Weather-Related Home Damage and Subjective Well-Being

Abstract

Climate change is causing weather-related natural disasters to become both more frequent and more severe. We contribute to the literature on the economic impact of these disasters by using Australian data for the period 2009 to 2019 to estimate the effect of experiencing weather-related home damage on three measures of subjective well-being. Overall, we find little evidence of a statistically significant or sizable negative effect, on average, of weather-related home damage on subjective well-being.

1 Introduction

Climate change is causing an increase in both the frequency and the severity of weather-related natural disasters (World Meteorological Organization 2014). In fact, the number of weather-related natural disasters per decade rose by over 370% between 1971 and 2010 (World Meteorological Organization 2014).

The traditional approach used to estimate the economic impact of such disasters focuses on their impact on income-based measures.¹ Unfortunately, this approach has well-known weaknesses in the context of disaster impact measurement. Firstly, income-based measures may increase after a disaster due to the rebuild effort, which would be captured by this measurement approach; however, the associated opportunity costs would be overlooked. Secondly, income-based measures ignore many of the impacts of natural disasters, such as psychological costs, environmental harm, and loss of life.

Due to these factors, it is commonly agreed that most estimates substantially underestimate the true impact of natural disasters (Hochrainer 2009; Hallegatte and Przyluski 2010; Kousky 2014; Hallegatte 2015; Deloitte 2016; Sheng and Xu 2019). It is thus natural to look for alternative techniques that can address at least some of these issues and complement the more traditional income-based approach.

A relatively new but rapidly developing strand of economic literature focuses on measuring the impact of disasters by analysing the change in subjective well-being (SWB) (including Carroll et al. 2009; Luechinger and Raschky 2009; Kountouris and Remoundou 2011; Calvo et al. 2015a; Berlemann 2016; Sekulova and van den Bergh 2016; von Möllendorff and Hirschfeld 2016; Van Ootegem and Verhofstadt 2016; Hudson et al. 2019; Lohmann et al. 2019; Ahmadiani and Ferreira 2021; Berlemann and Eurich 2021; Frijters et al. 2021).

In this paper, we contribute to this literature by estimating the impact of experiencing weather-related home damage on subjective well-being in Australia, using a dataset that is much richer than those used in prior studies. While most papers in this literature use individual-level data on subjective well-being, few use measures of impact at the individual level; thus, they measure the impact of living in an area where a disaster took place rather than the effect of being affected personally by a disaster. In the survey that we use, however, the respondents were asked whether their household had experienced any weather-related housing damage.

While a few other studies use similar measures of direct exposure, these studies have small datasets of about 1,000 respondents (or fewer) and lack data from before the disasters.² Instead, our panel dataset spans an 11-year period, with over 20,000 people who were interviewed on average six times.

In addition to using a better dataset, we make two methodological contributions to the existing literature. Firstly, Bond and Lang (2019) argue that SWB regressions are highly dependent on the way in which ordinal happiness data are transformed into cardinal data, showing that a monotonic transformation that is rank preserving at the individual level can lead to coefficient estimates of the opposite sign. Following the method suggested by Bloem (2021), we perform a series of data transformations to investigate empirically whether our results are robust to changes in the assumed underlying cardinal scale. Second, the recent literature (Borusyak and Jaravel 2018; de Chaisemartin and D’Haultfœuille 2020; Callaway and Sant’Anna 2021; Goodman-Bacon 2021; Sun and Abraham 2021) shows that the standard two-way fixed-effect (difference-in-difference) analysis can be misleading when the impact of a treatment changes over time. We therefore present the results of both the traditional two-way fixed-effect analysis and the more robust approach suggested by Callaway and Sant’Anna (2021b).

Finally, we compare the impact on subjective well-being we estimate, to existing impact estimates on income-based measures, within the same context of natural disasters in Australia.

We find that experiencing weather-related home damage has a well-being effect that is relatively small in magnitude and is frequently statistically insignificant. More specifically, our main analysis shows that experiencing weather-related home damage explains at most 12% of the within-person standard deviation of subjective well-being. Our conclusions are generally robust to three different measures of well-being and a variety of assumed underlying cardinalizations and econometric methods. We further show that our finding about the limited impact on subjective well-being is consistent with some of the findings of studies that compute the impact of natural disasters in Australia on other outcome measures and discuss several explanations for the limited impact on well-being. These possible explanations include government assistance, insurance coverage, social comparisons, and the limited ‘average’ intensity of the disasters.

The rest of the paper is structured as follows. Section II discusses the literature on well-being and the well-being effect of weather-related disasters. Section III presents the data used in the analysis, section IV the methodology, and section V the results. Section VI provides a broader discussion of these results, while section VII concludes.

2 Well-Being and Weather-Related Disasters

Economic theory often assumes that people are rational expected utility maximizers. It is not possible to measure utility itself directly, but it is often assumed that income can be used as a proxy. The classical view among most economists is that not only does an increase in income increase utility but absolute income alone can be used to measure utility (Stutzer 2004). This view of absolute income being the sole proxy measure of utility is criticized for being too narrow and for ignoring the importance of other factors affecting well-being.

A relatively new but rapidly increasing strand of literature focuses on the use of self-reported subjective well-being as a direct measure of utility (Clark et al. 2008). A variable has to meet two requirements to be used as a proxy measure of utility: a strong theoretical basis showing why it should be related to utility and a reliable way to measure it. While there is a widespread consensus that people treat well-being, rather than income, as their life objective (Frey and Stutzer 2002; Rehdanz and Maddison 2005), whether subjective well-being can actually be measured accurately and reliably is, perhaps not surprisingly, a more controversial issue. Many argue, however, that most individuals can evaluate their own subjective well-being accurately and that self-reported subjective well-being is a reliable measure of true subjective well-being (Di Tella and MacCulloch 2006; Hayo 2007; Ferrer-i-Carbonell 2013).³

There is also some debate regarding whether well-being is measured completely by any one variable or whether different variables measure different aspects of overall well-being. It is generally agreed that there is a difference between hedonic well-being and eudaimonic well-being (Proctor et al. 2015). Hedonic well-being is attained through activities that generate happiness or pleasure, while eudaimonic well-being is gained through activities that generate fulfilment or a sense of purpose (Ryan and Deci 2001). As such, it is possible that different activities or events may not affect hedonic and eudaimonic well-being in the same way.⁴ Due to this difference between hedonic and eudaimonic well-being, different ways of measuring well-being may generate different results, depending on the degree to which hedonic and eudaimonic well-being are measured. As a result, happiness and life satisfaction may measure (at least slightly) different concepts. However, some authors view these differences as being relatively unimportant. For example, Waterman (1993) notes a strong empirical correlation between eudaimonic and hedonic subjective well-being. A similar conclusion is drawn by Di Tella and MacCulloch (2008), who conclude that happiness and life satisfaction are very strongly correlated and that empirical results are similar irrespective of which measure of subjective well-being is chosen. However, it should be noted that, for a single event (such as weather-related home damage), there is always the possibility for the choice of subjective well-being measure to make a material difference to the conclusion even if the choice is usually unimportant. This is especially true in our case as HILDA (Department of Social Services; Melbourne Instituse of Applies Economic and Social Research 2019) measures happiness as ‘how much time in the last four weeks a respondent has been happy’, while life satisfaction is measured with the answer to the question ‘all things considered, how satisfied are you with your life?’ In our analysis below, we therefore use various subjective well-being measures.

An enormous wealth of factors that have the potential to influence happiness are identified in the academic literature (see Table 1 and Dolan et al. (2008) for a comprehensive review of what influences subjective well-being).

Table 1 Categories of Factors Influencing Well-Being (Dolan et al. 2008)

While there are many factors that may influence happiness, it is often unclear how long-lasting any effect is. The potential for seemingly major events and variables to have only short-term, as opposed to long-lasting, influences on an individual’s well-being (utility) is widely accepted within the field of psychology but often disregarded by economic theory (Oswald and Powdthavee 2008). One empirical finding that suggests the stability of long-term happiness is the Easterlin Paradox. This is an empirical phenomenon that shows that, while absolute income has a significant influence on happiness at a given point in time, long-term increases in income are not associated with lasting increases in happiness (Easterlin 1974; Easterlin 2003; Paul and Guilbert 2013; Bartolini and Sarracino 2014).⁵ Two main theories, comparison theory and setpoint theory, help to explain the time dynamics of events in happiness and thus the stability of happiness, which we will discuss next.

Comparison theory suggests that happiness (and utility) is affected not by absolute quantity but by relative quantity and thus ranking against others (Clark et al. 2008). As a result, comparison theory proposes that adaption will occur to rank preserving quantity changes. The importance of rank and relative quantity is suggested as an explanation for the Easterlin Paradox and the broader trend of happiness stability over time (Di Tella and MacCulloch 2006; Clark et al. 2008; Paul and Guilbert 2013; Bartolini and Sarracino 2014).⁶ Empirical evidence suggests that income is especially affected by comparisons, while non-material variables, such as marital status and health, are far less influenced by changing aspirations (Di Tella and MacCulloch 2006). Additionally, evidence implies that comparisons matter more in rich countries and absolute quantities matter more in poor countries (Clark and Oswald 2002; Asadullah and Chaudhury 2012). In our context of Australia, a relatively rich country, comparisons could thus matter to the extent to which weather-related house damage has, for example, an uneven financial impact, affecting some households considerably more than others. If many people are affected relatively equally by the same shock, the relative position of individuals might remain more or less stable in the pre- and post-shock periods.

Setpoint theory, also known as adaption theory, argues that individuals have a natural level of happiness that they always return to in the long run (Brickman and Campbell 1971). Setpoint theory suggests that all events will have a short-term impact on happiness as opposed to a long-lasting effect and that gradual improvements to objective living standards and quality of life in the long term will have no effect on happiness as individuals will remain at their natural level of happiness, which is determined by their individual personality (Luhmann and Intelisano 2018). The empirical support for setpoint theory is mixed and suggests that the level of adaption is strongly heterogeneous, depending on the event type, event severity, and individual characteristics (Oswald and Powdthavee 2008; Baryshnikova and Pham 2018). In our context, setpoint theory would predict an impact if the disaster were severe enough to lead to a sudden and substantial deterioration of objective living standards, but also, in such a case, the impact would fade over time.

There is a growing literature focusing on the effect of weather-related events on subjective well-being. While these studies typically find a negative effect, there is little consensus about the size, duration, and statistical significance of this effect, with some studies finding large or lasting negative effects and others identifying no significant impact or only a short-term impact. Table 2 provides an overview of the studies in this area. Most studies, lacking information about who has been directly affected by a disaster, compare the life satisfaction of people living in areas affected by a disaster with the life satisfaction of people living in unaffected areas.

Table 2 Summary of the Literature on Weather-Related Disasters and Subjective Well-Being

Berlemann (2016) and Döpke and Maschke (2016) both measure disasters at the level of the country. Berlemann (2016) finds that countries that are affected by hurricanes tend to experience a decrease in life satisfaction, though often the estimated impact is statistically impossible to distinguish from zero. Similarly, Döpke and Maschke (2016) find no significant effect of natural disasters on the life satisfaction in a country.

More studies measure disasters at the regional level (Carroll et al. 2009; Luechinger and Raschky 2009; Kountouris and Remoundou 2011; von Möllendorff and Hirschfeld 2016; Ahmadiani and Ferreira 2021; Berlemann and Eurich 2021; Frijters et al. 2021). While these studies typically find negative and significant effects of various weather-related disasters, there is considerable variation in the size and duration of the effect. For example, while Frijters et al. (2021) document ‘small-to-moderate sized effects’ of climate and environmental disasters only within the first 2 weeks after the disaster, Luechinger and Raschky (2009) conclude that flood disasters ‘have large consequences’, with effects lasting for at least a year and a half.

Studies that use regional disaster data typically rely on large datasets with tens of thousands or even millions of observations, but studies that have access to information about which individuals have been directly affected by events often rely on small datasets, with hundreds of observations (Calvo et al. 2015; Sekulova and van den Bergh 2016; Van Ootegem and Verhofstadt 2016; Hudson et al. 2019; Lohmann et al. 2019). There is substantial variation in the findings of those studies: Van Ootegem and Verhofstadt (2016) find no significant effects of floods, while Sekulova and van den Bergh (2016) report large effects for those experiencing major damage but no significant impact for those experiencing only moderate damage. Similarly, Hudson et al. (2019) and Lohmann et al. (2019) find substantial negative effects on those experiencing disaster-related damage but smaller or insignificant effects on those experiencing the disaster but suffering no damage. Finally, Calvo et al. (2015) discern no effects of experiencing property damage but large negative effects of knowing people who died because of the disaster.

Like this study, studies that have individual-level measures of experiencing a weather-related disaster typically focus on an indicator that reflects whether (or the extent to which) a respondent’s home has been damaged by the weather event, though some also have access to broader indicators, like bodily damage (Sekulova and van den Bergh 2016) or bereavement (Calvo et al. 2015).

Note further that, in this literature, only two studies rely on panel datasets that follow the same individuals over time and hence can control for individual fixed effects. Von Möllendorff and Hirschfeld (2016) have a large panel dataset of German respondents but do not have information about who is affected directly by the extreme weather events. Calvo et al. (2015) do have information about who is affected directly by the extreme weather events but only have data for about 500 US women, each interviewed at three different points in time.

Finally, several papers study the impact of natural disasters in Australia, one of the countries most affected by disasters (Ladds et al. 2017). Focusing on the insurance costs of disasters in Australia, Ladds et al. (2017) find that most of the impact comes from storms, floods, tropical cyclones, and bushfires. Non-weather-related natural disasters, like earthquakes and landslides, are relatively rare in Australia.⁷ Like this study, Baryshnikova and Pham (2018) and Johar et al. (2020) use HILDA data, but Johar et al. (2020) focus on income-related measures, while Baryshnikova and Pham (2018) concentrate on mental health effects (measured using the Mental Health Component Summary Score).

As far as we are aware, the study by Carroll et al. (2009) is the only one to focus on the impact on subjective well-being of natural disasters in Australia and to estimate the impact of drought in Australia between 2001 and 2004. They find that a spring drought causes a large loss of well-being for rural communities but has no effect on urban communities, while droughts in other seasons have no significant effect at all. The authors note that this relatively short time period includes an especially severe drought that occurred in 2002 (Carroll et al. 2009). It is thus at least possible that their estimate is more an estimate of that particular event than of general drought.

Summarizing, there is a growing literature that investigates the impact of extreme weather events on subjective well-being. Our research differs from the existing literature in several ways. First, our study uses a large-scale, nationally representative, individual-level panel dataset that allows us to control for individual-level fixed effects and contains information about whom among the respondents has been directly affected by extreme weather events. The study by Calvo et al. (2015) is the only other study that focuses on SWB and uses three waves of individual-level panel survey data for about 500 women. This paper uses 11 waves of an individual-level panel survey of about 24,500 individuals, a sample representative of the Australian population. Other studies that focus on SWB use either cross-sectional data at a single point in time or repeated cross-sectional data. They can thus control for fixed effects only at the regional level, estimating the impact of living in an area affected by an event rather than the impact of being directly affected by the event. Second, as explained in the methodology section below, our paper is the first to address two recent methodological criticisms of the existing literature that the results of life satisfaction regressions are sensitive to the way in which ordinal happiness data are transformed into cardinal data and that the traditional two-way fixed effects used in several papers can result in misleading estimates. Finally, while there is only one estimate of the impact of floods on subjective well-being in Australia, there are several studies that estimate either the insurance cost or the income loss associated with various types of weather-related disasters. Our study will thus contribute to the comparison of so-called objective and subjective impacts.

3 Data and Data Transformations

We use data from the Household, Income and Labour Dynamics in Australia (HILDA) Survey to estimate the impact of weather-related home damage on people’s happiness in Australia. HILDA is an annual household-based longitudinal survey that was first collected in 2001 and currently consists of 19 years of data from 2001 to 2019. HILDA contains a wide variety of information and is designed to facilitate research into three primary areas: incomes, labour markets, and family dynamics (Department of Social Services; Melbourne Institute of Applied Economic and Social Research 2019). According to two of the lead survey designers,

The Household, Income and Labour Dynamics in Australia (HILDA) Survey is one of only a small number of well-established, large, nationally-representative household panel studies conducted in the world. (Watson and Wooden 2012, p. 369)

HILDA contains several variables of particular relevance to this research. Since 2001, HILDA has asked respondents to evaluate their overall life satisfaction. More specifically, HILDA asks ‘all things considered, how satisfied are you with your life?’ Respondents are asked to pick a number between 0 and 10 to indicate how satisfied they are, with 0 meaning totally dissatisfied, 5 meaning neither satisfied nor dissatisfied, and 10 meaning totally satisfied.

There is some debate within the literature surrounding the distinction between happiness and life satisfaction, with the resulting possibility that superficially similar questions may lead to differing results. Because of this, we also include two alternative dependent variables collected by the HILDA Survey, which we use to check the robustness of our findings. These alternative variables are happiness, measured as the amount of time in the last four weeks that the respondent has been happy,⁸ and satisfaction with the home in which the respondent lives.⁹ Since 2009, HILDA has also asked each respondent whether, within the last 12 months, ‘a weather-related disaster (flood, bushfire, cyclone) damaged or destroyed’ the respondent’s home.¹⁰ About 1.5% of observations report weather-related home damage (WRHD).

Using HILDA data, Johar et al. (2020) find that people who report damage are more likely to have spent money on repairs, renovations, and maintenance of homes, have higher expenditures on those items, and spend more on non-medical insurance. The authors conclude that their findings suggest that the ‘exposure measure reflects true significant housing damage and destruction’ (Johar et al. 2020, p. 10). They are also able to match approximately two-thirds of the HILDA reports to known events.

We use these responses as our primary dependent variable and independent variable, respectively. The dataset contains 160,452 observations of 24,500 people, implying that each person is interviewed approximately 6.5 times on average. The interviews were primarily conducted between September and October, with over 80% taking place during this period and over 90% between September and November.

It is possible that different groups of people are affected differently by a weather-related disaster. Due to this possibility, we create four binary variables to divide the data into subsets.¹¹ These binary variables are gender, homeownership status, family prosperity, and insurance status. Gender is 1 if the person is male and 0 if the person is female. Homeownership status is 1 if the person is a homeowner and 0 otherwise. Family prosperity is 1 if the person self-reported as being ‘prosperous’, ‘very comfortable’, or ‘reasonably comfortable’ and 0 if the person self-reported as either ‘just getting along’ or being ‘poor’ or ‘very poor’. We also construct a dummy variable to represent the respondents’ insurance status, the details being provided in Section b. The descriptive statistics are presented in Table 3.

Table 3 Descriptive Statistics

In the past, the reliability and validity of self-reported subjective well-being among individuals was assumed to validate the use of this measure in groups, the underlying assumption being that validity at the individual level ensured validity at the aggregate level.¹² Effectively, it was believed that the aggregated well-being score of Group 1 could be compared with the aggregated well-being score of Group 2. However, more recently, this has become a point of contention, with the discussion centring on the legitimacy of ranking aggregated ordinal data.

It is widely accepted that subjective well-being data are ordinal, rather than cardinal, in nature. Despite this, a significant proportion of happiness researchers treat subjective well-being data as cardinal and thus use ordinary least squares to perform their analysis (Schröder and Yitzhaki 2017). Researchers who are unwilling to assume cardinality instead usually use ordered probit or logit models in their research. Typically, though, ordinal models return similar estimates to cardinal methods (Ferrer-i-Carbonell and Frijters 2004). Overall, the similarity of empirical results between cardinal and ordinal models leads many to conclude that, for subjective well-being data, the distinction between cardinality and ordinality is unimportant (Schröder and Yitzhaki 2017).

The recent literature has challenged this assumption, showing that common models may produce unreliable results when applied to ordinal data. Bond and Lang (2019) show that it is effectively impossible to rank two groups in order of happiness. A monotonic transformation that is rank preserving at the individual level but will change the rank at the group level can almost always be found.¹³ This also means that two monotonic transformations that are individually rank preserving can provide coefficient estimates of the opposite sign. We provide a simple two-person example of this in Table 4. At the original scale, the group means of Group 1 and Group 2 are identical. However, this is fragile to a change in the underlying scale. Applying a monotonically increasing convex transformation results in Group 1 having a higher mean than Group 2, while a monotonically increasing concave transformation has the opposite effect.

Table 4 Data Transformation Example

This strong criticism of the reliability of subjective well-being data for use in comparing groups also casts doubt on the conclusions of the existing studies. In response to this, several authors attempt to defend the robustness of the existing literature and to put forward alternative methods of analysis for future research using subjective well-being data. Chen et al. (2019) suggest focusing on the median instead of the mean when analysing subjective well-being data due to the robustness of the median to monotonic rank-preserving transformations. Similarly, they suggest using the equality of mean and median in symmetric distributions to reinterpret existing ordered logit and probit models as median, rather than mean, regressions. This reinterpretation makes the conclusions of many existing papers robust to the criticism of Bond and Lang (Chen et al. 2019).

An alternative approach is taken by Bloem (2021), who suggests using a series of data transformations to check the robustness of results obtained using cardinal models. His research presents two key insights. The first is that passing the theoretical criteria developed by Schröder and Yitzhaki (2017) and Bond and Lang (2019) is not a guarantee of reliable results.¹⁴ While it may be possible that there is no monotonic rank-preserving transformation that will reverse the sign of an estimated coefficient, this is no guarantee that valid transformations will not substantially change both the size and the significance of an estimated coefficient. The second insight is that failing to meet the theoretical criteria does not automatically invalidate the results obtained if only relatively extreme transformations substantially influence the results (Bloem 2021). Following on from these two insights, Bloem (2021) suggests applying a series of transformations to the underlying data to check the robustness of results to different underlying cardinalizations. We follow this suggestion, which ensures that the endpoints of the data remain numerically unchanged. The transformation is given by Eq. (1):

$$T\left(Y\right)={Y}_{MAX}\text{*}{\left(\frac{Y}{{Y}_{MAX}}\right)}^{\sigma }$$

(1)

When σ is exactly equal to one, the transformed scale exactly equals the reported scale. This means that the ‘true’ distance between reported values of 1 and 2 is assumed to be the same as the distance between reported values of 9 and 10. For values of σ greater than 1, the scale is assumed to be convex. This means that the gap between reported values of 1 and 2 is assumed to be smaller than the gap between reported values of 9 and 10. For values of σ less than 1, the scale is assumed to be concave, which implies that the gap between reported values of 1 and 2 is greater than the gap between reported values of 9 and 10. The transformation process is shown in Fig. 1, where the σ = 1 line reflects the original reported values:

Fig. 1

Transformed Variable Example

Intuitively, low levels of sigma (top lines) can be interpreted as meaning that people are pessimistic and need relatively high levels of latent life satisfaction (transformed value) not to choose to report a low number on the life satisfaction scale (the original value). In addition, high levels of sigma can be interpreted as meaning that people are optimistic; even at relatively low levels of latent life satisfaction, they will choose a high number on the life satisfaction scale.

We apply this transformation to all three dependent variables (life satisfaction, recent happiness, and home satisfaction), using 100 evenly spaced values of σ between 0.1 and 10. The descriptive statistics for a sample of the transformed variables are presented in Table 5. We display within-person standard deviations as they are the most relevant.¹⁵

Table 5 Transformed Dependent Variables and Within-Person Standard Deviations

Table 5 illustrates that, while a sigma of 1 uses the actual answers given by the respondents, small values of sigma transform almost all the answers into the top score (10) and hence suggest that almost everybody is very happy and satisfied with life. In contrast, big values of sigma transform most answers into low scores for happiness and life satisfaction. In short, sigma values that are close to 1 are closer to the respondents’ answers, with small or large sigma values implying a substantial transformation from respondents’ answers. Note that Bloem (2021) suggests that near-linear transformations are more plausible for life satisfaction scales as three experimental studies (Banks and Coleman 1981; van Praag 1991; Oswald2008) all failed to reject linear transformations of life satisfaction scales. Bloem (2021) identifies the interval [0.4, 2.5] as the most plausible range for a study that uses a 0–10 life satisfaction scale like the one used here.

4 Methodology

To examine the impacts of weather-related disaster events on subjective well-being, we use a two-way fixed-effect (TWFE) difference-in-difference (DiD) estimation approach. Equation (2) presents the TWFE estimation, where the measure of subjective well-being (SWB) is alternatively LifeSat(isfaction), HomeSat(isfaction), or Happiness. Individual and time fixed effects are denoted by ${\alpha }_{i} \text{and }{t}_{t}$, respectively.

$\begin{array}{c}{SWB}_{it}={\alpha }_{i}+{t}_{t}+{{\beta }_{1}WeatherDamage1}_{it}+{{\beta }_{2}WeatherDamage2}_{it}+\\ {{\beta }_{3}WeatherDamage3\text{+}}_{it}+{\beta }_{4}{age}_{it}+{\beta }_{5}{age}_{it}^{2}+{u}_{it} \#\end{array}$

(2)

We estimate three different TWFE treatment effects. Any impact of weather-related home damage (WeatherDamage) on subjective well-being that occurs within 12 months of treatment (Year 1) is captured by the term ${\beta }_{1}$. Similarly, ${\beta }_{2}$ is the estimate of any impact between 13 and 24 months following treatment (Year 2). Finally, we allow for a ‘long-run’ treatment effect that occurs in all later time periods (Year 3+).

As previously discussed, it is possible that different groups of people experience different subjective well-being effects from weather-related home damage (see Eq. 3). Other studies that use subjective well-being find mixed evidence of the existence of such heterogeneous effects. We analyse possible heterogeneous effects using four additional categories: gender, wealth, homeownership status, and insurance.

$$\begin{array}{c}{SWB}_{it}={\alpha }_{i}+{t}_{t}+{{\beta }_{1}WeatherDamage1}_{it}+{{\beta }_{2}WeatherDamage2}_{it}+\\ {{\beta }_{3}WeatherDamage3\text{+}}_{it}+ {\beta }_{4}{age}_{it}+{\beta }_{5}{age}_{it}^{2}+{\beta }_{6}{Subset}_{it}+\\ {{\beta }_{7}WeatherDamage1}_{it}{Subset}_{it}+{{\beta }_{8}WeatherDamage2}_{it}{Subset}_{it}+\\ {{\beta }_{9}WeatherDamage3+}_{it}{Subset}_{it}+{\beta }_{10}{age}_{it}{Subset}_{it}+{\beta }_{11}{age}_{it}^{2}{Subset}_{it}\\ +{\beta }_{12}{t}_{t}{Subset}_{it}+{u}_{it} \\ \#\end{array}$$

(3)

A key insight from the recent DiD literature is that the common TWFE approach can result in biased estimates in the presence of treatment heterogeneity either across units or over time. While it is well known that the OLS (and thus the TWFE) approach uses changes in treatment status to generate estimates, it was not fully recognized until recently that this would implicitly result in newly treated units being compared with previously treated units rather than only with units that have not yet been treated. This comparison of treated units with other treated units can result in the TWFE estimates being biased (Borusyak and Jaravel 2018; de Chaisemartin and D’Haultfœuille 2020; Callaway and Sant’Anna 2021; Goodman-Bacon 2021; Sun and Abraham 2021). The approach developed by Callaway and Sant’Anna (2021b) bypasses this issue by only comparing treated units with units that have not yet been treated. This creates a series of group–time estimates, which can then be aggregated flexibly into treatment parameters.^16,17

Both the standard TWFE approach and the Callaway and Sant’Anna (2021b) approach impose a substantively similar set of assumptions on the underlying data-generating process to obtain causal estimates.¹⁸ In addition to the assumptions of the ordinary least squares (OLS) technique, four additional assumptions are required for a DiD approach to be valid: the exogeneity assumption, the no pre-treatment effect assumption, the stable unit treatment value assumption (SUTVA), and the common trends assumption (Lechner 2010). If these four assumptions are met by the underlying data-generating process, the DiD estimator will result in causal estimates.

The exogeneity assumption states that the covariates, or control variables, must be independent of the treatment status (Lechner 2010).¹⁹ It is easy to see why this must be the case. Consider the relationships between happiness, weather-related home damage, and income, where weather-related home damage is the dependent variable and income is a covariate. It might seem reasonable to include income as a control variable in the regression as it would be expected to influence happiness. Unfortunately, this would violate the exogeneity assumption. Experiencing weather-related home damage may reasonably be expected to change income, resulting in endogeneity and invalid estimates. For this reason, only control variables that cannot be influenced by the treatment status may be included in a DiD regression. Clearly, fixed-effect variables pass this criterion, as does the respondent’s age. However, for other time-variant variables this assumption of independence of the treatment status is harder to defend. To avoid for ‘over-controlling’ (Wooldridge 2020, p. 199) or ‘controlling for post-treatment variables’ (Gelman and Hill, 2006, p. 188), in our main specification, we only control for age²⁰ as well as both individual and time fixed effects. As a robustness check, we also run TWFE regressions with a larger set of control variables including dummy variables for various levels of household income, rental payments, geographic region, education, marital status and health.

The assumption of no pre-treatment effects states that the treatment must have no effect in the pre-treatment period (Lechner 2010). While this may appear to be trivially true, it is not actually the case in some situations. Primarily, this assumption is violated whenever a person can anticipate treatment or is aware of how they can change their treatment status. We view it as unlikely that this assumption poses a major issue in this analysis as it would involve people being aware of when they will suffer weather-related home damage and pre-emptively changing their happiness to reflect this.

SUTVA states that an individual’s outcome must only be affected by their individual treatment status and not by the treatment status of others around them. SUTVA is satisfied if there are no spillover or interaction effects between individuals (Lechner 2010). Two potential violations of SUTVA exist in this analysis. Firstly, it is possible that people’s happiness partially depends on the happiness of the people whom they know. This is intuitively a very reasonable proposition – it seems likely that one’s happiness may be influenced by the happiness of one’s family, friends, or neighbours. Secondly, utility may be affected by relative quantity rather than absolute quantity. This would result in a person being affected by the treatment that other people receive. Suppose that an individual’s home was not damaged by a weather-related disaster but that the homes across the road were damaged.²¹ The impact of such ‘near misses’ on people’s happiness is unclear. Happiness may decline, with the intuition being that people care about those around them and thus become less happy when they experience material loss due to a weather-related disaster. Their own happiness may remain unchanged as they have personally suffered no loss or no gain. Alternatively, their happiness could increase. The charitable explanation of this would be that they feel lucky to have emerged unscathed. A considerably less charitable explanation is that they are now considerably better off than those who experienced damage and that has increased their happiness. Both happiness interdependence and the influence of spillover treatment effects would violate SUTVA. The literature mentioned in Sect. 2 suggests that the significance and size of these issues are unclear; however, they seem unlikely to be overly large in comparison with the main effect (of experiencing weather-related home damage personally).

The common trends assumption is widely considered to be the most important DiD assumption (Angrist and Pischke 2008; Friedman et al. 2013). For the common trends assumption to be satisfied, the difference in pre-treatment outcomes between the treatment and the control group must remain constant over the time period during which treatment occurs. This key assumption is not testable as the required counterfactual (the outcomes of the treated group had no treatment) is clearly not observable. However, when researchers have several time periods worth of data available, they can check whether the treatment and control groups have common trends in the periods when treatment does not occur using placebo regressions (Lechner 2010). While this does not ensure that the common trends assumption is met, it can provide the researchers with a degree of confidence regarding the likely validity of the assumption. Of course, the researcher should also consider whether the treatment coincides with any other shock that is likely to violate the common trends assumption. We view that as unlikely to be the case here as the treatment (experiencing weather-related home damage) is ‘messy’, with different people experiencing treatment in different years.²²

We conduct a common trends placebo test using both estimation approaches. Following the TWFE approach, we regress life satisfaction on a dummy variable that takes the value of 1 the year before weather-related home damage is experienced²³ for a range of σ values between σ = 0.1 and σ = 10. Using the Callaway and Sant’Anna (2021b) technique, we further perform an event study that includes estimated placebo treatment effects for all the years before treatment occurs and find support for common trends. The results are presented in Fig. 2 and suggest that the common trend assumption is reasonable, except possibly at very large or small values of sigma, which, as discussed above, are likely to be implausible.

Fig. 2

Placebo Regression of Weather Damage on Life Satisfaction

Summarizing, there are four important considerations when analysing the effect of a staggered treatment event on subjective well-being. Firstly, the extent to which different subjective well-being variables measure individual elements of well-being rather than measuring overall well-being remains unclear. A common example of this is the distinction between happiness and life satisfaction. In the dataset that we use, there is some evidence to suggest that they measure different things, with relatively low correlations between the two variables. The average within-year correlation of life satisfaction and happiness is 0.47, while the average within-person correlation is only 0.22. Secondly, the effect of an event on subjective well-being may be heterogeneous, with different groups being affected by treatment differently. Thirdly, it has been shown that the use of TWFEs can produce biased estimates in the presence of treatment heterogeneity. Finally, it is problematic to treat subjective well-being data as cardinal for the purposes of analysis (which is commonly the case) given that almost all authors agree that subjective well-being data are ordinal in nature. As shown by Bond and Lang (2019), this can lead to misleading results that are fragile to monotone transformations of the underlying subjective well-being data.²⁴

We employ four techniques in our estimation strategy to mitigate these issues and provide more robust estimates. Firstly, we use three different dependent variables in our analysis: life satisfaction (LifeSat), satisfaction with housing (HomeSat),²⁵ and happiness (Happiness).²⁶ The use of both life satisfaction and happiness reduces the risk of a misleading result arising from measuring a subset of subjective well-being rather than overall subjective well-being. We also include satisfaction with housing as it is likely to be the most sensitive to weather-related home damage.²⁷ Secondly, we check for heterogeneous treatment effects based on insurance status, gender, homeownership status, and self-reported prosperity level. Finally, we use two estimation techniques to generate more robust estimates: we apply a series of data transformations, as suggested by Bloem (2021), to moderate the concerns raised by Bond and Lang (2019), and we supplement the typical TWFE estimation approach with the novel technique developed by Callaway and Sant’Anna (2021b).

The model specifications that we employ include control variables for age and age squared as well as both individual and time fixed effects. In addition, all the TWFE models are run using individual-cluster and time-cluster robust standard errors.²⁸ The Callaway and Sant’Anna (2021b) models are implemented using bootstrapped simultaneous confidence intervals.²⁹

5 Results

5.1 Weather-Related Home Damage’s Effect on Subjective Well-Being

Table 6 presents the results obtained from the TWFE model in Eq. (2), while Table 7 contains the TWFE estimates as a percentage of the within-person standard deviation.

Table 6 Weather-Related Home Damage’s Effect on Subjective Well-Being (TWFEs)

Table 7 Effect of Experiencing a Weather-Related Natural Disaster as a Percentage of the Within-Person Standard Deviation of Subjective Well-Being (TWFEs)

Using the common TWFE model (Table 6), we estimate a 0.08-point decrease in life satisfaction in the 13–24-month period (Year 2) after experiencing weather-related home damage. This estimate is statistically significant at the 1% level; however, the estimated effect is relatively small and is less than 12% of a within-person standard deviation in life satisfaction (Table 7). Using happiness (rather than life satisfaction) as the dependent variable, we estimate a 0.032-point decrease in the 0–12-month period following a weather event. This estimate is statistically significant at the 5% level; however, it is again small at 5.65% of a within-person standard deviation. We also estimate a similarly sized decrease in long-run happiness (from the period more than 2 years ago), which is statistically significant at the 10% level. Ex ante, home satisfaction may reasonably have been the measure of subjective well-being that is most sensitive to weather-related home damage, but this is shown not to be true ex post. We estimate a 0.08-point decrease in home satisfaction in the 0–12-month period following a weather event. This estimate is again relatively small (8.57% of a within person standard deviation) and only statistically significant at the 10% level.

Table 8 further reports the results of a robustness check where a larger set of control variables are included in regression Eq. (2), including dummies for various levels of household income, rental payments, geographic region, education, marital status and health.

Table 8 Weather-Related Hom-e Damage’s Effect on Subjective Well-Being (TWFEs) – Full Controls

Comparing the estimates in Tables 6 and 8 shows that coefficient estimates are similar whether or not a large set of control variables is included. This suggests the results we obtain are robust to changes in control variables included.³⁰

Table 9 presents the results obtained from the CSA1 model applied to Eq. 2. Figure 3 shows the event study estimates obtained from the CSA2 model.

Table 9 Overall Effect of Weather-Related Home Damage on Subjective Well-Being (CSA1)

Fig. 3

Event Time Effect of Weather-Related Home Damage on Subjective Well-Being (CSA2)

Using the more advanced and theoretically appropriate Callaway and Sant’Anna (2021) technique, we find no evidence that weather-related home damage has any effect on subjective well-being. With the overall ATT estimate from the CSA1 model, we estimate that the overall effect from experiencing weather-related home damage is a 0.04-point decrease in life satisfaction. We further estimate a 0.02-point decrease when using happiness as the dependent variable and a 0.02-point increase in home satisfaction. None of these estimates are statistically significant. Similarly, using the CSA2 model to estimate the effect at each point in the event time, we find that all the estimates are consistently small and statistically insignificant.

Overall, we find that experiencing weather-related home damage has limited effects on subjective well-being. While the estimates are generally negative, they are relatively small in magnitude and often statistically insignificant.³¹ Additionally, the statistically significant results that we obtain from the TWFE approach are not robust to alternative measures of subjective well-being. Using the more advanced Callaway and Sant’Anna (2021) techniques, we find evidence to suggest that, on average, weather-related home damage has no effect on subjective well-being. The estimates are small and statistically insignificant across all three measures of subjective well-being and both aggregation approaches.

We now carry out a robustness check of our main model estimates by applying the data transformation suggested by Bloem (2021). It is important to note that any such transformation will affect both the mean and the distribution of the data and thus the resulting coefficient estimates and p-values. Section III above provides an overview of how the mean and within-person standard deviation change when the transformation is applied. We present the regression results of the transformed data below.

Fig. 4

Weather-Related Home Damage’s Effect on Transformed Life Satisfaction (TWFEs)

Fig. 5

Weather-Related Home Damage’s Effect on Transformed Home Satisfaction (TWFE)

Fig. 6

Weather-Related Home Damage’s Effect on Transformed Happiness (TWFEs)

The TWFE estimates of the effect of weather-related home damage on transformed life satisfaction, satisfaction with housing, and happiness are displayed in Figs. 4 and 5, and 6, and the CSA1 estimates are presented in Fig. 7.³² All the figures display both the point estimate and the robust 95% confidence interval for different values of sigma.³³

Fig. 7

Weather-Related Home Damage’s Effect on Transformed Subjective Well-Being (CSA1)

In Figs. 4, 5, 6 and 7, we show that the estimates tend to be robust to transformations of the assumed underlying cardinal scale and that the assumption of a different scale would typically not change the conclusions drawn above. The estimates are generally negative, small in size, and often statistically insignificant. In Fig. 7, we combine the Callaway and Sant’Anna (2021b) technique with the Bloem (2021) data transformation. We find estimates that are statistically insignificant at the 5% level for all the values of sigma and all the measures of subjective well-being.

5.2 Robustness Checks

We next analyse the extent to which different groups of people experience different subjective well-being effects from weather-related home damage. We find no significant differences in impact by gender or insurance status, while the differences by homeownership status vary widely across specifications.³⁴ In this section, we thus focus on the effects of weather-related home damage (WeatherDamage) on subjective well-being (SWB) for both the wealthy and the poor.

People whom we define as wealthy are all those who self-reported as being at least ‘reasonably comfortable’, while those whom we define as poor reported being either ‘just getting along’ or worse.³⁵ The TWFE estimates are presented in Table 10, while the CSA1 sub-group estimates are presented in Table 11.

Table 10 Weather-Related Home Damage’s Effect on Subjective Well-Being by Wealth (TWFE)

Table 11 Weather-Related Home Damage’s Effect on Subjective Well-Being by Wealth (CSA1)

Across both the TWFE and the CSA1 model, the coefficient estimates for the wealthy are small and statistically insignificant even at the 10% level. In addition, five out of the nine estimated coefficients are positive. In general, the wealthy do not experience any loss of well-being from a weather-related disaster and may even experience a small gain

There is mixed evidence regarding whether the poor experience a decline in subjective well-being following a weather-related natural disaster. The coefficient estimates are always more negative than for the rich irrespective of which model is used. The TWFE estimates suggest a statistically significant decline in life satisfaction across all the time periods, with coefficient estimates between − 0.12 and − 0.15. There is also evidence to suggest a decline in home satisfaction in the 0–12-month period following a weather event as well as a 0.06-point decline in happiness in both periods of 0–12 months and 25 + months. However, the Callaway and Sant’Anna (2021) models suggest that even the poor do not experience a statistically significant decline in subjective well-being. It should be noted that the estimates are still always more negative for the poor than for the rich.

Considering the Callaway and Sant’Anna technique is theoretically more appropriate than considering two-way fixed effects, and it seems likely that even the differences between the wealthy and the poor are minor and that that, on average, neither the wealthy nor the poor experience a significant decrease in subjective well-being from weather-related home damage.

6 Discussion

We have found little evidence to suggest that weather-related home damage leads, on average, to a sizeable decline in subjective well-being. Using the common two-way fixed-effect approach, we estimate a statistically significant (at the 1% level) decline in life satisfaction in the 13–24-month period following weather-related home damage; however, this decline is limited in size, explaining less than 12% of a within-person standard deviation. In addition, the estimated coefficients using alternative measures of subjective well-being, like happiness or home satisfaction, tend to be smaller and/or less statistically significant. Using the more advanced Callaway and Sant’Anna approach, we further find no evidence of any statistically significant effect across all three measures of subjective well-being. These conclusions are robust to a wide variety of data transformations.

Furthermore, we find that the estimates are similar for men and women and for people with and without home contents insurance. The estimates remain small in size, are frequently statistically insignificant, and are not robust to alternative measures of well-being. However, we find somewhat more mixed results when considering heterogeneity based on prosperity. In particular, the TWFE model shows that the poor experience a statistically significant and long-lasting decline in life satisfaction following weather-related home damage but that the wealthy experience no effect. The more advanced CSA model finds that neither the wealthy nor the poor experience a statistically significant decline in subjective well-being following weather-related home damage, although it should be noted that the estimates for the poor are always more negative than those for the rich.

Thus, in general, subjective well-being was not found to be substantially affected by weather-related home damage. This might seem to be a surprising result at first glance, especially given that the insignificant effect of experiencing weather damage extends not just to life satisfaction and happiness but also to satisfaction with housing.

How do our results compare to other estimates of the impact of natural disasters in Australia? A number of studies estimate the ‘average annual cost of natural disasters’ in Australia by focusing on insured costs. McAneney et al. (2019) estimate that ‘the annual average insured cost of natural disasters is approximately AUD 2 billion with an standard error of the same magnitude’. Ladds et al.’s (2017) review finds four studies with average annual estimates of $1.22 billion, $1.65 billion, $1.75 billion, and $3.26 billion.³⁶ They too point to the uncertainty around these estimates related to both the quality of the data used and the specific factors that are included in the estimates. The different studies do not even agree on the number of fatalities or the number of disasters.

While these dollar amounts might seem sizable, they are relatively small on a per capita basis³⁷ or as a fraction of the GDP.³⁸ At the same time, disasters do not affect everybody to the same extent and, hence, costs are typically not uniformly distributed across the population. In our data, roughly 1.5% of the sample experiences weather-related damage. Hence, while the overall ‘average’ loss might be relatively small, the average loss across those directly affected, which is what our life SWB estimates measure, could be more substantial.

Baryshnikova and Pham (2018) and Johar et al. (2020) also employ Australian HILDA data but focus on other impact measures. Johar et al. (2020) use the HILDA dataset to estimate the impact of weather-related house damage on variables like income and employment. They find that there is ‘little evidence of any significant impacts on employment or household income’ (Johar et al. 2020, p. 19). They do find some evidence of increased financial hardship among those affected by weather-related housing damage. More specifically, their basic specification suggests that direct disaster exposure leads to a 4.9% point increase (or 160% increase) in the chance of reporting a ‘major worsening in financial situation’ during the past 12 months. Using our specification and dataset, we find a similar, albeit somewhat smaller, effect, with those affected by weather-related housing damage being 2% points (from about 3% to about 5%) more likely to report a major worsening of their financial situation. While such worsening of one’s financial situation is likely to have a negative effect on one’s life satisfaction,³⁹ given that the vast majority of those experiencing weather-related damage do not report such worsening, the average impact on life satisfaction across those affected by weather-related damage might well be hard to distinguish from noise in the data.

Baryshnikova and Pham (2018) use HILDA data and a quantile approach to estimate mental health effects (measured using the Mental Health Component Summary Score). While they find substantial heterogeneity across quantiles, they conclude that experiencing weather-related housing damage in the previous 12 months is associated with a 0.37 drop in the mental health score. Note that the mental health score ranges between 0 and 100, is on average about 50, and has a standard deviation of around 10. Hence, the 0.37 drop represents about 4% of a standard deviation, which is not unlike some of the effect sizes that we found.

In summary, our finding of limited effects on subjective well-being are in line with the findings of other studies that use other ways of measuring the impact of natural disasters in Australia.

There are several plausible explanations for why we do not find a sizeable impact of weather-related home damage on subjective well-being. Of course, the first is that most people’s well-being is simply unaffected by experiencing weather-related natural disasters, and therefore a null result is almost guaranteed. However, there are several alternative explanations.

One possibility is that there is selection into treatment status by the size of the treatment effect. Assuming that certain locations are more prone to weather-related disasters, it is a reasonable expectation that these disaster-prone locations primarily contain people who are relatively less affected by these disasters than those who choose to live elsewhere. If there is indeed selection into treatment, the conclusion (of a null result) remains valid providing that those weather-related natural disasters remain ‘predictable’ by location and thus allow selection into treatment to continue. If climate change results in weather-related natural disasters becoming more randomized so that selection into treatment could no longer occur, the average affect (of weather-related home damage on subjective well-being) would be likely to increase.

In addition to selection into treatment, there are several uncontrolled-for factors that may also help to explain our results: the severity of damage experienced, social comparisons, insurance coverage, and government assistance. First, the results may be explained by a low average severity of experienced damage. When studying the impacts of weather events on subjective well-being in Papua New Guinea, Lohmann et al. (2019) find that suffering more severe damage resulted in a larger loss of well-being than suffering moderate damage. Unfortunately, data on the severity of the weather-related home damage experienced was not collected as part of the HILDA Survey, making this theory untestable. However, it is at least a possibility that the majority of people who experienced weather-related home damage only experienced mild damage, consistent with Johar et al. (2020), and that this dominated when deriving the estimated effect. Interestingly, a fixed dollar value level of damage would appear bigger (in relative terms) for the poor than for the wealthy. Thus, even if the majority of experienced damage appeared minor to most people, it may have appeared large to the poor and as a result affected their well-being more than that of other groups.

Social comparisons may also help to explain our findings. It is widely accepted in the well-being literature that well-being may depend just as much on relative factors as on absolute factors. This social channel may lead to a null result found. Firstly, weather disasters are naturally clustered. A given disaster will be confined geographically. Thus, people who are affected by a weather disaster are almost certain to know other people who are affected by a weather disaster. If comparisons are strongly localized, people may adapt quickly to a weather disaster because everyone around them is also affected. In reality, this is related to the conclusion drawn by Calvo et al. (2015), who find that social connectedness played an important role in well-being recovery following Hurricane Katrina, with the socially isolated experiencing a long-term drop in well-being.

Insurance coverage may also help to explain the lack of any significant impact on well-being. Australia is a wealthy developed country with a functioning insurance market. Using European data over a 25-year time span, Luechinger and Raschky (2009) find that, while experiencing a weather-related disaster did have a negative effect on well-being, this was mostly mitigated by the presence of risk transfer mechanisms. While we tried our best to control for this, HILDA has limited data on insurance coverage. Data on whether respondents had home contents insurance was only collected in two years: 2014 and 2018. Among people whose insurance status did not change between these two years, the insurance coverage was high at just under 75%. It is certainly possible that a highly developed insurance market and high rates of insurance coverage may help to explain the results that we obtain. At the same time, using the limited information on insurance available in HILDA, we did not find any evidence that would suggest that the impact of weather-related home damage was bigger for the uninsured.

Finally, the presence of government disaster relief efforts may help to explain the results. The Australian Government has numerous disaster relief programmes that provide financial assistance after a disaster. For example, both one-off payments and short-term income support payments are available for individuals who are affected by disasters (Portillo-Castro 2019). As of 2018, it was estimated that the direct cost of these disaster recovery programmes was $2.75 billion per year (Portillo-Castro 2019). Given that the aim of disaster recovery aid is to minimize the harm caused by a disaster and to aid the recovery, finding small and/or insignificant impacts may at least partially be caused by government disaster relief efforts.

Given the limitations of our particular dataset, it is difficult to know the degree to which each of these four factors helps to explain the overall result and hence to present clear policy recommendations. However, it certainly seems reasonable that low average disaster severity, social comparisons, a high level of insurance coverage, and government disaster relief efforts all play a role in explaining the results. It is also not clear whether the impact of future weather-related disasters will be similar to that of previous ones. Climate change is likely to make extreme weather events increase in both severity and frequency, exposing more people to weather-related home damage and making it harder for selection into treatment to occur. It is worth noting that climate change also has the potential to reduce the availability of insurance and the ability of governments to provide post-disaster aid. All these possibilities could reasonably be expected to result in future weather disasters having a larger impact on well-being than past ones.

7 Conclusion

As the rate of climate change continues to increase, so too will both the number and the severity of weather-related natural disasters. As a result, an increasing amount of attention is being paid to the damage associated with weather-related natural disasters. This damage has certainly increased; however, it is often difficult to quantify accurately. The traditional economic approach to measuring the cost of a disaster focuses on the flow of the GDP due to its wide availability and cross-country consistency. However, the GDP has several weaknesses when used to measure the impact of disasters. While it measures the economic activity that occurs due to the rebuilding effort, it ignores both the loss of capital stock and the opportunity cost of diverting resources to help with the rebuilding. It also ignores many non-income-based impacts, such as increased stress, environmental harm, and even loss of life.

An alternative approach to using traditional economic techniques to measure the impact of disasters is to use subjective well-being. In this paper, we contribute to this emerging strand of literature by estimating the impact of experiencing weather-related home damage on subjective well-being in Australia using 11 years’ worth of panel data. To increase the robustness of our findings, we use two novel approaches in our analysis. Firstly, we supplement the common two-way fixed-effect technique with the more advanced and theoretically appropriate Callaway and Sant’Anna (2021b) technique. Secondly, we apply a series of data transformations to determine empirically whether our results are robust to changes in the assumed underlying cardinal scale and find that they are.

In general, we conclude that weather-related home damage has a well-being effect that, on average, is negative, small in size, and often statistically insignificant. Our main analysis shows that no coefficient estimate is greater than 12% of a within-person standard deviation. We find no major difference in the well-being effect among male and female respondents and insured and uninsured people.

We obtain mixed evidence of an effect on the poor, with the TWFE model finding that the poor experience a statistically significant and long-lasting decline in life satisfaction following weather-related home damage while the more advanced CSA model shows that neither the poor nor the wealthy experience a statistically significant impact.

There are several possible reasons for our finding that weather-related home damage does not have a sizeable impact on well-being. It may be that weather-related home damage truly has no effect on well-being. However, there are other factors that may explain our result. Firstly, selection into treatment may occur so that the people who experience weather-related home damage are also those who are least affected by it. Secondly, our findings may be caused by low average damage severity, meaning that most people who suffered weather-related home damage experienced only minor damage. Social comparisons may also explain a null result as affected people may compare themselves with other affected people rather than people who did not experience weather-related home damage. Finally, direct financial aid through both high levels of insurance coverage and generous government financial assistance may also explain what we found.

As climate change causes disasters to become larger and more widespread, the effects on well-being may change, a reasonable assumption being that there will be a greater impact on well-being. However, our research suggests that, at the moment, weather-related home damage does not seem to have a large average impact on subjective well-being.

Other studies that focus on alternative measures estimate the annual costs of natural disasters in Australia to be several billions of dollars and state that, while people who experience weather-related damage are not more likely to report being unemployed or having lower incomes, they are somewhat more likely to report a major worsening of their financial situation and tend to have lower mental health scores. This result highlights that impact estimates can vary significantly because of what exactly is measured and how it is measured. To gain a full picture of the impact of disasters, it is thus important to consider a wide range of indicators, objective as well as subjective, as well as a wide range of studies.