68829***	11.7087***	11.67813***	7.959063***	7.923163***	9.577748***	9.316872***
BEDRM	0.0274607**	0.026216**	0.038291**	0.039109**	0.026583*	0.025387**	0.041165**	0.042119**
BATHRM	0.0451245***	0.045375***	0.108704***	0.107054***	0.041307***	0.041634***	0.103538***	0.101106***
GARAGE	0.0552711***	0.054394***	0.017109	0.016837	0.054131***	0.05331***	0.018963	0.019058
AREA	0.0002302***	0.000229***	0.000485***	0.000493***	0.000222***	0.000221***	0.00047***	0.000477***
HOUSETYPE	0.0143539**	0.015474**	0.005277	0.005848	0.01339**	0.014326**	0.005024	0.005845
FLOOR_CARP	0.045492**	0.046148**	−0.02102	−0.01936	0.037792**	0.03827**	−0.0216	−0.01929
FLOOR_TILE	0.0003183	−0.00263	0.042293	0.04471	−0.0052	−0.00814	0.043673	0.046645
FLOOR_WOOD	0.0652864***	0.065564***	0.005018	0.005683	0.057759***	0.057795***	0.009459	0.010952
ROOF_TILE	0.0120244	0.013189	−0.00991	−0.00608	0.005845	0.006617	−0.01414	−0.01024
ROOF_ASBES	0.0212661	0.024845	0.01748	0.013895	0.014104	0.017271	0.018705	0.015376
WALL_BRICK	−0.0210556	−0.02323	−0.02678	−0.03264	−0.01594	−0.01795	−0.02603	−0.03289
WALL_WOOD	0.0212269	0.022266	−0.00658	−0.00479	0.01515	0.015821	−0.0025	−0.00125
POOL	0.0546497**	0.05555**	0.142154***	0.144515***	0.050192**	0.051316**	0.135236***	0.136784***
GARDEN	0.0570809***	0.060746***	0.071464***	0.072951**	0.054932***	0.058189***	0.073143**	0.074851***
STORIES	0.0650049	0.0652***	0.098362***	0.105767***	0.067397***	0.067501***	0.096244***	0.10372***
RIVERVIEW	−0.1145237	−0.12135	0.322308***	0.343242***	−0.14387	−0.1498	0.298448***	0.318181***
INCOME	0.0000773**	7.34E-05**	3.75E-05	3.32E-05	4.45E-05	4.02E-05	0.000026	1.89E-05
PARKDIS	−0.0000146	−1.7E-06	−8.1E-05	−0.00021**	−2.2E-05	−8.9E-06	−0.00013	−0.00027**
HIGHWAY	0.0000819***	0.000086***	−0.00014***	−0.00019***	0.000055***	0.000059***	−9.7E-05***	−0.00015***
FLOOD	−0.0626698**		−0.07155**		−0.05916**		−0.06885**
DISSHOP	0.0000944***	8.71E-05***	4.79E-05	−2.3E-07	7.43E-05***	6.71E-05***	0.000027	−2.4E-05
RAILWAY	−0.000093***	−0.0001***	0.000251***	0.000484***	−6.8E-05***	−7.5E-05***	0.000212**	0.000458***
DISCBD	0.0000199	2.06E-05	0.000151**	0.000152**	1.71E-05	1.86E-05	0.000125**	0.000118**
DIS_ELEV		1.89E-07		1.78E-05***		1.37E-07		1.89E-05***
D_2006	0.0678472**	0.067146*	−0.11235**	−0.11696**	0.065377**	0.065063**	−0.11634**	−0.12146**
D_2007	0.2339419***	0.232405***	0.094414*	0.093579*	0.227376***	0.226137***	0.085767	0.084006
D_2008	0.2753989***	0.273881***	0.089733	0.084176	0.271701***	0.270431***	0.086296	0.079724
D_2009	0.2385291***	0.235701***	0.086793	0.083688	0.231178***	0.22877***	0.082407	0.077962
D_2010	0.2064737***	0.201132***	0.073006	0.072126	0.201254***	0.196656***	0.070368	0.068749
D_2011	0.0896609**	0.086931**	−0.0913	−0.09461	0.091909**	0.089758**	−0.09419	−0.09822*
D_2012	0.0806578**	0.078817**	−0.05943	−0.06361	0.081836**	0.080249**	−0.06309	−0.06824
R 2	0.4608	0.4578	0.68	0.68
Adj. R2	0.449	0.4459	0.67	0.67
n	1,399	1,399	946	946
rho					0.300053***	0.301386***	0.171414***	0.19189***
Log-likelihood					298.902	295.0711	116.6622	112.3683
Wald test					51.206***	51.452***	10.142	12.805

Note

• ***, ** and * denote that coefficients are significant at 1%, 5% and 10% levels, respectively.

In order to isolate the impact on property values of flood risk map information released in October 2009 and the actual floods in 2011, the DID approach was employed. The outcome of the two groups (flood-affected and non-affected properties) over three time periods (before and after the release of the flood risk maps and actual floods) was observed. A comparison of the two groups provided the treatment effect. Here, the availability of flood risk information and the flood incidence for two groups is expressed as

(8)

where d₁ = 1 if the property was sold between when the flood risk information was made available and the actual flood incidence, and 0 otherwise; d₂ = 1 if the property was sold after the flood incidence and 0 otherwise; and d_f = 1 if the property was in a flood risk area and, 0 otherwise. The coefficient α_inf measured the first treatment effect (availability of flood risk information, D₁) and α_flood measured the second treatment effect (flood incidence in 2011, D₂). The results of the regression analyses are shown in Table 2. The OLS estimation results are shown in columns 2–5 of Table 2, and the ML estimation results in columns 6–9. The ML results are reported because the DID estimation was extended to identify and correct for spatial interactions. The spatial lag is defined as the spatially weighted average of the extent to which neighbouring housing prices affect the price of a particular house. For this study, the spatial lag DID model3 was defined as:

(9)

where ρ is the spatial autocorrelation parameter and W is the spatial weight matrix. W is an n × n matrix where n is the number of observations. The spatial lag DID model can capture both the spatial effect as well as the treatment effect.

Table 2. Difference-in-Differences Analysis to Isolate the Release of Flood Risk Map Information and the Effect of Actual Floods on Property Values

	OLS estimation				ML estimation
	Group 1		Group 2		Group 1		Group 2
	Model 1	Model 2	Model 1	Model 2	Model 1	Model 2	Model 1	Model 2
CONSTANT	12.15933***	12.1886***	11.65049***	11.7921***	8.170768***	8.195209***	9.298884***	9.305311***
BEDRM	0.028467**	0.029949**	0.039653**	0.038511**	0.02742**	0.027745**	0.042948**	0.041861***
BATHRM	0.052969***	0.050809***	0.109633***	0.107662***	0.048855***	0.048198***	0.103881***	0.101298***
GARAGE	0.05489***	0.055411***	0.010351	0.01122	0.053481***	0.053746***	0.012322	0.014597
AREA	0.000239***	0.000241***	0.000479***	0.000491***	0.00023***	0.000231***	0.000463***	0.000475***
HOUSETYPE	0.021338**	0.021551**	0.001985	0.0039	0.020122***	0.020149***	0.00165	0.003876
FLOOR_CARP	0.05412***	0.054669***	−0.01393	−0.00813	0.044963***	0.045397***	−0.01485	−0.00839
FLOOR_TILE	0.019814	0.019977	0.077428	0.080524	0.012689	0.013679	0.078802	0.082343
FLOOR_WOOD	0.071662***	0.071245***	0.001372	0.003777	0.062775***	0.062914***	0.006049	0.009065
ROOF_TILE	−0.01119	−0.00976	−0.00926	−0.00475	−0.01669	−0.01649	−0.01385	−0.00912
ROOF_ASBES	0.011	0.0135901	0.030523	0.029694	0.004001	0.003522	0.032054	0.031265
WALL_BRICK	−0.07255***	−0.07231***	−0.04282	−0.05282	−0.06599***	−0.06558***	−0.04094	−0.05204
WALL_WOOD	−0.01923	−0.017232	−0.01825	−0.02592	−0.02464	−0.02435	−0.01299	−0.02135
POOL	0.046564**	0.046092**	0.147046***	0.147711***	0.041853**	0.040747**	0.139487***	0.139709***
GARDEN	0.062282***	0.062071***	0.077649***	0.07692***	0.060053***	0.05998***	0.079771***	0.079027***
STORIES	0.064739***	0.065711***	0.100373***	0.104659***	0.067457***	0.06801***	0.09777***	0.102317***
RIVERVIEW	−0.1811	−0.18181	0.331865***	0.34863***	−0.20996*	−0.21008*	0.305676***	0.322543***
INCOME	4.35E−05	0.000045	3.37E-05	2.93E-05	1.09E-05	1.29E-05	2.17E-05	6.25E-06
PARKDIS	−3E-06	−6.8E-06	−6.1E-05	−0.00016*	−1.2E-05	−1.7E-05	−0.00011	−0.00022**
HIGHWAY	7.21E-05***	7.13E-05***	−0.00013***	−0.00016***	4.43E-05***	4.34E-05**	−8.6E-05***	−0.00012***
FLOOD	0.015731		0.011544		0.017288		0.014305
DISSHOP	7.16E-05***	0.000072***	5.48E-05	0.0000394	5.18E-05**	5.34E-05**	3.21E-05	0.000014
RAILWAY	−8.3E-05***	−8.1E-05***	0.000264***	0.000453***	−5.7E-05***	−5.6E-05***	0.000221**	0.000427***
DISCBD	2.69E-07	−2.37E-06	0.000161**	0.000136**	−1.6E-06	−3.7E-06	0.00013**	0.0001*
DIS_ELEV		1.04E-07		1.31E-05**		5.6E-08		1.45E-05**
D 1	−0.01445	−0.00644	−0.04388	−0.0223	−0.01584	−0.00146	−0.04131	−0.01638
D 2	−0.19779***	−0.2519***	−0.18701***	−0.1880***	−0.18894***	−0.18907***	−0.19128***	−0.19257***
R 2	0.3834	0.3875	0.6692	0.6716
Adj. R2	0.3722	0.3764	0.6602	0.6627
n	1,399	1,399	946	946
rho					0.318157***	0.31758***	0.189323***	0.201703***
Log-likelihood					206.0023	205.8612	134.6597	130.3533
Wald test					50.314	50.068	12.13	13.825

Note

• ***, ** and * denote that coefficients are significant at 1%, 5% and 10% levels, respectively.

3.5 Temporal Variation of Flood Risk

In this section we capture the temporal variation of flood risk. It was hypothesised that the disclosure of flood risk map information and the actual flood-incidence-induced effect on the property market disappeared with time or varied inversely with elapsed time (see Lamond, 2008; Bin & Landry, 2013). In order to test this hypothesis, the spatial HP model was extended to incorporate the elapsed time variable (see Hansen et al., 2006; Bin & Landry, 2013).

The time elapsed spatial lag model can be defined as

(10)

where MT is the elapsed time. The estimation of this model is shown in Table 3.

Table 3. Estimating the Temporal Variation of Flood Impacts

	OLS estimation		ML estimation
	Group 1	Group 2	Group 1	Group 2
CONSTANT	11.85259***	11.67174***	9.6964***	10.54767***
BEDRM	0.037972**	0.0213289	0.03642**	0.021991
BATHRM	0.072928**	0.132123***	0.071922***	0.133088***
GARAGE	0.05946***	0.0283755	0.057972***	0.029293
AREA	0.000138***	0.000539***	0.000138***	0.000538***
HOUSETYPE	0.0338124**	0.03361**	0.033288**	0.032713**
FLOOR_CARP	−0.02145	−0.015277	−0.025233	−0.01195
FLOOR_TILE	−0.08355*	−0.03751	−0.08481*	−0.04078
FLOOR_WOOD	0.02753	0.00498	0.0260779	0.006601
ROOF_TILE	0.03031	0.026397	0.030716	0.030372
ROOF_ASBES	−0.00113	0.03796	−0.003362	0.045079
WALL_BRICK	0.00492	0.0096	0.0066	0.008813
WALL_WOOD	0.03785	0.0351	0.038469	0.037328
POOL	0.04489	0.10881**	0.04251	0.104478**
GARDEN	0.077567**	0.11951**	0.07663***	0.11675**
STORIES	0.043288	0.12744**	0.049887	0.128276**
RIVERVIEW	0.152028	0.22393**	0.15047	0.215611**
INCOME	0.00017**	5.71E-05	0.000152**	0.000054
PARKDIS	−0.00011**	−0.00029**	−0.000118**	−0.00031**
HIGHWAY	2.05E-05	−0.00017**	1.49E-05	−0.00015**
FLOOD	−0.17237***	−0.15051**	−0.1657***	−0.14185***
DISSHOP	7.27E-05	−6.56E-05	6.86E-05*	−6.5E-05
RAILWAY	−9.46E-05**	0.000385**	−7.43E-05*	0.000373**
DISCBD	1.27E-05	0.000106	4.22E-05	8.77E-05
DIS_ELEV	5.88E-06	1.97E-05**	6.24E-07	2.04E-05**
MT	−0.003029**	0.0020667	−0.00285**	0.002006
R 2	0.5575	0.7552
Adj. R2	0.5203	0.7283
n	323	254	323	254
rho			0.1759***	0.092842***

Note

• ***, ** and * denote that coefficients are significant at 1%, 5% and 10% levels, respectively.

4 Results

For this analysis pooled cross-sectional data relevant to the selected flood-affected suburbs within the BCC region were used. Data cleaning was conducted and some were discarded due to missing attributes and outliers. It is noted that the application of the spatial technique discussed in Section III is not possible for a large dataset (e.g., Neill et al., 2007). Moreover, the observations need to be close enough to create a spatial impact. Thus, if some variables are dispersed, a weight matrix cannot be estimated. On the other hand, the dataset should be large enough to incorporate sufficient determinants to minimise the effect of outliers. Hence, selected suburbs were grouped after examining suburb characteristics and dispersion.

For example, Chelmer and Graceville are two adjacent suburbs and show similar characteristics. According to the census and statistics data (2011), both suburbs belonged to the high-income category (average weekly household median income was $A2,216 in 2011). These two suburbs were therefore merged into a one cluster (group 2) to obtain a reasonable number of transaction records for the study period. The first group, comprising of Oxley and Durack, which are also in close geographical proximity, showed an average median income of $1,329 and were therefore classified as low-income suburbs in line with Australian Bureau of Statistics census data (2011). The second group is close to the river, while the first group is much farther (the mean distances to the river are 3,259 m and 513 m for group 1 and group 2, respectively). A correlation matrix was used to examine the presence of multicollinearity – an econometric problem which may occur in using HP analysis.

Overall, the real price of property varied between $371,500 and $8 million. The differences in the mean value of property were observed across different groups. The high correlation (0.66) between house price and income is a good approximation for dividing submarkets based on house prices. The descriptive statistics of structural characteristics such as number of bedrooms (BEDRM), number of bathrooms (BATHRM) and number of garage spaces (GARAGE) are presented in the Appendix. These data indicate that the distribution of these variables is more or less similar for both groups. However, the variables of interest in this study, namely the dummy variable for flood affected area (FLOOD) and interaction of elevation and distance to the river (DIS_ELEV) did show differences between the groups.

5 Conclusions

This paper has examined the impact on property values of flood-affected and flood risk variables among other factors. The impact of floods on property values is shown to produce a significant drop in property prices in cases where the property was in the flood zone (see Table 1). For such properties the average property price decreased by 6 per cent in group 1 (low-income suburbs) and 7 per cent in group 2 (high-income suburbs). This analysis is followed by examining the impact of releasing flood risk map information and the effect of actual floods on property values. Application of the conventional HP function was not appropriate for comparing two consecutive events such as the release of flood risk maps (i.e., information on a potential disaster) and the actual flood incidence (i.e., disaster event). Therefore, in this study two treatments – release of flood risk map information and actual flood incidence – were compared using a quasi-experimental technique using DID estimation. More robust estimation was achieved by comparing the two incidents with a suitable control group. Previous studies have not distinguished between the effects of these incidences.

The results show that the average value of properties in the flood zones decreased after the 2011 flood event. In comparison to the actual flood event, the impact of the release of flood risk map information is found to be minimal. The release of flood maps decreased property values by 1–4 per cent (see Table 2). The release of flood risk maps undoubtedly reduced the level of information asymmetry among buyers, but was not expected to have a larger impact than the impact of the actual floods, which reduced property values by 18–19 per cent. There are good reasons why there is discounting of the possibility of properties in flood zones being affected in another flood. For example, the history of the occurrence of floods highlights the fact that extreme floods happen (on average) once in a lifetime. As well, after the 1974 floods in Brisbane, the Wivenhoe dam was built below the Somerset dam to prevent another repeat of the major floods (Atfield & Moore, 2011). Atfield and Moore (2011) quote Professor Trevor Grigg, who authored a report on the flood mitigation advantages of Wivenhoe Dam after the 1974 floods, as saying: ‘A lot of people believed we were protected by these dams but of course we never were.’ Furthermore, there is evidence in the academic literature to show that people who did not experience floods undervalued the risk (see Siegrist & Gutscher, 2008). It is also to be expected that properties located closer to a water body have both amenity as well as disamenity values. The results from Table 2 confirmed that the release of flood risk information in October 2009 was largely discounted. Apart from the reasons provided above, it is also possible that amenity values to households resulting from living close to a water body outweigh the potential disamenities. As well, there is the probability that ‘some consumers appear to not have a good understanding of their level of risk or of their insurance policy’ (Productivity Commission, 2014, p. 414). With respect to underinsurance, the Productivity Commission (2014, p. 414) states: ‘While a significant proportion of households appear to be underinsured, it is not known how many are making a fully informed choice and how many are underinsured due to market distortions (especially information asymmetry) or cognitive barriers.’

We also estimated the temporal variation of flood impacts on property values. Our hypothesis, based on past studies (see Lamond, 2008; Bin & Landry, 2013), was that property values tend to bounce back due to factors such as rehabilitation efforts undertaken after the floods (see Table 3). The regression results showed that in low-income suburbs property values are still in decline, while for high-income suburbs the sign is positive, although not statistically significant. In other words, the results indicate that property values in rich suburbs recover faster than those in poorer suburbs.

This research also highlights the importance of suburb characteristics for property values and the speed of recovery of property values after a flood event. These findings are important to guide post-disaster funding, such as response, relief assistance to households, recovery and mitigation undertaken by the state and local governments.

This study also provides additional information needed to create better price signals for effective risk management, as called for by the Productivity Commission (2014). As the Commission states: ‘For insurers to price premiums according to the risk the policyholder faces, they need information on natural disaster risk’ (p. 427). The Commission further states: ‘For insurers to efficiently price insurance products, good information is required for all these component elements of risk’ (p. 427). While, ‘historically, insurers have had limited information with which to price risk’ (p. 427), progress has been made with better information being made available. Nevertheless there is still room for improvement in areas such as ‘insurers’ knowledge about individual property characteristics’ (p. 428), distance to river and elevation of property and other suburb characteristics and amenity values.

In addition, this study provides further information that could assist insurers to price premiums based on the risk households face such as property, geographical as well as suburb characteristics. In short, the study identifies some of the main characteristics of ‘high-risk households’. As the Productivity Commission (2014, p. 414) states: ‘Governments should not address affordability concerns by providing subsidies, especially to high-risk households. Subsidies reduce the effectiveness of insurance in communicating and managing risk.’ The study results also provide an opportunity for insurers to narrow down risks beyond postcode levels in formulating insurance premiums where risks can be more accurately determined. This could avoid flat insurance premiums where ‘both high- and low risk policyholders are paying the same premium, [and] low-risk policyholders are likely to be cross-subsidising high-risk policyholders’ (p. 427). A lack of insufficient information ‘can also lead to insurers pricing defensively and charging higher premiums to all policyholders’ (p. 427).

Finally, there is the more general point that the availability of reliable information is a key precondition for competitive markets, given that investors need this information to make appropriate investment decisions. Awareness of the behaviour of property values is equally important for banks and other financial institutes in order to implement appropriate mortgage payments.