Energy-Saving Effects of Progressive Pricing and Free CFL Bulb Distribution Program: Evidence from Ethiopia

In Africa, about 70 percent of the total population still lives without electricity. Significant resources are needed to meet the gap. Demand-side management is crucial to curb the increasing demand even in developing countries. A traditional approach is to raise prices, but promoting energy-efficient products such as compact fluorescent lamp (CFL) bulbs is also a win-win proposition. While end-users can reduce their spending, power utilities can avoid costly investments in new generation capacity. This paper estimates the effects of progressive pricing as well as CFL distribution program in Ethiopia. It is found that the increasing block tariff structure reduced the demand: the price elasticity is estimated at 0.29. This is particularly useful to influence large-volume users, who are presumably the rich. The CFL program is also found effective to contain the electricity demand. The estimated impact is about 45 kWh per customer. This is significant energy savings particularly for low-volume users.

Keywords: impact evaluation; energy efficiency program; fixed-effect least-squares; fixed-effect quantile regression

Electric access is one of the most important development challenges in developing countries. Globally, it is estimated that about 1.3 billion people still do not have access to electricity ([20]). In Sub-Saharan Africa, electricity access is particularly limited: about 585 million people, or approximately 70 percent of the total population, are still living without power. In addition, even if electricity access is granted, the quality of the power supply often remains poor. Many households suffer from frequent, if not chronic, power outages every day. A significant resource gap exists to meet the growing energy demand in Sub-Saharan Africa. While the whole region needs US$40.8 billion per annum in the energy sector, only US$11.6 billion is currently available ([42]).

To this end, demand-side management is critical in developing countries, because it is particularly costly to develop new power generation capacity. In Africa, for instance, an additional 1 MW of installed capacity costs between US$0.3 million and US$0.5 million (figure 1).[1] A traditional way of curbing demand is, of course, to increase prices. A popular instrument is the increasing block tariff structure: higher marginal prices are applied when one consumes more. The literature is generally supportive of the significance of pricing: the elasticity is traditionally estimated at −1 to −2 (e.g. [37]; [40]). Recent studies found more modest elasticities from −0.1 to −0.4 (e.g. [10]; [2]; [31]). In developing countries, however, affordability may be an additional matter of concern because electricity is a basic infrastructure service.

Graph: Figure 1. Annualized Investment and Operation and Maintenance (O&M) Costs in Africa (US$ Million per MW) Source:[32].

Another policy option to manage the energy demand may be to promote energy-efficiency technologies, which is an inexpensive win-win proposition.[2] While power utilities can avoid costly investment in energy capacity, end-users can reduce their energy bills. Thus, affordability is not an issue here. Improved energy efficiency also contributes to mitigating global warming, although Africa is a minor CO2 emitter at present, accounting only for some three percent of the global emissions. In this context, many countries, including in the Philippines, Rwanda, Thailand, Uganda, and Vietnam, implemented compact fluorescent lamp (CFL) bulb distribution programs (e.g. [41]; [9]). Recent lifetime analyses indicate that total CO2 emissions could be reduced by about 75 percent by using CFLs instead of incandescent lamps (Ramroth 2008; [1]).

Of course, the long-term sustainability remains debatable in developing countries, especially because CFL lamps contain about 4–5 mg of mercury per bulb, while incandescent lamps use no toxic material (e.g. [1]). In developing countries, CFL bulbs are often not segregated out of the other waste stream and left in open dumpsites (e.g. [34]). Despite such a drawback, the CFL distribution programs are generally expected to save energy significantly. In Vietnam, for instance, a total of one million CFL bulbs were provided at favorable prices. From an engineering point of view, it is expected to reduce peak demand by 31 MW under the assumption that an old incandescent lamp would be replaced with a new CFL bulb, although it is still a small fraction of the total energy requirement ([9]).[3],[4]

The current paper aims at examining the effects of both pricing and a free CFL distribution program on residential energy consumption in Ethiopia. On one hand, the country uses an increasing block tariff structure to curb overconsumption of energy. On the other hand, Ethiopia carried out a series of CFL programs with more than nine million bulbs distributed in recent years.

Methodologically, a standard demand function for electricity is estimated with the effect of the CFL program incorporated. The CFL effect may be less obvious than it looks. First, enforcement is an issue. People may or may not use distributed bulbs as intended. There is general concern about CFL bulb quality. Incandescent lamps still outperform CFLs in terms of resemblance of original color, starting time, and frequency of switching ([1]). A survey indicates that many people do not believe that CFL saves money even after they used them ([30]).

Second, the literature discusses the rebound and backfire effects of energy efficiency programs. It is well known that fuel-efficient car users drive more, not less, because of improved energy efficiency.[5] For residential lighting, the rebound effect is estimated at 5–12 percent ([16]). On the other hand, a one-time intervention or temporary shock can change people's behavior fundamentally and make the change permanent. In Brazil, the temporary suspension of the electricity supply encouraged customers to purchase energy-efficient home appliances, making achieved energy savings sustained even after the supply constraint was eased ([5]). Finally, related to the above, sustainability is a matter of concern particularly in developing countries. CFL bulbs are still relatively expensive for the poor. Thus, the impacts of the CFL bulb distribution as well as electricity prices are potentially different depending on income level.

To address these issues, this paper uses monthly panel data and estimates the household demand for electricity by the fixed-effect regression with score-matching. To investigate the distributional impact, a two-step fixed-effect quantile regression ([4]) is also used. The following sections are organized as follows: Section 1 provides an overview of the Ethiopian electricity sector, including the recent CFL distribution programs. Section 2 develops our empirical strategy. Section 3 presents main estimation results and policy implications. Section 4 discusses robustness of the results. Finally, section 5 concludes.

Despite rapid economic growth in recent years, the Ethiopian electricity system remains largely underdeveloped. The national utility, Ethiopian Electric Power Corporation (EEPCO), has electrified about 45 percent of the country's towns and villages and currently serves about 2.1 million customers (or approximately 13 million people).[6] In the country, however, only 14 percent of the total population has access to grid electricity. Particularly in rural areas, only two percent have access to power. In addition, the quality of the electricity services remains poor. People are experiencing power outages on average 44 days per year ([11]).

In recent years, the Government of Ethiopia and EEPCO have been ramping up their efforts in two areas: universal electric access and demand-side management. EEPCO owns an installed capacity of about 2,000 MW and a distribution network of 126,000 km. In the early 2000s, EEPCO started expanding power access in rural and remote areas. Over the past five years, approximately 5,000 villages were electrified (figure 2). The total customer base of EEPCO increased from 800,000 in 2005 to two million in 2011. Still, about 11,000 villages remain to be electrified in the country. The national target is to achieve an electrification rate of 75 percent by 2015.

Graph: Figure 2. Number of Villages Electrified in Ethiopia Source: Authors' own calculation based on data provided by EEPCO.

Graph: Figure 3. Block Rates of Electricity Tariffs in Ethiopia Source: Authors' own calculation based on data provided by EEPCO.

As more people are connected, the demand for electricity is also increasing significantly. To curb the demand, EEPCO uses an increasing block tariff structure. The current tariff schedule has eight blocks where the marginal rate increases with power consumption from Birr 0.273 (or US$0.016) to Birr 0.694 (or US$0.04) per kWh (figure 3). Note that the real prices have been decreasing over time, because these nominal prices have not been adjusted for long time. Still, this progressive schedule is expected to contribute to controlling for the people's electricity demand.

To manage the increasing demand, EEPCO also implemented three CFL distribution programs in recent years. The first program was carried out in June to August 2009, distributing 350,000 CFL bulbs free of charge. The bulbs were distributed nationwide but practically focused on the capital city, Addis Ababa, because CFLs were allocated based on the number of existing customers per service center. The vast majority of the existing EEPCO customers live around Addis Ababa. Though the program was not targeted but implemented on a voluntary basis. Customers had to bring their old light bulbs to the nearest EEPCO service center. Thus, there is no enforcement issue in this case. Most of the program participants obtained a maximum of four CFL bulbs under the program.

Following the successful implementation of the first phase, the second and third CFL bulb distribution programs were implemented in July 2011 and January 2012, respectively. In total, 9.5 million CFL bulbs were distributed (table 1). The main difference from the first phase was that CFL bulbs were not distributed free of charge but sold at a significant discount (table 2). For instance, customers paid seven Ethiopian Birr (or US$0.40) for an 11 W CFL bulb while the market price was about Birr 25 (or US$1.42). The gap between the program and market prices was then financed by EEPCO and international development agencies.

Table 1. Engineering Calculations of Energy Savings Expected from the CFL Bulb Program

1 Source: Authors' own calculation based on data provided by EEPCO.

Table 2. CFL Bulb Market and Subsidized Prices

2 Source: Authors' own calculation based on data provided by EEPCO.

To estimate the effects of the increasing block tariff as well as the CFL bulb distribution program, the following conditional demand function is considered:

\begin{eqnarray} kw{h_{it}} = \alpha CF{L_{it}} + {\beta _{MP}}M{P_{it}} + {\beta _D}{D_{it}} + {c_i} + {c_t} + {\varepsilon _{it}} \end{eqnarray}

(1)where kwhit is the amount of electricity consumed by household i at time t. From the specification point of view, one can take the logarithm of variables in the equation. Both log and non-log specifications are examined in the following sections. CFLit is a dummy variable which takes one if household i received CFL bulbs distributed under the program and used them at time t, and zero otherwise. The analysis is focused on the first phase of the CFL bulb distribution program in 2009. As our data are panel, ci and ct are the individual- and time-specific fixed effect, respectively. To take into account the possible individual effects in the standard errors, the clustered standard errors are considered at the household level.

MP denotes the marginal tariff rate, and D represents the [26] difference variable. Neoclassical economic theory tells that the price variable in the demand function should be a marginal price that a consumer faces, rather than the average price (computed as the total bill divided by total consumption).[7] Taking the inflationary effect into account, the marginal rate is defined in real terms.[8] In addition, the increasing block tariff schedule causes positive implicit consumer surplus, which is referred to as the "Nordin's difference variable"—denoted by D. It is the difference between the actual bill and what would be paid if the final block rate were applied to total consumption ([37]; [26]).[9] This can be interpreted as implicit income, and, thus, the elasticity with respect to D should be the same as the income elasticity but opposite in sign.

There are two important empirical issues to estimate equation (1). First, the marginal price and Nordin's D are generally endogenous. To deal with this problem, the instrumental variable (IV) technique is used with MP and D instrumented by their predicted values based on expansion of the definition of the Nordin's difference variable. See appendix 1 for more details on these instruments.

Second, the program participation was voluntary. Thus, CFLit is also likely to be endogenous. Particularly in the case of the Ethiopian program, customers themselves had to bring old incandescent lamps to the nearest EEPCO customer center to obtain new CFLs. This seems to have created a self-selection mechanism. That is, CFLit is correlated with people's unobservable preferences, which are included in the error term ɛit. If this is the case, the ordinary least squares (OLS) estimator is likely to be biased.

In fact, the data show that there is a systematic difference prior to the distribution in energy use between the CFL program beneficiaries and nonbeneficiaries. The average energy consumption among the program participants is 187 kWh with the majority using less than 200 kWh (figure 4). On the other hand, the average energy consumption of nonbeneficiaries is 314 kWh, and many households consume more than 500 kWh per month (figure 5). It seems that low-volume electricity consumers, who are presumably the poor, were more willing to participate in the free CFL distribution program than high-volume users.

Graph: Figure 4. Energy Consumption by CFL Beneficiaries in May 2009 Source: Authors' own calculation based on data provided by EEPCO.

Graph: Figure 5. Energy Consumption by CFL Non-Beneficiaries in May 2009 Source: Authors' own calculation based on data provided by EEPCO.

Our sample data cover 4,000 customers living in the Bole-Kazanchis Service Area, Addis Ababa. This Service Area covers 13,000 customers in total, out of which about 3,500 customers received mostly four CFL bulbs under the first phase of the CFL program. The beneficiary identification numbers were matched with the EEPCO's customer database from which the historical power consumption records were derived in collaboration with EEPCO. 2,000 customers were randomly selected from the group of program participants (i.e. treatment group), and another 2,000 customers were also randomly selected from the nonparticipant (control) group. For each customer, monthly energy consumption and bill data were collected for the period: January 2007 to August 2012. This sample period was selected to cover at least two years before and after the program, respectively. In the sample, 60 percent of beneficiaries received CFL bulbs in June 2009, and the rest did in July 2009.

To mitigate the self-selection bias between the two groups, the fixed-effect regression model is used, which can remove common time- and group-specific effects as long as they are time-invariant (e.g. [19]; [22]; [23]). As shown in figure 6, there seems to be a time-invariant factor that differentiates program participants and nonparticipants and does not seem to change over time regardless of the program. The long-term trend shows a continuous increase in people's electricity consumption over time. This is because household income generally increases as the economy grows. There is also seasonality in demand: the electricity demand is generally high during summer (June to August). This holds for both program participants and nonparticipants. But the seasonality seems to be larger for nonparticipants, who presumably own more seasonal appliances such as air conditioning. This has not changed despite the program. Finally, though gradually increasing, Ethiopia's electricity supply capacity is still limited, causing chronic power outages. Ethiopia is highly dependent on hydropower, of which the capacity availability fluctuates depending on weather (e.g. [8]).

Graph: Figure 6. Average Electricity Consumption of CFL Beneficiaries and Non-Beneficiaries Source: Authors' own calculation based on data provided by EEPCO.

Still, there may remain a risk of self-selection bias. To ensure comparability between program participants and nonparticipants in a statistical manner, a conventional matching technique is also applied. To match the two groups, the EEPCO's contract number is used. Note that the current analysis only uses the utility's administrative data and the list of the CFL program beneficiaries. It is expected that the contract numbers for program participants are systematically greater, because the contract numbers are normally assigned chronologically. Therefore, the smaller contract numbers are older customers (figure 7).[10] Those who were connected earlier are relatively rich and likely to already own enough lightbulbs and thus less likely to participate in the free CFL program. This is supported by our data. On the other hand, the program participation conditional to the contract number is less relevant to the household electricity consumption: the correlation is found to be minimal at −0.026.

Graph: Figure 7. EEPCO Contract Number and Date of Connection Source: Authors' own calculation based on data provided by EEPCO.

Hence, our primary empirical model is the panel fixed-effect instrumental variable (FE IV) regression with score-matching. An alternative model is the random-effects model. The conventional specification test will be conducted to examine which model fits the underlying data better. In addition, a semi-parametric technique of quantile regression is also used to estimate the distributional impacts of the program. Quantile regression has the great advantage of capturing potential differences in the response of the dependent variable at different points. In our context, the program impacts can be different between low- and large-volume consumers. In addition, quantile regression is more efficient than least squares estimators if the error term is not normal (e.g. [3]; [24]).

The two-stage fixed-effect quantile regression (2SFEQR) model can provide a consistent estimate as T increases ([4]). A Monte Carlo simulation shows that with T = 20 possible bias is less than 0.04 percent. Thus, at the first stage, |${\hat c_i}$| and |${\hat c_t}$| are estimated through the standard IV fixed-effect regression. Then at the second stage |${\hat c_i}$| and |${\hat c_t}$| are subtracted from kwh, and quantile regression is performed for |${\overline {kwh} _{it}} = kw{h_{it}} - {\hat c_i} - {\hat c_t}$|⁠. Five quantiles are examined: 0.1, 0.25, 0.5, 0.75 and 0.9.

The summary statistics of our data are shown in table 3. An average customer (or household) uses about 250 kWh per month, paying on average Birr 62 or US$3.54. The monthly energy consumption is absolutely low but not minimal. This reflects the fact that the sample data are focused on the capital area where household income is relatively high in the country. Recall that our sample data are almost perfectly balanced for about 4,000 customers over the 67-month period (from January 2007 to July 2012). The program was implemented in June-July 2009. The dummy variable for CFL is set at one for one-fourth of the total observations, because program beneficiaries account for half of the sample, and they used the distributed CFLs in the second half of the sample period. The price variables are adjusted with inflation taken into account.

Table 3. Summary Statistics

3 Source: Authors' own calculation based on data provided by EEPCO.

First of all, the propensity score-matching is applied. One probit regression is performed for every cohort (year-month) set of participating customers since the group of potential control customers changes over time. As expected, customers with greater contract numbers, who had been connected more recently, were more likely to participate in the CFL program (table 4). Only partial results are shown for several selected cohorts. In total, 3,979 observations are found outside of the common support, which is an overlap between the propensity scores of the beneficiary group and those of nonbeneficiaries, and excluded from the following analysis. The lack of common support implies that the instrument does not work to find matches between the two groups.

Table 4. Probit Results for Propensity Score Matching

• 4 Source: Authors' own calculation based on data provided by EEPCO.
• 5 Notes: The dependent variable is the dummy variable for the treatment group. Robust standard errors are shown in parentheses. *, **, and *** indicate the statistical significance at ten, five, and one percent, respectively.

Then, the fixed-effect OLS regression is performed, although this is likely to be biased, as discussed above. Clustered standard errors are shown in parentheses. As expected, the estimated CFL bulb impact is found to be negative (table 5). However, the coefficient of the marginal price turned out insignificant, contradicting economic theory. This OLS result may have captured the characteristics of the increasing block tariff schedule under which higher block rates are applied to larger-volume consumers. Formally, the conventional exogeneity test indicates that the marginal price and Nordin's difference variable are not exogenous. The chi-squared test statistic with a degree of freedom of 69 is estimated at 25,708.35, well above the conventional threshold.

To address the endogeneity issue, the instrumental-variable fixed-effect (IV FE) regression is used. The impact of the CFL bulb distribution is estimated at 45.5 kWh per household, which is statistically significant. Hence, each bulb could save energy by 11 kWh per month.[11] Given an average monthly energy consumption of 250 kWh in the sample area, this indicates the program's significant impact on residential energy savings in Ethiopia. From the specification point of view, the fixed-effect model is more preferred than the random-effect model. The Hausman specification test statistic with a degree of freedom of 69 is estimated at 2,431.19, which is well above the usual threshold to reject the hypothesis that the error term ɛit is not correlated with the individual-specific fixed effect ci, though the estimated coefficients are broadly similar.

Table 5. Fixed- and Random-Effect Regression with Score Matching

• 6 Source: Authors' own calculation based on data provided by EEPCO.
• 7 Notes: The dependent variable is KWH. Clustered standard errors at the household level are shown in parentheses. *, **, and *** indicate the statistical significance at ten, five, and one percent, respectively.

When the price endogeneity is taken into account, the price elasticity turned out to be negative and significant. The elasticity is estimated at −0.29 in the FE IV model. This is relatively small in absolute terms compared with the existing literature but reasonable in our case, because many households are still using electricity for basic living needs. In theory, the elasticity with respect to the Nordin's D is a mirror of income elasticity. The estimated elasticity is relatively high at 0.77, indicating the increasing electricity demand as the economy grows. This reconfirms the importance of demand-side management.

Even if the logarithm is taken on both sides, the main results are broadly the same (table 6). The OLS result seems to be biased. The price elasticity is positive, and the effect of CFL is less significant. According to the FE IV, the coefficient of CFL is negative at −0.025, which is significant. The price elasticity is much higher than previously estimated, and, on the other hand, the income elasticity implied by the Nordin's D coefficient is very close to the previous estimate.[12] All the indications are that the CFL program contributed to energy savings, and the electricity consumption decreases with prices and increases with household income.

To assess the distributional effect, the two-stage fixed-effect quantile model is used with clustered standard errors used at the household level.[13] Note that the two endogenous variables—MP and D—are also replaced by their predicted values using the results of the first-stage FE IV regression. The results are broadly the same as the above results. However, there are marked differences among different quantiles. The energy savings are estimated at −48 kWh per month for the first quantile (i.e. lowest volume users [table 7]). For the fifth quantile (i.e. the largest volume users, the estimated impact is much larger [in absolute terms]). The coefficient is −87. Thus, large-volume power consumers, who are most likely the rich, benefited most from energy savings, presumably because they use more electricity for lighting.

Table 6. Fixed- and Random-Effect Regression with the Logarithmic Specification

• 8 Source: Authors' own calculation based on data provided by EEPCO.
• 9 Notes: The dependent variable is lnKWH. Clustered standard errors at the household level are shown in parentheses. *, **, and *** indicate the statistical significance at ten, five, and one percent, respectively.

Table 7. Distributional Effects: IV Fixed-Effect Quantile Regression

• 10 Source: Authors' own calculation based on data provided by EEPCO.
• 11 Notes: The dependent variable is KWH. Clustered standard errors at the household level are shown in parentheses. *, **, and *** indicate the statistical significance at ten, five, and one percent, respectively.

In relative terms, however, the program impact is larger for low-volume power users, who are probably the poor. For the first quantile, the estimated energy savings accounted for 60 percent of their predicted power consumption (figure 8). This is because lighting is their primary energy use. By contrast, the relative impact is much smaller for large-volume power consumers; the achieved energy savings represent only about 35 percent of their total power consumption. Households in this group may use a lot of electricity for other purposes, too.

Graph: Figure 8. Absolute and Relative Impacts of CFL Bulb Distribution by Quantile Source: Authors' own calculation based on data provided by EEPCO.

The price and income elasticities also vary among quantiles. Price elasticity is the lowest among the first quantile. This reflects the fact that the electricity demand by the poor is mostly to meet basic needs. As household income increases, customers may have more flexibility in energy consumption. For instance, people may afford more energy-efficient home appliances. As a result, price elasticity tends to be higher for the rich. The result is consistent with the literature, such as [27]: the largest part of welfare loss from the energy price increase is found to be associated with the change in purchasing power because of higher prices rather than deadweight loss.

Regarding income elasticity, it tends to be higher for low-volume customers. The elasticity is estimated at 1.76 among the first quantile. The elasticity is still positive at 1.23 but relatively low compared with low-volume customers. This can be interpreted to mean that electricity is not a luxury good but rather a necessity. Therefore, as the economy grows, the power demand is likely to increase disproportionally. The demand by the low-volume users would likely increase most vigorously.

The policy implications are straightforward. Both pricing and CFL bulb distribution programs are effective to manage the electricity demand. First, price incentives, such as increasing block rates, work well even in developing countries where prices are often kept relatively low because of the affordability issue. Still, the measured price elasticity is found to be significant and particularly high in absolute terms for large-volume users. Thus, pricing should be a main policy instrument to rationalize the increasing electricity demand among the rich.

The CFL bulb distribution program is also found to be cost-effective in Ethiopia. The marginal rate applied to those who consume the average amount of energy is Birr 0.55 or US$0.032 per kWh. A CFL bulb costs Birr 25–27 or US$1.42–1.54 in the market (see table 2 above). As shown above, the average energy saving is 45.5 kWh per household or 11 kWh per bulb; the payback period is estimated at about four months. In addition, the normal lifetime of CFL bulbs is 8,000–10,000 hours, which is much longer than that of incandescent lamps. Incandescent lamps usually last 1,000–2,000 hours (e.g. [1]). Therefore, the CFL distribution must be a cost-effective way for households to reduce energy costs.

From the utility perspective, the CFL program is considered to have contributed significantly to keep the required capacity low. In rural Ethiopia, lighting is used for four hours: one hour in the morning and three hours in the evening ([21]). In a more urban setting, people are likely to use more electricity for lighting. The local monthly "price" of indirect hookups assumes the daily use of 60-watt incandescent lamp for ten hours, which costs 10 Birr per bulb ([15]).[14] Thus, the estimated coefficient is translated to a peak reduction by about 150 watts per customer. If this is actually the case, the program is considered to have reduced the peak demand by 13.3 MW in total.[15] This is a cost-effective investment for EEPCO since the total cost of purchasing CFL bulbs (ignoring the bulb distribution or program administrative costs) is only about US$0.5 million.[16]

One might be concerned about the possible contamination effect of the second phase of the CFL program which was implemented in July 2011. Nonbeneficiaries from the first phase may have benefited from the second phase and vice versa. To check robustness of our estimation results, the FE IV regression was performed with the sample panel data limited up to June 2010 (table 8). The results are broadly the same as the main results shown above. The magnitude of the CFL effect is slightly smaller in both linear and logarithmic specifications. This implies that the first phase beneficiaries also benefited again from the second phase program, therefore intensifying the measured effect by the program variable, CFL. It also means that large-volume electricity users seem to remain nonparticipants in the second or third phase of the program; they are less interested in bringing old lamps to the EEPCO service centers to replace them with new ones.

Table 8. Fixed–Effect IV Regression with Limited Sample Data Before the 2nd Phase CFL Program

• 12 Source: Authors' own calculation based on data provided by EEPCO.
• 13 Notes: Clustered standard errors at the household level are shown in parentheses. *, **, and *** indicate the statistical significance at the ten, five, and one percent, respectively.

One might also think that average prices are more important than marginal rates when examining the residential electricity demand. As shown above, our data support the fact that, as economic theory predicts, the marginal price has an important role to play in curbing the electricity demand. Still, there is a practical view that households are only interested in average prices to consider their electricity consumption. In the sample, average real price is 0.227 Birr per kWh, with a wide variation from 0.081 to 9.21 Birr. The results are shown in table 9. In the linear specification, the price elasticity is estimated at 0.39, not negative. When the logarithm is taken on both sides, the coefficient (i.e. elasticity) is 1.09, still positive. These results are consistent with our prior expectation, because the underlying pricing system adopts an increasing block tariff structure where higher prices are applied when people use more electricity. The positive elasticities must capture this relationship.

Table 9. Fixed–Effect OLS Regression with Average Prices

• 14 Source: Authors' own calculation based on data provided by EEPCO.
• 15 Notes: Clustered standard errors at the household level are shown in parentheses. *, **, and *** indicate the statistical significance at ten, five, and one percent, respectively.

The CFL effect seems less conclusive with the average price variable. In the linear model, the coefficient of CFL is negative as expected, but the magnitude is much smaller than the estimate with the marginal price. Moreover, in the logarithmic specification, the coefficient turned out to be positive: CFL bulbs might have raised the electricity consumption. This is less convincing and seemingly indicating the validity of using the marginal prices in the demand estimation with our data.

Electricity infrastructure is one of the most important development challenges in Africa. In the region, about 585 million people, or roughly 70 percent of the total population, are still living without access to power. Significant resources are required to address the existing infrastructure gap. Demand-side management is crucial to curb the increasing demand. A traditional approach is to raise energy prices. Another is to promote energy efficiency. The latter can be an inexpensive, win-win proposition. While end users can reduce their energy costs, power utilities can avoid costly investments in developing new generation capacity. Furthermore, energy efficiency can contribute to mitigating global warming.

Ethiopia has been implementing both approaches: it adopts the increasing block tariff and carried out a series of energy-efficiency programs at the end user level to manage the increasing demand for electricity. The current paper evaluates both the increasing block tariff structure and the first CFL bulb distribution program implemented in June-August 2009. The evidence suggests that both are important to manage the demand. Price instruments are more effective to influence the demand of large-volume power consumers (i.e. the rich). The CFL effect is also significant. However, because the lighting demand is only a small fraction of their electricity demand, the CFL program has a relatively modest effect on these customers.

By contrast, the CFL program is effective to curb the electricity demand among low-volume users (i.e. the poor). It was shown that the CFL program mostly benefited the poor. The program was not targeted, but customers themselves had to bring old incandescent lamps to EEPCO service centers. As a result, the program turned out to be pro-poor. The evidence clearly shows that the CFL bulb distribution program contributed to reducing electricity consumption by about 45 kWh per customer. The payback period is four months, clearly suggesting the cost effectiveness of CFL bulbs at the end user level. Unlike the pricing mechanism, the free-of-charge or discounted distribution is also useful to address the affordability issue.

It was shown that the program had the distributional effects. In absolute terms, large-volume power consumers achieved the largest energy savings. However, in relative terms, lower-volume consumers benefited most from the program. For the poorest (i.e. first quantile customers), the estimated energy savings accounted for 60 percent of their total power consumption.

Although the potential self-selection bias is expected to be removed largely by the individual-specific fixed-effect ci, the conventional OLS estimation of the demand equation is potentially biased because price and quantity are jointly determined in a supply-demand context. Particularly under the block tariff framework, consumers may be able to choose the marginal price by changing their electricity consumption. The literature also indicates the possibility that the block pricing causes measurement errors. Consumers are often uncertain about which marginal rate would be eventually applied before they receive the bills ([6]).[17] Therefore, price and quantity are endogenous and interdependent.

To deal with this problem, we need extraneous information (i.e. instrumental variables). Following the traditional approach in the literature (e.g. [6]), MP and D are instrumented by their predicted values based on the simple expansion of the definition of the Nordin's difference variable. To obtain the predicted values, the actual bill payment is regressed on a constant term and the quantity of consumption:

\begin{eqnarray} Bil{l_{ij}} = {\gamma _{1j}} + {\gamma _{2j}}kw{h_{ij}} + {u_{ij}}, \end{eqnarray}

(2)because |${D_j} = Bil{l_{ij}} - M{P_j} * kw{h_{ij}}$|⁠.Billij is the amount of bill payment for customer i, who is faced with jth block marginal rate. Thus, the marginal price and Nordin's D can be instrumented by |${\hat \gamma _{2j}}$| and |${\hat \gamma _{1j}}$|⁠, respectively ([18]). Note that |${\hat \gamma _{2j}}$| and |${\hat \gamma _{1j}}$| are block-specific predictions estimated from a separate regression for each j. The two parameters are not exactly identified because there are several issues that cannot be observed by researchers. In theory, the difference variable is equal to the income variable in magnitude in a linear model. However, this often fails in reality, because of lack of consumers' information of price structure and the fact that the difference variable only accounts to a small fraction of the total household income ([28]). In our case, there are several possible measurement errors and unanticipated payment practices. Some households have arrears and monthly payments are postponed or paid through some installment arrangements. Some fixed charges are also incurred to some customers such as connection fees. These result in nonlinearity of individual household payments within each marginal tariff block.

Atsushi Iimi (corresponding author) is a Senior Economist in Transport & ICT Global Practice of the World Bank where he specializes in development economics in Africa; his email address is aiimi@worldbank.org. Raihan Elahi is a Lead Energy Specialist in Energy & Extractives Global Practice of the World Bank; his email address is relahi@worldbank.org. Rahul Kitchlu is a Senior Energy Specialist in Energy & Extractives Global Practice of the World Bank; his email address is rkitchlu@worldbank.org. Peter Costolanski used to be a consultant working at the World Bank; his email address is pcostolanski@gmail.com.