Content uploaded by Angeli Kirk
Author content
All content in this area was uploaded by Angeli Kirk on Jun 25, 2015
Content may be subject to copyright.
Chapter 20
Comparing Cookstove Usage Measured
with Sensors Versus Cell Phone-Based
Surveys in Darfur, Sudan
Daniel Lawrence Wilson, Mohammed Idris Adam, Omnia Abbas,
Jeremy Coyle, Angeli Kirk, Javier Rosa and Ashok J. Gadgil
Abstract Three billion people rely on combustion of biomass to cook their food,
and the resulting air pollution kills 4 million people annually. Replacing inefficient
traditional stoves with “improved cookstoves”may help reduce the dangers of
cooking. Therefore analysts, policy makers, and practitioners are eager to quantify
adoption of improved cookstoves. In this study, we use 170 instrumented cook-
stoves as well as cellphone-based surveys to measure the adoption of free-of-charge
Berkeley-Darfur Stoves (BDSs) in Darfur, Sudan where roughly 34,000 BDS have
been disseminated. We estimate that at least 71 % of participants use the stove more
than 10 % of days that the sensor was installed on the cookstove. Compared to
sensor-measured data, surveyed participants overestimate adoption both in terms of
daily hours of cooking and daily cooking events (p< 0.001). Average participants
overreport daily cooking hours by 1.2 h and daily cooking events by 1.3 events.
These overestimations are roughly double sensor-measured values. Data reported
by participants may be erroneous due to difficulty in recollection, courtesy bias, or
the desire to keep personal information obscure. A significant portion of sensors
was lost during this study, presumably due to thermal damage from the unexpected
commonality of charcoal fires in the BDS; thus pointing to a potential need to
redesign the stove to accommodate users’desire to cook using multiple fuel types.
The cooking event detection algorithm seems to perform well in terms of face
validity, but a database of cooking logs or witnessed accounts of cooking is absent;
D.L. Wilson (&)!J. Coyle !A. Kirk !J. Rosa !A.J. Gadgil
University of California, Berkeley, CA, USA
e-mail: dlwilson@berkeley.edu
M.I. Adam
Al-Fashir University, Darfur, Sudan
O. Abbas
Potential Energy, Berkeley, CA, USA
A.J. Gadgil
Lawrence Berkeley National Laboratory (LBNL), University of California,
Berkeley, CA, USA
©Springer International Publishing Switzerland 2015
S. Hostettler et al. (eds.), Technologies for Development,
DOI 10.1007/978-3-319-16247-8_20
211
the algorithm should be trained against expert-labeled data for the local cooking
context to further refine its performance.
20.1 Introduction and Purpose
Worldwide, 3 billion people rely on combustion of biomass to cook their food
(World Bank 2011). Resulting indoor and outdoor air pollution kills 4 million
people every year, making cooking smoke one of the world’s greatest environ-
mental health problems (Lim et al. 2013). Many efforts to reduce the dangers of
cooking smoke focus on replacing inefficient, high-emission, traditional stoves with
“improved cookstoves,”yet widespread and sustained adoption remains elusive,
and impacts are poorly understood.
Analysts, policy makers, and practitioners are eager to know which cookstoves
and marketing approaches can increase user-acceptance and long-term adoption.
Evidence remains limited in part because it has been difficult to collect accurate and
ample data on cookstove use. Because intensive direct observation can be expen-
sive, onerous, and alter user behavior (Landsberger 1958); most studies of cook-
stove adoption rely on infrequent survey-based self-reported data (Lewis and
Pattanayak 2012; Burwen 2011). These data are subject to two types of error: First,
respondents may overreport normative behaviors, generating “social desirability”or
“courtesy”bias (Edwards 1957; Nunnally 1978). Overreporting use leads to sys-
tematic overestimates of adoption, which leads to systematic downward bias on
estimates of any positive impacts of adoption. Even without intentional misrep-
orting, respondents can struggle to recall and aggregate over large periods, but more
frequent visits are prohibitively expensive at scale and onerous for participants.
Even if recalled information is unbiased on average, any measurement error in the
dependent variables (say, fuel expenditures) reduces the precision of estimates
leading to reduced statistical significance (Das et al. 2012).
Given these challenges and needs, the ability to monitor cookstoves’use with
objective and unobtrusive means is critical to improving products, process quality,
and properly measuring impact. Some studies have begun to implement time-and-
temperature logging stove use monitors (SUMs) which can measure cooking events
without input from the user (Berkeley Air Monitoring Group 2013; Burwen and
Levine 2012; Ruiz-Mercado et al. 2011,2013; Thomas et al. 2013). Partnering
sensors with traditional surveys allows estimates of the magnitude of error in self-
reported usage. In Guatemala, Ruiz-Mercado et al. (2013) instrumented improved
cookstoves with SUMs in 80 households every other month for 32 months. They
find high usage (around 90 %) and relative consistency between survey and sensor-
generated data. The Berkeley Air Monitoring Group (2013) found similarly accu-
rate self-reported data for 25 cookstove users in Kenya. In Rwanda, Thomas
et al. (2013) installed 27 SUMs on cookstoves and rotated them through 97
households for an average of 9.8 days each, staggered over 5 months. They find
212 D.L. Wilson et al.
SUMs-measured usage (73 % of users adopt) to be significantly lower than
self-reported data (90 %).
Primary weaknesses in the current SUMs literature fit into four categories: (1)
Small sample sizes or short experimental durations, (2) high exposure of users to
interactions with research staff leading to potential behavior modifications, (3) a
lack of SUMs-supported studies measuring factors influencing adoption, and (4) the
general dearth of SUMs data needed to understand a problem as complex and
contextually heterogeneous as cookstove adoption. The Berkeley Air (2013) study
employed only 25 participants, and the Thomas et al. (2013) study follows stove
users for only 2 weeks. In the Ruiz-Mercado et al. (2011,2013) and the Thomas
et al. (2013) studies, relatively high levels of adoption were observed, but this was
in the context of significant exposure of study participants to field staff. Thomas
et al. (2013) found small but statistically significant declines in usage over the
course of the 2 weeks following sensor implementation, potentially due to the
saliency of observation [as seen for exercise in Prestwich et al. (2009) and savings
in Karlan et al. (2010)]. To date, SUMs have been implemented at small scales and
in relatively few contexts. Cookstove adoption is a contextually heterogeneous
problem owing to the high variability of cooking environments (e.g., cultural,
economic, and culinary factors). Therefore, regardless of shortcomings of previous
work, the literature contains a dearth of objective information on cookstove
adoption objectively measured by sensors.
In this study, we seek to improve upon previous work while adding critical
information to the small pool of studies objectively measuring cookstove adoption
with SUMs. We add significantly to the existing literature by measuring: (1)
Adoption of cookstoves in an internally displaced people’s (IDP) camp context and
(2) the correlation between user-reported and sensor-measured cookstove adoption
both in terms of number of cooking events and hours spent cooking per day. The
Berkeley-Darfur Stove (BDS) is the subject cookstove of this study. Scientists and
students at University of California, Berkeley and Lawrence Berkeley National
Laboratory (LBNL) developed the BDS for assembly, dissemination, and use in the
Darfuri cooking context. Potential Energy, a nonprofit headquartered in Berkeley,
manages the implementation of the BDS. Going by the Arabic nickname “5-Minute
Stove,”the BDS is valued by customers for being a fuel-efficient and fast-cooking
stove. Between 2009 and 2013, more than 34,000 BDSs were distributed in North
Darfur to rural, urban, and internally displaced households. About 85 % of these
cookstoves, including those employed in this study, were disseminated free of charge.
20.2 Design and Methods
The Darfur SUMs experiment involved 180 women within the Al-Salam IDP camp
just outside of Al-Fashir, North Darfur. Al-Salam is made up of five “administrative
units”that represent the geographical origin (before moving into the camp) of the
20 Comparing Cookstove Usage Measured …213
residents within the units. The 180 women were selected by the typical means of
BDS dissemination in the IDP camps: a coordinating meeting was held with the
chief Omdas (leaders) of the each of the five administrate units, and each Omda was
asked to select women for the study from a master list of residents within their unit.
Each Omda was instructed to select 36 women that would participate in the study
and that would be available during the period of surveys (i.e., would not be moving
away from the camp soon). Chief Omdas were provided with the schedule of the
surveys and informed selected participants of the dates of baseline and follow-up
surveys. Both baseline surveys and follow-up surveys took place in a women’s
center within Al-Salam camp. Enumeration teams performed a baseline survey of
household demographics and cooking practices at the time of BDS dissemination.
Dissemination took place over 5 days (one administrative unit per day) between 28
July and 1 August 2013. Subsequently, starting in late August, one administrative
unit was followed up with and given a follow-up survey roughly every 2 weeks
until late October 2013. At the time of the follow up, SUMs data were downloaded
from instrumented stoves. No training on stove use, urging, or inducement to use
the stove was administered during the follow up; women simply answered ques-
tions such as “how many times per week do you use your stove?”while in the
presence of roughly 35 peers who were queued to individually answer the same
questions with one of three enumeration teams. Participants were interviewed
individually, but potentially within earshot of one another.
Of the 180 participants, 170 women had instrumented cookstoves. Ten cook-
stoves were left uninstrumented in the case that some participants declined to
consent to participate in the research at the time of dissemination (none declined).
SUMs fit into one of three functional categories: Primary SUMs that sample as fast
as possible while still collecting data over the full deployment period (i.e., not
running out of memory), moderate sampling rate SUMs that sample once every
3 min, and fast sampling “piggyback”redundant sensors that validate data from a
primary SUM. 190 SUMs were mounted to the 170 instrumented cookstoves with
20 cookstoves having two SUMs each (one primary and one “piggyback”) and the
remaining 150 cookstoves having one SUM and one “dummy SUM”that appeared
identical to a true SUM, but contained no sensor. Following previous work by Ruiz-
Mercado, Berkeley Air, and Levine, the sensor within the SUM, shown in Fig. 20.1,
Fig. 20.1 SUM assembly showing the primary case, iButton, spring, and cap
214 D.L. Wilson et al.
was Maxim Integrated’s DS1922E high-temperature iButton with 8 kB of memory
and a temperature logging range of 15 °C–140 °C in 0.5 °C increments.
As pictured in Fig. 20.2, baseline and follow-up surveys were administered by
teams of two enumerators who redundantly recorded responses on paper and on a
cell phone platform running Open Data Kit (ODK) (Hartung et al. 2010). ODK
would send original survey data back to Berkeley in real time while enumerators
performed data entry at a later date (also using ODK) for paper surveys. One
peculiar issue encountered during the study was the inability to access and use the
Google Cloud Computing-based variant of ODK Aggregate from the study loca-
tion. As Sudan is on the list of sanctioned nations managed by the United States
Office of Foreign Assets Control, many cloud-based technologies are inaccessible.
Our solution was to use an Aggregate instance deployed at Berkeley to commu-
nicate and collect data from our devices in the field. This scheme also allowed us to
retain complete control over the survey data and to reduce network performance
issues occasionally encountered with Google AppEngine-based instances of ODK
Aggregate. Other possible solutions could have included deploying a local server in
country or using a virtual private network (VPN), however this would have required
substantial additional training. Upon retrieving survey data, entries were hand-
checked for obvious typographical errors (e.g., if the original ODK survey listed a
woman’s age as 23 and the data entry version listed it as 233, the data entry version
was manually corrected). Nonobvious incongruences were not manually corrected.
SUMs data were analyzed over the period from the first midnight after a Unit’s
last baseline survey until three midnights before a Unit’sfirst follow up survey. The
padding preceding the follow up survey is to ignore pre-survey courtesy uses.
Fig. 20.2 Women participate
in enumeration activities
while BDSs wait to be taken
home. Human subjects’faces
have been hidden;
enumerators’faces are shown
20 Comparing Cookstove Usage Measured …215
SUMs data were labeled as cooking or not cooking using the following algorithm:
First, minimums and maximums were calculated and labeled using a moving
window of approximately 30 min. Maxima not more than 5 °C above ambient were
eliminated. Points not more than 1 °C above the local minima were also labeled as
minima. The initial set of extrema identified by these rules can be seen on an
example day in Fig. 20.3a1. Next, each sensor was processed going forward
through time to eliminate spurious extrema with the following rules: (1) To elim-
inate maxima due to heating not caused by cooking in the stove, maxima less than
5°C above the previous minima, and maxima with ramp rates less than 0.2 °C/min
relative to the previous minima were eliminated; (2) To eliminate minima due to
fluctuations in cooking temperature expected when cooking with biomass, minima
that were within 20 % of the previous maxima (relative to the previous minima)
were eliminated. The remaining extrema are shown in Fig. 20.3a2. Third, runs of
consecutive minima and runs of consecutive minima were consolidated to identify
the start and end of cooking events, respectively. Runs of maxima were consoli-
dated by selecting the last maxima in a run not less than 80 % of the highest
maxima (relative to the previous minima), and runs of minima were consolidated by
selecting the last minima in a run that was either not more than 10 °C above the
Fig. 20.3 Boxes (a1–a3)
illustrate the event detection
algorithm over a one-day
period for a particular SUM.
Ambient temperature is
shown as a black line,5°C
above ambient is shown as a
dashed black line. For fully-
processed data shown in
figure (a3), non-cooking
events are shown as in purple
dots, and cooking events in
orange dots
216 D.L. Wilson et al.
lowest maxima or not more than 10 °C above ambient. Finally, all points from each
minimum to the next maximum were labeled as cooking. The final labeling of
cooking and not cooking is shown in Fig. 20.3a3.
Generally speaking, these algorithm labels event when the cookstove’s tem-
perature is increasing rapidly due to a fire or times where the cookstove is cooling
intermittently before a refuel. This algorithm is conservative in terms of cooking
duration and may overestimate the number of individual cooking events especially
for cooking behaviors where the stove is allowed to cool substantially (>80 % of the
difference between peak temperature and ambient) before fuel is reloaded. Applying
this algorithm to data from all sensors, the resulting labels of cooking events had
good face validity and no obvious systematic errors.
20.3 Results
Primary cooks in the SUMs experiment had ages ranging from 18 to 75 with a
median age of 34.5. Household size, defined as “number of people who eat from the
same pot,”ranged from 3 to 20 with a median of 7 members. Most households had
similar numbers of men/women and boys/girls (under 14) with a median household
having 2 women, 2 men, 1 girl, and 1.5 boys.
180 women participated in the baseline survey and 171 participated in the first
follow up. Of the 190 sensors sent to Darfur, 137 survived to the first follow
up. These 137 SUMs represent 122 unique cooks; 15 of the surviving sensors were
redundant “piggyback”sensors. Of the 122 unique cooks who had instrumented
stoves and returned functioning SUMs, 117 were administered follow-up surveys.
Figure 20.4 illustrates the fate of the 190 sensors. By far, the greatest contributor to
loss of SUMs was unrecoverable data due to thermal damage. 31 SUMs from 29
unique cooks showed likely signs of thermal damage: 28 SUMs were lost due to
dead batteries or unreadable iButton memory chips (Maxim states that overheating
can substantially shorten iButton battery life, destroy memory chips, or both) and an
additional 3 sensors showed signs of physical thermal damage including charring or
Fig. 20.4 The fate of the 190 SUMs deployed to Darfur. Blue represents SUMs successfully
recovered at the first follow up. Shades of red represent SUMs lost to various causes. The most
significant cause of SUMs loss was unrecoverable data due to overheating
20 Comparing Cookstove Usage Measured …217
rupturing of the iButton housing. The team had performed a substantial number of
experiments to determine appropriate placement of the SUMs to prevent over-
heating, therefore the loss of so many SUMs was puzzling. However, our team later
determined that many cooks invert the BDS and fill the bottom with charcoal to
cook. This user-led innovation allows cooks to utilize the BDS as an improved
charcoal stove, but this behavior also overheats the SUM. One ramification of
unrecoverable data from overheated SUMs is that users who exhibit behaviors that
overheat their SUMs are systematically underrepresented in SUMs data. We expect
that cooks using their cookstoves often, at high temperature, and inverted are likely
to have a much higher SUMs burnout rate than cooks who do not use their stove
regularly. For this reason, we believe that the data we show herein generally
underestimates the adoption of the BDS by the population. Furthermore, we are
able to put an upper bound on the magnitude of underestimation resulting from the
higher SUMs burnout for adopter cooks than non-adopter cooks. Unless mentioned
otherwise, the rest of the analysis in this paper assumes no bias in the estimates
obtained from the SUMs.
Using the definition of a cooking event discussed in the methods section
(Sect. 20.2), a determination was made between cookstove “users”and “non-users.”
The distinction between a “user”and “non-user”was determined by the proportion of
days that the stove was used. In order to account for “courtesy”uses immediately
before a follow up, a 2-day period preceding the first follow-up survey was ignored.
For the purposes of this analysis, a demarcation between “user”and “non-user”was
made at 10 % of possible stove use days during the observed period (i.e., if a par-
ticipant used the stove less than 1 in 10 days, she is categorized a “non-user”). This
classification is arbitrary and used only as a metric by which to separate women who
regularly use the stove and those who use it very little or not at all. Using this
classification, 71 % (87 participants) are categorized as “users”and 29 % (35) as “non-
users.”To obtain an upper bound on the bias effect from higher SUMs failure rates
with “user”cooks, we recalculate this percentage assuming that all the thermally-
failed SUMs were with the “users”group. This leads to an upper-bound estimate of
77 % (116) users and 23 % (35) nonusers. For those participants with surviving
sensors, a summary of SUMs-measured cooking behavior is tabulated in Table 20.1.
The algorithm employed in this analysis detected 7,358 individual cooking
events with average event duration of 39 min (SD = 26 min) over the timeframe
SUMs were analyzed. A histogram of event durations is shown in Fig. 20.5.
Participants in this study overestimated their BDS usage in surveys both in terms
of events per day and total cooking time per day. Relative to algorithmic estimates,
84 % of participants overestimate cooking hours, and 81 % overestimate cooking
Table 20.1 Summary data for all participants, users, and nonusers
All subjects Users Nonusers
n Mean SD n Mean SD n Mean SD
Hours 122 1.1 1.1 87 1.5 1.0 35 0.028 0.063
Events 122 1.5 1.4 87 2.0 1.3 35 0.036 0.076
218 D.L. Wilson et al.
events. The average participant overreports daily cooking hours by 1.2 h with 95 %
CI [1.0, 1.4] and overreports her daily cooking events by 1.3 events with 95 % CI
[1.1–1.6]. Nonusers overreport their daily use more substantially: 1.7 h with
CI [1.4, 2.1] and 2.2 events with 95 % CI [1.9, 2.6].
For the two survey questions analyzed, 160 surveys were available that had
copies of both original ODK data and data entry versions of paper surveys. Of the
320 total answers between the two questions, there was only one disagreeing pair
between original ODK data and pencil and paper data later entered via ODK.
20.4 Conclusions
The combination of SUMs and an algorithm for cooking event detection has
provided a detailed record of the frequency, timing, and duration of cooking events
from a large number of users. The experimental population is a representative proxy
for the population that has received the BDS free of charge in Darfuri IDP camps.
Although the BDS units distributed in this study were free of charge, 71 % (con-
servative) to 77 % (upper bound) of participants were categorized as “users”by the
definition used in this study indicating that this stove is valued by the women who
own it. Qualitatively, BDS adoption in the Al-Salam IDP camp is high considering
that women receiving the stove were selected at random and had no personal
monetary investment in the product.
The juxtaposition of paper and cell phone-based surveys against SUMs data has
highlighted the discrepancies self-reported versus sensor-measured usage patterns.
Data from both the paper and cell phone-based surveys was consistent, and in
surveys participants overestimate adoption both in terms of hours and events
(p< 0.001) relative to the SUMs algorithm. Average participants overreport daily
cooking hours by 1.2 h and daily cooking events by 1.3 events, and nonusers
overreport daily cooking behaviors more dramatically than the average participant:
1.7 h and 2.2 events. For the average participant, these overestimations are roughly
double the SUMs-measured values for hours and cooking events, and for nonusers
0
500
1000
1500
2000
2500
0 100 200 300 400
Event Duration (minutes)
Count
Fig. 20.5 Histogram of the
duration of 7,358 individual
cooking events
20 Comparing Cookstove Usage Measured …219
overestimations are 60-fold higher than SUMs-measured values for hours and
events cooked per day. Data reported by the participants may be erroneous due to
difficulty in recollection, courtesy bias, or the desire to keep personal information
obscure (e.g., as reported by field staff, respondents may try to obscure true
household size by misreporting cooking hours) These findings herein indicate that
data from SUMs are more detailed, accurate, and meaningful than self-report use
data, and should be preferred as a gold standard for future studies of stove use.
A portion of SUMs was lost during this study, presumably due to thermal
damage from charcoal fires. For future studies, SUMs placement must be altered to
avoid damage. Additionally, Potential Energy may consider a redesign of the BDS
to better accommodate users’desire to burn charcoal in the BDS.
A major question unanswered in this work is the correlation between the start
and stop times of algorithm-determined cooking events and “true”witnessed
cooking events. While the event detection algorithm seems to perform well in terms
of face validity, absent a database of cooking logs or witnessed accounts of
cooking, the algorithm should be trained against expert-labeled data for the local
cooking context to further refine its performance.
Acknowledgments Authors gratefully acknowledge financial support from multiple sources for
this work. The primary funding for this work comes from a DIV Phase-1 grant by the United States
Agency for International Development (USAID) to Potential Energy. Daniel Wilson and Angeli
Kirk are grateful for support from the National Science Foundation (NSF) Graduate Research
Fellowship. Additional funding for personnel and materials for this project has been generously
provided by funding agencies including the Development Impact Lab (USAID Cooperative
Agreement AID-OAA-A-13-00002) which is part of the USAID Higher Education Solutions
Network, the Blum Center for Developing Economies, a Behavioral Sensing Grant from The
Center for Effective Global Action (CEGA), and Department of Energy Contract DE-AC02-
05CH11231 to Lawrence Berkeley National Laboratory (LBNL), operated by the University of
California.
References
Berkeley Air Monitoring Group (2013). Monitoring and evaluation of the Jiko Poa cookstove in
Kenya Report. http://www.washplus.org/sites/default/files/kenya-jiko_cookstove.pdf.
Accessed March 1, 2014.
Burwen, J. (2011). From technology to impact: Understanding and measuring behavior change
with improved biomass cookstoves. UC Berkeley: Energy and Resources Group.
Burwen, J., & Levine, D. I. (2012). A rapid assessment randomized-controlled trial of improved
cookstoves in rural Ghana. Energy for Sustainable Development, 16(3), 328–338.
Das, J., Hammer, D., & Sanchez-Paramo, C. (2012). The impact of recall periods on reported
morbidity and health seeking behavior. Journal of Development Economics, 98(1), 76–88.
Edwards, A. L. (1957). The social desirability variable in personality assessment and research.
New York: Dryden.
Hartung, C., Lerer, A., Anokwa, Y., Tseng, C., Brunette, W., & Borriello, G. (2010). Open data
kit: Tools to build information services for developing regions. In Proceedings of the 4th ACM/
IEEE International Conference on Information and Communication Technologies and
Development. Article No. 18.
220 D.L. Wilson et al.
Karlan, D., McConnell, M., Mullainathan, S., & Zinman, J. (2010). Getting to the top of mind:
How reminders increase saving. New Haven: Working Paper, Department of Economics, Yale
University.
Landsberger, H. A. (1958). Hawthorne revisited. Ithaca, New York: Cornell University.
Lewis, J. J., & Pattanayak, S. K. (2012). Who adopts improved fuels and cookstoves? A systematic
review. Environmental Health Perspectives, 120(5), 637–645.
Lim, S. S., Vos, T., Flaxman, A. D., Danei, G., Shibuya, K., et al. (2013). A comparative risk
assessment of burden of disease and injury attributable to 67 risk factors and risk factor clusters
in 21 regions, 1990–2010: A systematic analysis for the global burden of disease study 2010.
The Lancet, 380(9859), 2224–2260.
Nunnally, J. C. (1978). Psychometric theory (2nd ed.). New York: McGraw-Hill.
Prestwich, A., Perugini, M., & Hurling, R. (2009). Can the effects of implementation intentions on
exercise be enhanced using text messages? Psychology and Health, 24(6). doi:10.1080/
08870440802040715.
Ruiz-Mercado, I., Canuz, E., Walker, J. L., & Smith K. R. (2013). Quantitative metrics of stove
adoption using stove use monitors (SUMs). Biomass and Bioenergy, 57:136–148. http://dx.doi.
org/10.1016/j.biombioe.2013.07.002.
Ruiz-Mercado, I., Masera, O., Zamora, H., & Smith, K. R. (2011). Adoption and sustained use of
improved cookstoves. Energy Policy, 39, 7557–7566.
Thomas, E. A., Barstow, C. K., Rosa, G., Majorin, F., & Clasen, T. (2013). Use of remotely
reporting electronic sensors for assessing use of water filters and cookstoves in Rwanda.
Environmental Science and Technology, 47(23), 13602–13610.
World Bank. (2011). Household cookstoves, environment, health, and climate change.
Washington, DC: The World Bank.
20 Comparing Cookstove Usage Measured …221
