Successful greenhouse gas mitigation in existing Australian office buildings

Frequent site energy consumption auditing is a potential strategy to mitigate greenhouse gas (GHG) emissions from existing buildings. Such a strategy has been practised in Australia for nearly 15 years. This paper documents and analyses the effect of repetitive audits on measured site energy consumption. Using a self-constructed database of over 3500 audited disclosures representing over 800 unique office buildings, empirical models demonstrate that measured site energy consumption declines, on average, over the first five re-certification periods. The results also suggest a market average post-certification equilibrium in Australia of approximately 430 MJ/m2/year (120 kWh/m2/year) within approximately six years, if all else – including green management strategy – is held constant. Since GHG emissions from buildings in Australia are highly correlated with site energy consumption, such a result is comparable with meeting 50-year GHG mitigation targets reliant on the implementation of existing technologies. This suggests that repetitive auditing is a successful approach for motivating owners to invest in existing energy efficiency technologies.


Introduction
Greenhouse gas (GHG) mitigation in urban environments is sensitive to energy consumption in existing buildings because many existing buildings predate the earliest forms of energy efficiency regulation in statutory building codes. These buildings can last a long time. Building stock replacement rates in developed countries range from 0.66% to 3% per year (Eichholtz, Kok, & Quigley, 2010;Holness, 2008;Jowsey & Grant, 2009;United Nations Environment Programme (UNEP), 2007), meaning that a complete transition to current building code energy performance standards could take somewhere between 30 and 130 years if the potential to retrofit existing buildings is ignored. 1 As such, simulations of GHG emissions for an entire building stock conclude that existing buildings have a disproportionate effect on the total (Coffey et al., 2009;Kohler & Hassler 2012;Seo & Foliente, 2011).
In light of the importance of existing buildings to urban GHG mitigation, the framework of intervention in property markets needs reconsideration because energy efficiency, a popular strategy for GHG mitigation, requires attention to design and to user behaviour. Green building assessment traditionally began as a means to segment the market for new construction (Crawley & Aho, 1999). Public policymakers found emerging market differentiation helpful to assess compliance with various statutes, policies and standards designed to promote building thermal efficiency (Cole, 1999;Kontokosta, 2011;Simons, Choi, & Simons, 2009). With regard to the process of market differentiation, the property industry has generally been interested in dissociating the environmental performance potential of design with the observed performance in-use. But scholars (writing in this journal and elsewhere) have long recognized the interdependency of design and user behaviour with regard to environmentally beneficially outcomes, particularly energy efficiency and GHG mitigation (Bordass, Leaman, & Ruyssevelt, 2001;Schweber & Leiringer, 2012). By implication, successful urban GHG mitigation policies must implement socio-technical solutions that optimize design and usability.
One potential policy is building energy performance disclosure, which has caught the attention of policymakers seeking to mitigate GHG emissions in existing buildings. The policy problem is that, with the exception of major redevelopment activities, existing buildings are not typically subject to revisions of building codes. Market uptake of voluntary energy consumption disclosure schemes has opened up the intervention of mandatory disclosure as a potential solution to the problem (Kontokosta, 2013). In 2010, the federal Australian government mandated measured energy consumption disclosure in private sector non-domestic buildings via the Building Energy Efficiency Disclosure (BEED) Act. 2 Similar measured performance rating disclosure laws have since been enacted at the local and state level in the United States using the Energy Star methodology (Hsu, 2014;Kontokosta, 2013). Will a mandate to disclose measured energy performance have an effect on operational energy consumption from the existing building stock?
To understand better existing building intervention, this study contributes the first model of a shift in energy consumption outcomes purely as a result of repetitive energy performance disclosure schemes in existing non-domestic buildings. In markets where there is a strong relationship between energy consumption and GHG emissions, such as Australian cities that predominantly consume electricity in non-domestic buildings, energy consumption acts as an effective proxy for the management of GHG emissions. This study also addresses a lack of benchmark data on non-domestic building energy performance by showing how the implementation of repetitive performance disclosure creates a more unbiased benchmark: prior audit results from the same property.
The core research task is estimation of an effect per building as a result of participation in National Australian Built Environment Rating System (NABERS) Energy, the Australian site energy consumption and GHG emission disclosure scheme. Maintaining NABERS Energy certification requires annual renewal with a fresh third-party site energy consumption audit. The result is that hundreds of buildings across Australia have undergone repetitive certification. The question of calculating the effect on site energy consumption per building will be answered with a set of statistical models examining site energy consumption as a function the depth of participation in NABERS Energy (measured as the number of certificates obtained by a particular building).

Background
The objective of rating tools is to create product differentiation and thereby encourage private sector innovation. In a paper on the general practice of using ratings to segment markets, Chatterji and Toffel (2010) argue that firms will adapt their practices in order to improve external ratings, particularly if the market perceives them to be rated poorly. Fuerst and McAllister (2011) and Kontokosta (2013, p. 35) apply this general theory of behaviour change through differentiation to the property (real estate) sector, with the latter arguing that 'the potential for energy disclosure policies to shift market awareness of building energy efficiency is substantial'. By increasing market awareness and enabling differentiation, rating systems create a market for energy efficient buildings. As evidence of this market, Warren-Myers (2012) reviews a wide range of literature arguing that building energy efficiency creates value in real estate, leading to theories that energy efficiency is positively associated with measures of building value. 3 Borck and Coglianese (2009) review the general environmental management literature on product differentiation and produce a helpful framework to understand the outcome as measured by environmental performance metrics. The effectiveness of environmental differentiation is defined as follows: where 'effect per participant' measures the average environmental performance outcome for the market segment that participates in a rating scheme. For building energy consumption, effect per participant can be measured by the quantity of energy saved as a result of participating in a disclosure scheme. The 'spillover effect' represents the influence of participants on the behaviour of non-participants, such as the development of new energy efficient technologies by participants that diffuse throughout the entire sector. Spillover effects can vary depending on the context of the intervention, the accounting framework adopted and on the assumptions of the researcher. These choices are complex and therefore this paper will only consider the direct effect of intervention. 4 The literature implicitly assumes that energy-efficient buildings are the outcome of a robust certification process (for examples, see Eichholtz et al., 2010;Fuerst & McAllister, 2011;and Miller, Spivey, & Florance, 2008). Using the formula above, the implication is that the participation rate in a certification scheme determines the effectiveness of a voluntary certification programme. Hence Fuerst (2009) andQuigley (2013) attribute increasing participation in voluntary green building assessments as evidence of their effectiveness in mitigating energy-related GHG emissions emitted by the building sector.
The argument that participation is the critical metric has spawned interest in the determinants of voluntary Successful GHG mitigation in o⁄ce buildings participation. Kok, McGraw, and Quigley (2012) and Fuerst, Kontokosta, and McAllister (2014) argue that higher incomes and traditional indicators of a 'healthy' property market (i.e. low vacancy rates and high capital values) are positively associated with the local adoption of energy efficient technologies. Kok, McGraw et al. (2012) also argue that energy prices influence the diffusion of buildings certified as energy efficient, while Fuerst et al. (2014) find only limited support that public policies promoting certification (not mandating) affect adoption.
However, knowing how to affect participation rates is only useful if effect per participant is non-zero. The mechanism that leads to the possibility of zero effect per participant is the widespread use of asset ratings, which simulate potential performance as opposed to measuring actual performance. Asset ratings are preferred by architects, engineers and consultants in the property development sector because their objective is to isolate the effect of decisions made in design by excluding variation caused by human factors in building operation and management. Unsurprisingly, as evidence of the potential gap between building design and usability, there are many detailed case studies of individual buildings that underperform expectations for energy efficiency once human factors are introduced (for examples, see Bordass et al., 2001;Gabe, 2008;and Scofield, 2002).
But is this underperformance systematic? Two studies of green-certified buildings in the United States argue that while variance at the individual building level is high, certified green buildings are more energy efficient than comparable uncertified buildings on average (Fowler, Rauch, Henderson, & Kora, 2011;Turner & Frankel, 2008). Data on energy consumption from the Turner and Frankel study have been subjected to additional statistical analyses, one of which confirms the original conclusion (Newsham, Mancini, & Birt, 2009) while the other finds evidence of systematic underperformance in large buildings (Scofield, 2009). Oates and Sullivan (2012) gathered data from 19 office buildings in Arizona, finding that all but one building underperformed its asset rating and 15 buildings underperformed the baseline code specification for energy efficiency.
These three studies are largely the extent of empirical knowledge on the systematic effect of green building certifications on measured building energy consumption. Data availability is one reason for this lack of research; one of the studies reports that very few green buildings actually measure energy consumption post-occupancy (Oates & Sullivan, 2012). All three assess the impact of a rating system designed for new construction, leaving no empirical evidence with regard to the effect of existing building interventions. Additionally, while the causal path between green building certification and energy efficiency is strong (Newsham et al., 2009), it is not guaranteed, since building owners can qualify for Leadership in Energy and Environmental Design (LEED) certification in the United States and similar new construction green building schemes elsewhere with minimal investment in energy efficiency because of the elective nature of that certification process.
A knowledge gap that may be responsible for the lack of study into the effect of interventions on existing building energy consumption is the absence of benchmark data. Newsham et al. (2009) uses a five-yearold survey of commercial building energy consumption across the United States to extract statistically the most comparable non-certified building for each LEED building in the Turner and Frankel (2008) dataset. Both Turner and Frankel (2008) and Oates and Sullivan (2012) attempt to compare a simulated asset rating with post-occupancy measured performance, but discuss how the simulated asset ratings are not meant to measure total consumption. 5 This journal recently devoted a special issue looking at the challenge of measuring, understanding and improving energy performance in existing non-domestic buildings (Isaacs & Steadman, 2014).
The next section briefly examines the NABERS Energy assessment system for readers unfamiliar with its use in Australia. This is followed by a discussion of data collection, the methods used to construct the quantitative models and the estimation of these models.

NABERS Energy
The NABERS Energy scheme was developed by the New South Wales state government and has been used across Australia since 1999 for building owners choosing to advertise their energy performance credentials. Prior to 2008, NABERS Energy was branded as the Australian Building Greenhouse Rating (ABGR), but no change in the underlying audit method has occurred. 6 Third-party auditors assess 12 months of site energy consumption and produce a rating from 0 to 6 stars based on GHG emissions resulting from that measured site energy consumption. 7 Star ratings are calibrated regionally such that a 2.5-star rating is assigned to a building with average GHG emissions from site energy consumption in each metropolitan area. Certificates are freely available via the programme website. 8 To control for vacancy rates, the effective size of the building being rated is reduced by a pro-rata calculation of occupancy for the year being audited. Readers interested in a thorough description of the NABERS Energy auditing process are encouraged to consult the Department of Environment Climate Change and Water NSW (2010).
The boundaries of the site energy consumption audit are typically 'Base Building' services, which exclude tenant power consumption (computers and plug-in appliances). Included in the Base Building rating are common area lighting, space conditioning, hot water production and all common area power consumption. By design, Base Building ratings conveniently mimic the boundaries of energy and GHG costs paid by the party liable for operating expenses in an office lease contract.
Owners wishing to improve NABERS Energy ratings have three options: invest in on-site operational energy efficiency, purchase Green Power offsets for electricity consumption or fuel-switch to maintain site energy consumption while reducing source GHG emissions. This paper measures the first optioninvestment in on-site operational efficiency -which is the most common approach. Green Power is a national Australian scheme administered by the federal government that allows an electricity consumer to pay a rate premium for electricity that goes to renewable energy producers in exchange for certification that the consumer's electricity was generated by zero-emission renewable energy. 9 When an owner elects to purchase Green Power to improve his NABERS Energy rating, the certificate includes star ratings with and without the Green Power purchase. Because Green Power offsets must be excluded from mandatory disclosure under the BEED Act, it is not a common option.
Fuel switching is also rare. The correlation from first to final certification of the ratio of GHG per unit of site energy is above 0.9, which most likely reflects minor variability in electricity production. In Australia, electricity is the source energy used for nearly 85% of total commercial building energy demand (Department of Climate Change and Energy Efficiency, 2012).

Data
Information on building energy consumption is extracted directly from a comprehensive collection of ABGR and NABERS certificates gathered by the author over the past 14 years. All publicly available certificates using the Base Building scope described above have been obtained from the certification agency via the internet since the commencement of ABGR in 1999 up until April 2012. Additional NABERS Energy certificates issued between April 2012 and the end of October 2013 were obtained for buildings complying with BEED Act disclosure regulations from the regulator's website. 10 Thus, the full NABERS Energy dataset spans nearly all certifications between August 1999 and October 2013.
Multiple certificates for the same building with the same expiry date are removed to eliminate duplicates, with the chosen certificate having the highest NABERS ID Number (a proxy for the issue date). A small number of certificates (71) are removed owing to missing data that clearly identify the certified building. In total, the cleaned dataset contains 3661 unique NABERS Energy certificates. The certificates are then organized in issue sequence for each individual building in the database based on ascending NABERS ID numbers. Table 1 shows there are 1153 unique buildings in the energy dataset, with 818 having been certified at least twice.

Energy consumption and GHG emissions
To describe environmental performance data captured in an audit, each certificate includes a disclosure of overall building energy use intensity (EUI) (MJ/m 2 / year) that is unaffected by Green Power offsets. The only accounting adjustment is that energy sourced from on-site zero-emission sources are excluded from the disclosed EUI. To ensure valid comparisons, energy performance in this study is measured using these raw consumption data. EUI has been consistently produced on every certificate using a static methodology. Star ratings are calibrated separately for each Australian city and thus unsuitable for comparison across regions.
Despite GHG mitigation being a key objective for investment in operational building energy efficiency, emission figures from NABERS Energy certificates are not used in this study. Early NABERS Energy certificates only report emissions that take Green Power offsets into account, leading to a number of 'zeroemission' buildings. Later certificates switch between  Successful GHG mitigation in o⁄ce buildings different GHG accounting protocols. This leads to the discard of many valid certificates in an attempt to compare data only within the same accounting framework. Finally, over 14 years, GHG accounting has been very dynamic; even if accounting scopes were consistent, conversion factors between the raw data and GHG emissions are unknown and have varied over time. The non-disclosure of energy fuel sources for each building further complicates the ability to compare GHG emission totals. Thus, it would be difficult to differentiate trends in NABERS Energy-reported GHG emissions between operational management and accounting changes. As indicated above, EUI is highly correlated with GHG emissions in Australia.
However, it is possible to measure the effect of the Green Power offset purchasing decision on operational efficiency. All building owners electing to purchase over 1% of their electricity via the Green Power scheme in every NABERS re-certification are identified using a binary variable. This variable will enable the model to differentiate whether Green Power offsets act as a substitute or complement to operational energy efficiency.

Green management strategy
A second binary variable identifies 'green owners', which will be defined as building owners who are explicitly differentiating their assets as green or sustainable in the commercial property market. In the establishment of a benchmark rating system measuring the depth of green strategy present in global property investment firms, Bauer, Eichholtz, Kok, and Quigley (2011) rated three Australian-based firms -Stockland, GPT and the Commonwealth Property Office Fundas three of the top five 'global environmental leaders' for publicly listed property companies. In addition, GPT and a fourth Australian-based firm, Investa, were identified as the top two global environmental leaders for private property holding companies. Buildings owned by these four firms are considered to have 'green owners'.

Location
The process of assigning NABERS certificates to an individual building makes it possible to generate variables based on location. In particular, Australian four-digit postcodes convey two useful pieces of data. One is the state or territory in which each building is located. This is important because Australia has three distinct levels of government -federal, state and local -and certain states, including New South Wales, the Australian Capital Territory, Western Australia and Victoria, were early supporters of NABERS Energy. Hence, the particular state location can proxy fixed state effects, such as local government policies and climate, that may influence the decision to pursue operational resource efficiency.
The second useful variable that can be generated from a postcode is whether or not a building is located in a capital city central business district (CBD). Office markets in a CBD offer prospective tenants greater choice than smaller provincial or suburban centres. Competition between owners may lead to greater investment in resource efficiency in major cities as part of an asset positioning strategy. Postcodes are used to identify buildings located in each capital city CBD: 0800 for Darwin, 2000 for Sydney, 2601 for Canberra, 3000 for Melbourne, 4000 for Brisbane, 5000 for Adelaide, 6000 for Perth and 7000 for Hobart.

Building size
The existing literature identified that building energy consumption, when normalized by building area, may be affected by building size (Scofield, 2009). The net lettable area (NLA) of each multi-certified office building was obtained from property reports and owner disclosures. A consistent measure of NLA could not be obtained for 12 of the 818 multi-certified buildings, so these 12 are omitted from further analysis. 11

Other characteristics
Data on other hedonic characteristics of all the buildings in this dataset, such as building age, were not available on a consistently measured scale for all buildings and is therefore omitted from the analysis in this paper to avoid further omissions of observed buildings. To give readers a brief illustration of building age distributions in NABERS Energy-certified buildings, the author has excellent data on building age in one market, central Sydney, which contributes 119 of the 818 multi-certified office buildings to the dataset. In 2012, these 119 buildings have an average age of 27.7 years (median of 24), with a range between 3 and 76 years and standard deviation of 16.2 years. Anecdotal evidence gathered by the author suggests similar distributions in other cities. Buildings are not eligible for NABERS Energy certification until they have been in operation for 24 months.

Descriptive statistics
Table 2 provides a descriptive overview of the entire energy dataset with analysis by number of certificates obtained. To ensure sufficient sample sizes, the number of multiple certificates is capped at eight. This means that 31 buildings with more than eight NABERS Energy certificates are not analysed beyond their eighth certificate. Note that the aggregate column on the far right only includes multi-certified buildings; the column of buildings with only one NABERS Energy certificate is excluded from the totals. The descriptive statistics suggest that repetitive participation in NABERS Energy is associated with a measurable improvement in building energy efficiency on average. Mean energy consumption indicators decrease between a building's initial certification and its final certification. Population variance also decreases from initial certification to final certification.
A key variable of interest, change in energy consumption, shows a clear trend of increasing energy savings over time and a decrease in variance. Box plots in Figure 1 demonstrate the reduction in energy consumption and variance, particularly the reduction of outliers, as the number of certifications increase. These box plots also suggest that after five certifications, mean energy consumption begins to stabilize while variance continues to decrease. The robustness of these observations will be subjected to further tests.
Besides change in consumption, four variables are also associated with the number of certifications. As would be expected, the number of certificates earned is related to the year a building first sought assessment; early adopters are the only buildings with the highest numbers of certifications. Second, the percentage of buildings managed by green owners increases as the number of re-certifications increase. Unsurprisingly, this means green owners are likely to be early adopters of NABERS. Finally, there are also positive associations between building NLA, the percentage of buildings in a CBD and the number of certifications. Table 3 presents the correlation matrix between these related variables. The strongest correlation is between the year of entry into NABERS and the number of certificates. There are cross-correlations between green ownership, building size and the number of certificates. Green owners begin NABERS certification early and are likely to own large properties. Hence the interpretation of the green owner variable needs caution because it could be measuring green strategy as intended, or it could represent unmeasured characteristics of large institutional property owners, such as greater access to capital or the involvement of professional property managers. In this study, number of certifications (depth of participation) is the key variable of interest, so the exact interpretation of the green owner variable is not important.
The strong correlation between year of certification and number of certificates, however, does lead to a potential interpretation problem. The improvement as depth of participation increases seen in Figure 1 could be an artefact of fixed time effects. To control for fixed time effects, it is necessary to fix the number of certificates. Figure 2 demonstrates that if change in EUI is captured at every intermediate certification, central tendencies and variance follow a similar pattern as if change in EUI is only captured at the final certification (Figure 1).
Although time between certificates is not uniform, it is often slightly more than a year. Looking at Table 2, the median number of days between certificates is typically around 400 days, indicating that most buildings are recertified soon after an existing certificate expires. Nevertheless, it will be necessary for a multivariate  (Table S1), each additional certificate produces a consistent 5-7% reduction in EUI on average. After the sixth certificate, the marginal decline is not statistically significant at conventional levels. One interesting observation is the similarity in final energy consumption between the sixth period of re-certification (sequence 6 to 7) and the seventh period of re-certification (sequence 7 to 8). The market appears to settle on a limit in average office building EUI of approximately 430 MJ/m 2 /year (120 kWh/m 2 /year).

Successful GHG mitigation in o⁄ce buildings
The cumulative tests (Table S2) shed some light on why eight-certificate buildings take longer to reach this equilibrium. Mean initial consumption rises as the depth of participation increases. This leads to two potential interpretations: late adopters may be more energy efficient at the time of NABERS entry as a result of spillover effects or a subsample bias could be responsible for inflating the average benchmark EUI in early adopters.

Multivariate analysis
Multiple regression models are run to include other exogenous variables -location, capacity for improvement and willingness to purchase Green Power offsets -in order to test the robustness of the descriptive statistics above.
Consider building j, in which EUI has been observed in the dataset from period s ¼ 1, the initial benchmark certificate, to period s ¼ max j , representing the most recent re-certification for building j. Subtracting the energy consumption benchmark from the energy consumption of the most recent re-certification gives the total change in EUI, DP j , which will be the dependent variable: P j(s=max) would be an equivalent choice for the dependent variable in the multiple regression model. With regard to the variable of interest -depth of involvement in NABERS Energy -the two approaches will give identical independent variable coefficients because DP j is a linear transformation of P j(s=max) . But with regard to the overall model, the two approaches model slightly different outcomes. Modelling P j (s=max) is less interesting than DP j because P j (s=1) and P j (s=max) are highly correlated (coefficient ¼ 0.737), so while the overall explanatory power of the model will be high when P j (s=1) is inserted as an independent variable, much of the explanation is due to the uninteresting correlation between initial and final energy consumption. What is more interesting is how strongly one can explain the change in energy consumption using a variety of independent control variables in addition to the depth of NABERS Energy certification. Hence DP j is chosen as the dependent variable.
The multiple regression model takes the form: where a is an intercept; b is an estimated coefficient; and 1 represents stochastic error. These three variables are estimated with ordinary least squares regression using the observed dataset of 806 multi-certified buildings. Each independent variable or vector variable (the latter indicated by bold typeface) in the observed dataset measures: LOC j ¼ fixed effects associated with the location of building j NLA j ¼ size of building j (net lettable area) CAP j ¼ capacity for building j to improve its energy performance OWN j ¼ fixed characteristics associated with the owner of building j AVGDAYS j ¼ average days between certificates for building j

CERT j ¼ depth of participation by building j in NABERS Energy (number of certificates)
A list of all variables contained within each independent vector variable is shown in Table S3. Note that the year of certification is not included in this model because of its high correlation with the variable of interest. Later, a test is run to assess the impact of omitting this variable.
The independent variable of interest is the depth of NABERS Energy participation, measured using a flexible functional form in the vector variable CERT j . The number of certificates obtained proxies the depth of participation. Since diminishing returns to performance outcomes are expected as the number of re-certification periods s increases, the vector variable includes a series of binary variables measuring depth of participation. If a building has obtained s certificates, then it assumes a value of 1 for the s-certificate variable and 0 for the remaining CERT j variables. Note that interpretation of this specification is similar to the cumulative univariate t-test -the coefficient of each variable measures the cumulative influence of s certifications, not the marginal influence. The two-certificate variable is omitted from all specifications as a reference category.
The capacity for improvement, CAP j , is a critical control variable given the use of an initial certificate to benchmark energy performance. The best variable to estimate capacity to improve is this initial EUI benchmark (P j s=1 ). An alternative measure for capacity to improve the star rating from the initial NABERS certificate was considered, but the continuous distribution of initial EUI was found to differentiate potential better than the categorical measure of initial star ratings. All else being equal, buildings with high initial EUI are expected to improve more than those with low initial EUI. Another reason to include this variable is to relax the assumption that an initial NABERS certificate is acceptable as a pre-intervention benchmark. Buildings with low initial EUI are more likely to have invested in operational energy efficiency prior to certification, so including this variable allows the model to control for this.
In OWN j , green ownership will be assessed as an interaction between the green owner binary variable and the variable for improvement capacity. The interacted variable allows for a difference in building energy improvement potential between green and non-green owners. It can be interpreted as the excess energy savings that a green owner will pursue beyond that which a normal owner would pursue having accounted for what both groups would pursue given an initial energy consumption benchmark. The alternate specification where green ownership is only included as a binary variable does not produce as good of a fit to the observed data.

Test for ¢xed time e¡ects
High correlation between year of the benchmark NABERS Energy certificate and the number of total NABERS Energy certificates obtained by each building means that year of entry cannot be reliably controlled for in the specified model because of concerns with collinearity. But there is a chance that the depth of participation variable is acting as an instrument for fixed time effects. To test for the influence of fixed time effects, three additional models will be run that fix the number of certificates earned. Fixed time effects become the variables of interest in these models.
Models with the number of re-certifications fixed are constructed using a similar approach as described in the previous section, with DP j as the dependent variable. However, instead of measuring DP j using P j(s=max) for each building j, these models will measure DP j at the fixed period u of NABERS Energy certification: Additional models are not run for the fifth through eighth certificates because of diminished sample sizes.
Each model investigating fixed time effects follows the general specification similar to equation (3): Ordinary least squares regression is used to estimate the intercept, coefficients and stochastic error. The control variables LOC j , AST j , CAP j and OWN j are identical to those described in the first multiple regression model.

Results
The estimation of the multiple regression model (equation 3) is presented in Table 4. The dependent variable, change in EUI, is negative if a building reduces its energy consumption, so the significant negative coefficient as the number of certificates increase reveals a strong association between depth of participation in NABERS Energy and energy efficiency outcomes.
The univariate t-test and multivariate results are in good agreement, with the multiple regression model showing a slightly lower reduction in energy consumption as a function of participation depth. Figure 3 compares the univariate and multivariate results. 12 The one notable deviation from agreement begins at the sixth Figure 3 Comparison between the t-test and multivariate results

Successful GHG mitigation in o⁄ce buildings
certificate. The multivariate analysis shows little change in energy savings after the sixth certificate while the univariate trend continues to show a decrease in energy consumption. This divergence suggests that the multivariate approach is better able to attribute any additional increase in energy savings past the sixth certificate to other factors besides participation in NABERS Energy, such as green ownership characteristics. The multivariate curve supports the conclusion that post-certification equilibrium is reached approximately five to six years after a building enters NABERS Energy.
The control variables in Table 4 confirm a number of expected outcomes. A building with higher initial EUI is likely to experience higher energy savings. The interaction between green ownership and initial EUI reveals that buildings with a green management strategy are successful in reducing energy consumption beyond the model's expectation for an average owner. Buildings with owners that purchase Green Power regularly have significantly more energy consumption reductions in all specifications. This observation suggests that Green Power offsets are used to complement operational energy efficiency, not substitute for it.
The locational controls in this model reveal that, in nearly all cases, operational energy efficiency outcomes are not significantly influenced by unobserved factors unique to a building's location, such as local policies, interstate economic differentiation, or climate. Suburban Queensland appears to be one exception that could perhaps be understood through a unique economic boom occurring in the state at the time of this study. During this period, vacancy rates in suburban Brisbane office space were unusually low (1-2%), so high demand for suburban buildings means there was little scope for tenants to demand energy conservation and likewise, little need for owners and developers to invest in energy efficiency.
The only other control variable that has an effect on change in energy consumption is the average number of days between certificates. But it is in an unusual direction: buildings with more time between certificates reduce energy consumption less than buildings with fewer days between certificates. This suggests that the variable may be acting as in instrument for future expectations; owners committed to undertake regular NABERS Energy audits are more concerned with energy efficiency than owners that certify infrequently. Hence it can be concluded that expectations of future audits are an important component of building energy efficiency.
There are some potential missing variables in the multiple regression model. Data availability limits the scope of independent variables measured systematically for the entire population. The author has run multiple regression models on subsamples of the data that include average hours of occupation, building service quality ratings (Premium, A, B and C grade), and building age. The only variables that added any further value to the model were binary variables representing lower building service quality ratings (B and C grades) and building age as a continuous variable. Low service quality was associated with less interest in energy efficiency relative to high service quality. This supports the literature arguing that energy efficiency is a 'luxury good' (Fuerst et al., 2014). The coefficient for building age in these subsamples is negative, suggesting that older buildings are more likely to improve energy consumption than newer buildings. Much of this age effect is likely captured in the variable for the capacity to improve (initial EUI). What the addition of an age variable likely contributes is the fact that older buildings are more likely to be scheduled for major renovations and there is usually a wider scope for investment in energy efficiency in these cases (e.g. replacing an entire mechanical conditioning system).
The model explains just over one-third of variability in energy efficiency improvements. Keeping in mind that energy savings are not automatically correlated with investment -many buildings with large investments in energy efficiency fail to perform efficiently in practice (Newsham et al., 2009) -the explanatory power of this model is comparable to estimations of factors that influence the construction of new energy efficient buildings (Fuerst et al., 2014;Kok, McGraw, et al., 2012).
However, there is a concern that number of certificates is measuring a fixed time effect as opposed to depth of participation. This concern was tested using the fixed time effects model specified in Equation 5. Table 5 summarizes the fixed time effect results from the models where the number of certification periods is fixed. At the time of a building's second certificate there are weakly significant fixed time effects in 2006, 2011 and 2012 -13, as measured in reference to the earliest adopters. Buildings commencing NABERS Energy certification in these years are statistically more likely to have greater reductions in energy consumption at the time of a building's second certificate. But once these buildings obtain a third or fourth certificate, there are no longer significant fixed time effects. Hence it can be concluded that number of certificates is a good proxy for depth of participation and is not likely to be an instrument for omitted time effects.

Conclusions
The results of this study point to a consistent relationship between depth of NABERS Energy participation and operational energy efficiency outcomes in Australian office buildings. Initially, the more NABERS Energy audits undertaken by a building owner, the more operational energy conserved on average, all else being equal. However, after the sixth audit, owners appear to reach an apparent post-intervention equilibrium energy consumption intensity, which, for the Australian sample in this study, measures approximately 430 MJ/m 2 /year (120 kWh/m 2 /year) on average for core building services. 13 Multivariate analysis revealed some differentiation within the population; owners with green asset management strategies obtain marginally higher levels of energy efficiency, Note: * , * * and * * * p-values less than 0.10, 0.05, and 0.01 respectively.
Successful GHG mitigation in o⁄ce buildings as do owners purchasing Green Power offsets regularly. This latter observation suggests Green Power offsets are a complement, not a substitute, to operational energy efficiency. Location is generally unimportant in relation to operational energy conservation, although energy efficiency does not appear to be a concern in a booming property market defined by a scarcity of supply, such as the Brisbane suburban market during the period of this study.
Additional tests were needed to establish the robustness of the relationship between depth of certification and energy savings because of high correlation between the year a building commences NABERS Energy certification and the depth of its participation. Fixed time effects are weakly important at the second NABERS Energy audit, but these fixed time effects disappear as the buildings undergo further certification.
There is another possible explanation for the relationship between depth of participation and energy efficiency that has little to do with private investment: the possibility that rising vacancy rates during the global financial crisis are responsible for the energy consumption improvements observed in this study.
The fixed time effect test provides some evidence against this possibility: the global financial crisis, which began in 2008, is notably absent from the time effects. Furthermore, the guidelines for NABERS Energy auditors (Department of Environment Climate Change and Water NSW, 2010), specify that vacant space is to be excluded from the denominator of any area-normalized metric such as EUI. The author has no evidence auditors are systematically failing to adhere to these standards, thus vacancy is not likely affecting these results.
For building owners and tenants, this study reinforces the effectiveness of integration between behaviour and design in building energy management or GHG emission outcomes. One direct implication is that owners and tenants may face increased scrutiny with regard to energy management practices if policymakers decide to mandate regular energy or GHG emission audits. This increased oversight could encourage owner-tenant partnerships for energy management, such as the inclusion of so-called 'green lease' clauses that specify the responsibilities of each party in meeting energy or GHG emission targets.
For policy-makers, the results of this study add clarity with regard to outcome expectations of using repetitive auditing of energy consumption as an operational component of a market-based regulation. Nearly all buildings in this dataset are privately owned, so there is strong evidence that repetitive disclosure motivates the private sector to invest in operational energy efficiency. It is also possible to conclude, using the control variable representing time between certifications, that expectations of future auditing play an important role in motivating owners to manage energy efficiency and maintain high performance.
Finally, the overall justification for the introduction of NABERS Energy is GHG mitigation. How effective has NABERS Energy been at mitigating GHG emissions in Australia? Policy targets are nearly always stated as a percentage reduction relative to an annual benchmark; for example, Australia's federal government has committed to an unconditional 5% reduction on GHG emissions measured in the year 2000 by the year 2020. How does the intervention of NABERS contribute? Depending on how one accounts for outliers, potential sub-population bias and spillover effects, pre-NABERS Energy consumption averaged between 580 and 626 MJ/m 2 /year (between 161 and 174 kWh/m 2 /year) for the entire building stock in this study (Table S2). It was then observed in the t-tests that six NABERS Energy audits or more delivered an average stock consumption of 430 MJ/m 2 / year (120 kWh/m 2 /year), meaning a reduction in GHG emissions from base building services somewhere between 26% and 32%. The multiple regression model attributes some of the reduction to green asset management strategies potentially unrelated to the presence of NABERS Energy, so it is best to use the asymptote around an average energy reduction of 120 MJ/m 2 /year (33 kWh/m 2 /year) observed in Figure 3. Using this, NABERS Energy audits can be held responsible for approximately 20% of the observed decline in building energy consumption. Further work on this project is assessing whether the motivation to participate -voluntarily or under mandate -affects these results.
This conclusion integrates well with the findings of Pacala and Socolow (2004), who argued that deployment of existing technology could reduce GHG emissions from energy consumption in the built environment by 25% relative to 2004 global emissions. The reductions seen in this study indicate that NABERS Energy, as implemented in Australia, may be an effective tool for introducing these existing technologies to the market more rapidly. Pacala and Socolow (2004) proposed a 50-year timeframe for their 25% reduction; if the conclusion in this study is accurate, then stronger targets are likely to be achievable within a 50-year timeframe.

Supplemental data
Supplemental data for this article can be accessed at http://dx.doi.org/09613218.2014.979034 Endnotes 1 This statement refers to the performance standards of current statutory building codes. A change to the high-performance expectations of most voluntary green building assessment tools would take much longer. Kok, Miller, and Morris (2012) estimate that only 10% of new construction in the United States has sought certification under its LEED assessment system. 2 BEED applies to office buildings greater than 2000 m 2 and the required energy disclosure is a NABERS Energy certificate (discussed in detail below). The NABERS Energy rating from the certificate must be displayed prominently on any advertisement for lease or sale. The European Union mandates building performance ratings for existing buildings, but these have all been based on simulated, not measured, performance. European governments are now considering the introduction of Display Energy Certificates (DECs), which have been a legal requirement since 2008 in the UK (for public sector buildings above a certain size). DECs are based on actual measured consumption and thus are comparable with Energy Star in the United States and NABERS Energy in Australia (Fuerst, van de Wetering, & Wyatt, 2013). 3 The statement 'energy efficiency is positively associated with measures of asset value' means that reduced energy consumption (i.e. increased energy efficiency) is associated with increased measures of asset value. An equivalent statement would be that nominal energy consumption is negatively associated with asset value. 4 As Borck and Coglianese (2009) report in their review, little is known about the existence or size of spillover effects. In the buildings sector, Simcoe and Toffel (2013) find evidence of a spillover effect resulting from government procurement policies that have nudged private building producers to adopt LEED construction standards as a 'de facto' construction standard for non-government projects. However, their conclusion comes with a number of caveats typical of studies looking for spillover effects, namely the possibility of reverse causation (i.e. environmentally conscious municipalities are those most likely to develop green procurement policies). An extreme example of spillover effects is the use of voluntary standards as mandates in public policy (Kontokosta, 2011;Simons et al., 2009). 5 Asset ratings ignore behavioural energy demand, such as plug loads (computers and other devices that tenants plug into a wall socket), so a researcher receiving total energy consumption data post-occupancy must estimate how much of that energy is consumed by the services included in an asset rating. This introduces the potential for significant bias and error. 6 Some minor adjustments have been made, such as increasing the top star rating to 6 from 5, but these have not affected the core assessment methodology or the audit process. 7 For clarity, a NABERS Energy audit only measures site energy directly. The process to calculate star ratings involves estimating GHG emissions from site energy consumption at the source (i.e. including generation and transmission losses for electricity). This paper predominantly uses the site energy audit result to define 'operational energy' because it is the most consistent and not subject to accounting protocol variations. Audit results indicating the fuel mix are not made public. Since purchased electricity is the dominant fuel in Australia, the site energy audit is a good proxy for GHG emissions in this study. 8 See http://www.nabers.gov.au/. 9 This Green Power purchasing scheme is the only GHG emission offset scheme recognized by NABERS Energy. As such, the terms 'Green Power purchasing' and 'Green Power offsets' are used interchangeably in this paper. Other forms of GHG emission offsets, such as re-forestation credits, will not improve a NABERS Energy rating. 10 See http://www.cbd.gov.au/. 11 To test whether the omission of these 12 buildings biases the results, all the statistical models in this paper have also been run with all 818 buildings, omitting building size as a variable when necessary. No changes to these results occur, most likely because the models in this paper find that building size is not a factor in energy efficiency outcomes in Australia. 12 Using the data in Table 5, the expected value of the change in energy consumption from each t-test for each recertification period s is calculated as such: E(DP s ) = e ln(P1) E(ln P s ) E(ln P 1 ) − 1 A further adjustment is to modify E(DP s ) so it is relative to the second certificate because the above equation calculates the expected change relative to the initial benchmark certificate. Thus, each value of E(DP s ) for s . 2 is reduced by E(DP 2 ).