Daylgt Retail 2.3.7
Daylgt Retail 2.3.7
Daylgt Retail 2.3.7
COMMISSION
TECHNICAL REPORT
October 2003
P500-03-082-A-5
Managed By:
New Buildings Institute
Cathy Higgins, Program Director
White Salmon, Washington
CEC Contract No. 400-99-013
Prepared For:
Donald Aumann,
Contract Manager
Nancy Jenkins,
PIER Buildings Program Manager
Terry Surles,
PIER Program Director
Robert L. Therkelsen
Executive Director
DISCLAIMER
This report was prepared as the result of work sponsored by the
California Energy Commission. It does not necessarily represent
the views of the Energy Commission, its employees or the State
of California. The Energy Commission, the State of California, its
employees, contractors and subcontractors make no warrant,
express or implied, and assume no legal liability for the
information in this report; nor does any party represent that the
uses of this information will not infringe upon privately owned
rights. This report has not been approved or disapproved by the
California Energy Commission nor has the California Energy
Commission passed upon the accuracy or adequacy of the
information in this report.
ACKNOWLEDGEMENTS
This report is a part of the Integrated Energy Systems - Productivity and Buildings Science
program, a Public Interest Energy Research (PIER) program. It is funded by California
ratepayers through California's System Benefit Charges administered by the California
Energy Commission (CEC) under (PIER) contract No. 400-99-013, and managed by the
New Buildings Institute. Heschong Mahone Group would like to acknowledge the support
and contributions of the individuals below:
Heschong Mahone Group, Inc.: Principal in Charge: Lisa Heschong. Project Director: Lisa
Heschong. Project staff: Cynthia Austin, Sean Denniston, Charles Erhlich, Carey Knetch,
Douglas Mahone, Mudit Saxena, Heschong Mahone Group.
Subcontractors: RLW Analytics, Inc: Dr. Roger Wright, RLW Analytics, Inc. statistical
metholodies; Ramona Peet, analyst. Wirtshafter Associates: Dr. Robert Wirtshafter, census
and GIS analysis
Review and Advisory Committee: Dr. Jed Waldman, California Department of Public
Health; Dr. Gage Kingsbury, Northwest Evaluation Association; Dr. Judith Heerwagen,
private consultant; Abby Vogen, Wisconsin Energy Center; Dr. Cliff Federspiel, Center for
the Built Environment; Barbara Erwine, Cascadia Consulting; Dr. Robert Wirtshafter,
Wirtshafter Associates, Inc.
Participant support: This project was supported by many managers and employees of the
participant retail corporation. We are greatly appreciative of their time, facilitation and
review of this study. Studies such as this are highly unusual, in that they require a leap of
faith on the part of corporations that are traditionally highly private and protective of internal
data. We have promised anonymity to this corporation, as we did with the previous study,
and request that everyone involved will endeavor to respect that request and honor the trust
placed in the study team.
PREFACE
The Public Interest Energy Research (PIER) Program supports public interest energy
research and development that will help improve the quality of life in California by
bringing environmentally safe, affordable, and reliable energy services and products to
the marketplace.
This document is one of 33 technical attachments to the final report of a larger research
effort called Integrated Energy Systems: Productivity and Building Science Program
(Program) as part of the PIER Program funded by the California Energy Commission
(Commission) and managed by the New Buildings Institute.
As the name suggests, it is not individual building components, equipment, or materials
that optimize energy efficiency. Instead, energy efficiency is improved through the
integrated design, construction, and operation of building systems. The Integrated
Energy Systems: Productivity and Building Science Program research addressed six
areas:
♦ Productivity and Interior Environments
♦ Integrated Design of Large Commercial HVAC Systems
♦ Integrated Design of Small Commercial HVAC Systems
♦ Integrated Design of Commercial Building Ceiling Systems
♦ Integrated Design of Residential Ducting & Air Flow Systems
♦ Outdoor Lighting Baseline Assessment
The Program’s final report (Commission publication #P500-03-082) and its attachments
are intended to provide a complete record of the objectives, methods, findings and
accomplishments of the Integrated Energy Systems: Productivity and Building Science
Program. The final report and attachments are highly applicable to architects,
designers, contractors, building owners and operators, manufacturers, researchers, and
the energy efficiency community.
This Daylighting and Retail Sales Report (Product # 2.3.7) is a part of the final report
within the Productivity and Interior Environments research area and presents the results
of a study into relationships between daylighting and sales at a retail outlet store.
The Buildings Program Area within the Public Interest Energy Research (PIER)
Program produced these documents as part of a multi-project programmatic contract
(#400-99-413). The Buildings Program includes new and existing buildings in both the
residential and the non-residential sectors. The program seeks to decrease building
energy use through research that will develop or improve energy efficient technologies,
strategies, tools, and building performance evaluation methods.
For other reports produced within this contract or to obtain more information on the
PIER Program, please visit www.energy.ca.gov/pier/buildings or contact the
Commission’s Publications Unit at 916-654-5200. All reports, guidelines and
attachments are also publicly available at www.newbuildings.org/pier.
ABSTRACT
This study presents evidence that a chain retailer is experiencing higher sales in daylit
stores than in similar non-daylit stores. Statistical models, using up to 50 explanatory
variables, examine the relationship between average monthly sales levels and the
presence of daylight in the stores, while simultaneously controlling for more traditional
explanatory variables such as size and age of the store, amount of parking, local
neighborhood demographics, number of competitors, and other store characteristics.
The study included 73 store locations in California, of which 24 stores were daylit
primarily by diffusing skylights. Statistical regression models found that increased
annual hours of useful daylight per store were strongly associated with increased sales,
but at a smaller magnitude than a previous study. No season variation in the
relationship of daylight to sales was found. The study also included interviews with store
managers and surveys of employees, along with an analysis of the energy savings due
to automatic control of the electric lights.
TABLE OF CONTENTS
1. INTRODUCTION ____________________________________________________ 1
2. SELECTION OF STUDY PARTICIPANT _________________________________ 3
2.1. Selection Criteria _________________________________________________ 3
2.2. Participant Search ________________________________________________ 4
2.3. Participant Description _____________________________________________ 5
3. DATA COLLECTION _________________________________________________ 7
3.1. Corporate Databases and Records ___________________________________ 7
3.2. Corporate Interviews ______________________________________________ 7
3.3. On-site Visits ____________________________________________________ 8
3.3.1. Surveyors and Training _______________________________________ 8
3.3.2. Survey Equipment ___________________________________________ 8
3.3.3. Survey Protocol _____________________________________________ 9
3.3.4. Manager Interview __________________________________________ 10
3.3.5. Lighting Quality Survey ______________________________________ 10
3.4. Census Demographic Data ________________________________________ 10
3.5. Local Market Analysis ____________________________________________ 11
3.6. Data Verification _________________________________________________ 12
3.7. Parking Data Verification __________________________________________ 12
4. DATA PROCESSING AND VARIABLE DEFINITION_______________________ 13
4.1. Dependent, or Outcome, Variables __________________________________ 13
4.2. Independent, or Explanatory, Variables: ______________________________ 14
4.2.1. Daylight Variable Definition ___________________________________ 16
4.2.2. Energy Observations ________________________________________ 17
5. STATISTICAL METHODOLOGY ______________________________________ 19
5.1. Variable Testing Method __________________________________________ 19
5.2. Preliminary Investigations _________________________________________ 21
5.2.1. Defining the Core Model _____________________________________ 21
5.2.2. Simple Yes/No Daylight Model ________________________________ 21
5.2.3. Daylight Hours Analysis ______________________________________ 22
5.2.4. Demographic Test __________________________________________ 25
5.3. Final Regression Models __________________________________________ 25
5.3.1. Comparison of Linear versus Log Models ________________________ 26
6. ANALYSIS FINDINGS _______________________________________________ 29
6.1. Log Models_____________________________________________________ 29
6.1.1. Daylight Effect Interaction with Parking __________________________ 31
i
RETAIL AND DAYLIGHTING TABLE OF CONTENTS
ii
RETAIL AND DAYLIGHTING TABLE OF CONTENTS
TABLE OF FIGURES
APPENDICES
Figure 22: Summary Statistics for Natural Log Models _________________________ 63
Figure 23: Log Model of 10-Month Sales, 2001 _______________________________ 64
Figure 24: Log Model of 24-Month Sales, 1999-2000 __________________________ 64
Figure 25: Order of Entry and Partial R2, Log 10 Month Sales, 2001 ______________ 64
Figure 26: Order of Entry and Partial R2, Log 24 Month Sales, 1999-2000 __________ 65
Figure 27: Summary Statistics for All Variables Considered in Linear Models _______ 66
Figure 28: Linear Model of 10 Month Sales, 2001 _____________________________ 67
Figure 29: Linear Model of 24 Month Sales, 1999-2000 ________________________ 67
Figure 30: Order of Entry and Partial R2, Linear 10 Month Sales, 2001 ____________ 68
Figure 31: Order of Entry and Partial R2, Linear 24 Month Sales, 1999-2000 ________ 68
Figure 32: Linear Model of 10 Month Transactions, 2001 _______________________ 69
Figure 33: Linear Model of 24 Month Transactions, 1999-2000 __________________ 69
Figure 34: Glossary of Statistical Terminology________________________________ 71
iii
RETAIL AND DAYLIGHTING TABLE OF CONTENTS
iv
RETAIL AND DAYLIGHTING EXECUTIVE SUMMARY
EXECUTIVE SUMMARY
This study presents evidence that a major retailer is experiencing higher sales in daylit
stores than in similar non-daylit stores. Statistical models were used to examine the
relationship between average monthly sales levels and the presence of daylight in the
stores, while simultaneously controlling for more traditional explanatory variables such
as size and age of the store, amount of parking, local neighborhood demographics,
number of competitors, and other store characteristics. The retailer, who will remain
anonymous, allowed us to study 73 store locations in California from 1999 to 2001. Of
these, 24 stores had a significant amount of daylight illumination, provided primarily by
diffusing skylights.
This study was performed as a follow-on to a similar study completed for Pacific Gas
and Electric in 19991, which found that for a certain retail chain, all other things being
equal, stores with skylights experienced 40% higher sales than those without skylights.
This study, on behalf of the California Energy Commission, examined a second retail
chain, in an entirely different retail sector, to see if the original findings would hold in a
new situation, and if we could learn more about any daylight effect that might exist.
As a first step in this process, a simple model with daylight as a yes/no variable, and
using basically the same format and inputs as the previous study, did not find a
significant correlation between the presence of daylight, and increased sales. We then
pursued the study in greater detail, adding more information to the model and describing
daylight on a continuous scale by the number of daylit hours per year in each store.
The retailer in this study had a less aggressive daylighting design strategy and also
more variation in both the range of daylight conditions and the range of store designs
than the retailer in the first study. For this study, we collected much more detailed
information about the characteristics of each store, and verified all information on site.
Neighborhood demographics and retail competition were described using detailed, site-
specific GIS analysis. Store managers were interviewed and employees were surveyed
about their observations and preferences. For the final analysis, the amount of daylight
in each store was described as the number of hours per year that daylight illumination
levels exceeded the design electric illumination level.
Statistical regression models of average sales for the stores, using up to 50 explanatory
variables, and both linear and natural log descriptions of the variables, found that
increased hours of daylight per store were strongly associated with increased sales, but
at a much smaller magnitude than the previous study. In addition, for this chain, the
daylight effect on sales was found to be constrained by the amount of parking available
at the store site. Sites with parking lots smaller than the norm experienced decreased
sales associated with daylight, while stores with average and ample parking experienced
increased sales as both the amount of daylight and parking increased. The statistical
models were also more comprehensive, explaining about 75% of the variation in the
data (model R2=0.75), compared to 58% in the previous study.
1
Heschong Mahone Group (1999). Skylighting and Retail Sales. An investigation into the relationship
between daylight and human performance. Detailed Report for Pacific Gas and Electric Company. Fair
Oaks, CA.
v
RETAIL AND DAYLIGHTING EXECUTIVE SUMMARY
vi
RETAIL AND DAYLIGHTING INTRODUCTION
1. INTRODUCTION
The Skylighting and Retail Sales study1 completed in 1999 by the Heschong Mahone
Group on behalf of the California Board for Energy Efficiency found a compelling
statistical correlation between the presence of daylighting in a chain retail store and
higher sales for those stores.
The study was reviewed by a panel of experts, recruited by Lawrence Berkeley National
Laboratory, involving a wide range of disciplines related to the study. In general, the
review panel was satisfied with the soundness of the basic methodology and the rigor of
the statistical analysis.
There were, however, some weaknesses to the original study and lingering peer review
questions,2 that could only be addressed in follow-up studies.
1. Replicating findings: The biggest weakness in the original study was that
the participant remained anonymous, making it impossible for anyone else to
verify the findings. Anonymity was difficult to overcome, since it was unlikely
that any retailer would be willing to reveal their identity in a study that publicly
discussed sales effects. However, a second study, of another retailer, would
increase confidence that such a skylighting effect could be replicated.
2. Controlling for other influences: The original study controlled for twelve
potential influences on sales. Not all stores in the study were visited to verify
conditions. It was highly probable that there were other factors affecting sales
that were collinear with skylighting that the original research team could not
determine. A more detailed study, including verification visits to all sites in the
study, and collection of more information about store characteristics, should
be able to reduce the uncertainty that other factors collinear with skylighting
might actually be responsible for the original findings.
3. Bounding the effect: The 40% increase in sales associated with skylighting
seemed to be improbably high. At best it could be assumed to be an upper
bound of an effect. If we found positive sales associated with daylighting in
another chain, could we establish upper and/or lower bounds to the effect?
1
Heschong Mahone Group (1999). Skylighting and Retail Sales. An investigation into the relationship
between daylight and human performance Detailed Report for Pacific Gas and Electric Company. Fair
Oaks, CA.
2
Heschong Mahone Group (1999). Daylighting and Productivity. An investigation into the relationship
between daylight and human performance. Review Report. Fair Oaks, CA.
1
RETAIL AND DAYLIGHTING INTRODUCTION
The study described in this report, supported through the California Energy
Commission's Public Interest Energy Research (PIER) program, was designed to
address these concerns, while also expanding other areas of our knowledge about the
interaction of retail sales and daylighting.
In this study, a second retailer was identified who had appropriate conditions for such a
study, and who was willing to participate in the study. As with the original retail study,
strict anonymity was requested and observed. The retailer provided us with
dimensionless monthly sales index data for each store for a 34-month period. The
research team then identified 73 store sites appropriate for the study, one-third with
daylighting, and collected extensive data about each site.
The research team’s information about each store site was used in a statistical
regression model, and for secondary analysis. This information included:
1. Information about the size, age, history and monthly sales volumes of the
stores (from corporate sources)
2. Population characteristics within a radial distance of the stores (from U.S.
Census 1990 and 2000)
3. Number of competitors within a radial distance of the stores (from public
databases)
4. On-site observations about the neighborhood and about the stores’
architectural features, skylighting system, lighting, mechanical systems, and
other site-specific conditions.
5. Interviews and surveys with store managers and employees.
This data was processed and put into a multivariate regression model. A number of
modeling approaches were investigated. Monthly data allowed us to look for seasonal
patterns. Two different electric lighting conditions during the study period allowed us to
examine illumination intensity issues. Although we were not allowed to interview
customers, interviews with store managers and surveys of store employees allowed us
to examine attitudes and perceptions associated with daylighting. A range of daylighting
conditions within the participant’s store sites allowed us to probe for a dose/response
relationship between daylighting and sales.
Finally, the analysis results were studied, and conclusions were drawn about the role of
daylighting in the sales of this retail chain. This report describes the data and analysis
methodologies in greater detail. Conclusions are then presented in Chapter 8. The
Appendices include the data collection forms and other study details, as well as a brief
glossary of statistical terms to assist readers who are less familiar with the statistical
methods utilized by this study.
2
RETAIL AND DAYLIGHTING SELECTION OF STUDY PARTICIPANT
Selection of a participant for a large statistical study is a very important and strategic
investment. We invested a considerable amount of time and effort to find the most
promising retail corporation to study for our second effort at understanding how daylight
might effect retail sales patterns.
3. Have some daylit and some non-daylit stores so that daylighting effects
could be compared between otherwise identical environments. It would be
ideal to have a continuous range of daylighting conditions so that a dose-
response relationship between daylighting and sales could be studied.
3
RETAIL AND DAYLIGHTING SELECTION OF STUDY PARTICIPANT
8. Have little variation between daylit and non-daylit stores other than the
amount of daylight available. This would minimize the number of other factors
that needed to be controlled for.
4
RETAIL AND DAYLIGHTING SELECTION OF STUDY PARTICIPANT
Each of the four chain stores had very different characteristics of skylight distribution and
store design issues. The corporate history of the chains also varied widely. Two
participants, Retailer A and B, showed the most potential for study.
The team discussed the possibility of doing two studies, each with less depth, in order to
increase the diversity of the study. We received initial site characteristics data from both
candidates and visited a sample of sites from both retailers. From the initial site
reconnaissance, we concluded that Retailer B had some particularly confounding
variables that we would not be able to fully control for in our statistical analysis. Retailer
A seemed to have fewer confounding issues, and more data available about store
performance, history and design characteristics. At this point, Retailer A allowed us to
review store plans, and provided additional information that reassured us about the
feasibility of the study. We finally decided to work with Retailer A on the study and
expressed our appreciation to the others for their interest.
5
RETAIL AND DAYLIGHTING SELECTION OF STUDY PARTICIPANT
6
RETAIL AND DAYLIGHTING DATA COLLECTION
3. DATA COLLECTION
Data collection for the study proceeded from a number of sources. First we collected as
much information as possible from the retail participant directly, from corporate records,
plan rooms, and interviews with corporate managers. We then conducted on-site
surveys of every store to be included in the study, to confirm information from the
corporate records and to collect new, detailed data about the physical conditions at each
store. Next we collected and processed Census and market conditions information from
various public databases, using GIS analysis to create site specific information. Finally,
we analyzed interviews with store managers and surveys of store employees to gain a
more qualitative understanding of the conditions at each store.
7
RETAIL AND DAYLIGHTING DATA COLLECTION
8
RETAIL AND DAYLIGHTING DATA COLLECTION
Flicker Checker: a spinning tool from Motorola for checking for the presence of
electronic flicker in the lighting.
Tape measure: to measure dimensions not found in the plan review.
1
Heschong Mahone Group, Lighting Baseline Study, for Southern California Edison, 2000. Presented at the
IESNA conference 2001, Ottawa Canada.
9
RETAIL AND DAYLIGHTING DATA COLLECTION
1
Heschong Mahone Group, Lighting Baseline Study, for Southern California Edison, 2000. Presented at the
IESNA conference 2001, Ottawa Canada.
10
RETAIL AND DAYLIGHTING DATA COLLECTION
Our goal was to create a reliable comparison between sites in our study as a control
variable, rather than pre-determine the most favorable location for a new store. An initial
comparison of a few store sites in our study showed that the simpler fixed radius
analysis captured population effects within 85 to 90% accuracy of the far more complex
drive-time analysis. In addition, we noted that many of our store sites were located in
rapidly developing areas where street information was not up-to-date in the GIS
databases. Thus, we determined that using a fixed radius Census analysis would give us
sufficient accuracy, and perhaps also a more reliable comparison between our sites. We
interviewed the real estate manager for the chain and determined the appropriate radius
to use in the census analysis. This is henceforth called the “standard radius.”
We reviewed 34 possible census characteristics with the real estate manager for the
retailer, and together selected twelve characteristics that represented a range of
population, economic, ethnic, housing and transportation information, and that were
considered most relevant to this particular chain’s target customer. We will not identify
the specifics of the census variables considered in order to protect the retailer’s identity.
A GIS consultant processed this information into ten census variables for each study
site. Since each variable was based on census data within the area determined by a
standard radius, the census variables also became density indicators for each site.
At the time of our data collection, the 2000 US Census data was just becoming
available. Population and ethnic characteristic data was available for the 2000 census,
but for housing, economic and transportation data we had to use 1990 data. The
difference between the 1990 and 2000 population data determined growth rates for the
sites.
11
RETAIL AND DAYLIGHTING DATA COLLECTION
local customer awareness of the stores. To account for this effect, we create a “sister
store” scalar. We mapped out the stores in the study and counted the number of stores
sharing a similar media market. The store locations were rated, on a scale of 1 to 5, for
the density of other sister stores nearby from the same chain. A store with a rating of 1
was alone in its media market, while a store with a rating of five had the highest density
of sister stores nearby.
12
RETAIL AND DAYLIGHTING DATA PROCESSING AND VARIABLE DEFINITION
Upon completion of data collection and verification, the data was processed into useful
variables for analysis. If a store characteristic did not exhibit sufficient variation between
stores, we could not use it in the analysis. For example, the variety of signage was not
found to vary much between survey sites, and so signage type was not considered in the
analysis. Likewise, whenever fewer than four stores exhibited a particular characteristic,
that characteristic was dropped from the analysis.
In some cases data was combined in order to increase the range of variation in the data
for analysis. For example, the acoustic properties of the stores were originally collected
according to five different properties, but they were subsequently combined into one
acoustic scalar indicating overall noise levels in the stores. Similarly, we were given data
on the square footage and number of products sold for a variety of sales areas. We first
collapsed this information into three types of sales areas, and eventually collapsed it into
a variable named “total sales area scalar.”
Forty-one explanatory variables and two dependent variables were ultimately defined
and included in the preliminary analysis. These variables took the form of binary
variables (yes/no) or scalar variables (a range of values indicating relationships from
small to large). In order to preserve the anonymity of the participant, not all information
about the variable definitions or ranges can be revealed. For reporting purposes, most
variables were transformed into a dimensionless scalar in order to mask identifying
information about the retailer.
Descriptive statistics for the variables considered in the analysis are included in the
Appendix. These include the minimum, maximum, range, mean and standard deviation
for that variable. When a scalar variable is used, the minimum is a dimensionless unit of
one, and the maximum illustrates the relative range of that variable.
13
RETAIL AND DAYLIGHTING DATA PROCESSING AND VARIABLE DEFINITION
14
RETAIL AND DAYLIGHTING DATA PROCESSING AND VARIABLE DEFINITION
15
RETAIL AND DAYLIGHTING DATA PROCESSING AND VARIABLE DEFINITION
16
RETAIL AND DAYLIGHTING DATA PROCESSING AND VARIABLE DEFINITION
and duration of daylight due to climate location, daylight system and store interior
design. When only a sub-area of the sales floor had useful daylight, the daylit hours
were calculated for that sub-area, then proportioned relative to the size of the store.
Thus, if only one half of the sales area was daylit, the annual daylight hours were
reduced by half.
Number of daylit hours per year per store was predicted by running computerized hourly
simulations of each store, based on building design variables, local climate using typical
meteorological year data (TMY2), type of glazing, amount of glazing area, dimensions
and surface reflectances within the store. This was fairly easy for the standard skylit
stores, using our automated spreadsheet SkyCalc. It was more difficult for the few
stores using non-standard daylighting systems, such as clerestory windows or roof
monitors. For those, we used an annual DOE-2 model, which could account for the
effects of vertical glazing.1
The SkyCalc daylight hour calculations are limited by the granularity of TMY weather
data available for each site that could be used to generate input. There are 16 climate
files available for SkyCalc in California, based on the 16 climate zones defined by the
California Energy Commission for the Title 24 Building Energy Standards. Thus, the
daylight availability analysis is for typical weather in a nearby city representing the
appropriate California climate zone for each site, rather than actual yearly weather for a
specific city or store site.
1
For more information on SkyCalc, see: Heschong, Lisa and Jon McHugh, "Skylights: Calculating
Illumination Levels and Energy Impacts," Journal of the Illumination Engineering Society, Winter 2000,
Vol. 29, No. 1, pp. 90-100, and Skylighting Guidelines, 1999, a web-based publication on skylighting
design, downloadable from www.energydesignresources.com.
17
RETAIL AND DAYLIGHTING DATA PROCESSING AND VARIABLE DEFINITION
18
RETAIL AND DAYLIGHTING STATISTICAL METHODOLOGY
5. STATISTICAL METHODOLOGY
The heart of this study was the statistical analysis of the data collected. This analysis
entailed developing statistical models that seek to explain the factors that affect retail
sales in this particular chain. Our goal was to control for other influences on sales in
order to isolate the effect of our key variable of interest: daylighting. Developing these
models requires both science and insight. It requires reasonable experience with what is
likely to influence sales, a thorough understanding of how reliable the available data is,
and a certain amount of trial-and-error looking for mathematical models which best fit the
data. A variety of statistical tests are used to determine which modeling approach
provides the most mathematically accurate representation of the data. It is important to
remember that the statistical models are a mathematical abstraction of reality. They do
not so much provide true or false answers, as provide a way to simplify a very complex
retail environment and start to quantify the relative magnitude and certainty of various
influences on sales performance.
Regression models try to fit lines that best describe a plot of data points. Multivariate
models consider more than one dimension at once. Linear models try to fit straight lines
through the data. It is also possible, but far more complex, to consider curved, or non-
linear, relationships, as we did with models using a natural log function.
All of the analysis was pursued using multivariate regression models run in SAS using a
variant of backwards step-wise regression to eliminate the least significant variables. F-
tests1 were performed on groups of variables to insure that they could be dropped as a
group as well as individually. The analysis used p≤0.10 as the threshold criteria for
inclusion of explanatory variables in the models, meaning that for a variable to be
considered significant in determining sales, there must be no greater than a 10% chance
of error in making this decision, or 90% certainty. All statistical terms are explained in
Section 9.2 in the Appendix.
Models were judged based on their R2 (the percentage of variation in the data explained
by the model), the parsimony (minimum explanatory variables for maximum explanatory
power), and consistency between the models. Ultimately, models predicting more
moderate effects for daylight were also judged to be more realistic than those with wildly
diverging values.
1
See Appendix Section 9.2 for an explanation of “F-test” and other statistical terms used in this report.
19
RETAIL AND DAYLIGHTING STATISTICAL METHODOLOGY
The backward elimination method first calls for fitting a model using all potential
explanatory variables and calculating the t-statistic associated with each variable. The
explanatory variables are then deleted from the model one-by-one, until all variables
remaining in the model are associated with a significant t-statistic. During each iteration,
the variable with the least explanatory power is identified and deleted from the model.
The RLW variable selection method1, used in this study, is a variant of the backward
elimination method. Similar to the backward elimination method, the RLW variable
selection method begins with calculating a model using all potential explanatory
variables and the associated t-statistics. However, the RLW method allows for the
deletion of multiple variables during each iteration, whereas the backward elimination
method does not. This procedure helps to identify co-linearities between insignificant
variables, which might otherwise be dropped without first understanding how such co-
linearities could potentially influence results. Specifically, the RLW method consists of
the following steps:
1. Calculate a “full” linear regression model including all potential explanatory
variables.
2. Identify all insignificant variables from the model resulting from step 1.
3. Perform an F-test to test whether the set of individually insignificant variables are
statistically significant as a group. Specifically, the null hypothesis of the F-test is
that the beta coefficients of each of the variables in the group are zero, while the
alternative hypothesis is that there is at least one variable in the group where the
beta coefficient is not zero. If the F-test shows the set of variables are not
statistically significant as a group, all variables identified in step 2 are also
identified for deletion. If the set of variables tested is statistically significant as a
group, this indicates there is a collinear relationship between the variables that is
affecting the model. In this case, a reduced set of variables is defined for the F-
test and deletion from the model.
4. Calculate a reduced model including all explanatory variables that were not
identified for deletion.
5. If any previously significant variables become insignificant in the reduced model,
calculate an F-test for all variables previously deleted from the model and the
newly insignificant variables under the guidelines provided in step 3.
1
The RLW variable selection methodology was developed by Dr. Roger Wright, lead statistician of this
study.
20
RETAIL AND DAYLIGHTING STATISTICAL METHODOLOGY
The models were tested for “outliers,” or store sites that were performing significantly
different from all the others, and therefore unduly influencing the findings. One store
tested as an outlier and so was isolated from the equation. This store was a very high
selling daylit store.
These two core models were then used to test various ways of defining the daylight
variable and other physical conditions of the individual stores.
21
RETAIL AND DAYLIGHTING STATISTICAL METHODOLOGY
however, include physical comfort characteristics about each store. Figure 2 shows the
p-values for significant variables in this model.
Daylight Yes/No Model, Significant Variables (p≤0.10)
10 month 24 month
Sales Area 1 0.018 0.069
Sales Area 2 0.036 0.019
Longer Hours 0.011 0.051
Store Age 0.000 0.001
Population Growth 0.008
Population Density 0.078
Housing 0.085
Education 0.029
Transportation 0.037
# Sister Stores 0.036 0.019
Competitors, Radius 1 0.016 0.011
Frontage Height 0.027
Parking 0.000 0.000
Model R2 68.5 68.7
Figure 2: Daylight Yes/No Model, List of Significant Variables
In these initial models, the yes/no daylight variable was not significant. The R2 of the
models was higher than the previous study (R2 went from 58% to 69%), suggesting the
other variables we included were increasing our precision in predicting sales. The new
census variables were significant, but were not consistent for both models. Likewise, the
height of the storefront was significant in one model, but not both.
22
RETAIL AND DAYLIGHTING STATISTICAL METHODOLOGY
accounted for interaction effect between the presence of daylight and other significant
variables in predicting the sales index.
The good news from this exercise was that none of variables that had shown up as
predicting sales in the “core” models were significant in predicting the daylight hours,
thus they were determined not to be collinear with daylighting.
The bad news from this exercise was that there were clearly many potential explanatory
variables that were collinear with daylight hours. In addition to those shown above, a
quick test told us that there were many other variables that were collinear with daylight
hours, but below the model threshold 90% significance level. All of these collinear
variables could potentially cause problems in our subsequent sales analysis,
confounding the effect of the daylight variable on the sales index. This was likely to be
the greatest concern for explanatory variables that were also significant predictors of
sales.
A number of these collinear variables had a logical relationship to the presence of
daylight, such as the ceiling type, which was almost a yes/no indicator for daylight, or
higher vertical illuminance levels, which was almost universally higher in daylight stores.
When there were such obvious dependencies between variables related to daylight
levels, we ran the variables in separate sales index models, testing which of the
competing variables had more significance and predictive power. In these tests, the
daylight hours variable stayed in the model as significant and the other descriptors
dropped out as not significant. Thus, we concluded that daylight hours, rather than other
correlated conditions, were more useful predictors of sales. Therefore, we left the other
variables out of our subsequent models.
While we can never be certain that excluded collinear variables are not influencing the
daylight hour results, we have higher confidence in the daylight hour variable for two
23
RETAIL AND DAYLIGHTING STATISTICAL METHODOLOGY
reasons. First, per above, it provided more precision in predicting a sales effect.
Secondly, there was a more obvious hypothesis for a causal relationship.
For similar reasons, at this point we also dropped the electric lighting scalar, type of
lamps, and cleanliness of store variables. We did not have confidence these variables
were reliable. The electric lighting scalar was based on observations at the time of the
survey, but the lighting could easily have been different during the other times of the
study period. The type of lamps variable was based on conflicting evidence from a
number of sources, and may also have changed during the study period. The
cleanliness variable had a very limited range of conditions between stores, and was also
rather subjective. We have seen in our previous study that shoppers tended to associate
daylighting with a cleaner store, and it seemed possible that our surveyors may have
made the same assessment.
24
RETAIL AND DAYLIGHTING STATISTICAL METHODOLOGY
skylight area, etc, and the daylight interaction variables included in that model, such as
parking area or age of the store.
This is a much more nuanced approach to studying the effects of daylight. Sometimes
the daylight effect for an individual store is predicted to be positive and sometimes it is
negative. The key issue of interest is whether the net effects of daylight across the fleet
of stores in the chain is positive or negative. Using the interaction variables, we
calculated the predicted sales for each store according to the models, and then
summarized the net effect on the chain. When we calculated this net effect for the chain
for each model, we found a net positive effect for three of the four models. These values
are reported in the Findings, Section 6.1.
1
a logarithmic function based on the natural number “e”
25
RETAIL AND DAYLIGHTING STATISTICAL METHODOLOGY
end of the scale. A log model thus makes sense when one expects there to be a case of
diminishing returns, where a unit increase in an explanatory variable at the bottom of the
scale is expected to have a proportionately bigger effect than at the top. For example,
one might hypothesize that doubling the size of a parking lot will have a relatively greater
effect for a small 50 space lot than a large 500 space lot.
450 1000
400
300 100
250
200
150 10
100
50
0 1
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
We tested models using natural logs of both the dependent variable (sales) and of some
of the explanatory variables that seemed appropriate for a logarithmic scale. A
logarithmic function requires that the variable be defined on a continuous scale, and also
that it does not include any values less than or equal to zero, since the natural logarithm
function is only defined for positive numbers. Thus, only some variables can be
converted to natural logarithms. In addition to meeting the mathematical criteria for
taking the log of a variable, a logged variable should also have a logical explanation for
why a diminishing effect might be expected as the scale of the variable increases.
Using these criteria, we took the natural log of the following explanatory variables and
included them in a “log” model of the Sales Index: Sales Area, Store Age, and Parking.
When we did this, the number of significant interaction variables was also reduced,
suggesting that taking the log of these variables was doing a better job of accounting for
the interaction effects than the explicit interaction variables. In the log models, only the
parking*daylight hours interaction variable remained significant.
1
See Appendix 8.3 for an explanation of the R2 expression.
26
RETAIL AND DAYLIGHTING STATISTICAL METHODOLOGY
A Box-Cox comparison was run between the linear and the natural log models. Using
this method, it was found that there was virtually no difference in the explanatory power
of the two sets of models. Thus, they were judged equally good at explaining the data.
We then applied secondary criteria to comparing the models including the parsimony of
the models, the consistency between the two time periods, and the reasonableness of
the predicted effects.
• PARSIMONY: The log and linear models were found to be equally parsimonious. In
both cases the 10 month models used 11 variables and one outlier and the 24 month
models used 10 and the same outlier. The log and linear models had degrees of
freedom of 62 and 59 respectively.
• CONSISTENCY: The log and linear models were found to be equally consistent in
their predictions of a daylight effect per store site between the two time periods. This
is illustrated visually in Figure 6, which shows a very consistent pattern of predicted
daylight effects between the two time periods for both types of models.
Daylight Effect on Sales Index per Store, Daylight Effect on Sales Index per Store,
Comparison of 24m & 10m Linear Models Comparison of 24m & 10m Log Models
100% 100%
24m linear log 24m
80% 80%
Percent Change in Sales
Percent Change in Sales
40% 40%
20% 20%
0% 0%
-20% -20%
-40% -40%
-60% -60%
40% 40%
20% 20%
0% 0%
-20% -20%
-40% -40%
-60% -60%
Stores Ranked by 24m Log Value Stores Ranked by 10m Log Value
Figure 6: Comparsion of Daylight Effects per Store – Linear and Log Models
27
RETAIL AND DAYLIGHTING STATISTICAL METHODOLOGY
28
RETAIL AND DAYLIGHTING ANALYSIS FINDINGS
6. ANALYSIS FINDINGS
In this section we report on the findings primarily from the log models and secondarily,
the linear models. The full equation for each model, and the descriptive statistics for all
the variables considered in the model, are detailed in the Appendix.
As mentioned earlier, some of the variables have been converted into “scalars” in order
to preserve the anonymity of the participant. In all cases, these scalars were created
using simple division or multiplication as appropriate, so that they are statistically
consistent.
To calculate the overall effect of more daylight on corporate sales, we first calculated the
effect of daylight on each individual store, considering all interaction effects specific to
each store. We then summed the effects on all daylit stores, and divided by the sum of
all sales for those stores, to calculate the “net daylight effect,” or the average predicted
effect on sales for adding daylight to any store in the chain.
It is not really appropriate to calculate a standard deviation for these findings, since they
are not based on one yes/no variable, but a multi-dimensional group of variables. In
order to express the range of the potential effect of daylight on an individual store, we
have plotted the range of predicted effects for the two models in Figure 9 and Figure 17,
below. It is important to note that the models do not predict a positive effect for every
individual store. Some stores are predicted to have lower sales associated with
daylighting, based on the effects of the interaction variables.
We performed an additional statistical test to consider the certainty of the net effect
predicted by the combination of interaction variables. We used an F-Test to test the null
hypothesis that the beta coefficients of each of the interaction variables in the group are
simultaneously zero. The alternative hypothesis is that there is at least one variable in
the group where the beta coefficient is not zero. The groups of interaction variables were
found to be significant in all models, and so remain in our final models.
29
RETAIL AND DAYLIGHTING ANALYSIS FINDINGS
Figure 8 lists the variables this model found to be significant in predicting the sales
index, along with their magnitude and significance. The variable with by far the
strongest positive effect on sales was the size of the sales area. Other variables with
positive effects include the age of the store, the number of sister stores nearby, and a
more educated local population.
In a linear model, a one-unit change in the explanatory variable (X) predicts an
approximate constant, or fixed, unit change (B) in the sales index (Y). So for example, if
the size of the store increases by one square foot, then the sales index will go up by B.
In these log models, a one-unit change in a non-logged explanatory variable predicts an
approximate percentage change in the sales index. And for logged explanatory variables
in the log model, a one-percent increase in the explanatory variable predicts an
approximate percentage change in the sales index. So for example, in these log
models, since square footage was also logged, as the size of the store increases by one
percent then sales index is increased by approximately 6-7 percent.
While the B-coefficient for the Daylight Hours variable appears negative in these models,
the actual net daylight effect turns out to be positive, once the effects of the interaction
variables are taken into account. In order to express the range of the potential effect of
daylight on an individual store, we have plotted the range of predicted effects for the two
Log models in Figure 9 below.
30
RETAIL AND DAYLIGHTING ANALYSIS FINDINGS
Figure 9: Graph of Predicted Range of Daylight Effect per Store, Log Models
The Log Models find that adding daylight to stores (based on the norm of the corporate
design, or about 1090 hours of useful daylight per year, or about ¼ of the total yearly
daytime hours) will be associated with a “net daylight effect” showing an increase in
sales per Figure 10 below:
Figure 10 shows that the average net effects for the daylight interaction variables as a
group are positive for both models, and that the interaction variables are all significant as
a group (group F-test). This average effect is, however, not large enough in either case
to give certainty that it would not dip down below zero if we considered a different
population of stores in our analysis. A larger population of study sites (for example,
doubling the number of sites from 73 to 150) would have provided greater statistical
power, and would have likely provided greater certainty in the analysis.
Thus, the log models predict a chain-wide average increase in sales associated
with the presence of daylight of 1% to 6%.
31
RETAIL AND DAYLIGHTING ANALYSIS FINDINGS
60%
50%
Percent Daylight Effect
40%
30%
20%
10%
0%
-10% 0% 50% 100% 150% 200%
-20%
-30%
-40%
Store Parking Area relative to Norm
In order to understand the theoretical impact of the daylighting effect independent of the
parking interaction variable, we held parking constant. We performed this exercise at
three levels—the norm for the daylit stores, and the norm plus or minus the standard
deviation of parking for the daylit stores—and then predicted the net daylight effect for
the group of daylit stores, as shown in Figure 12. When parking is greater than norm,
the daylight effect jumps up dramatically to +20%. When parking is restricted, the
daylight is seen to have a negative effect. The nature of the linear equations used in the
regression models forces one end of the range to go negative when the other end is
strongly positive, something like a see-saw. Thus, there is less certainty about the high
or low ends of the predicted effect than the norm.
Condition 2001
Parking @ norm +5.6%
Parking @ 1 std. dev. below norm -8.7%
Parking @ 1 std. dev. above norm +19.7%
Figure 12: Daylight Effect Independent of Parking, Log 10 Month Sales, 2001
Figure 12 suggests that the daylighting effect may have its greatest advantage when
there is sufficient parking to take advantage of the additional demand created.
32
RETAIL AND DAYLIGHTING ANALYSIS FINDINGS
The 2001 model suggests that, when parking is held constant at the mean, for every
increase in 100 daylight hours per year per store, the daylight effect increases by 1%,
ranging from a low of -2% to a high of +14%. When parking is held constant at a high
level (the mean plus one standard deviation), then for every increase in 100 daylight
hours per year per store, the daylight effect increases by 2.4%, ranging from +2% to
+37%. These predictions are illustrated in Figure 13.
In the 1999-2000 model, when parking is held constant at the mean, an increase in 100
hours of daylight increases the daylight effect by 0.1%, ranging from +0.4% to +2.0%.
When parking is held constant at a high level (the mean plus one standard deviation),
then for every increase in 100 daylight hours per year per store, the daylight effect
increases by 2%, ranging from +5% to +27%. These predictions are illustrated in Figure
14. (Note that in Figure 14 the scale for one graph, “1999-2000 parking at mean model,”
has a different vertical scale, 1/10 the size of the others, in order to show detail in the
smaller effects seen in this example.)
The results for these plots suggest a “bounds” for a daylight effect. When parking is held
at norm, the daylight effect varies from a low prediction of –2% to a high of +14%
increase in sales per store. In these equations, the amount of parking becomes a
limiting factor. When we allow parking to increase up to a higher level, one standard
deviation above the norm, the prediction of the daylight effect per store ranges from +2%
to +37%. In all cases, as the number of useful daylight hours per year increases, the
relative daylight effect on sales also increases.
2001 Ln Daylight Effect, Parking Held Constant (mean) 2001 Ln Daylight Effect, Parking Held Constant (high)
50.0% 50.0%
y = 0.0002x - 0.0426
40.0% 40.0%
R2 = 0.566
30.0% 30.0%
Daylight Effect
Daylight Effect
20.0% 20.0%
y = 0.0001x - 0.0429
R2 = 0.2162
10.0% 10.0%
0.0% 0.0%
0 500 1000 1500 2000 0 500 1000 1500 2000
-10.0% -10.0%
-20.0% -20.0%
Daylight Hours Daylight Hours
Figure 13: Daylight Effect as a Function of Daylight Hours, Log 10 Month Sales, 2001
9900 Ln Daylight Effect, Parking Held Constant (mean) 9900 Ln Daylight Effect, Parking Held Constant (high)
5.0% 50.0%
4.0% 40.0%
y = 1E-05x + 0.0012 y = 0.0001x + 0.0071
R2 = 0.7946 R2 = 0.8134
3.0% 30.0%
Daylight Effect
Daylight Effect
2.0% 20.0%
1.0% 10.0%
0.0% 0.0%
0 500 1000 1500 2000 0 500 1000 1500 2000
-1.0% -10.0%
-2.0% -20.0%
Daylight Hours Daylight Hours
Figure 14: Daylight Effect as a Function of Daylight Hours, Log 24 Month Sales, 1999-
2000
33
RETAIL AND DAYLIGHTING ANALYSIS FINDINGS
Here the B-coefficients for the Daylight Hours variable are again negative, but once the
effect of the interaction variables are taken into account, the net daylight effect becomes
positive in the 10 month model. In the 24 month model it is found to be effectively zero,
as described in Figure 16 below. In order to express the range of the potential effect of
daylight on an individual store, we have plotted the range of predicted effects for the two
linear models in Figure 17 below.
34
RETAIL AND DAYLIGHTING ANALYSIS FINDINGS
The linear models find that adding daylight to stores (based on the norm of the corporate
design, or about 1090 hours of useful daylight per year) will be associated with the
following increase in sales:
Figure 16 shows that the predicted net effects are positive for the ten month model, but
slightly negative (or essentially zero) for the twenty-four month model. The interaction
variables are all significant as a group (group F -test) in both models, and so were
retained as a group. The linear and log models thus have essentially the same
prediction: that during the 10 month period during 2001 daylit stores increased their
sales relative to non daylit stores by 5-6%. During the 24 month period during 1999-
2000 the daylit stores were found to be selling at very similar levels to the non-daylit
stores (i.e., 0% to 1%).
In order to express the range of the potential effect of daylight on an individual store, we
have plotted the range of predicted effects for the two linear models in Figure 17 below.
Figure 17: Graph of Predicted Range of Daylight Effect per Store, Linear Models
Given the similarity of results, we did not attempt to describe the effect of daylight
independent of parking, as we did above with the log models.
Again, per the discussion in the log models, this average effect is not large enough to
give certainty that it would not dip down below zero (in the case of the ten month model)
or rise above zero (in the case of the twenty-four month model) if we considered a
different population of stores in our analysis. A larger population of study sites (say
doubling the number of sites from 73 to 150) would have been helpful to provide greater
certainty in the models’ predictions.
35
RETAIL AND DAYLIGHTING ANALYSIS FINDINGS
36
RETAIL AND DAYLIGHTING ANALYSIS FINDINGS
The attempt to replicate the format of the previous model was most likely unsuccessful
because of two characteristics of the stores in this study. First of all, the average
daylight effect was observed to be much smaller, i.e. closer to zero, and therefore less
likely to be found significant in a simple yes/no model. The greater sensitivity of a scalar
description of daylight, describing the number of hours per year of daylight above a
certain threshold helped to provide better resolution of the relationship between daylight
and sales.
Secondly, for this particular chain there seems to be an important interaction between
the amount of parking and any daylight effect. We do not know why this is so – we can
only hypothesize based on common sense why such an interaction might occur. There
is also always the possibility that this statistical interaction is simply a random
occurrence of how the types of stores are distributed in the data. In the previous model
we had no information about the amount of parking available at the various stores.
Parking was not considered as an explanatory variable in the original model. Thus, we
have no way to compare this parking-daylight effect with information from the previous
participant.
The use of the daylight hours variable in this set of models has allowed us to detect a
more subtle effect of daylight on sales, and also to describe a dose/response
relationship between more daylight and more sales. This dose/response relationship is
inherently more useful information to designers. It basically says that “more is better.” It
is not simply the presence of daylight at some threshold level, but progressively
increased exposure that is most useful. It is unclear, however, from this analysis
whether the increase has to do with more hours of useful daylight or higher levels of
daylight illumination, since the two go hand in hand and we cannot distinguish between
the two characteristics in the stores we studied.
37
RETAIL AND DAYLIGHTING ANALYSIS FINDINGS
The contrast hypothesis focuses on the differences between sister stores within the
same chain. But, from the corporation’s perspective, each store site is more importantly
competing with the competitor’s stores. Differences between this chain and other chains
are more likely to be important determinants of corporate success than differences within
the chain. During the California power crisis of 2001 almost all retailers in the state
agreed to operate their stores at reduced lighting power in order to conserve energy and
reduce peak loads on the state electric system. As a result, not only the study
participant but most of their competitors were also operating at reduced electric lighting
levels. Under these conditions, the daylit participant stores were even more successful
than the rest of the chain.
The competitive hypothesis suggests that under favorable economic conditions, the
daylit stores in the study were even more attractive relative to the competitors’ options
than the non-daylit stores. Could it be that when shoppers are motivated to buy more
products from the participant, they are even more motivated by the daylit stores? During
the California power crisis of 2001 almost all retailers in the state agreed to operate their
stores at reduced lighting power in order to conserve energy and reduce peak loads on
the state electric system. As a result, not only the study participant but most of their
competitors were also operating at reduced electric lighting levels. Under these
conditions, the daylit participant stores were even more successful than the rest of the
chain.
Because both events—the overall favorable economic conditions and the reduced
lighting levels—seem to have happened simultaneously, we cannot distinguish between
possible effects. Research into economic conditions that may have supported one these
hypotheses was outside the scope of this report. Of course, other possible differences
may exist between the two time periods that we did not observe or consider.
38
RETAIL AND DAYLIGHTING ANALYSIS FINDINGS
39
RETAIL AND DAYLIGHTING ANALYSIS FINDINGS
Interaction Variables
The interaction variables are also of interest. It was clear from the start of the study that
this retailer had a much greater variety of store conditions than our previous study
participant, and that a number of store characteristics were strongly associated with the
presence of daylight. The interaction variables account for these interactions and
moderate the effect of daylight relative to the presence of these other influences on
sales that are correlated to daylight. The interaction variables certainly make for a more
complex model, but also help us describe a more nuanced reality.
The interaction of parking with daylight hours is perhaps the least expected. As a simple
variable, more parking, relative to all other variables, is seen to have a negative effect on
sales. However, when we add in the interaction variable parking*daylight, more parking
increases the positive effect of more daylight. This is true in both the log and linear
models. This suggests that the negative effect of more parking is slightly overstated in
the simple variable, and is moderated in the daylit stores.
40
RETAIL AND DAYLIGHTING ANALYSIS FINDINGS
Figure 18 shows that the chain-wide average net effect of daylight is positive for
both models, ranging from a 1% to 2% increase in number of transactions.
The interaction variables are all significant as a group (group F-test) in both models, and
so were retained as a group. The magnitude of the predicted increase in transactions is
modest, and somewhat less than the prediction for the increase in sales for the
comparable linear 10-month model. Thus it is likely that the “daylight effect” is working
both to increase the amount of traffic through the stores, as evidenced by the increase in
the number of transactions, and also to increase the value of each set of purchases, as
evidenced by the relatively greater sales effect than transaction effect.
Somewhat surprisingly, the models maintained an almost identical format as the sales
models, with all of the same explanatory variables being retained, with the exception of
the demographic variables. In the sales transaction models, the demographic variable
housing was significant for both the ten month and twelve month models, and the
variables for population growth and transportation were significant in only the ten month
model. The inconsistency of the demographic variables once again argues that they are
slightly less reliable predictors of sales (or in this case, number of transactions) than the
other variables which are consistent across all models. The consistency of daylight in
41
RETAIL AND DAYLIGHTING ANALYSIS FINDINGS
predicting a positive effect based on a different outcome variable once again increases
our confidence that it is likely to be a true effect.
We did go through the exercise of isolating the effect of daylight independent of its
interaction with size of parking, as we did above with the log sales models, and found a
similar pattern; when parking is at or above norm, an increase in the number of useful
daylight hours per stores is also associated with an increase in the percentage effect on
transactions.
Again, per the discussion in the log sales models, this average effect is not large enough
to give certainty that it would not dip down below zero if we considered a different
population of stores in our analysis. A larger population of study sites (say doubling the
number of sites from 73 to 150) may have provided greater certainty in the models’
predictions.
42
RETAIL AND DAYLIGHTING OTHER STUDY FINDINGS
In addition to the regression analysis of sales and number of transactions per store,
which form the core of this research, we also looked at the potential energy impacts of
the daylighting system and assessed employee and store manager satisfaction with the
daylighting design. We assessed the energy impacts to quantify the rather predictable
dollar value of energy savings due to skylights combined with automatic photocontrols,
which will reliably occur in addition to any sales impact. We looked at employee and
manager satisfaction with the daylighting as a way to try to get insight into the causal
mechanisms of any daylighting effect, and also to identify any problems that might be
associated with the daylighting systems in the stores.
The analysis of energy impacts were based on both interviews with the corporate
management and our own estimates of energy savings based on store characteristics
and operation schedules. The energy estimates are not based on monitoring. The
assessment of employee and manager satisfaction with the daylighting system was
based on interviews with the managers and a formal survey distributed to employees.
The following sections present our findings in these two areas, and discuss their
relevance to the overall study.
43
RETAIL AND DAYLIGHTING OTHER STUDY FINDINGS
We calculated the energy savings from the current design and operation and then
gradually increased the optimum performance of the skylight and photocontrol system
heading towards a theoretical maximum performance. We found that the current system
(good) is saving about $.24/sf for an average store in the chain, while an improved
system (better) using current best-practices could save about $.54/sf, and an optimum
system (best) using state-of-the-art performance could save about $.66/sf at current
energy prices. Thus, the current daylight design is saving about one-third of the
maximum amount of energy that could potentially be saved from daylighting.
1
Brooks, M. 2002 “California Electricity Outlook: Commercial Building Systems” Presentation at PIER
Buildings Program HVAC Diagnostics Meeting, Oakland, CA, April 16.
2
Armstrong Industries, 2002, private communication.
3
These values are based on SkyCalc® runs, which account for lighting, heating and cooling savings, and
combine the net annual value of electricity and gas impacts into a blended kWh value.
4
It is not possible to translate megawatthours into peak megawatt impacts, since the dynamics of climate
and electric peaks greatly complicate the equation. A separate study should be done to understand the
potential peak impacts of skylighting systems on state power demand.
44
RETAIL AND DAYLIGHTING OTHER STUDY FINDINGS
45
RETAIL AND DAYLIGHTING OTHER STUDY FINDINGS
The biggest specific differences reported between the two store types were that:
Daylit stores overall were less uncomfortably dim (8%)
Daylit stores had fewer problems with uniformity (9%)
When these surveys were taken, the daylit stores tended to have higher illumination
levels than the non-daylit stores, and indeed employees reported fewer problems with
the stores being “uncomfortably dim”.
It is especially interesting that the employees reported that the daylit stores had
better uniformity by 9% (“There is too much light in some areas and not enough in
others”), even though we measured substantially bigger variation in illumination
levels in the daylit stores than in the non-daylit stores (the standard deviation of
both horizontal and vertical illumination was four times greater in the daylit
stores).
46
RETAIL AND DAYLIGHTING OTHER STUDY FINDINGS
47
RETAIL AND DAYLIGHTING OTHER STUDY FINDINGS
48
RETAIL AND DAYLIGHTING CONCLUSIONS AND DISCUSSION
This study represents evidence that a major retailer is experiencing higher sales in daylit
stores compared to similar non-daylit stores. In addition to replicating the findings of the
previous study, this study added other important dimensions of information:
• Average effect of daylighting on sales for all daylit stores in this chain was variously
calculated at 0% to 6%, depending on the type of model and time period considered.
• A dose/response relationship was found, whereby more hours of useful daylight in a
store are associated with a greater daylight effect on sales.
• A bound of an empirical daylight effect for this chain was detailed, with a maximum
effect found in the most favorable stores of about a 40% increase in sales. This
upper bound is consistent with our previous finding.
• Daylight was found to have as much explanatory power in predicting sales (as
indicated by the partial R2) as other more traditional measures of retail potential,
such as parking area, number of local competitors, and neighborhood
demographics.
• Along with an increase in average monthly sales, the daylit stores were also found
to have 1-2% increase in the number of transactions per month.
• No seasonal patterns to this daylight effect were observed.
• The value of the energy savings from the daylighting is far overshadowed by the
value of the predicted increase in sales due to daylighting: by the most conservative
estimate of at least 19 times, and more likely, under current conditions, by 45-100
times.
• During the California power crises, when almost all retailers in the state were
operating their stores at half lighting power, the stores in this chain with daylight
were found to benefit dramatically, with an average 5.5% increase in sales relative
to the other stores in the chain (which also increased their sales compared to the
previous period).
• Employees of the daylit stores reported slightly higher satisfaction with the lighting
quality conditions overall than those in the non-daylit stores. Most strikingly, they
perceived the daylit stores to have more uniform lighting than the non-daylit stores,
even though direct measurements showed the daylight stores to have much greater
variation in both horizontal and vertical illuminance levels.
• Store managers did not report any increase in maintenance attributable to the
skylights.
• The chain studied was found to be saving about $0.24/sf per year (2003 energy
prices) due to use of photocontrols, which could potentially increase up to $0.66/sf
per year with an optimized daylighting system.
We were allowed extraordinary access to the store sites and employees, and have had
the opportunity to review the methodology, findings and conclusions with the retailer
management. Thus, we have high confidence in the validity of source data, and the
reasonableness of these findings.
49
RETAIL AND DAYLIGHTING CONCLUSIONS AND DISCUSSION
50
RETAIL AND DAYLIGHTING CONCLUSIONS AND DISCUSSION
1
N Eklund, P Boyce, SN Simpson, “Lighting and Sustained Performance,” Journal of the IESNA, Vol. 29, No
1, Winter 2000, p 116-130.
2
Heschong Mahone Group, Daylighting and Student Performance, PG&E 1999.
51
RETAIL AND DAYLIGHTING CONCLUSIONS AND DISCUSSION
Since we had more information about each store, in this study, we were able to test a
number of different ways of defining the daylight presence in the stores. In addition to
the Daylight indicator variable (yes/no), we tried using indicator variables for Vertical
Glazing (yes/no), and Partial Daylight (yes/no). Ultimately, putting all of the stores on a
single scalar of daylight hours per year seemed to be the most robust approach, and the
most useful in providing design guidance.
The presumption here is that, if the presence of daylight is a positive attribute in a store,
then more hours of daylight per year will provide a greater benefit. As discussed earlier,
in the skylight design typically used in these stores, hours of daylight per year and
intensity of daylight at peak periods are strongly related: i.e. the more intense the
daylight at peak periods, the more hours of useful daylight per year. However, this
relationship is not necessarily a given. It is possible to design skylighting systems that
increase useful daylight during morning and evening hours by increasing the penetration
of low angle sunlight, while reducing peak intensities by reducing the transmission of
high angle sun. While it is our hypothesis that hours of useful daylight per year, rather
than average or peak daylight intensity, will likely be a more useful metric for evaluating
daylighting systems, we cannot distinguish between the two in this analysis.
52
RETAIL AND DAYLIGHTING CONCLUSIONS AND DISCUSSION
of the sales space and products, as evidenced in the perceptions of the shoppers
discussed above.
Two findings of this study suggest that higher illumination levels may be partially
responsible for the daylight effect. First of all, more daylight hours per year was directly
associated with greater sales. In skylit stores, more daylight hours per year will also be
directly associated with higher average illumination levels—since as daylight illumination
levels increase there are more hours per year above the design illumination threshold.
Secondly, we found that the relative daylight effect was greater during the 10-month
period when the average electric illumination levels were lower for all stores, thus the
difference in daytime illumination between daylit and non-daylit stores was greater.
During this period, since the electric lighting levels were constant for all stores, the daylit
stores tended to have slightly higher horizontal illumination levels during all daylight
hours.
Illumination levels in the daylit stores is a dynamic condition. Our measurements were
based on hand held illumination readings taken during the late winter months generally
between 10 am and 2 pm. We observed daylit stores to average 66% higher horizontal
illumination than the non-daylit stores during the February-March mid-day period of our
measurements. SkyCalc projections suggest that it would average about 133% higher
during the summer months.
Figure 20 compares the SkyCalc estimated average horizontal illuminance during the
two time periods, 10-month and 24-month, and during peak daylight conditions (July)
and minimum daylight conditions (December) for two average stores in the chain, one
daylit (dashed lines) and one not daylit (solid lines). During the 10-month time period, all
stores operated their electric lights at about ½ power, so that became the minimum
illumination in daylit stores, and the additional daylight contribution always increased
illumination levels during the day. During the 24-month period, all stores operated at full
power, but daylit stores turned off some of their lights during the daytime when threshold
daylight illumination conditions were met.
Hourly Illumination Levels, 10m Hourly Illumination Levels, 24m
Relative Illuminance
Relative Illuminance
Figure 20: Hourly Illumination Patterns of Average Daylit and Non-Daylit Stores, 10-
month (left) and 24-month (right) periods
Based on our understanding of the photocontrol operations controlling the electric lights,
and SkyCalc analysis of hourly daylight levels, we believe that during the 24-month
period the daylit stores operated above the non-daylit stores’ illumination level for about
20% of the store hours. Another 40% of the time, they would have been operating at
slightly less than the non-daylit stores’ illumination level, and the remaining 40% of the
time at the same level. Thus, increased horizontal illumination may not be solely
responsible for the increased sales due to a daylight effect.
53
RETAIL AND DAYLIGHTING CONCLUSIONS AND DISCUSSION
Improved Color Rendition: Daylight has the greatest range of spectral wavelengths of
any of our light sources, and is the source our eye has naturally adapted to through
millenniums of evolution. With a continuous spectrum of light, all colors will be more
vivid and have a more naturally rendered appearance under daylight. Any store which
sells products distinguished by color, or where color is a key selection criteria, is likely to
benefit from improved color rendition. This particular retailer, as well as the previous
study participant, does sell some products that would benefit from better color
discrimination. It is possible that there are fewer returns due to poor color selection and
greater attraction to products due to more vivid color rendition. In this study, we did not
collect evidence on this issue.
Greater Depth Perception: There is a second possible effect of the spectrum of
daylight, due to its greater richness in the blue end of the spectrum. One current theory
suggests that people perceive a space to be more brightly lit, and that the resulting size
of their pupil is smaller, under so-called “scotopic” sources of light, those rich in the blue
end of the spectrum. The smaller pupil size is likely to increase the depth of field of the
viewer, allowing a shopper to see greater detail over a wider range of focal distances.
Improved Lighting Quality: The daylit stores had consistently higher illumination levels
on vertical surfaces. During our daytime observations, average vertical illuminance
readings were 62% higher in the daylit stores, and the average minimum readings
(average-standard deviation) were 50% higher. The lowest vertical illumination reading
found in any daylit store was four times higher than the lowest reading in any non-daylit
store.
We examined a set of photographs of a sample of comparable daylit and non-daylit
stores looking for obvious changes in appearance or visual quality that we could not
capture with our surveyors’ light meter readings. The appearance of the product in the
aisles was remarkably similar between the two types of stores. Shadows and highlights
on products appeared very similar.
The most obvious differences between the two store types were in the appearance of
the floor, ceiling and upper walls. The floor in daylit stores was more likely to have
subtle shadows and changes in color, due to alternating patterns of skylight versus
electric light dominating the aisles, whereas the floors in the non-daylit stores were very
uniform in appearance. Similarly the ceiling in the daylit stores might best be described
as “lively,” with daylight bouncing off of many surfaces, compared to a much more
uniform appearance of the ceilings of the non-daylit stores. In the non-daylit stores the
uniform pattern of bright electric lights tended to accentuate the contrasting darkness of
the non-illuminated ceiling. Even when ceilings were the same height, the daylit stores
seemed to have higher ceilings since the ceiling surfaces were more brightly lit,
attracting the viewer’s eye upwards. The daylit stores also tended to have higher
illumination levels on the upper walls and high-mounted signage, creating a bright
horizon and perhaps helping with shopper navigation through the store.
Greater Variability: The range of illumination levels in daylit stores is considerably
greater than in non-daylit stores, and yet the employees judged the daylit stores to have
slightly better lighting quality, and substantially greater lighting uniformity, than the non-
daylit stores. The standard deviation for both vertical and horizontal illumination readings
was four times greater in the daylit versus the non-daylit stores. While uniformity of
illumination is often a key goal of lighting designers, the non-uniformity of the daylit
stores may actually be a positive feature.
54
RETAIL AND DAYLIGHTING CONCLUSIONS AND DISCUSSION
Daylight variability is a function of both space and time. The variation in illumination
levels for daylit stores over time is even greater than the variation in space described
above. Highlighted areas move around the store as the sun moves through the day and
seasons. A daylit store looks different in the morning than in the afternoon, and different
in the summer than in the winter, with slightly changing color of light and patterns of soft
shadows. Such variation in appearance may help to stimulate shopper interest over
repeated visits, and may also help to keep employees alert and mentally engaged over
time.
Connection to the Outdoors: Daylit stores offer some information about the weather
outside of a store. Most of the stores in this chain had diffuse skylights that offered no
view of the sky, but the intensity of daylight or the passing of clouds is clearly
discernable. Rain and hail can be heard through the skylights. This information about
weather conditions outside may also have a stimulating effect, as discussed above
under “Greater Variability.”
Biological Effects: There is a growing body of research being undertaken to
understand the biological effects of light. To date it is clear that bright light exposure
during the day helps reinforce the natural circadian rhythms of various hormones and
neural transmitters. Bright light, most likely in the natural daylight spectrum favoring blue
wavelengths, suppresses melatonin and increases the production of neural transmitters
such as serotonin and dopamine. Healthy circadian rhythms have been implicated in
better immune function and activity patterns. The illumination intensity, duration, timing
and spectral sensitivity of these effects are still being researched, and may likely vary by
time of day. We believe these are fairly slow biochemical effects, unlikely to be
detectable in less than 30 minutes, and most likely to be reinforced over days of
exposure. Thus, short store visits by shoppers are not likely to involve biological effects
from daylight exposure. Store employees, on the other hand, may experience some
health benefits from working in a daylit environment that reinforces their circadian
rhythms. We did not study any indicators that would identify any of these effects. We
hypothesize, however, that positive feelings generated from daylight exposure could
possibly reinforce general positive feelings towards the store for both shopper and
employees.
55
RETAIL AND DAYLIGHTING CONCLUSIONS AND DISCUSSION
56
RETAIL AND DAYLIGHTING APPENDICES
9. APPENDICES
57
RETAIL AND DAYLIGHTING APPENDICES
58
RETAIL AND DAYLIGHTING APPENDICES
Illuminance Measurements
Measurement Time Horiz @ L vert @ L vert @ L vert @ R vert @ R vert @ R vert @
4 feet
location 2 feet 4 feet 6 feet 2 feet 4 feet 6 feet
Wide aisle u/s
Wide aisle b/s
Narrow aisle u/s
Narrow aisle b/s
Back aisle u/s
Back aisle b/s
Checkout 1 2 3
n/a n/a n/a n/a
Legend: u/s=under skylights or if no skylight exists, b/s = between skylights
Reflectances
Floor A B Other
Upper Walls A B Other
Ceiling A B Other
Structure A B Other
Cleanliness very cln. clean normal dirty very dirty
1 2 3 4 5 Comment
Luminaires
Verify Verify mounting height and spacing with plan survey.
Layout A B C D Other
Flicker Yes No Hum Yes No
Cleaning Clean Dirty
Controls A B C D Other
Skylights no skylights (if no skylights, ignore the following 5 questions)
Type A B Other
Well Depth A B C D E Other
Cleaning Clean Dirty Discoloring Yes No
Obstructions A B C D Other
Thermal Environment
Heating A B Dry bulb temp. °C or °F
Air movement A B Air velocity ft/min or m/s
Cooling A B Cycling Yes No
Aural Environment silent quiet normal loud very loud
Ambient 1 2 3 4 5 Decibels
PA 1 2 3 4 5 Decibels
Air 1 2 3 4 5 Decibels
Primary Noise Source HVAC Street Chckout Patrons Other Location
Olfactory Environment very pleasant neutral very unpleasant
Pleasantness 1 2 3 4 5 Describe
59
RETAIL AND DAYLIGHTING APPENDICES
60
RETAIL AND DAYLIGHTING APPENDICES
For question 12, please circle the number that most closely matches your opinion of the lighting in this store
compared to other stores with which you are familiar.
l) How does the lighting in this store compare to lighting in similar stores?
worse the same better
1 2 3 4 5 6 7
61
RETAIL AND DAYLIGHTING APPENDICES
62
RETAIL AND DAYLIGHTING APPENDICES
Percent Population Growth, 2000-1990 PopGrow 1.00 13.18 12.18 5.09 2.55
Number of sister stores within certain radius Co-mktg 1.00 5.00 4.00 3.93 1.26
Number of competitor stores within radius 1 Compet 1 0.00 10.00 10.00 2.40 2.22
Storefront height scalar Height 1.00 3.57 2.57 1.91 0.47
ln(Parking) logPark 1.00 1.29 0.29 1.17 0.07
Daylit hours per year greater than threshold DayHrs 270.00 1800.32 1530.32 1090.55 408.86
Parking * DayHrs ParkDH 447.95 4557.73 4109.77 2654.69 1289.44
Number of sister stores within certain radius Co-mktg 0.07 0.02 3.38 0.001
Number of competitor stores within radius 1 Compet 1 -0.05 0.02 -2.81 0.007
Storefront height scalar Height -0.01 0.00 -2.27 0.027
ln(Parking) logPark -0.41 0.08 -4.94 0.000
Daylit hours per year greater than threshold DayHrs 0.00 0.00 -2.50 0.015
Age * DayHrs AgeDH 0.00 0.00 -1.71 0.092
Parking * DayHrs ParkDH 0.00 0.00 3.63 0.001
Model Summary:
RMSE 0.17
R^2 75.7%
63
RETAIL AND DAYLIGHTING APPENDICES
Number of sister stores within certain radius Co-mktg 0.091 0.02 3.85 0.000
Number of competitor stores within radius 1 Compet 1 -0.056 0.02 -2.97 0.004
Storefront height scalar Height -0.161 0.07 -2.34 0.023
ln(Parking) logPark -1.823 0.41 -4.41 0.000
Daylit hours per year greater than threshold DayHrs -0.001 0.00 -3.06 0.003
Parking * DayHrs ParkDH 0.00024 0.00 3.20 0.002
Model Summary:
RMSE 0.19
R^2 74.7%
Model Name: LN 01
Variable Description Variable Order of Entry Partial r^2
ln(Total Area) logArea 2 0.069
ln(Age) logAge 1 0.379
Figure 24: Order of Entry and Partial R2, Log 10 Month Sales, 2001
64
RETAIL AND DAYLIGHTING APPENDICES
Figure 25: Order of Entry and Partial R2, Log 24 Month Sales, 1999-2000
65
RETAIL AND DAYLIGHTING APPENDICES
Figure 26: Summary Statistics for All Variables Considered in Linear Models
66
RETAIL AND DAYLIGHTING APPENDICES
Number of sister stores within certain radius Co-mktg 180.89 52.71 3.43 0.001
Number of competitor stores within radius 1 Compet 1 -122.49 42.79 -2.86 0.006
Storefront height scalar Height -416.25 150.17 -2.77 0.007
Parking scalar Parking -579.25 105.31 -5.50 0.000
Daylit hours per year greater than threshold DayHrs50 -1.41 0.43 -3.25 0.002
Age * DayHrs AgeDH -0.08 0.05 -1.73 0.089
Parking * DayHrs ParkDH 0.73 0.18 4.22 0.000
Model Summary:
RMSE 439.95
R^2 76.5%
Number of sister stores within certain radius Co-mktg 217.31 56.38 3.85 0.000
Number of competitor stores within radius 1 Compet 1 -129.86 45.18 -2.87 0.006
Storefront height scalar Height -388.92 161.60 -2.41 0.019
Parking scalar Parking -516.95 111.11 -5.61 0.000
Daylit hours per year greater than threshold DayHrs -1.57 0.45 -3.47 0.001
Parking * DayHrs ParkDH 0.64 0.18 3.47 0.001
Model Summary:
RMSE 474.65
R^2 75.3%
67
RETAIL AND DAYLIGHTING APPENDICES
Figure 29: Order of Entry and Partial R2, Linear 10 Month Sales, 2001
Figure 30: Order of Entry and Partial R2, Linear 24 Month Sales, 1999-2000
68
RETAIL AND DAYLIGHTING APPENDICES
Number of sister stores within certain radius Co-mktg 2.80 1.05 2.65 0.010
Number of competitor stores within radius 1 Compet 1 -3.87 0.83 -4.65 0.000
Storefront height scalar Height -9.35 2.80 -3.34 0.001
Parking scalar Parking -11.77 2.04 -5.77 0.000
Daylit hours per year greater than threshold DayHrs50 -0.0255 0.01 -3.07 0.003
Age * DayHrs AgeDH -0.0018 0.00 -1.96 0.054
Parking * DayHrs ParkDH 0.0131 0.00 3.95 0.000
Model Summary:
RMSE 8.44
R^2 77.2%
Number of sister stores within certain radius Co-mktg 3.61 1.16 3.10 0.003
Number of competitor stores within radius 1 Compet 1 -3.95 0.91 -4.33 0.000
Storefront height scalar Height -7.40 3.31 -2.24 0.029
Parking scalar Parking -10.87 2.24 -4.84 0.000
Daylit hours per year greater than threshold DayHrs -0.04 0.01 -3.88 0.000
Parking * DayHrs ParkDH 0.01 0.000 3.87 0.000
Model Summary:
RMSE 9.51
R^2 75.2%
69
RETAIL AND DAYLIGHTING APPENDICES
70
RETAIL AND DAYLIGHTING APPENDICES
71
RETAIL AND DAYLIGHTING APPENDICES
72