You are on page 1of 16

Resources, Conservation & Recycling 175 (2021) 105856

Contents lists available at ScienceDirect

Resources, Conservation & Recycling


journal homepage: www.elsevier.com/locate/resconrec

Water savings of LEED-certified buildings


Kaifang Luo a, John H. Scofield b, Yueming (Lucy) Qiu a, *
a
School of Public Policy, University of Maryland College Park, College Park, MD 20742, United States
b
Department of Physics and Astronomy, Oberlin College, Oberlin, OH 44074, United States

A R T I C L E I N F O A B S T R A C T

Keywords: The LEED rating system awards credits for a variety of criteria, including water efficiency. LEED certification
Water savings indicates that the building is designed to conserve water. Given freshwater supply shortages and sustainable
LEED buildings goals, the purpose of this paper is to investigate whether LEED buildings suffer from the water performance gap.
Water performance gap
The water performance gap is the difference between planned and actual water savings. Compared with previous
studies, we analyze a substantially larger sample of 10,557 buildings in six U.S. cities covering 162 million square
meters; 354 of those buildings are LEED certified. In particular, we focus on three types of buildings: offices, K-12
schools, and multifamily housing. First, we identify comparable non-LEED buildings with propensity score
matching. Second, we apply weighted regression to estimate the effect of LEED certification on water savings
using matched samples. Statistical analysis indicates no significant differences in water consumption between
LEED and non-LEED buildings. LEED-certified buildings use no less water than non-LEED buildings, suggesting a
water performance gap for LEED buildings. A 10-fold cross-validation method has been applied to validate the
generalization accuracy of the established models. Based on our results, we propose that LEED system should
secure more points for water efficiency and apply weights based on local conditions and building attributes.
Moreover, the government should incentivize facility managers to reduce technology failure, promote long-term
monitoring, and raise public awareness regarding water conservation.

1. Introduction save 18%-39% site energy on average, some LEED buildings consume
more site energy than non-LEED buildings. Scofield (2009) uses
The building sector consumes a substantial amount of water. About 4 area-weighted averaging to criticize Newsham’s work and shows that
million m3 of water was consumed by buildings greater than 18,000 m2 LEED buildings demonstrate no source energy savings. These studies
in 2012 in the U.S., which accounted for 2.3% of the country’s total bring the energy performance gap to our attention (Liang et al., 2019),
water supply (Dieter et al., 2018). Hence, water sustainability is crucial which means the difference between the actual and predicted energy
for buildings given the worsening water pollution challenges and savings (Qiu and Kahn, 2019; De Wilde, 2014; Fedoruk et al., 2015) of
growing population (United Nations, 2014). The Leadership in Energy energy-saving technologies or designs.
and Environmental Design (LEED) program is a mainstream green The latest version, LEED v4, was launched in 2013 (Long, 2013). It
building rating system in promoting a sustainable construction envi­ aims to correct problems through the linkages between credits and
ronment (Richter et al., 2008). More than 24,500 buildings are certified outcomes (Greer et al., 2019) and keeps revising to meet such objectives.
by LEED in the commercial sector, containing 381 million m2 of floor Compared to the prior version, LEED v3, the credit categories were
area as of January 2020—accounting for about 5% of the U.S. com­ changed, and the credit point allocations were redistributed. Thus, LEED
mercial floorspace (USGBC LEED Project Database, 2020). The LEED v4 introduces seven impact categories as goals and assigns points for
rating system creates an explicit standard for multiple constructional different categories. The total credits of all categories will determine
dimensions. In order to earn more points in different categories, building whether a building could be certified. The water resource is one of the
owners are incentivized to adopt energy- and water-efficient technolo­ seven credit categories (Owens et al., 2013). Through point allocation,
gies. Thus, LEED certification indicates advanced energy and water the LEED rating system guides construction decisions to comply with
performance. Newsham et al. (2009) find that though LEED buildings system goals. LEED v4 has three prerequisites to earn water efficiency

* Corresponding author. University of Maryland, 3135 Van Munching Hall, 7699 Mowatt Ln., College Park, MD 20742, United States.
E-mail address: yqiu16@umd.edu (Y.(L. Qiu).

https://doi.org/10.1016/j.resconrec.2021.105856
Received 17 May 2021; Received in revised form 5 August 2021; Accepted 6 August 2021
Available online 19 August 2021
0921-3449/© 2021 Elsevier B.V. All rights reserved.
K. Luo et al. Resources, Conservation & Recycling 175 (2021) 105856

credit points—indoor and outdoor water use reduction and water, the second-lowest percentage among all categories. BREEAM has
building-level water metering. Reduction prerequisites refer to building the advantage that it scores based on absolute values. In contrast, LEED
design that must reduce water consumption. Building-level water emphasizes the percentage reduction from baseline levels (Horvat and
metering requires installing permanent meters, documenting data, and Fazio, 2005). Additionally, BREEAM provides specific country-specific
agreeing to share the data. If these three prerequisites are satisfied, a schemes tailored for the local conditions, and international schemes
project can be awarded additional points by decreasing landscape irri­ for nations that are not included in local schemes (Suzer, 2015). The goal
gation, saving water for cooling tower makeup, and installing is to develop a framework that can be used to address environmental
sub-meters for water end-use. Thus, LEED-certified buildings are ex­ concerns under local conditions (Suzer, 2015). Green Star also has a
pected to have reduced water usage as compared with non-LEED significant share of the global market for Green Building Rating Systems
buildings. Considering the energy performance gap for LEED build­ and is popular in Australia (Roderick et al., 2009). A total of nine cat­
ings, it is also worth investigating whether there is a water performance egories are included: Management, Indoor Environment Quality, En­
gap for LEED buildings, namely, the gap between the predicted and ergy, Transport, Water, Materials, Land Use & Ecology, Emissions, and
actual water savings of water-saving technologies or designs. Innovation. In constructions without applicable programs, Green Star
Recent LEED studies have concentrated primarily on energy savings, offers a ‘custom’ tool to help the construction team collaborate with the
but water conservation is often overlooked. Compared to prior studies, Green Star authorities to create a specific program (Suzer, 2015). In
few studies use quantitative methods to gauge the actual water perfor­ addition to this, the rating system of Green Star is more flexible in that
mance of LEED buildings. There is a lack of quantitative studies using its score for each category is calculated based on the ratio of fulfilled
empirical methods to estimate LEED buildings’ water use. Elkhapery points to the maximum available points (Roderick et al., 2009; Suzer,
et al. (2021) examine nine schools in Dubai and estimate that 27% of 2015).
water can be saved at LEED Platinum tier buildings. Most of the LEED In this paper, we seek to determine whether there is a water per­
water studies utilize small samples of less than 100 LEED buildings. formance gap for LEED buildings. For this purpose, we incorporate
Some examine the linkages between LEED water credits and water propensity score matching with weighted regressions. The first step is to
savings (Alawneh et al., 2018; Greer et al., 2019). A small-sample use propensity score matching to identify similar non-LEED buildings.
descriptive analysis is less convincing and might be criticized for Second, we use weighted regressions weighted by area involving prop­
external validity issues. Moreover, the majority of previous studies rely erty types as dummy variables to eliminate differences in floor area and
on descriptive analysis to demonstrate that LEED buildings save a sig­ city development in matched samples. Finally, a k-fold cross-validation
nificant amount of water (Al-Qawasmi et al., 2019; Awadh, 2017; Attiya approach is used in order to evaluate model performance. The accurate
et al., 2020; Wu et al., 2020). One strength of LEED is that among the performance and generalized capabilities of the established models are
green building rating systems, the design water model is only mentioned validated. The research is vital for policymakers to optimize the rating
in LEED v4. Plus, and LEED’s interior water minimum requirements are system and for LEED buildings to achieve their intended water
more stringent than Building Research Establishment Environmental sustainability.
Assessment Method (BREEAM)’s (Awadh, 2017). However, LEED also
has its drawbacks. For example, in addition to the main “water category, 2. Material and methods
” LEED evaluates water efficiency and management under other cate­
gories, such as “sustainable sites-SS”, which incorporates a “Rainwater 2.1. Data
management” factor (Al-Qawasmi et al., 2019). It might cause inaccu­
rate measurement of water efficiency. Compared with the Green Pyra­ Benchmarking data of 28,478 buildings for 2016 are organized by
mid Rating System (GPRS), another rating system in Egypt, LEED fails to Energy Star Portfolio Manager1, covering ten cities in the United States
prioritize and award more points for water efficiency given Egypt’s (Boston, Chicago, Denver, Los Angeles, Minneapolis, New York City,
water supply and demand issues (Attiya et al., 2020). Philadelphia, Portland, Seattle, Washington DC). We employ building
Several studies provide cross-sectional analyses of LEED 2009 (Wu characteristics such as property type, floor area, site EUI, source EUI,
et al., 2017; Pushkar, 2018). The LEED NC v3 provides a maximum of total greenhouse gas, water use, built year, and Energy Star score in the
four bonus points for priorities concerning project locations (Suzer, benchmarking data. In addition, “Financial Office” and “Office with
2015). A study by Wu et al. (2017) shows that LEED-2009-certified Data Center” are incorporated as “Office”, which is the same as how
projects differ in their water efficiency at the country and the state Energy Star Portfolio Manager organizes the property type when
level because before LEED 2009, local factors were not taken into ac­ calculating the Energy Star scores. We also collect the LEED project
count. In contrast, the implementation of regional priority points at the database with certification date before July 1, 2016, from the U.S. Green
state level can effectively address regional variations in water efficiency Building Council (USGBC)2, and merge with benchmarking data through
(Wu et al., 2017). Moreover, Pushkar (2018)’ study in Turkey, Spain, the building addresses in both benchmarking and LEED database.
and Italy finds that regional priority points motivate water savings.
The LEED rating system has been critiqued for not effectively ful­
filling its intended sustainability goals. LEED is a typical criteria-based 2.2. Empirical strategy
tool (CBT) for assessing environmental building performance. CBTs
assign points based on a checklist, which indicates corresponding This paper analyzes whether LEED certification influences water
environmental impacts (Suzer, 2015; Ali and Al Nsairat, 2009). Build­ consumption in buildings to identify the water performance gap. By
ings are rated by giving points to multiple categories, and the overall meeting certain construction specifications and receiving LEED points
score is calculated to determine the certification level. Horvat and Fazio for water efficiency, LEED-rated buildings are predicted to consume less
(2005) point out that other CBTs may be more preferred in rectifying water than conventional ones. The “water performance gap” hypothesis
problems through connections between credits and behaviors. The would be confirmed if LEED-certified buildings did not conserve or used
Building Research Establishment Environmental Assessment Method more water than their comparable counterparts.
(BREEAM) is one example (Horvat and Fazio, 2005). It was the first and
strongest environmental assessment system developed in the U.K.
(Suzer, 2015; Doan et al., 2017; Roderick et al., 2009). BREEAM UK New 1
U.S. EPA’s Energy Star Portfolio Manager. https://www.energystar.gov/bui
Construction 2018 consists of ten categories: Management, Health & ldings/facility-owners-and-managers/existing-buildings/use-portfolio-manage
Wellbeing, Energy, Transport, Water, Materials, Waste, Land Use & r.
Ecology, Pollution, and Innovation. A 7% of the weighting is given to 2
USGBC projects. https://www.usgbc.org/projects.

2
K. Luo et al. Resources, Conservation & Recycling 175 (2021) 105856

Table 1
Summary statistics.
Variables Unit Mean Area-weighted means Sd N Min Max
3 2
Water use intensity m /m 1.9 1.5 1.4 10557 0.1 14.4
Site energy intensity kBtu/m2 789.4 779.7 311.9 10557 172.2 2631.8
Source energy intensity kBtu/m2 1453.4 1596.1 607.9 10557 360.6 6455
Greenhouse gas intensity metric ton CO2/m2 57.3 59.8 24.2 10557 12.3 2.44E+02
Floor area m2 15504.3 42963.9 20634.5 10557 223 337858.8
Built year 1955.8 1963 39.6 10557 1000 2019
Energy Star scores 63.1 67 28.5 9567 0 100

2.2.1. Data preprocessing accounting for about 97% of LEED properties in the dataset.
Benchmarking data uses energy use intensity (EUI) to calculate a In this section, we use the propensity score matching (PSM) method.
building’s total energy consumption divided by floor area. Site energy is First, the Stata command psestimate is used to select a linear function of
the amount of energy used at the construction site. In contrast, source covariates that will later be used in PSM. Here treatment indicator be­
energy refers to the footprint of all associated energy losses, both on-site comes the dependent variable, whereas covariates for matching become
and off-site, incurred during generation and transmission. Likewise, we the independent variables. The base model is treatment along with site
calculate water use intensity (WUI) and the greenhouse gas intensity EUI. The remaining covariates, source EUI, floor area, built year, and
(GHGI) by dividing annual water use by floor area and by dividing total Energy Star score, are included step-by-step. The command compares the
greenhouse gas by floor area. Water intensity is the dependent variable base model with the new model, adding one additional covariate each
in this study, while LEED certification is the independent dummy time, and selects the covariate with the highest likelihood ratio test
variable. (LRT).
When preprocessing the data, we begin by eliminating records with Second, PSM runs a logit model to determine the predicted proba­
missing or zero values, resulting in a water use dataset for 16,248 bility of LEED certification for 9,567 properties. Control buildings are
buildings in six U.S. cities (Boston, Los Angeles, Minneapolis, New York assigned weights in predicted probabilities based on the distance to each
City, Philadelphia, Washington DC), including 79 different types of LEED building. When determining the shortest distance, we use the
properties. Multifamily housing, offices, and K-12 schools make up 71% single nearest neighbor matching with replacement to select the closest
of all properties. Thus, we focus on these three subsets in the following control building, conditional on the property type and city. Building
analysis. Data from different property types enable us to investigate the attributes that distinguish LEED-certified buildings from non-LEED
heterogeneity of water consumption over property types. Then, we buildings can be narrowed down using the PSM approach. Following
logarithmically transform water use intensity, site EUI, source EUI, total that, a balancing check is necessary to capture whether covariates are
greenhouse gas, and floor area. Next, we remove outliers outside the significantly different between LEED and non-LEED buildings. Balanced
range of the two standard deviations for the mean logarithm values by data should have insignificant p-values for each covariate, indicating
property. This step removes 975 records (roughly 8% of buildings). In that control groups from matched samples are not significantly different
total, the filtered water consumption dataset contains 10,557 buildings from treatment groups in covariates (D’Agostino Jr, 1998).
in six U.S. cities, including 354 LEED-certified buildings. Table 1 pre­
sents summary statistics, including the area-weighted mean values. The 2.2.3. Multicollinearity test
area-weighted mean values are more precise because it eliminates the Before modeling, we obtain the variance inflation factor (VIF) value
influence of floor area on measurement. The weighted average water use to check for multicollinearity. VIFs measure the correlation between two
intensity, site EUI, source EUI, and greenhouse gas intensity are 1.5 m3/ predictors (Gatignon, 2013). First, we perform an aggregate weighted
m2, 779.7 kBtu/m2, 1596.1 kBtu/m2, and 59.8 metric ton/m2. Our regression of all variables of interest using matched samples and prop­
sample buildings were built on average in 1963. Their weighted average erty types as dummy variables, weighted by floor area. Then, we
Energy Star score is 67. compute VIF values for the dependent and control variables.

2.2.2. Propensity score matching 2.2.4. Weighted regression


In screening comparable control groups, matching is a commonly Our main models are to run a series of area-weighted regressions
used quasi-experimental design. The matching process is based on based on matched samples with property types as dummy variables,
property type and city, since water use varies heavily by space type, weighted by floor area. It analyzes the impact of LEED certification by
local economic development, and physical environment. For each LEED city on water use intensityi (water use divided by floor area) with the
building, we identify a comparable non-LEED building of the same type following equation:
in the same city that has similar attributes as the LEED building, such as
WUIi = β0 + β1 Treatmenti + β2 Multifamily Housingi + β3 Officei
site and source EUI, built year, floor area, and Energy Star score. For
example, one Boston LEED office is matched with a Boston non-LEED + β4 Site EUIi + β5 Source EUIi + β6 Built yeari
office with similar building attributes. These matching criteria ensure + β7 Greenhouse gas intensityi + β8 Energy Star scorei + εi (1)
unbiased measurement of LEED certification’s impact on water con­
sumption without interference from other underlying causes. As Here we use Floor areai as the weight wi for building i because water
mentioned previously, there are approximately 10,557 records with 354 consumption increases with floor area. Typically, buildings with greater
LEED buildings in the cleaned water consumption dataset, but not all of floor area usually have higher water consumption. Therefore, we use
them have an Energy Star score. Thus, we use 9,567 records with full area-weighted regressions to prevent estimation bias. In addition, we
data in the matching model, of which 339 are LEED-certified, use property-type dummy variables to analyze heterogeneity in
weighted regression. The inclusion of three types of LEED buildings

3
K. Luo et al. Resources, Conservation & Recycling 175 (2021) 105856

Table 2
City-level matched samples.
Property Type City Total
Boston Los Angeles Minneapolis New York City Philadelphia Washington

K-12 School 0 0 0 0 10 0 10
Multifamily Housing 14 4 0 32 10 12 72
Office 36 70 20 90 24 202 442
Total 50 74 20 122 44 214 524

Fig. 1. Propensity score changes for offices at the city level before and after matching. Notes: This figure depicts the kernel density of propensity scores for each city’s
offices using matched and unmatched samples. Those with “Before matching” are kernel density plots based on unmatched offices (2,453 records). Those with “After
matching” are kernel density plots using matched samples, namely, 1 to 1 matching of LEED offices to conventional counterparts (442 records). Solid lines illustrate
treatment groups (LEED offices), while dashed lines delineate control groups (non-LEED offices). Each sub-graph shows the city-level discrepancy between LEED and
non-LEED offices\.

allows us to compare water performance across building types. Greenhouse gas intensityi denote the intensity of site energy, source en­
Treatmenti is the indicator of LEED certification of building i. ergy, and greenhouse gas, respectively. Built yeari denotes the year when
Multifamily Housingi and Officei are the dummy variables for property building i was built. Energy Star scorei is the Energy Star score of
types. Here Multifamily Housingi = 1 if building i is a multifamily hous­ building i in 2016. εi is the error term of the unobserved factors.
ing and 0 otherwise; Officei = 1 if building i is an office and 0 otherwise.
When Multifamily Housingi and Officei are both 0, it indicates building i is 2.2.5. 10-fold cross-validation
a K-12 school. The coefficients of Multifamily Housingi and Officei mea­ Model performance is commonly assessed through k-fold cross-
sure the heterogeneity of water use intensity across building types with validation (Zhou et al., 2016), where data are randomly divided into k
K-12 schools as the reference group. Site EUIi , Source EUIi , and partitions with basically equal numbers of incidents. The K partition is

4
K. Luo et al. Resources, Conservation & Recycling 175 (2021) 105856

Table 3 2.2.6. Robustness check


Multicollinearity test results. This section uses fewer covariates—site EUI, floor area, and built
Variable VIF (variance inflation 1/ year. Source EUI is removed due to potential correlation with site EUI.
factors) VIF Energy Star score is also eliminated because 999 buildings have missing
Treatment 1.01 0.99 values, occupying about 9% of properties. After that, we proceed in the
Property type same manner as for the primary model: (1) conduct propensity score
Multifamily Housing 7.27 0.14 matching with the single nearest neighbor matching method; (2) use the
Office 7.53 0.13 matched samples to conduct weighted regression with property types as
Site energy use intensity (kBtu/m2) 4.52 0.22
Source energy use intensity (kBtu/m2) 6.2 0.16
dummy variables, weighted by floor area. In this step, we use the same
Built year 1.15 0.87 weighted regression procedure as in the main model, presenting out­
Greenhouse gas intensity (metric ton CO2/ 2.92 0.34 comes at both the aggregate and city levels.
m2 )
Energy Star score 2.34 0.43
3. Results
Mean VIF 4.12

3.1. Matching

the holdout sample for testing, and the K-1 partitions are the training Firstly, the psestimate results indicate that all covariates–site EUI,
sets (Grimm et al., 2017). The training sets fit the specified model to source EUI, floor area, built year, and Energy Star score–should be
predict parameters. Then, each partition in turn is used as the testing included in the linear function. Secondly, we run propensity score
data to evaluate the model induced by the training sets. 10-fold matching and determine the shortest distance using the single nearest
cross-validation is recommended by Kohavi (1995). Our data are neighbor method with replacement. A total of 524 buildings are
randomly divided into 10 equal subsets, and the above process will run included in the principal model as matched samples. Half of them are
10 times, each time with a different holdout sample. In k-fold LEED buildings, accounting for 77% of 339 LEED buildings in the
cross-validation, the training data and testing data are assumed to be filtered water dataset. Thus, the lost LEED data is only a small fraction of
independent (Roberts et al., 2017). As shown in Eq.(2), the model per­ the total after matching. The matched samples in six cities are shown in
formance is determined by the mean prediction error of k partitions Table 2. Take offices as an example. Fig. 1 demonstrates the change in
(Shaban et al., 2021). kernel density of propensity scores over offices in six cities before and
after matching. Kernel density plots are employed to illustrate the dis­
1∑ n
tribution of propensity scores. As shown in Fig. 1, matched control of­
Fitness = RMSE (2)
k i=1 fices are more similar to LEED offices in terms of propensity scores than
unmatched samples. Consequently, the PSM approach ensures that the
Statistical index can also be used to evaluate model performance.
matched non-LEED buildings are comparable to LEED buildings in any
Root means squared error (RMSE) and correlation coefficient (R2) are
metric other than water use intensity. As shown in Table A1 in the
common metrics (see Eqs. (3) and (4)) (Nguyen et al., 2021; Pal, 2017).
Appendix, there is no statistically significant difference between treat­
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
∑n
1 ment and control groups in any covariates, confirming the balance of
RMSE = (pi − qi )2 (3) matched samples. According to Fig. A1 in the Appendix, the probability
n
of LEED certification conditional on observed covariates is within a
i=1

∑n range of zero-to-one, which supports the common support assumption.


i=1 (pi − qi )2
R2 = 1 − ( )2 (4)
∑n
i=1 pi − qi 3.2. Test for multicollinearity

Before modeling, we include a multicollinearity test using the VIF


where n is the total number of samples, qi is the estimated value, pi is the
value. Based on matched samples, we first run an aggregate weighted
actual value, qi is the mean estimated value. The high values of R and
regression of water use intensity and all dependent and control variables
low values of RMSE indicate better performance (Shah et al., 2021).
with a property type dummy variable, weighted by floor area. Following
that, we run the Stata command vif to obtain the VIF values. Values of

Fig. 2. The treatment effect of LEED


certification estimates by city
comparing matched and unmatched
outcomes. Notes: Treatment estimation
is categorized in groups by the city and
in aggregate. Each group uses all
matched samples or corresponding
cross-city data. The error bars show the
95% confidence intervals. The solid
rectangles represent the coefficients’
values of β1 that measure the change in
annual water use intensity in response
to LEED certification. Red rectangles
use matched samples, 524 records. Blue
rectangles use unmatched samples,
9,567 records.

5
K. Luo et al. Resources, Conservation & Recycling 175 (2021) 105856

Table 4
The aggregate and city-level results of weighted regressions with property types as dummy variables.
All matched samples Boston Los Angeles Minneapolis New York City Philadelphia Washington
Model number (1) (2) (3) (4) (5) (6) (7)

Treatment -0.023 -0.142 0.051 -0.178 -0.092 -0.001 -0.047


(0.05) (0.16) (0.13) (0.09) (0.10) (0.15) (0.06)
Property type
Multifamily Housing 1.211*** N/A N/A N/A N/A 1.229* N/A
(0.28) N/A N/A N/A N/A (0.48) N/A
Office 0.407 -0.847 -1.435* N/A -0.569* -0.068 -0.503
(0.26) (0.48) (0.68) N/A (0.24) (0.41) (0.30)
Site energy use intensity (kBtu/m2) -0.0001 0.004 -0.001 -0.0004 0.0003 0.0004 -0.0002
(0.0001) (0.004) (0.001) (0.0005) (0.0006) (0.001) (0.0006)
Source energy use intensity (kBtu/m2) 0.0004*** 0.002 0.002* -0.0004 0.0001 0.0003 0.0004
(0.0001) (0.001) (0.001) (0.0004) (0.0002) (0.0005) (0.0002)
Built year -0.002* 0.003 -0.003 -0.002 -0.002 -0.003 -0.004*
(0.0009) (0.003) (0.01) (0.001) (0.002) (0.004) (0.002)
Greenhouse gas intensity (metric ton CO2/m2) -0.001 -0.087 -0.017 0.00005 -0.008 0.013 0.001
(0.002) (0.07) (0.03) (0.002) (0.01) (0.01) (0.002)
Energy Star scores -0.005* 0.005 0.008 -0.036 -0.016*** 0.008 -0.001
(0.002) (0.01) (0.01) (0.02) (0.004) (0.01) (0.005)
Constant 4.576* -5.729 7.051 9.018* 6.671 4.266 8.331*
(1.86) (6.13) (11.51) (3.14) (3.42) (7.46) (3.58)
Observations 524 50 74 20 122 44 214

Notes: Standard errors in parentheses; * p < 0.05 ** p < 0.01 *** p < 0.001

Fig. 3. Kernel density plots of water use intensity for matched LEED and non-LEED buildings by city.

VIF all fall below 10 (see Table 3), indicating no issue with multi­ response to LEED certification. Fig. 2 illustrates the aggregate and city-
collinearity (Hair et al., 2010). level treatment estimation of LEED certification in Table 4 graphically.
Fig. 2 shows the estimation of β1 using matched and unmatched sam­
3.3. Impact on water use intensity ples, while Table 4 illustrates the regression results for all variables. For
Fig. 2, solid rectangles refer to the coefficient values of β1 , red for out­
The primary strategy is to look at how water use intensity changes in comes using matched samples, and blue for outcome using unmatched

6
K. Luo et al. Resources, Conservation & Recycling 175 (2021) 105856

Fig. 4. 10-fold cross-validation results. (a) Main model (weighted regression plus matching) based on R2; (b) Main model (weighted regression plus matching) based
on RMSE; (c) weighted regression only based on R2; (d) weighted regression only based on RMSE.

samples. The error bars represent 95% confidence intervals. If a bar the “energy and atmosphere” carries the most significant weight, 17
intercepts with the zero line, then it suggests an insignificant coefficient. points, and “water efficiency” the least, 5 points (USGBC, 2005). Later,
We can observe that the coefficients of treatment effects are not statis­ LEED v3 and v4 were introduced in 2009 and 2013. Similarly, they
tically significant at city-level and aggregate outcomes, no matter using assign energy and atmosphere aspects with over 30 points while water
matched or unmatched samples. It indicates that LEED buildings do not efficiency is assigned a maximum of 10 points (USGBC, 2010; USBGC,
save more water than non-LEED buildings. Thus, the actual water per­ 2013). Studies speculate that buildings may attain LEED certification by
formance of LEED buildings deviates from the designed water perfor­ maintaining the cheapest LEED points (Scofield, 2009). Moreover, LEED
mance – the water performance gap is confirmed. points have witnessed bunching around the certification thresholds
The failure of water savings may be associated with the LEED rating (Matisoff et al., 2014). Construction decision intuitively prioritizes en­
priority. LEED rates buildings in several categories and assigns them ergy and atmosphere credits to achieve certain low-level thresholds
different certification levels according to the sum of category credits. without valuing water efficiency credits. If owners are incentivized to
LEED v2.2, introduced in 2005, ranks buildings in six categories where prioritize water efficiency, LEED buildings’ water savings ability will be

7
K. Luo et al. Resources, Conservation & Recycling 175 (2021) 105856

upgraded (Alawneh et al., 2018). kernel density of water use intensity for matched LEED and non-LEED
Another possible reason for the inconsistency between water per­ buildings (see Fig. 3). We notice an apparent difference between com­
formance and green certification is occupants’ unpredicted behaviors parable LEED and non-LEED buildings in terms of the distribution of
(Gunay et al., 2013; Suh et al., 2014; Warren and Parkins, 1984). After water use intensity, except for Minneapolis, indicating a significant
all, in addition to technology innovation, occupants’ daily behaviors difference in water usage. However, our empirical analysis reveals
carry significant weight to maintain a sustainable built environment insignificant results in all cities. We use weighted regression at the city
(D’Oca et al., 2015), including underlying causes such as longer oper­ level, which successfully eliminates the bias resulted from confounding
ational hours and technology failure (Bordass, 2004; Menezes et al., variables. If we were to rely solely on descriptive figures, we might
2012). drastically misestimate the water savings of LEED buildings.
We now discuss the coefficients of other covariates in the regression
models. Detailed results are listed in Table 4. Model (1) in Table 4 3.4. Model cross-validation results
applies weighted regression to the aggregate matched samples. The
aggregate results in building types suggest that distinct kinds of prop­ We perform 10-fold cross-validation on our established model to
erties have significantly different water performance. Here the reference assess its performance. Since the established model incorporates
group is K-12 schools. Holding other factors fixed, the average water use weighted regression and propensity score matching, we use weighted
intensity of offices is not different from K-12 schools, whereas the regression without any matching as the comparison model. Fig. 4 shows
average water use intensity of multifamily housing is 1.211 m3/m2 more the results of 10-fold cross-validation using R2 and RMSE as the
than K-12 schools. The greater water use in multifamily housing lies in assessment criteria, respectively, for the established model and com­
the extra water demand in showerheads for bathrooms and kitchen and parison model. Comparatively, the comparison model has a higher
laundry-related usage. Besides, results demonstrate that with source EUI average R2 of 0.3 than the established model of 0.2. However, the
increasing by one kBtu/m2, a building’s water use intensity will corre­ established model provides the maximum R2 value of 0.5. Furthermore,
spondingly increase by 0.0004 m3/m2, suggesting an inter-related per­ our established model has a much lower average RMSE value of 0.5,
formance in water and energy savings. In addition, the results of model while the average RMSE value of the comparison model is over 0.9.
(1) indicate that if a building is constructed a year later, its water use Moreover, the established model gives us the lowest RMSE value of 6.4.
intensity will decrease by 0.002 m3/m2, which is consistent with tech­ A high value of R2 and a low value of RMSE means better performance
nological developments. Furthermore, the aggregate results suggest that (Shah et al., 2021). According to cross-validation results, our established
a point increase in the Energy Star score is associated with a 0.005 m3/ model with propensity score matching and weighted regression captures
m2 reduction in water use intensity. generalized capacity and accurate performance.
Models (2)-(7) utilize city-level weighted regressions to learn city
heterogeneity in LEED certification. First, results indicate that water
3.5. Robustness check
consumption varies with different building types in Los Angeles, New
York City, and Philadelphia. The default reference group is K-12 school.
Running Stata command psestimate, site EUI, floor area, and built
Due to the lack of K-12 school samples in Los Angeles and New York
year are all included in a linear function step by step. Next, we perform a
City, Stata automatically selects multifamily housing as the reference
propensity score matching and 574 buildings are matched. 287 of them
group. Compared to multifamily housing, the water use intensity of of­
are LEED properties (shown in Table B1 in the appendix). Fig. B1 in the
fices averagely is lower by 1.435 m3/m2 and 0.569 m3/m2 in Los
appendix presents the kernel density of changes in propensity score over
Angeles and New York City, respectively. Compared to K-12 schools,
offices before and after matching at the city level. In matched samples,
multifamily housing consumes 1.229 m3/m2 more on average than K-12
control offices are much closer to LEED offices than in unmatched
schools in Philadelphia, whereas offices show no difference. Heteroge­
samples. Fig. B2 in the appendix confirms the common support
neity in water savings over building types reflects that multifamily
assumption. Then, we use matched samples to carry out a weighted
housings consume the largest amount of water on average than offices
regression using a property type as the dummy variable, weighted by
and K-12 schools given the extra water demand for daily life such as
floor area. Fig. B3 in the appendix compares the treatment estimation
showering, cooking, and laundry. Second, Table 4 shows that for
results of the primary model and robustness check in a graphic form. We
buildings in Los Angeles, a one kBtu/m2 increase in source EUI will lead
can observe that the treatment coefficients are statistically insignificant
to a 0.002 m3/m2 increase in water use intensity. It demonstrates the
both at the city and aggregate levels. According to the results, LEED
energy-water nexus (Rothausen and Conway, 2011) in Los Angeles that
buildings generally do not conserve more water than conventional ones,
water and energy use are intrinsically linked (Copeland and Carter,
which is consistent with the main outcomes of the study. Table B2 in the
2014). It is possible because buildings in Los Angeles use electricity for
appendix provides detailed results. Both aggregate and city-level results
heating water (Movahedi and Derrible, 2020). Third, Table 4 shows that
validate the robustness of the treatment estimation in the main model.
as Energy Star scores increase by one point in New York City, the water
Our results confirm a water performance gap between designed and
use intensity of buildings will decrease by 0.016 m3/m2, further
actual water savings for LEED buildings.
strengthening the energy-water link in building consumption. Fourth,
we see from Table 4 that buildings constructed one year later in
4. Discussion
Washington will have a water use intensity decrease by 0.004 m3/m2,
which is in line with technological innovation over time. New con­
In light of the ongoing concern over freshwater scarcity, this study
struction is probably in better operating condition.
adds three contributions. Firstly, we collect the largest sample of LEED
It is crucial to use adequate statistical methods to control for con­
buildings in the U.S., enabling more convincing and representative
founding factors. Based on propensity score matching, we present the
findings. We collect water consumption data on 10,557 buildings to

8
K. Luo et al. Resources, Conservation & Recycling 175 (2021) 105856

determine whether LEED certification leads to water savings. 354 of assess the water performance gap if we can determine how many water
them are LEED-certified buildings. As LEED certifications are widely efficiency points each LEED building earned; however, our study cannot
recognized in rating green buildings, understanding the actual water provide such analysis.
performance of LEED buildings is crucial for policymakers and water
analysts to refine credit allocation. The greatest sample size of LEED 5. Conclusion
buildings ensures more reliable and generalizable results from LEED
water-related studies. Using a sample of 10,557 buildings in six U.S. cities, we find that
Secondly, this paper provides a more systematic and comprehensive LEED-certified buildings use no less water than non-LEED buildings,
approach on water performance for LEED buildings than previous work suggesting a water performance gap for LEED buildings. Our findings
providing only descriptive statistics. Matching identifies the treatment hold critical implications for policymakers on building water manage­
estimations by comparing them to comparable counterparts in a more ment. To bridge the water performance gap, we propose the following
scientific manner. For each LEED building, we conduct a propensity policy recommendations regarding LEED adjustments in light of our
score matching to find a comparable non-LEED building of identical type findings and the differences between LEED and other criteria-based
and with similar attributes located in the same city. In addition, we use measurement tools.
weighted regressions that are weighted by floor area and include First, from a structural aspect, the water performance gap should be
dummy variables to control for property type differences, eliminating mitigated by securing more credits to water efficiency. Our main model
the influence of floorspace and type. Based on empirical models and explores the water use intensity across six cities in the U.S., but none of
statistical results, we are able to determine buildings’ response to LEED them show water savings due to LEED certification. The reason mainly
certification concerning water consumption. Our findings suggest that lies in the low credits available for water efficiency under LEED. At
the water consumption of LEED buildings is of no difference from their present, LEED v4.1 BD+C assigns 7 credits for water efficiency, lower
conventional counterparts and thus confirms the water performance gap than “energy and atmosphere”, 10 credits, and “indoor environmental
among LEED-certified projects. quality”, 11 credits. LEED’s credit allocation reflects environmental
Thirdly, our empirical findings expand on the previous discussion of priorities for buildings. If LEED increases the water credits available for
LEED buildings. According to our results, LEED certification does not construction, more attention may be paid to water efficiency. Second,
guarantee water savings, confirming LEED buildings have a water per­ when assessing points for each building, it is better to apply weights
formance gap. There are multiple underlying causes, including struc­ according to local conditions and building attributes. We observed
tural, behavioral, and operational factors. For structural aspects, first, heterogeneity in water consumption based on building type. According
other rating factors outweigh water efficiency. Buildings cannot earn the to our research, multifamily housing consumes 1.22 m3/m2 more water
same energy and water points for similar impacts (Greer et al., 2019). than offices and K-12 schools. Adjustments regarding the attributes of
LEED v4.1 BD + C gives 7 credits to water efficiency, fewer than “energy different property types should be considered. Third, efforts should be
and atmosphere” (10 credits) and “indoor environmental quality” (11 made to promote better management practices. Aside from creating
credits). New constructions tend to prioritize other credits such as en­ incentives for facility managers to reduce technology failure, the gov­
ergy and atmosphere over water conservation. Studies have found that ernment can also promote long-term monitoring to improve water
LEED points cluster around certification thresholds in order to maintain management. Lastly, from the behavioral aspect, the government should
the lowest certification level (Scofield 2009; Matisoff et al., 2014). raise public awareness about water savings. Behavioral encouragement
Therefore, the certified buildings cannot guarantee their water effi­ should be incorporated into conservation practices in the future.
ciency. Second, before 2013, LEED building rating was a design-based or
construction-based process rather than performance-based (Chokor and CRediT authorship contribution statement
Asmar, 2017). As a result, predicted water savings of LEED-certified
buildings might be overstated without taking actual water usage into Kaifang Luo: Formal analysis, Investigation, Software, Validation,
account. For the operational aspect, the problem is the water-efficient Visualization, Writing – original draft, Writing – review & editing. John
technology failure. Both prolonged operation and ineffective manage­ H. Scofield: Data curation, Resources, Writing – review & editing.
ment result in water performance gaps (Bordass 2004; Menezes et al., Yueming (Lucy) Qiu: Conceptualization, Funding acquisition, Meth­
2012). For the behavioral aspect, a major contributor to the perfor­ odology, Project administration, Supervision, Writing – review &
mance gap in water is personal behaviors such as wasteful behaviors editing.
(Gunay et al. 2013; Suh et al. 2014; Warren and Parkins 1984; D’Oca
et al., 2015). Declaration of Competing Interest
Although our study uses more rigorous methods and a much larger
sample size than existing studies on LEED water efficiency, it has limi­ The authors declare that they have no known competing financial
tations. First, the limited number of LEED K-12 schools makes it less interests or personal relationships that could have appeared to influence
convincing to confirm the water performance gap in this building type. the work reported in this paper.
Only the aggregate model and Philadelphia model integrate matched
samples of K-12 schools. More K-12 schools and other building types Appendix A
should be incorporated into evaluating water savings heterogeneity
among multiple types for LEED buildings. Second, though the city-level Figure A1
main models control critical characteristics such as building size and Table A1
building year, future research should also validate other underlying
attributes such as occupancy. Third, it would be more straightforward to

9
K. Luo et al. Resources, Conservation & Recycling 175 (2021) 105856

Fig. A1. Check for common support assumption for PSM results by property type and city.

Table A1
Balancing check for propensity score matching results.
(a) For Multifamily Housing in Boston.
Variable Unmatched Matched Mean %reduct |bias| t-test V(T)/V(C)
Treated Control %bias t p>t

Site energy use intensity (kBtu/m2) U 599.28 769.09 -66.8 -1.49 0.138 0.24
M 599.28 646.24 -18.5 72.3 -0.60 0.556 1.08
Source energy use intensity (kBtu/m2) U 1337.7 1312.1 5.6 0.13 0.893 0.46
M 1337.7 1468.7 -28.7 -411.9 -0.68 0.509 0.77
Energy Star score U 54.75 61.854 -21.0 -0.62 0.537 1.21
M 54.75 46.5 24.4 -16.1 0.51 0.615 1.54
2
Floor area (m ) U 24204 11817 83.1 2.14 0.033 0.69
M 24204 16812 49.6 40.3 1.08 0.300 0.93
Built year U 2005.5 1953.3 71.0 1.44 0.150 0.03*
M 2005.5 2006.8 -1.7 97.6 -0.15 0.881 1.89
* if variance ratio outside [0.20; 4.99] for U and [0.20; 4.99] for M
(b) For Offices in Boston.
Variable Unmatched Matched Mean %reduct |bias| t-test V(T)/V(C)
Treated Control %bias t p>t
2
Site energy use intensity (kBtu/m ) U 751.25 852.65 -34.2 -1.39 0.165 0.18*
M 751.25 741.3 3.4 90.2 0.17 0.863 0.40*
Source energy use intensity (kBtu/m2) U 2013.4 2085.2 -9.1 -0.38 0.703 0.26*
M 2013.4 1940.2 9.3 -1.8 0.48 0.636 0.58
Energy Star score U 83.103 71.887 56.3 2.40 0.017 0.31*
M 83.103 82.034 5.4 90.5 0.30 0.768 1.01
Floor area (m2) U 67357 17569 168.1 9.41 0.000 1.70
M 67357 63412 13.3 92.1 0.37 0.711 0.51
Built year U 1977.2 1935.7 73.7 3.03 0.003 0.20*
(continued on next page)

10
K. Luo et al. Resources, Conservation & Recycling 175 (2021) 105856

Table A1 (continued )
(a) For Multifamily Housing in Boston.
Variable Unmatched Matched Mean %reduct |bias| t-test V(T)/V(C)
Treated Control %bias t p>t

M 1977.2 1981.1 -6.9 90.7 -0.51 0.609 1.81


* if variance ratio outside [0.47; 2.13] for U and [0.47; 2.13] for M
(c) For Multifamily Housing in Los Angeles.
Variable Unmatched Matched Mean %reduct |bias| t-test V(T)/V(C)
Treated Control %bias t p>t
Site energy use intensity (kBtu/m2) U 404.01 462.97 -41.9 -0.54 0.591 0.10
M 404.01 421.59 -12.5 70.2 -0.46 0.669 5.73
Source energy use intensity (kBtu/m2) U 969.83 969.96 -0.0 -0.00 1.000 0.17
M 969.83 994.59 -8.4 -19397.4 -0.27 0.804 1870.28*
Energy Star score U 46.333 62.498 -46.4 -0.87 0.386 1.34
M 46.333 34.667 33.5 27.8 0.43 0.690 1.67
Floor area (m2) U 21166 12806 77.6 1.44 0.151 1.30
M 21166 13687 69.5 10.5 1.13 0.322 253.71*
Built year U 2008 1983.3 132.5 1.63 0.104 0.01*
M 2008 2006.3 9.0 93.2 1.39 0.238 12.00
* if variance ratio outside [0.03; 39.00] for U and [0.03; 39.00] for M
(d) For Offices in Los Angeles.
Variable Unmatched Matched Mean %reduct |bias| t-test V(T)/V(C)
Treated Control %bias t p>t
Site energy use intensity (kBtu/m2) U 546.33 622.27 -32.9 -1.66 0.097 0.38*
M 546.33 545.56 0.3 99.0 0.02 0.983 1.44
Source energy use intensity (kBtu/m2) U 1575.7 1733.1 -26.0 -1.37 0.173 0.50*
M 1575.7 1577.7 -0.3 98.7 -0.02 0.985 1.57
Energy Star score U 87.947 75.567 66.0 3.16 0.002 0.22*
M 87.947 87.263 3.6 94.5 0.23 0.821 0.59
2
Floor area (m ) U 67556 22576 91.7 7.85 0.000 3.85*
M 67556 65529 4.1 95.5 0.14 0.889 0.94
Built year U 1982.4 1978 24.5 1.44 0.149 0.95
M 1982.4 1986.4 -21.9 10.4 -1.23 0.222 4.22*
* if variance ratio outside [0.52; 1.92] for U and [0.52; 1.92] for M
(e) For Offices in Minneapolis.
Variable Unmatched Matched Mean %reduct |bias| t-test V(T)/V(C)
Treated Control %bias t p>t
Site energy use intensity (kBtu/m2) U 708.27 889.91 -65.2 -2.09 0.039 0.51
M 713.49 669.67 15.7 75.9 0.61 0.550 3.14*
Source energy use intensity (kBtu/m2) U 1728.7 1879.9 -29.1 -0.91 0.364 0.42
M 1733 1624.6 20.9 28.3 0.82 0.419 2.35
Energy Star score U 92.2 71.912 97.0 2.66 0.009 0.01*
M 92.071 93.071 -4.8 95.1 -0.69 0.496 0.76
2
Floor area (m ) U 74665 24068 139.4 5.73 0.000 2.12
M 69141 77653 -23.5 83.2 -0.47 0.641 0.46
Built year U 1973.3 1947.8 68.6 2.29 0.024 0.69
M 1971.9 1973.4 -4.0 94.1 -0.10 0.917 0.70
* if variance ratio outside [0.34; 2.98] for U and [0.32; 3.12] for M
(f) For Multifamily Housing in NYC.
Variable Unmatched Matched Mean %reduct |bias| t-test V(T)/V(C)
Treated Control %bias t p>t
Site energy use intensity (kBtu/m2) U 663.57 873.23 -72.0 -3.22 0.001 1.11
M 663.57 650.99 4.3 94.0 0.10 0.923 0.39
Source energy use intensity (kBtu/m2) U 1265.2 1321.4 -12.6 -0.60 0.545 1.43
M 1265.2 1151.6 25.5 -101.8 0.67 0.505 0.76
Energy Star score U 67.211 59.684 25.0 1.14 0.254 1.20
M 67.211 73.895 -22.2 11.2 -0.69 0.492 1.28
2
Floor area (m ) U 18534 11115 65.5 3.05 0.002 1.29
M 18534 22735 -37.1 43.4 -0.78 0.438 0.36*
Built year U 2001.8 1948.9 160.8 8.11 0.000 1.70
M 2001.8 2000.7 3.2 98.0 0.11 0.916 2.73*
* if variance ratio outside [0.39; 2.60] for U and [0.39; 2.60] for M
(g) For Offices in NYC.
Variable Unmatched Matched Mean %reduct |bias| t-test V(T)/V(C)
Treated Control %bias t p>t
Site energy use intensity (kBtu/m2) U 872.45 886.43 -4.9 -0.29 0.769 0.41*
M 872.45 906.08 -11.8 -140.5 -0.52 0.601 0.28*
Source energy use intensity (kBtu/m2) U 2157.7 2019.1 21.8 1.27 0.203 0.32*
M 2157.7 2283.7 -19.9 9.1 -0.78 0.438 0.16*
Energy Star score U 79.075 72.835 37.9 2.15 0.032 0.25*
M 79.075 79.811 -4.5 88.2 -0.31 0.760 0.54*
2
Floor area (m ) U 77831 26663 108.2 9.17 0.000 2.18*
M 77831 65791 25.5 76.5 1.09 0.280 0.89
Built year U 1957.3 1935.9 70.3 4.93 0.000 1.01
M 1957.3 1963 -18.6 73.5 -0.92 0.357 0.88
* if variance ratio outside [0.58; 1.73] for U and [0.58; 1.73] for M
(h) For K-12 Schools in Philadelphia.
Variable Unmatched Matched Mean %reduct |bias| t-test V(T)/V(C)
(continued on next page)

11
K. Luo et al. Resources, Conservation & Recycling 175 (2021) 105856

Table A1 (continued )
(a) For Multifamily Housing in Boston.
Variable Unmatched Matched Mean %reduct |bias| t-test V(T)/V(C)
Treated Control %bias t p>t

Treated Control %bias t p>t


Site energy use intensity (kBtu/m2) U 698.58 752.58 -23.5 -0.53 0.599 1.05
M 698.58 720.54 -9.6 59.3 -0.15 0.886 0.96
2
Source energy use intensity (kBtu/m ) U 1342 1295.2 9.9 0.26 0.797 1.80
M 1342 1518.8 -37.4 -277.5 -0.48 0.643 0.75
Energy Star score U 70 55.374 46.5 1.23 0.220 1.91
M 70 62 25.4 45.3 0.41 0.691 2.23
Floor area (m2) U 20334 10794 78.4 2.33 0.021 2.72
M 20334 16164 34.3 56.3 0.49 0.637 1.51
Built year U 1947.4 1948.3 -2.2 -0.07 0.945 2.98
M 1947.4 1975 -65.9 -2842.9 -1.06 0.319 3.54
* if variance ratio outside [0.10; 9.60] for U and [0.10; 9.60] for M
(i) For Multifamily Housing in Philadelphia.
Variable Unmatched Matched Mean %reduct |bias| t-test V(T)/V(C)
Treated Control %bias t p>t
Site energy use intensity (kBtu/m2) U 565.75 668.95 -43.7 -0.75 0.455 0.18
M 565.75 602.78 -15.7 64.1 -0.49 0.641 1.36
2
Source energy use intensity (kBtu/m ) U 1532.1 1399.5 26.0 0.55 0.583 0.82
M 1532.1 1701.1 -33.1 -27.4 -0.55 0.597 0.99
Energy Star score U 47.6 54.512 -20.4 -0.48 0.628 1.29
M 47.6 29.2 54.4 -166.2 0.91 0.390 1.71
Floor area (m2) U 12417 17535 -37.8 -0.64 0.521 0.16
M 12417 12771 -2.6 93.1 -0.09 0.927 2.62
Built year U 1991.6 1964.2 65.4 1.61 0.108 1.47
M 1991.6 1975.8 37.8 42.2 0.60 0.567 1.47
* if variance ratio outside [0.10; 9.60] for U and [0.10; 9.60] for M
(j) For Offices in Philadelphia.
Variable Unmatched Matched Mean %reduct |bias| t-test V(T)/V(C)
Treated Control %bias t p>t
Site energy use intensity (kBtu/m2) U 701.12 862.35 -57.3 -1.57 0.119 0.08*
M 682.25 668.17 5.0 91.3 0.31 0.759 0.59
2
Source energy use intensity (kBtu/m ) U 1990.7 2243.6 -34.8 -0.96 0.337 0.11*
M 1944.9 1950.6 -0.8 97.8 -0.04 0.970 0.57
Energy Star score U 83.214 67.41 83.1 2.28 0.024 0.09*
M 82.333 84.25 -10.1 87.9 -0.51 0.614 0.61
Floor area (m2) U 71979 30816 101.7 4.09 0.000 1.68
M 60595 52503 20.0 80.3 0.53 0.601 1.06
Built year U 1986.7 1952.3 110.9 3.60 0.000 0.60
M 1983 1977.9 16.4 85.2 0.47 0.642 1.15
* if variance ratio outside [0.32; 3.12] for U and [0.29; 3.47] for M
(k) For Multifamily Housing in Washington.
Variable Unmatched Matched Mean %reduct |bias| t-test V(T)/V(C)
Treated Control %bias t p>t
Site energy use intensity (kBtu/m2) U 653.01 699.78 -13.8 -0.42 0.677 2.13
M 653.01 543.58 32.2 -134.0 0.63 0.542 6.80
2
Source energy use intensity (kBtu/m ) U 1625.4 1409.8 41.8 1.26 0.208 2.13
M 1625.4 1512.3 21.9 47.6 0.39 0.702 2.73
Energy Star score U 42.667 57.149 -56.9 -1.24 0.215 0.60
M 42.667 44.167 -5.9 89.6 -0.11 0.917 0.69
Floor area (m2) U 16571 15214 12.9 0.25 0.802 0.27
M 16571 14662 18.1 -40.7 0.51 0.623 1.30
Built year U 2010.5 1965.4 213.3 3.71 0.000 0.01*
M 2010.5 2011.2 -3.2 98.5 -0.38 0.709 1.11
* if variance ratio outside [0.14; 7.15] for U and [0.14; 7.15] for M
(l) For Offices in Washington.
Variable Unmatched Matched Mean %reduct |bias| t-test V(T)/V(C)
Treated Control %bias t p>t
Site energy use intensity (kBtu/m2) U 638.81 750.67 -62.8 -5.67 0.000 0.37*
M 638.38 638.38 -0.0 100.0 -0.00 1.000 0.80
2
Source energy use intensity (kBtu/m ) U 1932.6 2206.5 -51.4 -4.60 0.000 0.32*
M 1932.6 1927.5 1.0 98.1 0.11 0.915 0.78
Energy Star score U 82.532 70.75 77.3 6.88 0.000 0.28*
M 82.457 82.486 -0.2 99.8 -0.02 0.981 0.98
Floor area (m2) U 28846 19244 54.0 5.29 0.000 1.09
M 28001 27707 1.7 96.9 0.13 0.894 0.52*
Built year U 1985.4 1973.5 49.4 4.75 0.000 0.86
M 1985.2 1985.8 -2.4 95.1 -0.24 0.808 1.91*
* if variance ratio outside [0.72; 1.39] for U and [0.72; 1.40] for M

12
K. Luo et al. Resources, Conservation & Recycling 175 (2021) 105856

Appendix B

Figure B1
Figure B2
Figure B3
Table B1
Table B2

Fig. B1. Robustness check: propensity score changes for offices at the city level before and after matching. Notes: This figure still illustrates the matching results of
robustness check–the kernel density of propensity scores for each city’s offices using matched and unmatched samples. Those with “Before matching” are kernel
density plots based on unmatched offices (2,461 records). Those with “After matching” are kernel density plots using matched samples (478 records). Solid lines
represent treatment groups (LEED offices), while dashed lines denote control groups (non-LEED offices). Each sub-graph shows the city-level discrepancy between
LEED and non-LEED offices\. Apparently, with the matching approach, selected control offices are more similar to LEED offices in building attributes than un­
matched samples.

13
K. Luo et al. Resources, Conservation & Recycling 175 (2021) 105856

Fig. B2. Check for common support assumption of robustness check by property type and city.

Fig. B3. City-level and aggregate treatment


estimation of LEED certification comparing
results of the main model and robustness
check. Notes: Treatment estimation is cate­
gorized in groups by city. Each group uses
corresponding full samples or city-level
data. The error bars show the 95% confi­
dence intervals. The solid rectangles repre­
sent the coefficients’ values that measure
the change in annual water use in response
to LEED certification. Red rectangles
represent using matched samples generated
from the main model, 524 records. Blue
rectangles represent matched samples ob­
tained from robustness check, 539 records
(with full data of regression variables).

14
K. Luo et al. Resources, Conservation & Recycling 175 (2021) 105856

Table B1
Matched samples by city using robustness check.
Property Type city Total
Boston Los Angeles Minneapolis New York City Philadelphia Washington

K-12 School 0 0 0 0 10 0 10
Multifamily Housing 14 8 0 44 10 10 86
Office 48 72 18 104 24 212 478
Total 62 80 18 148 44 222 574

Table B2
Results of robustness check.
All matched samples Boston Los Angeles Minneapolis New York City Philadelphia Washington

Treatment -0.028 0.168 -0.101 0.302 -0.018 -0.234 -0.020


(0.05) (0.14) (0.12) (0.17) (0.09) (0.16) (0.07)
Property type
Multifamily Housing 1.239*** N/A N/A N/A N/A 1.004* N/A
(0.25) N/A N/A N/A N/A (0.44) N/A
Office 0.067 -0.595 -1.603** N/A -0.820*** -0.760 -0.533
(0.24) (0.36) (0.48) N/A (0.22) (0.42) (0.29)
Site energy use intensity (kBtu/m2) -0.0004* 0.005* -0.001 -0.001 0.0002 -0.001 -0.0004
(0.0002) (0.002) (0.001) (0.003) (0.001) (0.001) (0.0006)
Source energy use intensity (kBtu/m2) 0.001*** 0.001* 0.001* 0.018 0.0003 0.001* 0.0005
(0.0001) (0.001) (0.0004) (0.01) (0.0002) (0.001) (0.0003)
Built year -0.003*** 0.002 -0.002 0.005 -0.004* -0.006 -0.003
(0.001) (0.003) (0.01) (0.003) (0.002) (0.004) (0.002)
Greenhouse gas intensity (metric ton CO2/m2) 0.001 -0.093* -0.005 -0.297 -0.010 0.002 0.0003
(0.001) (0.04) (0.03) (0.20) (0.01) (0.01) (0.002)
Energy Star scores -0.002 -0.004 0.0002 -0.022 -0.003 0.013 -0.004
(0.002) (0.01) (0.01) (0.03) (0.004) (0.01) (0.005)
Constant 6.884*** -4.052 5.686 -9.023 8.855** 11.055 6.198
(1.82) (6.47) (10.80) (6.44) (3.03) (7.36) (3.29)
Observations 539 57 76 18 136 44 208

Notes: Standard errors in parentheses

References Greer, F., Chittick, J., Jackson, E., Mack, J., Shortlidge, M., Grubert, E., 2019. Energy and
water efficiency in LEED: How well are LEED points linked to climate outcomes?
Energy Build. 195, 161–167.
Alawneh, R., Ghazali, F.E.M., Ali, H., Asif, M., 2018. Assessing the contribution of water
Grimm, K.J., Mazza, G.L., Davoudzadeh, P., 2017. Model selection in finite mixture
and energy efficiency in green buildings to achieve United Nations Sustainable
models: a k-fold cross-validation approach. Struct. Equ. Model. 24 (2), 246–256.
Development Goals in Jordan. Build. Environ. 146, 119–132.
Gunay, H.B., O’Brien, W., Beausoleil-Morrison, I., 2013. A critical review of observation
Ali, H.H., Nsairat, Al, S., F, 2009. Developing a green building assessment tool for
studies, modeling, and simulation of adaptive occupant behaviors in offices. Build.
developing countries–case of Jordan. Build. Environ. 44 (5), 1053–1064.
Environ. 70, 31–47.
Al-Qawasmi, J., Asif, M., El Fattah, A.A., Babsail, M.O., 2019. Water efficiency and
Hair, J., Black, W.C., Babin, B.J., Anderson, R.E., 2010. Multivariate Data Analysis (7th
management in sustainable building rating systems: examining variation in criteria
ed.). Pearson Education International, Upper saddle River, New Jersey.
usage. Sustainability 11 (8), 2416.
Horvat, M., Fazio, P., 2005. Comparative review of existing certification programs and
Attiya, E.M., Shebl, M.A., Nasser, M.M., 2020. A comparative analysis of LEED and GPRS
performance assessment tools for residential buildings. Architect. Sci. Rev. 48 (1),
for the applicability in Egyptian office buildings. Int. J. Eng. Res. Technol. 9 (3).
69–80.
Awadh, O., 2017. Sustainability and green building rating systems: LEED, BREEAM,
Kohavi, R., 1995. A Study of Cross-Validation and Bootstrap for Accuracy Estimation and
GSAS and Estidama critical analysis. J. Build. Eng. 11, 25–29.
Model Selection. Ijcai, Montreal, QC, Canada.
Bordass, B. (2004). Energy performance of non-domestic buildings: closing the
Liang, J., Qiu, Y., Hu, M., 2019. Mind the energy performance gap: Evidence from green
credibility gap. In Proceedings of the 2004 Improving Energy Efficiency of
commercial buildings. Resour. Conserv. Recycl. 141, 364–377.
Commercial Buildings Conference.
Long, M., 2013. LEED v4, the Newest Version of LEED Green Building Program Launches
Chokor, A., El Asmar, M., 2017. Data-driven approach to investigate the energy
at USGBC’s Annual Greenbuild Conference. United States Green Building Council
consumption of LEED-certified research buildings in climate zone 2B. J. Energy Eng.
Media.
143 (2), 05016006.
Matisoff, D.C., Noonan, D.S., Mazzolini, A.M., 2014. Performance or marketing benefits?
Copeland, C., & Carter, N. T. (2014). Energy-water nexus: The water sector’s energy use.
The case of LEED certification. Environ. Sci. Technol. 48 (3), 2001–2007.
D’Agostino Jr., R.B, 1998. Propensity score methods for bias reduction in the comparison
Menezes, A.C., Cripps, A., Bouchlaghem, D., Buswell, R., 2012. Predicted vs. actual
of a treatment to a non-randomized control group. Stat. Med. 17 (19), 2265–2281.
energy performance of non-domestic buildings: using post-occupancy evaluation
De Wilde, P., 2014. The gap between predicted and measured energy performance of
data to reduce the performance gap. Appl. Energy 97, 355–364.
buildings: A framework for investigation. Autom. Constr. 41, 40–49.
Movahedi, A., & Derrible, S. (2020). Interrelated Patterns of Electricity, Gas, and Water
Dieter, C.A., Maupin, M.A., Caldwell, R.R., Harris, M.A., Ivahnenko, T.I., Lovelace, J.K.,
Consumption in Large-Scale Buildings. 10.31224/osf.io/ahn3e.
Barber, N.L., and Linsey, K.S. (2018). Estimated use of water in the United States in
Newsham, G.R., Mancini, S., Birt, B.J., 2009. Do LEED-certified buildings save energy?
2015: U.S. Geological Survey Circular 1441, 65 p., 10.3133/cir1441.
Yes, but…. Energy Build. 41 (8), 897–905.
Doan, D.T., Ghaffarianhoseini, A., Naismith, N., Zhang, T., Ghaffarianhoseini, A.,
... Nguyen, X.C., Nguyen, T.T.H., La, D.D., Kumar, G., Rene, E.R., Nguyen, D.D.,
Tookey, J., 2017. A critical comparison of green building rating systems. Build.
Nguyen, V.K., 2021. Development of machine learning-based models to forecast
Environ. 123, 243–260.
solid waste generation in residential areas: a case study from Vietnam Resour.
D’Oca, S., Corgnati, S., Hong, T., 2015. Data mining of occupant behavior in office
Conserv. Recycl. 167, 105381.
buildings. Energy Procedia 78, 585–590.
Owens, B., Macken, C., Rohloff, A., Rosenberg, H., 2013. LEED V4 Impact Category and
Elkhapery, B., Kianmehr, P., Doczy, R., 2021. Benefits of retrofitting school buildings in
Point Allocation Development Process. U.S. Green Building Council. Retrieved, 2
accordance to LEED v4. J. Build. Eng. 33, 101798.
(01).
Fedoruk, L.E., Cole, R.J., Robinson, J.B., Cayuela, A., 2015. Learning from failure:
Pushkar, S., 2018. The effect of regional priority points on the performance of LEED 2009
understanding the anticipated–achieved building energy performance gap. Build.
certified buildings in Turkey, Spain, and Italy. Sustainability 10 (10), 3364.
Res. Inf. 43 (6), 750–763.
Qiu, Y., Kahn, M.E., 2019. Impact of voluntary green certification on building energy
Gatignon, H., 2013. Identification of Multicollinearity: VIF and Condition Number.
performance. Energy Econ. 80, 461–475.
Retrieved from. https://faculty.insead.edu/hubert-gatignon/documents/HG%20Not
Richter, B., Goldston, D., Crabtree, G., Glicksman, L., Goldstein, D., Greene, D.,
es-Identification%20of%20Multicollinearity-VIF%20and%20Conditioning%
Kammen, D., Levine, M., Lubell, M., Savitz, M., Sperling, D., Schlachter, F.,
20Number_20140304.pdf.
Scofield, J., Dawson, J., 2008. Energy Future: Think Efficiency. American Physical
Society, College Park, MD. Retrieved from. http://www.aps.org/energyefficiencyre
port/index.cfm.

15
K. Luo et al. Resources, Conservation & Recycling 175 (2021) 105856

Roderick, Y., McEwan, D., Wheatley, C., & Alonso, C. (2009, July). Comparison of energy USGBC, 2005. LEED-NC Green Building Rating System for New Construction & Major
performance assessment between LEED, BREEAM and Green Star. In Eleventh Renovations Version 2.2.
International IBPSA Conference (pp. 27-30). USGBC, 2010. LEED 2009 for New Construction and Major Renovations Rating System.
Rothausen, S.G., Conway, D., 2011. Greenhouse-gas emissions from energy use in the USGBC, 2020. LEED Project Database. https://www.usgbc.org/projects. accessed
water sector. Nat. Clim. Change 1 (4), 210–219. November 22.
Scofield, J.H., 2009. Do LEED-certified buildings save energy? Not really…. Energy Warren, P.R., Parkins, L.M., 1984. Window-opening behaviour in office buildings. Build.
Build. 41 (12), 1386–1390. Serv. Eng. Res. Technol. 5 (3), 89–101.
Shaban, W.M., Yang, J., Elbaz, K., Xie, J., Li, L., 2021. Fuzzy-metaheuristic ensembles for Wu, P., Song, Y., Shou, W., Chi, H., Chong, H.Y., Sutrisna, M., 2017. A comprehensive
predicting the compressive strength of brick aggregate concrete. Resour. Conserv. analysis of the credits obtained by LEED 2009 certified green buildings. Renew.
Recycl. 169, 105443. Sustain. Energy Rev. 68, 370–379.
... Shah, M.I., Amin, M.N., Khan, K., Niazi, M.S.K., Aslam, F., Alyousef, R., Mosavi, A., Wu, Z., Wu, Z., Li, H., Zhang, X., Jiang, M., 2020. Developing a strategic framework for
2021. Performance evaluation of soft computing for modeling the strength adopting water-saving measures in construction projects. Environ. Geochem. Health
properties of waste substitute green concrete Sustainability 13 (5), 2867. 42 (3), 955–968.
Suh, S., Tomar, S., Leighton, M., Kneifel, J., 2014. Environmental performance of green Zhou, W.H., Garg, A., Garg, A., 2016. Study of the volumetric water content based on
building code and certification systems. Environ. Sci. Technol. 48 (5), 2551–2560. density, suction and initial water content. Measurement 94, 531–537.
Suzer, O., 2015. A comparative review of environmental concern prioritization: LEED vs
other major certification systems. J. Environ. Manage. 154, 266–283.

16

You might also like