Demonstrating the Intensive Benefit to the Local Implementation of a Statewide Building Code

Transcription

1 Demonstrating the Intensive Benefit to the Local Implementation of a Statewide Building Code Jeffrey Czajkowski Wharton Risk Management Center University of Pennsylvania Kevin M. Simmons Austin College James M. Done National Center for Atmospheric Research Boulder, CO April 2016 Working Paper # Risk Management and Decision Processes Center The Wharton School, University of Pennsylvania 3730 Walnut Street, Jon Huntsman Hall, Suite 500 Philadelphia, PA, USA Phone: Fax:

2 THE WHARTON RISK MANAGEMENT AND DECISION PROCESSES CENTER Established in 1985, the Wharton Risk Management and Decision Processes Center develops and promotes effective corporate and public policies for low probability events with potentially catastrophic consequences through the integration of risk assessment, and risk perception with risk management strategies. Natural disasters, technological hazards, and national and international security issues (e.g., terrorism risk insurance markets, protection of critical infrastructure, global security) are among the extreme events that are the focus of the Center s research. The Risk Center s neutrality allows it to undertake large scale projects in conjunction with other researchers and organizations in the public and private sectors. Building on the disciplines of economics, decision sciences, finance, insurance, marketing and psychology, the Center supports and undertakes field and experimental studies of risk and uncertainty to better understand how individuals and organizations make choices under conditions of risk and uncertainty. Risk Center research also investigates the effectiveness of strategies such as risk communication, information sharing, incentive systems, insurance, regulation and public private collaborations at a national and international scale. From these findings, the Wharton Risk Center s research team over 50 faculty, fellows and doctoral students is able to design new approaches to enable individuals and organizations to make better decisions regarding risk under various regulatory and market conditions. The Center is also concerned with training leading decision makers. It actively engages multiple viewpoints, including top level representatives from industry, government, international organizations, interest groups and academics through its research and policy publications, and through sponsored seminars, roundtables and forums. More information is available at

3 Demonstrating the Intensive Benefit to the Local Implementation of a Statewide Building Code Jeffrey Czajkowski, Ph.D. Wharton Risk Management and Decision Processes Center University of Pennsylvania Kevin M. Simmons, Ph.D. Austin College James M. Done, Ph.D. National Center for Atmospheric Research April 11, 2016 Abstract Ultimately, risk reduction from the implementation of building codes is due to not only the extent of the code as it applies to new construction, but also from the intensity of local adoption and enforcement. It is normally an open question as to how well a code is maintained and enforced at the local level, even for a relatively strong adopted statewide code such as the Florida Building Code. We test the importance of the intensity of building code implementation at the local level for reducing FL windstorm losses by utilizing Building Code Effectiveness Grading Schedule (BCEGS) rating data. BCEGS ratings provide a joint assessment of local building code effectiveness in terms of the strength of the adopted codes in addition to how well these adopted codes are enforced. We find that both components provide value in reducing windstorm losses in FL, with the extent of the statewide code being the dominant effect reducing losses on the order of 72 percent. Although not as substantial in terms of its loss reduction magnitude, intensively implementing building codes at the local level by ensuring codes are properly administered and enforced at this scale provides additional loss reduction value on the order of 15 to 25 percent. Understanding the relative value of these two implementation components is important to better inform building code policy and enforcement efforts given continuously updated codes. 1

4 I. Introduction In a world of steadily increasing costs and frequency of natural disasters (UNISDR, 2015) the implementation of effective risk reductions strategies is critical. Having strong building codes in place is frequently touted as a key risk reduction strategy in this regard, effective in reducing total property damage due to natural disaster occurrence (Mills et al., 2005; Kunreuther and Useem, 2010; McHale and Leurig, 2012; Vaughn and Turner, 2014) such as that from windstorms, including hurricanes. And a number of studies, typically by capturing the differences in losses from properties constructed before and after the implementation of a new and stronger building code, have demonstrated a reduction in windstorm losses due to stronger building codes (Fronstin and Holtmann, 1994; IBHS, 2004; Applied Research Associates, 2008; Deryugina, 2013; and Simmons et al., 2016). For example, an Institute for Business and Home Safety (IBHS) study commissioned following Hurricane Charley in 2004 (IBHS, 2004) found that homes built after 1996 (relating to the implementation of stronger building codes in Florida following the catastrophic losses from Hurricane Andrew in 1992) had lower claim frequency (60 percent less) and severity (42 percent less) as compared to homes built before Applied Research Associates (2008) and Simmons et al. (2016) expanded and further verified these Florida building code loss reduction results by using windstorm loss differences between pre and post 2002 year of construction (Y.O.C.) properties given the fully effective implementation of the statewide Florida Building Code (FBC) that year. As new building codes primarily apply to new construction, demonstrating windstorm loss reductions via Y.O.C. loss differences that correspond to a code change (e.g. post 2002 Y.O.C. for the FBC) captures the value of an improved building code at the extensive margin, i.e., the number (extent) of new residential properties built under the new code. Yet, the implementation of a stringent building code provision is ultimately dependent upon proper enforcement at the local jurisdiction through the permitting and inspection process (Schmith, 1999; AIR, 2010; Deryugina, 2013; Vaughan and Turner, 2014). Thus, it is always an open question as to how well a code is maintained and enforced at the local level, even for a relatively strong adopted 2

5 statewide code such as the FBC. Or, what is the intensity of building code implementation at the local level and how much does this matter? Here we test the importance of the intensity of building code implementation at the local level in Florida for reducing wind losses. We accomplish this by utilizing Building Code Effectiveness Grading Schedule (BCEGS) rating data at the zip code level, a joint assessment of local building code effectiveness in terms of the strength of the adopted codes in addition to how well these adopted codes are enforced. Czajkowski and Simmons (2014) have shown the BCEGS rating to be a statistically significant determinant of reduced hail related losses in the state of Missouri by the order of 10 to 20 percent on average. We are interested in understanding whether at the local level in FL more intense adoption and enforcement of a strong statewide code leads to lower windstorm losses? That is, obtaining the value of a new code at the intensive margin given the intensity of adoption and enforcement at the local level. Determining this thus allows for a better understanding of what portion of reduced windstorm losses is attributable to the extent of new construction coming on line under a new statewide building code regime versus the portion of reduced windstorm losses attributable to the intensity of adoption and enforcement at the local level. Given that both components are important in successful implementation of building code risk reduction efforts, it is important to understand their relative value to better inform building code policy and enforcement efforts. The paper proceeds as follows. Section 2 provides an overview of the BCEGS ratings and corresponding rating data in Florida. Section 3 is an overview of the statistical methodology and model data. Section 4 presents the initial regression results and Section 5 provides tests of the robustness of our results and evaluation of the regression model. Section 6 concludes the paper. II. Florida Building Code Effectiveness Grade Schedule Ratings BCEGS Rating Overview Since 1995 the Insurance Services Office (ISO) administers the Building Code Effectiveness Grading Schedule (BCEGS) ratings for the property/casualty insurance industry across the entire 3

6 country. 1 Today, the ISO BCEGS program evaluates more than 16,700 departments serving more than 25,000 communities. The BCEGS ratings assess the building codes in effect in a particular community and how well the community enforces its building codes, with special emphasis on mitigation of losses from natural hazards. Significantly for our study, building codes are often designed specifically for wind natural hazards. In order for a community to obtain a BCEGS rating, minimum building code requirements must be met. These minimum BCEGS requirements include: a building department must be permanently organized under state or local laws; a building code must be adopted; plan reviews must be conducted; field inspections must be made; and training of code enforcement personnel must be done. Beyond these minimum requirements, a community's BCEGS rating is based on their performance under three main classifications: 1) administration of codes; 2) review of building plans; and 3) field inspections. Specific criteria for the administration of codes includes amongst other items the building code edition in use, zoning provisions to mitigate natural hazards, the training and certification of code enforcers, the qualifications and licensing of building officials and contractors/builders, and public awareness programs. Specific criteria for both review of building plans and field inspections includes amongst other items the staffing levels, qualifications, and level of detail of plan reviews and inspections. Lastly, ISO collects underwriting information, including natural hazards common to the area, number of inspection permits issued, number of inspections completed, the building department's funding mechanism and date of establishment, size of the jurisdiction and population, and fair market value of all buildings. The BCEGS program delivers an overall grade for a jurisdiction representing a combination of elements related to the strength of code, level of code enforcement, quality of code administration, and the interaction of these features. Key components that are included in the development of the strength grade include the edition of the code adopted and any significant 1 Material in regard to the ISO BCEGS ratings is sourced from ISO s Building Code Effectiveness Grading Schedule (BCEGS ) For more detailed information on the BCEGS ratings please see the description on their website. In the states of Hawaii, Idaho, Louisiana, Mississippi, and Washington an independent rating bureau administers the BCEGS ratings. 4

7 amendments. The enforcement grade includes elements that provide insight into the level of plan review, inspection, and personnel. ISO s BCEGS methodology is unique in the level of detail it collects and considers related to code enforcement. Each jurisdiction/community is classified on a scale of 1 to 10, with a class of 1 representing exemplary enforcement of a model code and a class 10 indicating the jurisdiction has earned very few points on many evaluation criteria. A rating of 99 indicates that the jurisdiction/community is unclassified. Communities may have a class 99 if they have not participated in a BCEGS survey, or the community does not meet the minimum requirements of the BCEGS program. For Florida, all communities have either been evaluated or invited to be evaluated by the BCEGS program, nonetheless unclassified 99 communities persist. 2 From the overall 1 to 10 scale, ISO develops advisory rating credits that apply to BCEGS classifications ranges of 1 3, 4 7, 8 9, and 10. Points are accumulated by a particular jurisdiction/community for how well criteria are met for each of the above classifications. Each criterion has a certain number of points and the points are totaled to arrive at the BCEGS rating class. For example, BCEGS rating class 1 has a point range of to 100.0, while BCEGS rating class 5 has a point range of to A BCEGS rating of 1 to 3 is deemed to have a buildingcode department that: enforces the latest model code without amendments that would weaken the code s ability to reduce damage from natural hazards. The department has all the resources required to enforce its adopted code rigorously. The department also has a sufficient number of trained and certified staff to devote adequate time to plan reviews and inspections. (ISO, 2014) 3 Florida BCEGS Ratings BCEGS personal line rating classifications (i.e., for building code adoption and enforcement for one and two family dwellings) from 1995 to 2015 were provided by ISO for the state of Florida at the ZIP code level for 950 individual ZIP codes. BCEGS ratings are provided at the ZIP code 2 Individual risks in a community built prior to the implementation of BCEGS also receive the class As of the publication date, the 24 percent of ZIP codes rated 1 to 3 from our data methodology matches well to the FL BCEGS rating distribution on the ISO website available at 5

8 level in order to explicitly match to insured loss data which is tracked by ZIP code (to be described in Section 3). The BCEGS ratings we received from ISO were for surveys conducted in each ZIP code containing the aggregate personal line rating for each ZIP code along with the year when the BCEGS rating survey was conducted. For the survey year we apply the assigned BCEGS rating for that year. For years in between survey years, BCEGS ratings were applied as follows: The most recent data continues to be in place until the next survey. So if a survey were performed in 2003 and again in 2008, the years 2004 through 2007 would reflect the BCEGS rating from the survey of Years after 2008 would reflect the 2008 survey rating until a new survey year emerges. Finally, while the overall BCEGS rating data is from 1995 to 2015, a ZIP code rating does not begin until a survey year appears for the ZIP code. Continuing from the previous example, if 2003 was the first survey year, the rating from 1995 to 2002 would be classified as no rating. Additionally, an individual ZIP code may be intersected by multiple BCEGS rated jurisdictions (municipality or county) each having their own BCEGS rating. Thus, some ZIP codes had multiple surveys conducted in a given year stemming from the multiple rated jurisdictions comprising that ZIP code. We use a single rating for each ZIP code for any year derived from the non weighted average of the multiple county and municipality jurisdictions intersecting that particular ZIP code in that particular survey year (excluding any 99 or missing ratings). For example, ZIP code is comprised of the following five BCEGS rated jurisdictions: Suwannee (27 percent); Dixie (5 percent); Gilchrist (24 percent); Lafayette (43 percent); and Branford (1 percent). The land area of each BCEGS rated jurisdiction comprising the total land area of ZIP code is shown in parentheses with Suwannee, Dixie, Gilchrist, and Lafayette being BCEGS county level ratings and Branford being a municipality jurisdiction rating. In any one survey year more than one of these corresponding jurisdictions could have a BCEGS rating. As an example, Branford and Suwannee were the only two jurisdictions rated in 2005 therefore the average for these two jurisdictions (each rated as a 4) was taken as the average BCEGS rating for ZIP code in Of the 950 Florida ZIP codes, 273 of them (29 percent of total) have a single BCEGS rated jurisdiction (municipality or county) comprising the entire land area of the ZIP code, 547 of them (58 percent of total) have 1 jurisdiction comprising at least 90 percent of the total 6

9 land area of the ZIP code, and 901 of the 950 ZIP codes (95 percent of total) have at least one jurisdiction comprising at least 50 percent of the land area of the ZIP code. During the 1995 to 2015 timeframe, a total of 19,950 overall annual personal line BCEGS ratings are observed in the ISO data provided, including 99 ratings and no ratings. Figure 1 illustrates the BCEGS rating distribution from these 19,950 ZIP code observations. From 1995 to 2015, 21 percent of the rated ZIP codes had an overall BCEGS rating of 1 to 3, 60 percent had a rating of 4 to 7, and the remaining 19 percent had a rating of 8 or higher including no rating or 99 rating for that particular year. Insert Figure 1 Here To provide a snapshot of the rating data over time, Figure 2 illustrates the geographic distribution of the BCEGS ratings 1 to 3 as per the year of We see that the overall BCEGS ratings of 1 to 3 are not concentrated in any one area of the state but that these areas match well to a fair portion of the more densely populated areas of the state such as Southeast FL as well as the areas around Tampa, Orlando, and Tallahassee. Insert Figure 2 Here III. Florida Windstorm Losses, Statistical Methodology and Associated Data After Hurricane Andrew in 1992 it became clear that FL construction practices in place during the 1980s had not only been insufficient to withstand such a powerful wind storm (Sparks et al., 1994), but that inferior construction practices had unnecessarily magnified the extensive damage (Fronstin and Holtmann, 1994; Keith and Rose, 1994). In the aftermath of Hurricane Andrew, Florida began enacting statewide building code change that wrested away building code adoption control from individual localities. The Florida Legislature authorized a recommended statewide code in By 2002, all legal challenges to the code were resolved, and the statewide 2001 Florida Building Code (FBC) took effect on March 1, 2002 officially superseding all local codes (Dixon, 2009). The FBC is based on the national model codes developed by the International Code Council (ICC) and is among the strictest in the nation heavily emphasizing wind engineering standards and other additions for Florida s specific needs including for hurricane protection (Dixon, 2009). We quantify historical Florida wind event loss reductions due to the implemented 7

10 FBC statewide and at the local level through a statistically driven loss methodology that systematically accounts for relevant wind hazard, exposure, and vulnerability characteristics evolving over time from the adoption of the new uniform codes in Florida Insured Loss Data We collect annualized private market insured policy and loss data (number of claims and total damages for all represented earned house years in the insured portfolio) from the Insurance Services Office (ISO), aggregated at the ZIP code level for all Florida ZIP codes spanning the years 2001 to 2010, inclusive. ISO industry data represents a significant percent of total private property/casualty insurance annual market share in FL, 4 and we utilize aggregated policy data in any one year ranging from 669,000 to just over 1 million insured policyholders. From our 2001 to 2010 loss data, windstorm hazards are the largest cause of loss in FL, totaling $5.178 billion in losses, as well as being the most frequent source of a loss claim with 317,005 claims incurred. Of course, Florida windstorm losses vary over time and, as expected, are significantly linked to the occurrence of hurricanes. Table 1 provides a further detailed view of the ISO Florida windstorm incurred losses and claims over time. Across all years, an average of $517 million in losses and 31,701 claims are incurred each year, with an average windstorm claim being $10,089 incurred at the rate of 32.4 claims per 1,000 insured exposures (earned house years). However, excluding the significant hurricane years of 2004 and 2005, an average of $25 million in losses and 3,900 claims are incurred each year, with an average windstorm claim of $8,353 per claim incurred at the rate of 4.8 claims per 1,000 insured exposures (earned house years). Although windstorm losses and claims are considerably higher in significant hurricane years, they are still a substantial annual property risk. For example, 2007 had average windstorm claims of $25,399 per claim, and 2001 had 13.1 windstorm claims per 1,000 insured both outside the significant hurricane years of 2004 and Lastly, average annual premiums collected over this timeframe (data not shown) are just over $1 billion per year. While these premiums are sufficient to cover incurred loss amounts in non hurricane years, major windstorm year loss amounts (for example, 4 ISO personal communication places this around 40 percent market share 8

11 2004 windstorm losses are nearly 4 times higher than annual average premiums collected) indicate the critical role of further windstorm risk reduction measures in Florida. Insert Table 1 Here Figure 3 illustrates the geographic location of the 126 Florida ZIP codes that incurred a total of $10 million of losses or greater during 2001 to 2010 highlighted in blue, along with the ZIP codes having a total BCEGS rating of 1 to 3 in These 126 ZIP codes represent a total of $3.3 billion of the $4.8 billion of windstorm losses in the 911 identified ZIP codes with incurred losses used in the analysis, or 14 percent of the Florida ZIP codes represent nearly 70 percent of the losses. Overall during the 2001 to 2010 timeframe, losses ranged from $0 (in 9 ZIP codes) to $163 million in Punta Gorda and $118 million in Port Charlotte, where Hurricane Charley made landfall in The average loss per ZIP code was $5.2 million. Although there appears to be a fairly significant spatial correlation between highly rated BCEGS communities and windstorm losses, this view does not control for a number of other hazard, exposure and vulnerability factors that likely play a role in determining loss amounts by location and over time for which we now turn to in the regression methodology utilized. Insert Figure 3 Here Statistical Methodology and Data Description Our regression model is a semi log, ordinary least squares (OLS), fixed effects (time and space) model with the natural log of loss as the dependent variable. A further split of the ISO loss data obtained is by decade of construction. That is, for each year of ISO data from 2001 to 2010, each Florida ZIP code in that year contains a split of the losses, claims, premiums, and earned house years by the year of construction decade beginning in 1900 up to The base level of observation is ZIP code/year/decade of construction. Explanatory variables include insurance information (claims, exposures and premiums), construction type, demographic data based on the ZIP code, percentage of homes built since code adoption, measures of the ZIP code hazard 9

12 risk (how close to the coast the ZIP code is, etc.), hazard data concerning the wind speed and duration, and finally BCEGS data. The general form of the model is: Natural log of losses = β 0 + β 1 *Claims + β 2 *ln_ Premium + β 3 *Brick_Plus + β 4 *Mobile + β 5 *ln_income +β 6 *ln_value + β 7 *unit_fac_density + β 8 *pct_since_98_fac + β 9 *Coastal + β 10 *CCCL + β 11 *ln_near_dist + β 12 *Citizens + β 13 *Max_Wind + β 14 *wind_dur12 + β 15 *d_ Dummy Variables for the BCEGS ratings + Vector of dummy variables for year + Vector of dummy variables for 3 digit ZIP code. Florida was affected by 18 tropical cyclones over the period Spatial wind hazard data over Florida are sourced from the National Center for Environmental Prediction s (NCEP) North American Regional Reanalysis (NARR, 2015; Mesinger et al., 2006). NARR is a dynamically consistent historical climate dataset based on historical climate observations. Data are available 3 hourly on a 32km grid. Of importance to this study Mesinger et al. (2006) showed that the winds just above the surface compare well with surface station observations. The 32 km grid is too coarse to resolve high impact small scale features in the wind field such as thunderstorm winds or tornadoes. It is also too coarse to capture the intensity of the strongest hurricanes (as discussed in Done et al., 2015). Rather than downscaling the NARR data to obtain these small scale details using dynamical (e.g., Laprise et al., 2008) or statistical (e.g., Tye et al., 2014) methods (that could introduce further uncertainties) we choose to sacrifice the small scale details of the wind field and peak hurricane intensity for location accuracy of the NARR data. To account for these missing wind extremes, all wind speed values are normalized by the maximum value of wind speed in the dataset. Specifically, the 3 hourly wind data are interpolated from the 32 km grid to the ZIP code level and two wind field parameters are derived for use in the loss regressions: the normalized annual maximum wind speed, and the annual number of times the wind speed exceeds the Florida mean wind speed plus one standard deviation for at least 12 hours. A positive sign is expected for both of these variables indicating that as wind speeds increase and/or the ZIP code is exposed to high winds for an extended period of time, losses will increase. The choice of hazard 10

13 variables is based on recent work that highlighted the potential for wind parameters other than the maximum wind to drive losses (Czajkowski and Done, 2014; Zhai and Jiang, 2014; Jain, 2010). We have 2000 and 2010 demographic data from the decennial census at the ZIP code level for population, area (in square miles) of the ZIP, median household income, housing counts, housing value and percent of homes that are mobile homes. Intervening years are interpolated from decennial data for population, mobile home counts and total housing counts with an allocation factor based on the number of building permits for each ZIP and each year. For median household income and housing value a straight line interpolation method is used. Income and housing value are adjusted for changes in the consumer price index (CPI U) to CPI data are from the Bureau of Labor Statistics. Building permits are collected from census by place codes. To convert from place to ZIP code we use allocation factors based on 2010 housing counts provided by MABLE, a service of the Missouri Census Data Center (MABLE, 2015). For unincorporated areas we use allocation codes from county to ZIP from the same service. Several variables are used to measure vulnerability. Mobile homes are more likely to be damaged than permanent homes but cost less to replace. So a negative sign is expected. Income is the median household income for the ZIP code. Higher incomes could reflect a demand for better constructed homes but could also increase insured loss as wealthier communities have more expensive items to replace. The sign for this variable could be either positive or negative. Residential home value is another reflection of wealth. Again, this could increase loss as the home is more expensive and repairs would reflect the added cost. But it could also reflect better construction which would decrease loss A number of factors were utilized to represent the overall geographic hazard risk of a ZIP code. A ZIP code was identified as being a coastal ZIP code (dummy variable which equals 1) if the ZIP code border touches the Atlantic or Gulf Coast. Additionally, the distance of the centroid of the ZIP to the coast was calculated to account for the overall distance to the coast of each ZIP code. Following Dehring and Halek (2013) dummy variables that signifies whether a ZIP code contains a coastal construction control line (CCCL) were created (1 equals CCCL in place) to account for stricter building codes in these areas. Finally, following the 2005 hurricane season 11

14 there was a significant increase in the number of policies underwritten by Citizens, the state run wind pool and insurer of last resort. (Florida Catastrophic Storm Risk Management Center, 2013). Areas with large percentages of insured policies underwritten by Citizens could represent inherently higher windstorm risk. We spatially matched our Florida ZIP codes to the Florida house districts and took the percentage of Citizens policies of the number of occupied housing units as of December 31, 2011 (Florida Catastrophic Storm Risk Management Center, 2013). Given the potential for adverse selection, or offloading of high risk policies by the private market in these areas, it is unclear whether higher Citizens market penetration would lead to a positive relationship with losses due to the higher risk, or a negative relationship with private losses as many of the bad risks have been transferred to the residual wind pool. Regression models are limited by available data to understand how the dependent variable varies as explanatory variables change. If important variables are left out of the model, some bias can be expected. One way to minimize bias is to employ a fixed effects model. This is done by using binary (dummy) variables for cuts of the data that may be introducing bias. We use two sets of dummy variables for time (years) and geography (three digit ZIP codes). These dummy variables will contain all across group variation leaving the remainder of the model to contain the within group variation (Greene, 2003). Our primary variables of interest to capture building code value at the extensive margin are the percentage of homes in a given ZIP code that have been built since the statewide code was implemented and a dummy variable for an observation for homes built after The state legislature first adopted recommended code changes in So we track the percent of homes built since 1998 as one variable. To test for the effect of homes built after the introduction of the statewide building code, we construct dummy variables for those that are post We expect that observations for homes built after 2000 will decrease overall damage and that as the percentage of new homes increases, that damage will diminish as well. The primary variable to capture the value of building codes at the intensive margin are the BCEGS ratings. We also expect that ZIP codes actively participating in the BCEGS program achieving more favorable ratings will 12

15 have lower losses, all else being equal. Table 2 contains definitions for all variables used in the analysis. Insert Table 2 Here IV. Baseline Regression Results To include BCEGS participation in our regression model we take the ratings and create three dummy variables that are then included in the regression. The dummy variables comprise two ratings tiers for ZIP codes with ratings and a third tier for ZIP codes which do not have a BCEGS rating or have a rating of 99. The highest tier (More Favorable) has ZIP codes whose BCEGS rating is between 1 and 4 (i.e., the strongest and most well enforced building codes), the middle tier (Moderately Favorable) is ratings 5 through 7 and the low tier is ratings 8 and above plus missing or unclassified ratings. 5 For instance, if a ZIP code had a BCEGS rating of 2 the value of the dummy variable for the More Favorable tier would equal 1 while all other dummy variables for the BCEGS tiers would equal 0. The third tier representing no participation in the BCEGS program is the omitted category. Model results are shown in table 3. Insert Table 3 Here The overall performance of our regression model is satisfactory, R squared is 0.48, and consistent in terms of the performance of the explanatory variables. The variables to measure the effect of wind hazard are wind speed and duration. For both variables we have a positive sign and each is highly significant. Higher wind speed and higher duration of high wind speeds increases damage and thus loss. The remaining variables perform as expected with the one exception of our Coastal dummy. Coastal is a binary variable which equals 1 if the ZIP code borders the coast. In our model the sign is negative and significant. This is interesting as wind speeds from incoming hurricanes are higher at the coast than inland so an expectation that the sign would be positive is reasonable. But it may reflect that communities recognize this and undertake precautions that are not practiced inland. 5 The cut offs for these rating categories were determined through a series of model runs having varying cut off levels, e.g., 1 to 3 being more favorable, as well model runs with non grouped integer based ratings. In all model runs we find similar results as those presented here in that in general the more intensive ratings lead to lower losses as compared to lower levels. The final categories were determined through personal communication with ISO in light of the various model runs as well as the individual FL ratings with 2 being the highest. 13

16 Importantly, our results show the strong extensive effect that the statewide FBC had on losses from wind storms during this timeframe. In both variables that measure the implementation of the statewide codes at the extensive margin, the post 2000 dummy variable and the percent of homes built since 1998, losses are shown to go down, and both variables are highly significant. The coefficient on the post 2000 dummy suggests that homes built since the year 2000 suffer 72 percent lower losses than homes built prior to This number approximates well the results shown after Hurricane Charley in 2004 (IBHS, 2004) where newer homes (i.e., post 1996 construction) were 60 percent less likely to suffer damage at all and those that were damaged sustained 42 percent less damage than older homes. Our result of 72 percent lower damage reflects both those attributes as our data included ZIP code/year/yoc observations that suffered damage as well as those that did not. Our second variable that measures the extensive effect of the statewide codes percent built since 1998 has a similar story. Again, using the results from the base regression, for every 1 percent increase in the percent of homes in a ZIP code built since 1998, losses from wind storms decrease by 1.25 percent. For our data, the median percent of homes built since 1998 is slightly less than 10 percent. So if the state of Florida were to have a ZIP code average of 10 percent of homes built since the new codes, losses from wind storms comparable to those experienced in the past would be 12.5 percent lower, other things equal. Also notably, the signs on both BCEGS rating tiers (More and Moderately Favorable) in the base regression are negative and statistically significant. Therefore, in terms of windstorm loss reductions due to intensively implementing building codes at the local level we find that there is additional value in ensuring codes are properly administered and enforced at this scale. Compared to ZIP codes with low and missing ratings, ZIP codes with more favorable BCEGS ratings reduce losses by 15 percent. Likewise, ZIP codes with moderately favorable BCEGS ratings reduce losses by 24 percent compared to ZIP codes with low and missing ratings. These values are consistent with BCEGS rating related loss reductions of 10 to 20 percent on average found in related natural hazard contexts (Czajkowski and Simmons, 2014). While the magnitude of the moderately favorable BCEGS rating coefficient value is higher than that of the more favorable 14

17 BCEGS rating and therefore does not meet our prior expectations, the results certainly imply that at the least implementing moderately intensive adoption and enforcement at the local scale further reduces windstorm losses. Although, clearly here the extent of the statewide code given the post 2000 construction coefficient estimate has the dominant effect as compared to the intensive BCEGS ratings. We turn to some additional analyses to further verify these results. V. Robustness of results To test the robustness of our baseline results, we run five separate analyses: 1) we interact new construction with the BCEGS ratings; 2) we interact high wind values with the BCEGS ratings; 3) we estimate a statistical technique hurdle model to account for the possibility of a separate process occurring in the data that determines whether or not a loss is realized at all, which could then further affect the estimate of overall losses; 4) we split our data into equal BCEGS rating segmentations; and 5) we perform an out of sample validation to our model. New Construction Interaction Building codes primarily apply to new construction so to better isolate the relationship between building codes and new construction we interact the BCEGS ratings with new construction. In the first model we create three interaction variables which are found by the product of the dummy variables for each BCEGS rating tier and the dummy variable that indicates this observation is for homes built after Using the example from the baseline regression section, a ZIP code which has a rating of 2 and homes built after 2000 would have a value of 1 while all other interaction variables would equal 0. In the second model we create three interaction variables which are found by the product of with the actual percent of new homes built since 1998 and the dummy variables for each BCEGS rating tier. Results are presented in Table 4. From Table 4 model 1 the effect of this interaction on the dummy variables for the BCEGS tiers is nominal as both tiers still have negative signs and are statistically significant. Both interaction variables have negative signs although neither attains significance. We can infer from this that the intensive marginal effect of the ratings is not limited to new construction in any zip 15

18 code but rather absorbed across all construction in the local area. From Table 4 model 2 the post 2000 dummy still has a negative sign (lower losses) and continues to be highly significant. However, the percent of new homes since 1998 is now positive. The BCEGS dummy variables have two notable changes, a positive sign on the More Favorable rating dummy variable and the Moderately Favorable no longer attains significance despite its negative coefficient sign. But the interaction with percent of new homes built since 1998 and both BCEGS tiers are significant and both show negative signs. Insert Table 4 Here Wind Interaction It is plausible that the intensive value of building codes is most applicable in more extreme hazard conditions. To test for this, we interact the BCEGS ratings with our wind speed hazard variable. Results are presented in Table 5. Here our BCEGS rating variables are now positive and statistically significant in comparison to the omitted category with the coefficient value on the more favorable rating being lower than moderately favorable. From the wind speed interaction terms we see the negative and statistically significant coefficient value indicating lower losses in more extreme wind conditions as would be expected. The overall effect of the BCEGS rating coefficient combined with the wind interaction indicates lower overall losses for more favorable rated zip codes in comparison to the moderately favorable rated zip codes. Insert Table 5 Here Hurdle Regression Models One problem often encountered in attempting to model losses is that there may be a separate process occurring in the data that determines whether or not a loss is realized at all, which could affect the estimate of overall losses. Hurdle models are used to address this issue as they divide the analysis into two stages. The first stage models the probability that a loss occurs and the second stage models the loss using only observations that sustained a loss. We use a simple hurdle model and a second pioneered by James Heckman (1976, 1979). In both the 16

19 simple and Heckman model, the first stage is the same with a dependent variable which equals one if there was a loss and zero otherwise. This binary dependent variable is then regressed against variables that would affect the probability that a loss occurred. Its form is: Loss or No Loss = β 0 + β 1 * Max_Wind + β 2 * Wind Duration + β 3 * Population Density + β 4 * d_2000 The second stage in the simple hurdle model is the same as the regression model in Table 3 with the exception that only observations with a loss are included. The number of observations for the second stage is 19,107. In the Heckman model, one change is made in the second stage and that is the inclusion of the Inverse Mills Ratio (IMR) which is found from the probability distribution of whether or not losses were sustained. The IMR is included in the second stage as an explanatory variable. The Heckman model uses Maximum Likelihood for the second stage. The second stage for the simple hurdle model is: Natural log of losses = β 0 + β 1 *Claims + β 2 *ln_premium + β 3 *Brick_Plus + β 4 *Mobile + β 5 *ln_income + β 6 *ln_value + β 7 *unit_fac_density + β 8 *pct_since_98_fac + β 9 *Coastal + β 10 *CCCL + β 11 *ln_near_dist + β 12 *Citizens + β 13 *Max_Wind + β 14 *wind_dur12 + β 15 *d_ β 16 *More_Favorable + β 17 *Moderately Favorable + Vector of dummy variables for year + Vector of dummy variables for 3 digit ZIP code. The second stage for the Heckman model is: Natural log of losses = β 0 + β 1 *Claims + β 2 *ln_ Premium + β 3 *Brick_Plus + β 4 *Mobile + β 5 *ln_income + β 6 *ln_value + β 7 *unit_fac_density + β 8 *pct_since_98_fac + β 9 *Coastal + β 10 *CCCL + β 11 *ln_near_dist + β 12 *Citizens + β 13 *Max_Wind + β 14 *wind_dur12 + β 15 *d_ β 16 *More_Favorable + β 17 *Moderately Favorable + Vector of dummy variables for year + Vector of dummy variables for 3 digit ZIP code + IMR. Regression results are found in Table 6. Insert Table 6 Here In the first stage, variables such as Max Wind and Wind Duration significantly increase the probability that the ZIP code/year/yoc observation suffered a loss. The dummy variable for Post 2000 has a negative sign and is significant suggesting the probability of a loss is significantly lower for homes built after new building codes were adopted. In the second stage we see that our 17

20 variables of interest More Favorable BCEGS ratings and Moderately Favorable continue to have negative signs and are highly significant. The coefficients on both are now lower since only observations where a loss occurred are included but are very close to the results shown in Table 3. Equal Strata We have divided our observations into three strata based on the BCEGS rating more favorable, moderately favorable, and low or non rated. In our full baseline model this results in 54,662 observations in the More Favorable stratum, 12,243 in the Moderately Favorable stratum and 2,501 observations in the Low or Non Rated stratum. Large differences in the size of each stratum may influence the effect the ratings have on loss. So our final robustness check is to equalize the BCEGS rating strata and compare the resulting coefficients with those in our base regression. To test for any effect we force each stratum to the size of the smallest by drawing a random sample from each more and moderately favorable stratum giving each stratum 2,501 observations making the total number of observations in the model of 7,503. Insert Table 7 Here Five samples are drawn and the results of each are shown in Table 7. In the base regression the coefficient on the More Favorable stratum is suggesting losses for homes with BCEGS ratings from 1 to 4 have losses that are 15% lower on average than homes in zip codes with low or no rating. All five samples have coefficients on the More Favorable stratum are negative. The range of the coefficients range from to with an average value of For the Moderately Favorable stratum, the coefficient in the base regression is suggesting that losses for home with BCEGS ratings from 5 to 7 have losses that are 18% lower on average. Again, the coefficients for the Moderately Favorable stratum are negative and the range of the coefficients in the five samples is to with an average of the samples of Model Evaluation 18

21 To evaluate our model we used the second stage of the hurdle models and broke our data into two groups. The first group represents 90% of the data, randomly selected, and was used to run the model and collect parameter estimates. The second group is an out of sample control group to test the validity of the model. Parameter estimates from the first group are applied to the control group which gave us a predicted loss for each observation in the control group that can be compared to the actual loss for each observation in the control group. We then regressed the predicted loss from the control group against the actual loss. Insert Figure 4 Here Figure 4 plots the predicted loss against the actual loss and also provides the fitted line with 95% confidence limits. The R Squared for the regression is Our model appears to do a good job of predicting most losses until the size of the loss is very large. But those large observations represent less than 2% of the total observations in the out of sample control group meaning that we are predicting 98% of the observations actual losses very well. VI. Conclusion Florida is highly vulnerable to windstorm damages, as well as the oft referenced gold standard of a strong statewide building code (IBHS, 2015). Existing research has demonstrated the value of a strong FBC by quantifying windstorm loss reductions for new construction under the new code regime (IBHS, 2004; Applied Research Associates, 2008; and Simmons et al., 2016). While this captures the significant extensive marginal value of the FBC extent of new construction post the adoption of the code it is an open question as to how well a statewide code is actually administered and enforced at the local level? Understanding this is important as ultimately risk reduction from the implementation of building codes is due to not only the extent of the code but also from the intensity of local adoption and enforcement. Even for a relatively strong statewide building code such as the FBC. Here we show that both components provide value in reducing windstorm losses in FL, with the extent of the statewide code being the dominant effect reducing losses on the order of 72 percent. Although not as substantial in terms of its loss reduction magnitude, intensively 19

22 implementing building codes at the local level by ensuring codes are properly administered and enforced at this scale provides additional loss reduction value on the order of 15 to 25 percent. These results are robust across a number of specifications and in line with previous BCEGS rating research (Czajkowski and Simmons, 2014). Given the approximately $1.8 trillion of residential property exposure at risk (Hamid et al., 2011) and the $5.178 billion in windstorm losses we analyzed here, both the extensive and intensive marginal value of the FBC is significant. Further, understanding the relative value of these two implementation components is important to better inform building code policy and enforcement efforts as the FBC continues to evolve given ICC updates every three years. A number of caveats, however, need to be addressed. First, it is not completely clear as to why moderately favorably rated local codes provide more loss reductions than more favorably rated local codes? While we see some evidence suggesting the opposite finding as would be expected in regard to new construction and especially high wind environments (Tables 4 and 5 respectively), our results overall (Tables 3, 6, and 7) consistently suggest this to be the case. Of course, this still does illustrate the value of at the least implementing moderately intensive adoption and enforcement at the local scale as opposed to low or none at all. Future research could try to better understand the mix of strength and enforcement activities at the local scale embedded in the overall rating to see if this helps explain this result. Additionally, while we find the extent of the code to be the dominant effect in FL, in other states where a statewide code is not as universally as strong New York or Mississippi for example (IBHS, 2015) the opposite result may be true. Future research should try to investigate the extensive and intensive value of building codes in such statewide environments. 20

23 References AIR Worldwide (2010). Mississippi Insurance Department, Comprehensive Hurricane Damage Mitigation Program: Cost Benefit Study. Available: Applied Research Associates, Inc. (2008) Florida Residential Wind Loss Mitigation Study. Available: Czajkowski, J., J. Done (2014). As the Wind Blows? Understanding Hurricane Damages at the Local Level through a Case Study Analysis. Weather, Climate, and Society, 6(2): (DOI: /WCAS D ). Czajkowski, J., K. Simmons (2014). Convective Storm Vulnerability: Quantifying the Role of Effective and Well Enforced Building Codes in Minimizing Missouri Hail Property Damage. Land Economics, 90(3): Dehring, C. A., M. Halek (2013). Coastal Building Codes and Hurricane Damage. Land Economics, 89(4), Done, J.M., Holland, G.J., Bruyère, C.L., Leung, L.R., and Suzuki Parker, A., 2015: Modeling highimpact weather and climate: Lessons from a tropical cyclone perspective. Climatic Change, doi: /s Deryugina, T. (2013). Reducing the Cost of Ex Post Bailouts with Ex Ante Regulation: Evidence from Building Codes. Available at SSRN Dixon, R. (2009). Florida Building Commission Presentation. Available at _DixonFLBldgCode.pdf Florida Catastrophic Storm Risk Management Center, The State of Florida s Property Insurance Market 2nd Annual Report Released January 2013 for the Florida Legislature. Available from FSU%20Storm%20Risk%20Center.pdf Fronstin, P., A.G. Holtmann (1994). The Determinants of Residential Property Damage from Hurricane Andrew. Southern Economic Journal, 61(2): , Oct. Greene, William, (2003), Econometric Analysis, 5th Ed., Prentice Hall, Upper Saddle River, NJ. 21