Can economic incentives enhance adoption and use of a household energy technology? Evidence from a pilot study in Cambodia

The promotion of new technologies that promise welfare improvements in low-income settings has preoc cupied policymakers and researchers for decades (Besley and Case 1993). Interventions aimed at increasing uptake, however, often fail to meet expectations, and the developing world is littered with well-intentioned technologies that were either not adopted or eventually abandoned by their intended beneficiaries⁶. Much work in the past has wrestled with this challenge; in particular, researchers have drawn on insights from field experiments to identify policy levers that may improve uptake (Ashraf et al 2010, Cohen and Dupas 2010, Giné and Yang 2009, Miller and Mobarak 2015, Pattanayak et al 2009, Tarozzi et al 2014). These experiments aim to ensure widespread adoption of technologies by reducing their acquisition costs—monetary and otherwise. While this work has greatly advanced understanding of initial adoption (the extensive margin), the recent literature is relatively silent on what drives sustained high levels of use (the intensive margin). Needless to say, obstacles to sustained use continue to compromise the effectiveness of many technologies and, thus, represent a prominent knowledge gap (Bensch and Peters 2015).

The challenge of moving from traditional to clean technologies is particularly acute for energy-poor households in the developing world. Nearly three billion people cook and heat their homes with solid fuels in traditional stoves or open fires. The resulting household air pollution causes over four million deaths annually, a health burden borne largely by the poor, women, and children (World Health Organization 2016). Inefficient combustion of solid fuels also emits pollutants that are among the most significant contributors to climate change, including carbon dioxide, methane, and black carbon (Bond et al 2013), which is already impacting millions by disrupting weather patterns, accelerating glacial melt, and causing a range of damages that extend well beyond health (Menon et al 2002, Srivastava et al 2012)⁷.

Improved cookstoves (ICS) have long been envisioned as a solution to this multifaceted problem (Ruiz-Mercado et al 2011). Despite significant heterogeneity in cost, quality, and materials, generally speaking ICS are designed to reduce emissions by increasing combustive efficiency. In so doing, ICS are expected to yield health benefits, and also reduce the total amount of biomass required, easing stress on local forests (a common source of fuelwood) and the global commons (through lowered climate-altering emissions). Where fuelwood extraction is unsustainable, ICS interventions may also help maintain forest carbon stocks (Maes and Verbist 2012)⁸. These benefits, however, depend on sustained use⁹. Indeed, diminishing or improper use of ICS may entail little to no benefits (Hanna et al 2016), or even exacerbate an already inferior environmental equilibrium (Nepal et al 2011). This is not an insignificant problem, either; research increasingly highlights that 'stacking' of polluting and improved technologies is nearly ubiquitous (Brooks et al 2016, Masera et al 2000, Piedrahita et al 2016, Ruiz-Mercado and Masera 2015), and that this behavior can compromise emissions reductions (Aung et al 2016, Johnson and Chiang 2015).

How can the adoption and use of ICS be encouraged in ways that ensure expected environmental and health benefits are realized? We present results from a pilot study in rural Cambodia that tests whether and how economic incentives increase use of an ICS. Specifically, we augment capital-cost subsidies—traditionally employed to enhance adoption among the poor—with rebates linked to technology use, and evaluate the impact these economic incentives have on both initial adoption and sustained use. We find that households offered rebates used the intervention ICS more frequently and for longer periods than those who were not. We also observe initial reductions in consumption of solid fuels and time spent collecting or preparing fuels, though these impacts diminish over time. Our results suggest that the effectiveness of capital subsidies may be significantly enhanced by relatively simple monetary inducements. Nonetheless, policymakers must also consider the implications of behavioral responses to these incentives, as increased adoption and use do not necessarily translate into sustained environmental or livelihood benefits. Our study provides a promising launch pad for future research on these issues, given the growing interest in and use of incentive-based interventions such as carbon financing or other forms of results-based financing (RBF) for ICS¹⁰.

To motivate how economic incentives could enhance use of ICS, we draw on a model of household production of environmental quality and health (Grossman 1972, Pattanayak and Pfaff 2009). This framework helps explain (i) the detrimental effects of household air pollution exposure, and (ii) the effects of interventions that aim to reduce it (Jeuland et al 2015). Decisions to adopt technologies such as ICS are made by households on the basis of perceived private benefits (e.g. averting health risks or the drudgery of fuel collection)¹¹. These decisions involve trade-offs with consumption and leisure.

According to this model, households maximize utility by allocating resources—time and money—to consumption, health 'risk-averting behavior,' and leisure. Initial resource endowments and the supply-side context constrain the set of available alternatives, and thus affect investments in health-improving technologies. The choices households make also depend on preferences; the effectiveness of averting behavior in reducing utility-harming illness and pollution (which varies with technology and the pollution context); and the extent to which households value environmental quality, namely clean air. Illness declines with increased household environmental quality and averting behavior, and with the pollution-reducing actions of others. Environmental quality improves with averting behavior, but is harmed by consumption, which generates pollution.

This micro-level perspective offers insights about how to best target and test the effectiveness of policies to increase adoption and use of averting behavior. Given that investments in beneficial technologies are restricted by the availability of inputs of time, materials, and knowledge (Jeuland et al 2015), reductions in the costs of these inputs are predicted to increase averting behavior, all else being equal. Interventions aiming to increase uptake have traditionally seen costs as static, however, requiring a one-time subsidy or knowledge dissemination campaign. Dynamic incentives (namely, rebates linked to use over time) have received relatively less attention but may be a promising alternative.

The impacts of incentive-based interventions, however, are not always clear ex ante. For instance, monetary incentives may induce cheating by beneficiaries (Gravelle et al 2010). Others have investigated whether and how incentives have unintended consequences (Oxman and Fretheim 2009), such as inducing rebound by reducing the relative prices of pollution-generating activities (Davis et al 2014). In addition, the evidence on direct payments as a tool to enhance uptake of environmental health technologies is mixed (e.g. Krezanoski et al 2010). We add to the literature by evaluating the extent to which dynamic incentives can augment the effectiveness of more traditional one-time subsidies in the context of household energy interventions that deliver environmental and health benefits.

3.1. Study sites and sample selection

3.2. Description of control and treatment arms

3.3. Outcomes of interest

4.1. Sample description

4.2. Stove adoption

4.3. Frequency and duration of stove use

We next evaluate the impact of economic incentives on the frequency of stove use. Our linear regression model is specified as follows:

$$\begin{equation} Y_{i,j}=\beta_{0}+\beta_{1}LOW_{j}+\beta_{2}\,MEDIUM_{j}+\beta_{3}HIGH_{j}+{\bf X}_{i\gamma}+\epsilon_{i,j} \end{equation} \tag{ 1 }$$

where Y_i,j represents a SUMs-based measure of stove use for household j, and LOW, MEDIUM, and HIGH are indicator variables for the respective treatment arms²⁰. We also control for key household-level characteristics—namely, household size, monthly expenditure, access to private tap water, ownership of LPG and electric stoves, and awareness of improved stoves or fuels, represented by X_i. Finally, ɛ_i,j represents unobserved characteristics.

Results using a daily count of cooking events are reported in table 3. We do not detect aggregate differences between treatment and control households using the threshold-based measure of stove use (panel a, column 1). Disaggregating by treatment arm, however, we find a statistically significant effect of the intervention in the treatment-low and treatment-high communities, where households used the stoves up to two more times each day (panel a, column 2). In panel (b), we repeat the analysis with the slope-based measures, obtaining similar results.

Table 3. Mean daily stove-use count.

		(1)			(2)
VARIABLES		Daily stove-use count	Bootstrapped 90% CI		Daily stove-use count	Bootstrapped 90% CI
(a) Threshold approach (50 °C threshold)
Treatment		1.21	(−1.69, 4.11)
		[0.34]
Low					0.89*	(0.18, 1.59)
					[0.058]
Medium					0.74	(−0.11, 1.59)
					[0.14]
High					1.85**	(1.12, 2.58)
					[0.012]
Constant		1.42*	(0.39, 2.45)		1.45**	(0.66, 2.24)
		[0.051]			[0.02]
Household controls		Yes			Yes
Observations		59			59
R-squared		0.28			0.42
(b) Slope approach†
Treatment		1.28	(−0.53, 3.09)
		[0.18]
Low					1.01*	(0.22, 1.80)
					[0.06]
Medium					0.60	(−0.56, 1.76)
					[0.32]
High					1.89***	(1.40, 2.38)
					[0.001]
Constant		1.69*	(0.17, 3.20)		1.76**	(0.66, 2.85)
		[0.08]			[0.03]
Household controls		Yes			Yes
Observations		59			59
R-squared		0.25			0.33

Note: The dependent variable is mean daily stove-use count, estimated using the threshold approach in panel (a) and using the slope approach in panel (b). Column (1) presents results for all households assigned to any of the three treatment arms; column (2) presents results for each of the three treatment arms separately. Household-level controls include: household size; monthly expenditure; access to private tap water; ownership of LPG stoves; ownership of electric stoves or rice cookers; and awareness of the existence of improved stoves or clean fuels. Pairs cluster-bootstrapped p-values shown in brackets and 90% confidence intervals shown in parentheses, with clustering at the village level (Cameron et al 2008, Esarey 2016). ^***p < 0.01, ^**p < 0.05, ^*p < 0.1. ^†A cooking event is coded as having started when the mean increase in SUM-measured temperature over a ten-minute interval exceeds 0.5 °C per minute. A cooking event is coded as having ended once the weighted mean decrease in temperature over a ten-minute interval exceeds 0.4 °C per minute.

Table 4 presents results for stove use duration. In line with stove-use count results, we do not detect aggregate differences between treatment and control households (panel a, column 1). Disaggregating by treatment arms in column (2), however, we again find a significant effect of the intervention in the treatment-high community. These results are not robust; using stove-use duration measured via the slope approach in panel (b) yields estimates that are smaller and not statistically significant²¹.

Table 4. Mean daily stove-use duration, minutes.

		(1)			(2)
VARIABLES		Stove use duration	Bootstrapped 90% CI		Stove use duration	Bootstrapped 90% CI
(a) Threshold approach (50 °C threshold)
Treatment		103.2	(−558, 764)
		[0.58]
Low					23.0	(−13, 59)
					[0.21]
Medium					−19.8	(−66, 26)
					[0.41]
High					265.9***	(205, 327)
					[0.002]
Constant		108.1	(−27, 243)		115.5**	(42, 189)
		[0.16]			[0.04]
Household controls		Yes			Yes
Observations		59			59
R-squared		0.17			0.59
(b) Slope approach†
Treatment		13.6	(−30, 57)
		[0.48]
Low				17.2	(−17, 52)
				[0.32]
Medium				−1.0	(−53, 51)
				[0.97]
High				11.2	(−22, 44)
				[0.47]
Constant		97.5**	(49, 146)	100.5***	(61, 141)
		[0.02]		[0.009]
Household controls		Yes		Yes
Observations		59		59
R-squared		0.21		0.22

Note: The dependent variable is mean daily stove-use duration in minutes, estimated using the threshold approach in panel (a) and using the slope approach in panel (b). Column (1) presents results for all households assigned to any of the three treatment arms; column (2) presents results for each of the three treatment arms separately. Household-level controls include: household size; monthly expenditure; access to private tap water; ownership of LPG stoves; ownership of electric stoves or rice cookers; and awareness of the existence of improved stoves or clean fuels. Pairs cluster-bootstrapped p-values shown in brackets and 90% confidence intervals shown in parentheses, with clustering at the village level (Cameron et al 2008, Esarey 2016). ^***p < 0.01, ^**p < 0.05, ^*p < 0.1. ^†A cooking event is coded as having started when the mean increase in SUM-measured temperature over a ten-minute interval exceeds 0.5 °C per minute. A cooking event is coded as having ended once the weighted mean decrease in temperature over a ten-minute interval exceeds 0.4 °C per minute.

4.4. Biomass and solid fuel use

For logistical reasons, our baseline surveys (including fuel weighing) were conducted just after households were recruited and provided with improved stoves. We thus do not have objective data on fuel use prior to stove provision, i.e. the initial measures of fuel use already reflect changes that came with owning the improved stoves. Since we are interested in how rebates relate to trends in fuel use over time, we employ a difference-in-differences (DID) strategy, comparing differences in solid-fuel use between treatment and control households at the outset of the intervention and its conclusion. Our model is specified as follows:

$$\begin{equation} \begin{array}{*{20}c} {Y_{i,j} } & = & {\beta _0 + \beta _1 PERIOD + \beta _2 LOW_j + \beta _3 MEDIUM_j } \\ {} & + & {\beta _4 HIGH_j + \beta _5 (PERIOD \times LOW_j )} \\ {} & + & {\beta _6 (PERIOD \times MEDIUM_j )} \\ {} & + & {\beta _7 (PERIOD \times HIGH_j ) + \text{X}_{i,\gamma } + \varepsilon _{i,j.} } \\ \end{array} \end{equation} \tag{ 2 }$$

In equation (2), Y_i,j represents a measure of solid fuel use (self-reported or weighed); PERIOD is an indicator variable that equals one for data collected during the end-line survey, and zero for those collected during the baseline survey; β₂, β₃, and β₄—the coefficients on the indicator variables for each of the three treatment arms—represent the initial treatment effects in each treatment arm, relative to the control communities²²; and β₅, β₆, and β₇—the coefficients on the interactions between PERIOD and the treatment arm indicators—represent our DID estimates of the impact of the intervention in the treatment communities. β₁ isolates the change among the controls over the course of the intervention.

As shown in table 5, the coefficient on the treatment indicators in column (1) is negative and statistically significant; thus, the intervention led to large immediate reductions in weighed fuelwood use. The DID estimate in column (1) is positive but not statistically significant; disaggregating by treatment arms in column (2) also yields positive but statistically insignificant estimates. In addition, the coefficient for PERIOD is also negative, which suggests an average downward time trend in solid-fuel use among control households over the course of the intervention. The consistently positive DID estimates hint at a rebound in measured solid-fuel use among treatment households over the course of the intervention (i.e. the treatment effect observed at the outset of the intervention fades over time). Columns (3) and (4) report estimates obtained using data on self-reported solid-fuel use; the results are strikingly similar. In particular, the positive and statistically significant DID estimate for treatment-low households in column (4) is consistent with the rebound hypothesis.

Table 5. Solid-fuel use, kilograms per day.

		(1)			(2)			(3)			(4)
VARIABLES		Measured weight	Bootstrapped 90% CI		Measured weight	Bootstrapped 90% CI		Stated weight	Bootstrapped 90% CI		Stated weight	Bootstrapped 90% CI
Period		−1.90	(−4.60, 0.80)		−1.90	(−5.27, 1.48)		−2.01***	(−2.40, −1.62)		−1.98***	(−2.54, −1.42)
		[0.17]			[0.25]			[0.001]			[0.003]
Treatment		−3.04**	(−4.42, −1.66)					−3.81**	(−5.59, −2.03)
		[0.02]						[0.02]
Period × Treatment		1.12	(−1.65, 3.90)					1.23	(−1.25, 3.71)
		[0.37]						[0.31]
Low					−3.28**	(−5.46, −1.10)					−4.45**	(−6.27, −2.62)
					[0.045]						[0.011]
Medium					−2.78**	(−4.48, −1.07)					−3.93**	(−6.37, −1.49)
					[0.03]						[0.03]
High					−3.14**	(−4.44, −1.84)					−3.05**	(−4.99, −1.12)
					[0.011]						[0.04]
Period × Low					1.44	(−1.29, 4.17)					2.09***	(1.60, 2.59)
					[0.27]						[0.003]
Period × Medium					0.14	(−5.01, 5.29)					0.32	(−1.97, 2.60)
					[0.96]						[0.72]
Period × High					1.59	(−0.41, 3.61)					0.81	(−0.11, 1.72)
					[0.15]						[0.13]
Constant		3.31**	(1.51, 5.12)		3.36**	(1.48, 5.24)		4.08**	(1.42, 6.75)		4.15**	(2.02, 6.27)
		[0.02]			[0.02]			[0.04]			[0.02]
Household controls		Yes			Yes			Yes			Yes
Observations		106			106			110			110
R-squared		0.29			0.30			0.31			0.33

Note: The dependent variable is daily use of solid fuels in kilograms, estimated using an objective 24-hour weight measure in columns (1) and (2) and using household self-reports in columns (3) and (4). Columns (1) and (3) present results for all households assigned to any of the three treatment arms; columns (2) and (4) present results for each of the three treatment arms separately. Household-level controls include: household size; monthly expenditure; access to private tap water; ownership of LPG stoves; ownership of electric stoves or rice cookers; and awareness of the existence of improved stoves or clean fuels. Pairs cluster-bootstrapped p-values shown in brackets and 90% confidence intervals shown in parentheses, with clustering at the village level (Cameron et al, 2008; Esarey 2016). ^*** p < 0.01, ^** p < 0.05, ^*p < 0.1.

4.5. Time spent collecting or preparing solid fuel for combustion

Researchers and policymakers alike have struggled to ensure sustained use of a variety of household environmental, health, and energy technologies. Using data from Cambodia, we show that households' use of an improved cookstove is responsive to economic incentives. Households given incentives based on objective measures derived from stove-use monitors used the stoves more frequently and for longer periods of time than others. Economic incentives linked to use (e.g. carbon finance delivered in response to verified changes in behavior) may therefore be effective for spurring adoption of ICS technologies.

While the results were somewhat inconsistent and not always statistically distinguishable from zero, simple verification visits and household stove-use diary checks accompanied by incentives also appear to reduce the total amount of—as well as time spent collecting—solid fuel, suggesting that even relatively inexpensive monitoring approaches could enhance use of household technologies. Given the high relative cost of SUMs-based verification, these simpler approaches may prove cost-effective in resource-constrained contexts. That said, households in the treatment-low and treatment-medium communities do appear to have 'over-reported' stove use (perhaps to obtain use-based incentives), and more research is needed to understand the limits and potential problems with such verification procedures, whose ultimate success hinges on honest reporting of use.

Finally, our results point to the need to carefully consider how behavior responds to such incentives. Somewhat unexpectedly, we found that initial reductions in solid-fuel use—as well as in the fuel collection and preparation burden—fade somewhat over time. The 'substitution effect' associated with receiving a cleaner cooking technology may thus be partially offset by an 'income effect.' In other words, while households appear to initially respond to a cleaner technology's reduced price by switching away from traditional alternatives, over time the reductions may be offset by increased cooking and fuel consumption (perhaps from preparing more food), in contrast to results from other settings (e.g. Bensch and Peters 2013, Brooks et al 2016). In line with this phenomenon, the tendency of households to 'stack' multiple energy technologies and fuels reflects their desire for more services (Ruiz-Mercado and Masera 2015). While we do not have objective use data related to the other technologies that comprise the full energy mix used by our sample households, our survey measures do confirm that stove stacking was prevalent. Over eighty percent of respondents reported using two or more stoves for cooking and heating during the end-line survey.

The success of incentive-based interventions depends on how they are designed, how incentives are delivered, and how they interact with personal or societal norms and motivations (Gneezy et al 2011). Policymakers must carefully consider these elements, particularly when designing interventions that aim to promote use of environmental health technologies, whose use is frequently over-reported (Thomas et al 2013). Absent a complete and sustained substitution to cleaner technologies, many of the anticipated benefits—for the environment, health and livelihoods—of these solutions may not fully materialize. This is perhaps especially true for health (Burnett et al 2014), but rebound effects that increase energy use may also offset emissions or fuel savings benefits (Gillingham et al 2016, Greening et al 2000). Thus, while our study yields promising preliminary evidence that incentives can increase use of beneficial technologies in low-income contexts, future research—conducted in alternative contexts, with larger sample sizes that enable detection of smaller effects, objective measurements for all household energy devices, and greater variation in the levels of incentives—is sorely needed to better understand their net effects on overall social and environmental well-being.

This study was made possible by the generous support of SNV Cambodia. Faraz Usmani is also grateful for support from the Duke Global Health Institute. We acknowledge the excellent research assistance provided by Sushmita Samaddar and Hannah Girardeau, and the invaluable contributions of SNV Cambodia research staff (Kanika Pheap and Thary Sok) and our field team (Bunheng Tol, Channak Chun, and Khunny Khuth). We thank two anonymous reviewers for their helpful comments.