Inequities in Long-Term Housing Recovery After Disasters

Abstract

Problem, research strategy, and findings: Disaster impacts result from interactions between hazard exposure, physical vulnerability, and social vulnerability. We report empirical work from 1992′s Hurricane Andrew in Miami-Dade (FL) and 2008′s Hurricane Ike in Galveston (TX) to assess long-term trends in housing recovery. Longitudinal, parcel-level data on housing units along with neighborhood sociodemographic data permit analysis of the pace of recovery for different neighborhoods, populations, and housing types. Housing recovery is highly uneven for different population groups. Unsurprisingly, damage has major consequences; even after four years, the effects of damage are evident in the rebuilding process. Social vulnerability factors play differently in different settings. In Miami, income and race and ethnicity were critical determinants of higher losses and slower recovery rates, while in Galveston income was the more critical factor, with housing in lower-income areas suffering more damage and lagging significantly in the recovery process.

Takeaway for practice: Effective land use policy and building codes can reduce physical vulnerability and ultimately damage, thus enhancing resilience for all. Differentials in impact and recovery trajectories suggest that assessment and the monitoring of recovery is critical to target resources to areas that are lagging. Perhaps most important is having an effective plan in place that addresses housing recovery issues to help reduce long-term consequences. Pre-event planning for housing and social change can help support community vision and overcome inequities.

Keywords:

• housing recovery
• disaster
• inequality
• neighborhood transition

Notes

1. Buildings of better quality or built to stronger codes will fare better than others when facing similar destructive forces (Highfield, Peacock, & Van Zandt, 2014; Mileti, 1999). Planning and development policies can keep development out of more vulnerable areas (floodplains, surge zones, etc.), potentially reducing damage (Burby, 1998; Daniel & Daniel, 2003; Godschalk, Beatley, Berke, Brower, & Kaiser, 1999). These policies may also preserve natural resources such as wetlands or dunes that act as buffers, lessening damage (Brody, Highfield, & Kang, 2011; Brody, Zahran, Maghelal, Grover, & Highfield, 2007; Zahran, Brody, Peacock, Grover, & Vedlitz, 2008).

2. For a review of these and other examples in the literature, see Fothergill (1999), Fothergill, Maestas, and Darlington (1999), and Fothergill and Peek (2004).

3. Using improvement value, rather than total assessed value or sales value, allows us to better isolate both impact and rebuilding and repair activities related to the structure itself as opposed to changes in land value or demand for housing. The use of assessed value as an indicator for housing damage and recovery is well established in the literature (e.g., Bin & Kruse 2006; De Silva, Kruse, & Wang 2006; Fujita, 1989, Knaap, 1998; Zhang & Peacock 2010).

4. For additional discussion of the samples, please refer to Zhang and Peacock (2010) and Highfield et al. (2014).

5. Because most hurricane damage happened in southern sections of Miami-Dade County and also because the broad development and housing context was quite different in South Dade, the suburbs of the Miami-Dade County's urban core, from that in the remainder of the county (Dade County Planning Department, 1992; Portes & Stepick, 1993), only parcels south of Kendall Drive (SW 88th Street) are included in this study to preclude the assessment of recovery from being skewed by the underlying development trends.

6. Panel models include data from multiple points on each observation through time. In this case, the dependent variable for each set of observations consists of the logged improved assessed value prior to the event and subsequent logged values for four years after the event. For a discussion of panel models, see Wooldridge (2010).

7. In addition to addressing the skewed nature of the dependent variable, using logged values allows us to ignore variations in inflation between the two samples.

8. This approach is based on our previous analysis of housing recovery (Zhang & Peacock, 2010). In the case of Andrew, the period was from 1992 to 1996, while for Ike, it was 2008 to 2012. At first look, it may seem odd that the base year is the year that the two storm events occurred: August 24, 1992, and September 13, 2008. However, the assessed values for these years reflect the pre-impact structure's assessment, while the next year registers the impact in terms of a deflated assessment and subsequent years track recovery.

9. Damage = (pre-impact improvement assessed value – post-impact improvement assessed value) / pre-impact improvement assessed value.

10. For both models these include the age of the home and the number of bedrooms and baths in the Andrew model and square footage of the home in the Ike model. Ideally, we would have preferred numbers of bedrooms and baths for both, but the Ike assessor's data do not include these data.

11. Following the suggestions of a reviewer, we ran the models for Galveston with and without Bolivar Peninsula, but the results were not substantially changed.

12. In terms of the coefficients, a nonsignificant coefficient for a year dummy variable would indicate restoration on average net of other variables, a significant negative year dummy coefficient would indicate a failure to reach restoration levels, and a significant positive year dummy coefficients would indicate recovery beyond the pre-impact assessed values, net of other variables.

13. These interaction or net effect coefficients must be combined with the base coefficients to determine the overall effects of the variable at various years of the impact-recovery period; furthermore, while the significance test for the interaction coefficient assesses its significance from zero, the combine combined effect (baseline + net) were all tested for significance employing a Wald test.

14. Exponentiations for 10.83, 10.88, 11.33, and 9.90 are approximately ∃50,514, ∃53,104, ∃83,283, and ∃19,930, respectively.

15. 15. This may be surprising to some readers; however, southern sections of Miami-Dade County at the time of Hurricane Andrew (the early 1990s) were majority non-Hispanic White and non-Hispanic Black.

16. In other words, yearly averages for all damage categories have been divided by their respective pre-impact, base-year averages, yielding proportions where values of 1.0 are equal to base levels, values less than 1.0 indicate values below pre-impact levels, and values greater than 1.0 are greater than pre-impact levels. Hence, all values are 1.0 for the base year: Reaching 1.0 in subsequent years means that on average restoration levels are reached, and values below 1.0 indicate a failure to repair and rebuild to reach even restoration levels, while values greater than 1.0 indicates surpassing restoration levels.

17. Since the dependent variable is a logged value, the coefficients are semi-elasticities representing the percentage change in the dependent variable given a unit change in the independent variable. Rather than depending on simple rule-of-thumb transformations, our discussions will transform coefficients appropriately and present these appropriately transformed values; hence, in this case, the percentage change is 100(e ^−.0455 ‐ 1) = ‐4.5.

18. The positive, significant, and increasing values for the damage–year interaction coefficients are 0383,.0413, and.0423 for the damage ´ yr2, damage ´ yr3, and damage ´ yr4 coefficients, respectively.

19. The conversion for the combined effect is 100(e ^{(–.045475 +.0423107)} – 1) = –.31643. Also, the combined effect is statistically significant (Wald = –28.25; p <.001).

20. The combined effect is ‐0.08486 + 0.01688 =‐.06798, which is statistically significant (Wald = –21.53; p <.001), and the conversion is 100(e ^{(–0.08486 + 0.01688)} – 1) = –6.57, or –6.6%.

21. The conversion is 100(e ^(.03924) – 1) = 4.

22. The combined effect is.0392 +.035 =.07424, which is significant (Wald = 5.16; p <.001), and the conversion is 100(e ^{(.0392 +.035)} – 1) = 7.71.

23. The combined effect is.0392 +.4499 =.4891, which is significant (Wald = 21.23; p <.001), and the conversion is 100(e ^{(.0392 +.4499)} – 1) = 63.

24. This is reflected by the nonsignificant owner-occupied dummy coefficient in the Ike model.

25. The combined effect is.03652 +.49917 =.53569, which is significant (Wald = 6.61; p <.001), and the conversion is 100(e ^{(.03652 +.49917)} – 1) = 70.86.

26. The combined effects for 2011 and 2012 are.03652 +.64509 =.68161 and.03652 +.61618 =.6527, respectively; both are significant (Wald = 7.47; p <.001 and Wald = 6.86; p <.001), and the conversions are 100(e ^{(.3652 +.64509)} – 1) = 97.71 and 100(e ^{(.3652 +.61618)} – 1) = 92.07.

27. A reviewer suggested that in the case of Ike, there may have been interactions between race and ethnic composition and income such that the positive effects for minority composition were overcome by the more positive effects of income in high-income, low-minority areas. This seems unlikely due to nonsignificant effects for income in later years. Nevertheless, we also tested for this possibility by including interaction terms between race and ethnicity and median income, but they were not significant.

28. These have been created by using average values for each of the independent variables and then the appropriate coding for owner versus rental housing; furthermore, they have been indexed to their base values so they are more easily interpreted with respect to losses and gains toward restoration (1.0).

29. For example, buried in the multifamily numbers for Galveston are the major failures to build back six rather large public housing facilities (28.7% of multifamily housing stock valuation before Ike), which were zeroed out (torn down) in 2010. These were followed by complete or partial demolition of other multifamily structures in the following years, with four being torn down completely, resulting in nearly 10% of these parcels losing in excess of 40% of their assessed value.

30. For example, a critical and important element not included in these models is the nature of resources in the form of insurance, savings, grants, and other sources that were available and employed in the recovery processes. Indeed, one interesting and important difference between these two events is that post-Andrew homeowners insurance was a primary driver for recovery, and insurance redlining (based on race) may well have played a role in shaping minority differentials (Peacock & Girard, 1997; Peacock et al., 2007; Zhang & Peacock, 2010). However, recovery efforts after Ike, as a flooding event, was dependent upon flood insurance, of which nearly 50% of owners did not carry and which has caps on payouts.