Assessing stormwater control measure inventories from 23 cities in the United States

Since the 1987 Clean Water Act Section 319 amendment, the US Government has required and funded the development of nonpoint source pollution programs with about $5 billion dollars. Despite these expenditures, nonpoint source pollution from urban watersheds is still a significant cause of impaired waters in the United States. Urban stormwater management has rapidly evolved over recent decades with decision-making made at a local or city scale. To address the need for a better understanding of how stormwater management has been implemented in different cities, we used stormwater control measure (SCM) network data from 23 US cities and assessed what physical, climatic, socioeconomic, and/or regulatory explanatory variables, if any, are related to SCM assemblages at the municipal scale. Spearman’s correlation and Wilcoxon rank-sum tests were used to investigate relationships between explanatory variables and SCM types and assemblages of SCMs in each city. The results from these analyses showed that for the cities assessed, physical explanatory variables (e.g. impervious percentage and depth to water table) explained the greatest portion of variability in SCM assemblages. Additionally, it was found that cities with combined sewers favored filters, swales and strips, and infiltrators over basins, and cities that are under consent decrees with the Environmental Protection Agency tended to include filters more frequently in their SCM inventories. Future work can build on the SCM assemblages used in this study and their explanatory variables to better understand the differences and drivers of differences in SCM effectiveness across cities, improve watershed modeling, and investigate city- and watershed-scale impacts of SCM assemblages.


Introduction
Stormwater management is a necessary practice in every city with heightened investment driven by regulatory compliance, increased urbanization, aging infrastructure, and climate change [1]. In 2012, the US Environmental Protection Agency (US EPA) estimated that an investment of approximately $19.2 billion in stormwater management was needed to meet the national water quality objectives of the Clean Water Act [1]. Flood mitigation in urban settings, via water conveyance, has been a primary focus of stormwater management since its first implementation [2][3][4]. Newer types of stormwater infrastructure designed to clean, harvest, infiltrate, detain, or retain storm runoff (referred to here as stormwater control measures (SCMs)) have been regulated and implemented in US cities for decades [4][5][6]. More recently, it has become clear that to achieve the goals of stormwater management at site and city scales, SCM networks must be considered in addition to individual SCMs implemented in isolation [6][7][8][9][10][11]. Furthermore, cities wishing to develop their stormwater plans, especially small and midsize cities, can greatly benefit by learning from other cities that already have mature plans [9,12,13].
Understanding broad drivers and constraints of city SCM assemblages would enable cities to make more effective city-scale stormwater plans about the type and number of SCMs implemented. This is especially true for cities with less mature stormwater programs. As opposed to implementing individual SCMs in a piecemeal way, cities may begin with an understanding of what types of SCM assemblages have been implemented by cities in similar physical, socioeconomic, regulatory, and climatic settings. However, it is not clear if or how federal regulations requiring management of stormwater have interacted with city-scale physical, climatic, and socioeconomic factors to shape current SCM assemblages.
Comparison of stormwater management in multiple cities is required to understand what the strongest drivers and/or constraints of varying stormwater management approaches are. Yet, few cross-city comparisons of stormwater management exist, and those that do only compare a few cities at a time. For example, three cities in similar social and ecological settings in Utah had different designs and densities of implemented SCMs, and each city's stormwater infrastructure varied through time on its own trajectory [14]. While that study only considered storm sewers, detention basins, and canals, McPhillips and Matsler [15] compared eight types of SCMs. They found that the types of SCMs in Portland, OR; Phoenix, AZ; and Baltimore, MD have become more diverse over time, evolving from SCMs with large footprints and single-purpose functions to more SCMs with smaller footprints and multipurpose functions. Hale [14] explicitly highlighted the need for cross-city comparison studies of stormwater infrastructure with large sample sizes to understand the factors influencing SCM variation between cities.
Although there are no studies statistically analyzing SCMs among more than three cities, there has been extensive work to identify important considerations when selecting SCM types. A study in the Great Lakes area of the United States interviewed stormwater professionals from across the region [16]. Those interviews highlighted that local considerations, such as topography, soils, and climate, as well as citizen awareness were thought to drive stormwater management decisions. Constraints, such as tighter rules from state and local regulators, which are often driven by federal regulation, and more expensive land required for larger SCMs also drive the adoption of new SCM technologies [16]. A panel of experts in the field of stormwater identified that meeting permit compliance in a cost-effective way was the primary driver guiding decision-making for municipal stormwater infrastructure projects [17]. Other work also found that both federal regulations and individual stormwater manager values combine to determine what types of SCMs are chosen [18]. The potential benefits of SCM networks, however, are now recognized to go beyond permit compliance. The Water Environment Federation (WEF) and the American Society of Civil Engineers (ASCE) [11] identified, 'flood control, stream channel protection, groundwater recharge, water quality improvement, protection of public safety, health, and welfare' , and more as potential SCM benefits. As such, they suggested considerations when selecting SCMs, including physical, construction and maintenance, environmental, social factors, and permitting.
It is known that local stormwater design and criteria manuals have a strong influence on SCM selection at the site scale, but understanding current SCM assemblages using such manuals is challenging. For example, most cities have had manuals evolve over time [19], so comparisons with current SCM inventories would require a date of SCM installation to associate the specific manuals with specific SCMs. This would make it more difficult to obtain a large-enough sample size (e.g. enough cities that had SCM inventories with date of SCM installation) to reveal meaningful statistical relationships. There is also the fact that some cities use multiple manuals [19], as illustrated by the recent work by Grabowski et al [20], where they investigated 122 plans in just 20 cities to understand how cities define green infrastructure, a type of infrastructure most frequently associated with stormwater management. Furthermore, it is often difficult to obtain such documents, even in cities with combined sewers [21]. Even if manuals for all cities could be easily obtained, using data directly to reveal statistical relationships between SCM inventories and physical, climatic, socioeconomic, and regulatory variables has greater potential to gain insight to the underlying drivers and constraints of stormwater management-on which, local design and guidance manuals are based. Important considerations have been identified for the selection of individual SCMs, but it is not clear what factors have actually driven and/or constrained current city-scale SCM assemblages. To compare SCM assemblages and understand drivers of assemblages across cities, our goal was to use the database of implemented stormwater controls (DISCs; 22) to perform a rigorous statistical analysis on SCM assemblages testing a general hypothesis that physical, climatic, socioeconomic, and regulatory attributes of cities are governing their SCM assemblages. This hypothesis is predicated upon the idea that, like species, individual SCMs have niche environments, in which they will perform optimally, and their assemblages come from the interplay of environmental drivers and constraints. To meet this research goal, we asked the following questions: 1. How do SCM density and assemblages of SCMs differ among US cities? 2. Which physical, climatic, regulatory, and socioeconomic variables best explain differences in SCM assemblages between cities?

Methods
To address our research questions, we first collected SCM inventories from as many cities as reasonably possible [23]. After compiling a list of possible explanatory variables based on hypothesized relationships with SCM types and SCM assemblages, we collected and analyzed data representing the possible explanatory variables.

Data collection
We used SCM data from the DISC along with data from six additional cities for a total of 23 cities (figure 1; 23). All SCM data were spatial data except for dry wells in Phoenix, AZ; catch basins in New York City, NY; and green roofs, inlets, drains, and catch basins in San Francisco, CA, which came as lists. While we collected data for gross-pollutant traps (e.g. catch basins), we did not end up using gross-pollutant traps in our analysis, as discussed in more detail in the next section. Of the 23 cities we collected data for, fifteen were MS4 phase I cities, eight were MS4 phase II cities, nine had combined sewer systems within their boundaries, and five were under consent decrees with the US EPA (i.e. Baltimore, MD; San Diego, CA; San Francisco, CA; Seattle, WA; and Washington D.C.). Of the five cities under a consent decree, all were MS4 phase I cities and three had combined sewer systems. Eight Köppen climate regions [24,25] were represented by the cities, with humid subtropical being the most prominent (seven cities). Possible explanatory variables that were collected (table 1) included physical, climatic, socioeconomic, and regulatory variables that were investigated because they may either directly affect SCM assemblages or to be indicators of variables that directly affect SCM assemblages. For example, we hypothesized that shallow depth to water table (DTWT) would limit the implementation of infiltration-based SCMs and that older cities would have a less diverse composition of SCMs since available SCMs have become more diverse over time. Examples of indicator variables [26,27] used in this study are median housing age, which was used as an indicator of city and infrastructure age; population density, which was used as an indicator of development density and type (e.g. compact vs sprawl); and minimum, mean, and maximum DTWT, which were used to indicate groundwater conditions throughout each city. The only variables we hypothesized would affect SCM assemblages but that we did not include in our analysis were related to the subsurface and included soil properties and depth to bedrock. Soil variables were not included because there was poor coverage in the study cities from existing national-scale soil databases (e.g. Soil Survey Geographic Database (SSURGO); 28, 29). For example, SSURGO's coverage of saturated hydraulic conductivity data ranged from 0% to about 59% of city land area with an average coverage of only 28%. The only datasets of depth to bedrock that we were able to identify have relatively low accuracy, and their creators urge caution when applying them [30,31], so depth to bedrock was also not included in our analysis. Despite the exclusion of these subsurface properties, the analysis presented herein provides useful insights. The processing steps for each explanatory variable and example hypotheses are described in table S1, and the values of the explanatory variables are presented in table S2. All spatial data were converted to the North American Datum of 1983 geographic coordinate system and projected to the USA Contiguous Albers Equal Area Conic projected coordinate system. To be consistent between datasets and analyses between cities, all SCM and explanatory variable data were clipped to city boundaries when possible (i.e. lists could not be clipped).

Intercity SCM comparison and analysis
To compare SCM assemblages between cities (question 1), we used the definitions and classification systems identified by the ASCE and WEF in the manual of practice Design of Urban Stormwater Controls [11] (referred to here as Manual of Practice (MOP) and built upon by Choat et al [23]. Those classification systems included what we considered to be a fine resolution classification system (MOP-fine), which contained 27 SCM types, and a coarse classification system (MOP-coarse), which contained five SCM types (table 4.2 in [11]). Our analyses focused on SCMs falling under the four MOP-coarse categories of basins, swales and strips, filters, and infiltrators. Simple definitions for these SCMs can be misleading by ignoring the diversity of form and functions found within each type [23], but for the sake of clarity we provide short definitions in table 2, or see table 4.2 in [11] for a detailed breakdown of the functions provided by the various SCM types. Note that other classification systems may represent particular functions more accurately (e.g. separating different types of basins for water quality functions), but this classification system served as a generally applicable and well-documented way to classify SCMs. The fifth MOP-coarse category of gross-pollutant traps was excluded because they include SCMs, such as screens that are simple inline treatment devices with limited functionality other than gross-pollutant removal. They can also be expected to be found in every city yet were only included in 11 cities' inventories. In eight of the eleven cities' inventories in which they were listed, they accounted for greater than 80% of the listed SCMs and in one case as much as 99.5% of all SCMs. Including gross-pollutant traps would have greatly skewed our analysis.
We compared SCM density (total SCM counts per impervious area) and relative SCM abundance (count of each SCM type per total SCM count) under each of the classification systems. An unbiased Shannon diversity index, H ′ [45][46][47], was calculated as a measure of SCM assemblage diversity to better understand which cities listed a greater diversity of SCMs and what was influencing that diversity. Adapted to our analysis, H ′ was a function of the proportion of SCM type i out of all SCMs in a city (p i ), the number of unique SCM types in that city (S), and the total number of SCMs in that city (N), so it was a useful measure for understanding city SCM assemblages. The Shannon diversity index was calculated as, To address our second research question about which explanatory variables best explain differences in SCM assemblage, we used two statistical approaches. Our first statistical approach was to investigate if any variable explained the observed differences in SCM assemblages. SCM counts were Hellinger transformed (square root of relative abundance) such that variables with many zeros or very low counts would be given lower weight [48]. In this approach, we performed nonparametric tests on the Hellinger-transformed

SCM type
Definition [11] Basin Unit operations in which water is detained for a period that varies with the type of basin and the design requirements Swales and strips Unit operations with the distinct purpose of conveying stormwater from one point to another at very shallow water depths

Filters
Unit operations where stormwater flows through an engineered porous medium and into an underdrain Infiltrators Unit operations in which a design volume is infiltrated to the native soil to recharge aquifers MOP-coarse SCMs and each of the explanatory variables. For the continuous explanatory variables, we applied Spearman's correlation [49], and to further investigate the effects of regulation, we applied the rank-sum test [50,51] to the categorical variables (i.e. combined sewer overflow (CSO) presence, if the city is under a consent decree, and MS4 phase). Nonparametric tests were used because some of the data did not pass tests of normality [52] and/or equal variance [53]. In our single variate analysis of continuous variables, we included the Shannon diversity index as a response variable. The second statistical approach was used for the possibility that some continuous variables may show threshold relationships with the SCM assemblages or certain SCM types. For example, infiltrators may not be implemented below some threshold in DTWT. This is expected to be the case when a single infiltrator is being implemented, and we tested whether such thresholds appear when examining city-scale SCM implementation using summary statistics of the explanatory variable (e.g. mean DTWT over a city). Two approaches were taken to test if such relationships exist. First, segmented regression was performed to test if regression models produced smaller squared residuals with the inclusion of a breakpoint, where the data both below and above the breakpoint had their own line of best fit. For any regression models with improved squared residuals, Spearman's rank-order correlation [49] was used to test if statistical correlation existed between the response variables and the explanatory variables falling below the threshold or above the threshold independently. Second, the nonparametric Wilcoxon rank-sum test [51] was used to test for statistical differences in medians in the data below and above a given threshold. To identify statistically significant thresholds, each explanatory variable data point, except for the smallest and largest five, were tested as thresholds. If more than one statistically significant threshold was identified, only the one producing the smallest p-value was retained.

SCM assemblages and density (question 1)
We counted the number of SCMs per square mile of impervious area in each city to understand how SCM density differed between cities. SCM density varied over orders of magnitude with as little as 0.74 SCMs per square mile of impervious area in Los Angeles, CA and as much as 505 SCMs per square mile of impervious area in Washington D.C. (figure 2). MS4 phase I cities and especially those with combined sewers had the greatest SCM densities.
To better understand how SCM assemblages differed between cities, we examined the fraction of total SCM counts as each SCM type in each city and calculated the Shannon diversity index for MOP-fine SCMs. Diversity in MOP-fine SCMs showed large variability ( figure 3). Pocatello, ID only listed one SCM that was considered in our statistical analysis (infiltration basin) and represented the lowest MOP-fine SCM diversity out of all cities. Baltimore, MD had the highest MOP-fine SCM diversity, listing multiple types of basins, filters, and infiltrators. Overall, basins and infiltrators were common even when MOP-fine SCM diversity was low (left of figure 3), and swales and strips and filters drove the greater diversity in cities with high diversity (right of figure 3).

Analysis of explanatory variables for SCM assemblages (question 2)
To better understand if any single variable was most related to the relative abundance of a given MOP-coarse SCM class, Spearman's correlation was calculated between Hellinger-transformed MOP-coarse SCMs and explanatory variables (figure 4). Cities that were not limited by shallow water tables preferred stormwater infiltrators over basins, swales and strips, and filters (figure 4). Mean DTWT was correlated with minimum and maximum DTWT, mean slope, and standard deviation of slope (figure S1). Swales and strips and filters were implemented more often in the same cities as one another, and their implementation was correlated with the same explanatory variables (figure 4). Impervious percentage, population, and population density  were correlated with one another, but population density had the largest and most statistically significant correlation coefficients with swales and strips and filters. Filters were implemented less frequently when maximum DTWT was deeper, perhaps due to greater implementation of infiltrators, which can provide some degree of filtration. Filters were implemented more frequently with an increasing percentage of regulated waterways considered to be impaired ( figure 4). This can likely be attributed to greater implementation of bioretention facilities, which provide the greatest variety of pollutant control out of any MOP-fine SCM type [11]. Also, bioretention facilities were the most commonly listed MOP-fine SCM type considered to be a filter [23].
Basins were the most frequently listed MOP-coarse SCM type (table 2 in 23), but the relative abundance of basins was only positively correlated with one variable, the 2 yr and 24 h design depth, and only when that depth was about 2 in. or greater (figures 4 and S2). Older cities with greater impervious percentage and population density had smaller relative abundances of basins and implemented more filters and swales and strips. Older cities (greater median housing age) were positively correlated with population, population density, and impervious percentage, and they were negatively correlated with minimum DTWT. Of all explanatory variables, basins were most strongly and significantly negatively correlated with median housing age followed by imperviousness and standard deviation of slope. Overall diversity of MOP-coarse SCMs was positively correlated with socioeconomic variables of median household income and population density ( figure 4). Wilcoxon rank-sum analysis of categorical explanatory variables indicated that MS4 phase I cities listed higher rates of swales and strips than MS4 phase II cities (p ⩽ 0.1; figure 5). The presence of combined sewers and consent decrees led to greater differences in MOP-coarse SCM composition and diversity. In cities with combined sewers, basins (p ⩽ 0.05) were implemented less frequently in favor of swales and strips (p ⩽ 0.05) and filters (p ⩽ 0.05), leading to greater diversity in cities with combined sewers (p ⩽ 0.01). Similar trends were observed in cities that were under a consent decree, except basins did not show a significant relationship, and consent decrees were a more significant predictor of the presence of filters (figure 5).
Segmented regression revealed four relationships between SCMs and explanatory variables that had smaller squared residuals with the inclusion of a breakpoint compared to without one (figure S2). Of those, the only statistically significant Spearman's correlation observed was the increasing proportion of basins with increasing depth of the 2 yr and 2 h design storm, and that relationship was the only significant above a breakpoint of 1.97 in. (table S4). Several statistically significant thresholds (22 total; nine with p ⩽ 0.1, eight with p ⩽ 0.05, and five with p ⩽ 0.01) were identified using the Wilcoxon rank-sum test to test if SCM abundance was different below and above a given threshold (figure S3). Other than minimum, mean, and maximum DTWT, all thresholds were identified in climatic variables (e.g. aridity index, temperature, vapor pressure deficit, annual precipitation, and the 2 yr and 24 h design storm depth). These results generally suggest that infiltrators are favored over the other three SCM types in more arid (i.e. warmer and drier) climates that have greater depths to water table.

Discussion
SCM density and diversity in the study cities' inventories exhibited a large variability between cities (question 1). SCM density (i.e. SCM counts per impervious area) varied over four orders of magnitude over our 23 study cities (figure 2). One city only reported one MOP-fine SCM considered in our analysis, some cities reported three or less, while others reported more than ten ( figure 3). These results could reflect record-keeping practices of cities or actual SCM implementation or both. In an ideal world, these would be the same. However, some cities did not keep inventories of all considered SCM types. For example, Fort Collins, CO had a well-organized SCM inventory, but they did not include swales and strips in their inventory. Another example of inconsistent record keeping is that some of the cities' inventories did not include privately owned SCMs, while others did. It is also possible that other groups, such as counties, departments of transportation, or sewerage districts, sometimes have inventories of SCMs, but we were unable to collect such inventories. These limitations create a double-zero problem common in species composition data; while the presence of an SCM can be directly understood, the lack of an SCM is difficult to interpret. We do not know if an SCM that is not listed is simply not included in that city's inventory or if there actually are not any in the city. We addressed the limitation of the double-zero problem the same way it is commonly overcome in analyses of species assemblages-by using Hellinger-transformed SCM counts [48,54] instead of comparing the overall magnitude of SCMs implemented except for our analysis of SCM density (figure 2). Despite how representative cities' inventories were of the full array of implemented SCMs, they highlight SCMs that are prioritized within a city.
Of the continuous explanatory variables we considered, the physical and socioeconomic characteristics of cities best explained SCM assemblages (question 2). In particular, there were relatively more infiltrators when the water table was deeper ( figure 4). A greater fraction of SCMs were basins in cities with newer housing, and swales and strips were used with high population density. A high fraction of filters were used, where there was a large percentage of regulated waterway length considered impaired ( figure 4). Similarly, based on categorical explanatory variables of MS4 phase, presence of combined sewers, and whether a city was under a consent decree, infiltrators were the only SCM type that were not well explained by any of the three (figure 5). Identifying the unique contributions of combined sewers and consent decrees is difficult based on these results, but the results suggest cities under consent decrees are more likely to choose filters, while cities with combined sewers are more likely to choose a diverse array of SCMs. Although our analysis revealed correlations between the single explanatory variables and SCM types, some of the explanatory variables were strongly correlated with each other across classes (e.g. regulatory and physical classes). For example, cities with combined sewers tend to be older cities (i.e. older median housing age), and older cities had greater impervious percentage and population density, with these relationships being among the strongest and most statistically significant of all variables considered (figure S1).
Perhaps the most surprising result from our analysis was the lack of explanatory power of climatic variables when applied as continuous explanatory variables. However, our analysis of breakpoints and thresholds in the relationships between explanatory variables and Hellinger-transformed SCMs highlighted that climatic variables may be better used as categorical predictors (figure S3).

Limitations
The amount and types of SCMs implemented in cities change over time because of changes to impervious cover, economic activity, and policy [15]. In temporal studies, it has been found that installation of SCMs can lead to population displacement [55,56]. We focused on examining SCM inventories across cities at one time point without considering changes over time. There are no other rigorous statistical or general comparison studies that we are aware of that have compared SCM assemblages between more than three cities. Possible reasons for this are that many cities are still developing and refining their databases to store SCM information, and terminology differences make such comparisons challenging [23]. In a previous study, Hale [14] found that the development of stormwater infrastructure was decoupled from impervious cover. However, for the analysis, Hale included conveyance structures, which could have resulted in conclusions different from those of our study, which excluded traditional conveyance structures. Another important consideration highlighted by Hale was that new infrastructure is directly impacted by already-existing infrastructure. This is a possible constraining factor for SCM implementation that we were not able to capture, because we did not consider temporal information on SCM implementation, but could help explain some of the variability in SCM implementation that remained unexplained by our analysis. Further work may explore in greater detail the local, regional, and national factors in decisions made in cities and across cities for how selection of SCM types is made.

Implications
Despite significant expenditures by the federal government to help eliminate and/or control nonpoint source pollution, the US EPA found that four of 13 stream and river quality indicators showed statistically significant decreases between 2009 and 2013, with no indicators showing improvement [57]. Such data imply that current stormwater management approaches are not in general adequate to prevent adverse downstream impacts. The results from this work suggest that SCM implementation in the 23 study cities has been influenced by federal regulations related to the Clean Water Act (i.e. regulation of cities with combined sewers, larger cities with municipal separate stormwater and sewer systems, and the utilization of consent decrees for enforcement). For example, cities with combined sewers and/or under consent decrees had greater SCM density and diversity, but it is important to note that neither SCM density nor diversity is a measure of effectiveness, but rather they are useful measures for comparing SCM assemblages across cities. However, there is potential to provide a broader range of functions and system-wide resilience with a more diverse SCM inventory, but there is also the potential for greater maintenance time and workforce training requirements with more SCM types. In addition to federal regulations, factors specific to the cities, especially physical and socioeconomic factors, are also important drivers or constraints of SCM composition. While federal regulations are clearly having an influence on stormwater management decisions in cities, this does not mean that the magnitude and type of SCMs implemented are adequate to reach water quality targets in downstream water bodies.
As more SCMs are implemented in the United States to meet the goals of the Clean Water Act, knowing how other cities have approached similar problems under similar constraints can help inform cities planning to implement newer approaches to stormwater management. For example, cities looking to develop a stormwater plan can find partner cities in similar settings that have already addressed stormwater challenges from whom to learn. One example of stormwater sister cities are New York City and Copenhagen, which have been partnering since 2015. While there are numerous factors to be considered when developing a stormwater plan and selecting SCMs (e.g. construction and maintenance, environmental factors, and permitting [11]), this work has provided evidence that important indicators exist that may allow for prediction of SCMs in cities, which has implications for modeling and stormwater network design. Those same indicators may be used to identify good cities to partner with as all cities attempt to deal with the challenge of urban stormwater management. For example, if a city has combined sewers and relatively shallow depths to groundwater (e.g. mean depth to water table <∼7 ft), then that city may consider partnering with either Grand Rapids, MI or Sacramento, CA. Specifically, our work has shown that important indicators of SCM assemblages, which are impervious percentage, DTWT, land surface slope, median household income, regulatory factors (e.g. MS4 phase, combined sewer presence, and consent decrees), and thresholds in climatic factors (e.g. aridity index, annual precipitation, etc), are important indicators when identifying partner cities.

Conclusions
If stormwater management is to reach its potential of meeting site-, city-, and watershed-scale goals of not only mitigating the negative effects of urbanization but also providing additional services that improve the environment and quality of life for all, then challenges need to be explicitly addressed at each scale. Advancing the practice toward that vision will be greatly accelerated if stormwater asset management systems [58] include SCM functions and are made available, as suggested by Choat et al [23]. However, stormwater management is not a one-size-fits-all practice, so even if data are made available, understanding what factors are influencing SCM assemblages in different cities will be useful in allowing cities to learn from one another. To advance toward that goal, we performed robust statistical analyses on SCM inventories from 23 US cities to better understand SCM assemblages and what influences variation in SCM assemblages as we addressed our research questions: 1. How do assemblages of SCMs and SCM density differ among US cities? SCM assemblages and density varied wildly between cities (figures 2 and 3). Some cities listed only one or two SCM types, while others listed more than ten (figure 3). The four cities reporting the largest SCM density were each MS4 phase I cities with combined sewers (figure 2). Cities that implemented a low diversity of SCMs tended to be dominated by infiltrators and basins, but as cities implemented a greater diversity of SCMs swales and strips and filters were listed more frequently (figures 3 and 5). New York City was the only city to report swales and strips as their dominant MOP-coarse SCM type.
2. Which physical, climatic, regulatory, and socioeconomic variables best explain differences in SCM assemblages between cities?
Physical variables (i.e. topography, DTWT, and imperviousness) were the only class of variables to be significantly correlated to each MOP-coarse class of SCM. Infiltrators were the only MOP-coarse SCMs to be best explained by physical variables, while socioeconomic variables displayed greater correlation with filters, basins, and swales and strips. The most statistically important socioeconomic indicator variable was median household income because population density and median housing age were highly correlated with impervious percentage. Climatic variables were shown to be better treated as categorical indicator variables, where cities below and above a given threshold implement different SCM types at different rates. All federal regulatory variables appeared as significant predictors, but the presence of combined sewers and whether a city is under a consent decree or not were shown to be especially important indicators in understanding SCM assemblages.
We postulated a general hypothesis that physical, climatic, socioeconomic, and regulatory attributes of cities are governing their SCM assemblages. Our results generally support our hypothesis, as many SCMs were well explained by the explanatory variables considered in this study. It was surprising to find that climatic variables were perhaps the least important in explaining observed SCM assemblages. Their importance was highlighted, however, when they were considered as categorical predictors of individual SCM types. Other explanatory factors (topography, DTWT, imperviousness, income, combined sewers, and consent decrees) were found to be variables that affected the selection of SCM types more than climate characteristics, possibly because all cities are tasked with treating stormwater, regardless of the characteristics of the storms that generate that stormwater.
While one can make assumptions about SCM function based on design, it is critical that future work further explore implications of SCM assemblages on city-and watershed-scale functions. The International BMP database [12,59] is an important resource that has helped aggregate information on function of individual SCM types, but there is still a need to further our understanding of the emergent functions of SCM assemblages and understand broader suites of functions beyond the core water quantity and quality-related functions. This would also help us understand if there is a relationship between greater diversity of SCMs and meeting the more diverse goals now common in stormwater management, including coping with climate change improving environmental justice. Additionally, future qualitative work could further explore the development of SCM design guidelines in cities, which might provide some insight into SCM choices that could not be explained here. An important step toward these goals is the collection and aggregation of SCM assemblage data, such as expanding the database of SCMs used in this study [22] to include more cities and inventories from non-city entities.

Data availability statement
Most data that support the findings of this study are openly available. (Database of Implemented Stormwater Controls; DISC); https://tinyurl.com/HUB-DISC) [22].
The data that support the findings of this study are openly available at the following URL/DOI: www. hydroshare.org/resource/9b2572b9ee58484483d539051adc019a/#citation.