Power to Choose? An Analysis of Choice Frictions in the Residential Electricity Market

Transcription

1 Power to Choose? An Analysis of Choice Frictions in the Residential Electricity Market Ali Hortaçsu, Seyed Ali Madanizadeh and Steven L. Puller 1 April Hortaçsu: University of Chicago and NBER, hortacsu@uchicago.edu; Madanizadeh: University of Chicago, seyedali@uchicago.edu; Puller: Texas A&M University and NBER, puller@econmail.tamu.edu. We thank the University of Chicago Energy Initiative and EI@Haas for generous financial support. We are grateful for assistance with data and institutional questions from Kelly Brink, Robert Manning, Calvin Opheim, and Jess Totten. We received very useful comments from presentations at Camp UCEI, Department of Justice, East Anglia, the 2010 FTC Microeconomics Conference, Iowa State, Maryland-AREC, Michigan State, NBER Summer Institute, the Public Utility Commission of Texas, SWUFE, Texas A&M, and Washington University-Olin.

2 Abstract This paper studies the size and consequence of consumer inertia in a setting where retail choice is first permitted. Retail choice has been an important feature of the deregulation of many formerly regulated markets such as long-distance telecom, electricity, and natural gas. The expansion of consumer choice has been touted as a means to allow consumers to pay more competitive prices and enjoy more value-added services. However, choice frictions can diminish the benefits of retail choice. We use household-level data for the first four years of residential electric choice in Texas to study the choices of consumers. We find that the majority of customers continued to purchase from the incumbent despite large potential monetary savings from switching. We develop an econometric model that allows us to distinguish between three different sources of this observed inertial behavior (1) inattention due to status quo bias, (2) an incumbent brand advantage, and (3) switching costs. We find that the incumbent enjoys a brand effect that provides it with a market share advantage even when lower-priced newcomers enter the market. However, the incumbent s brand advantage appears to diminish over time. We also find that search and switching costs are important; only a fraction of households actively consider their options each month, and those who do consider other options are face switching costs. We use our choice parameter estimates to conduct a policy counterfactual experiment that informs consumers about the ability to choose and the nature of their options, and we find that such an information treatment could yield notable consumer surplus gains. Our findings have implications for the design of new markets when consumers are provided with choice. Keywords: consumer inertia, search costs, switching costs, electricity markets

3 1 Introduction Retail choice has been an important feature of the deregulation of a variety of markets that were formerly regulated, including long-distance telecom, electricity, and natural gas. Under retail choice, consumers who previously purchased utility services from a single utility at regulated rates are allowed to purchase from alternative suppliers at prices that are unregulated. Competitive firms are permitted to enter a market and offer retail services to consumers who could purchase only from a regulated utility in the past. This expansion of consumer choice has been touted to have several benefits. Creating competition for the provision of utility services can lead to more competitive pricing in the short-run. In addition, a market for services that were previously supplied only by a regulated utility may incentivize firms to develop more value-added services in the longer run. However, choice frictions can diminish the benefits of retail choice. Consumers who have never had the option of choosing a utility service provider will not necessarily exercise the option to choose a lower price alternative provider. For example, consumers may not actively acquire information on the options offered by other providers, even if that information would indicate that better options exist. Also, consumers may be inhibited from switching to another provider if they incur monetary or psychic costs of switching. Or households may value the brand name of the incumbent the old utility and this may reduce the amount of switching to other providers. Each of these sources of choice frictions can mitigate the consumer surplus gains of retail choice policy. This phenomenon allowing consumer choice in settings where consumers previously did not have options, and the associated choice frictions is not confined to formerly regulated utilities. In the healthcare sector, the prescription drug benefit program under Medicare Part D provides the elderly with multiple plan options provided by private insurance companies rather than a single plan specified by the government. 1 In primary education, parents in some jurisdictions are offered a menu of public schools that their children can attend rather than a single school that children are zoned to attend. 2 And in the arena of social security, the traditional role of government in pay-as-you-go retirement 1 For example, See Abaluck and Gruber (forthcoming) who study choice of Part D plans and find that many elderly do not choose the plan on the efficient portfolio of plans, and that restricting the set of choices can be utility increasing to the elderly. Kling et al. (2009) also study Medicare Part D choice behavior and show that information interventions with personalized cost estimates of each drug plan can affect plan choice. 2 See, for example, Hastings, Kane and Staiger (2009) and Hastings and Weinstein (2008) on the school choice and the effects of information provision. 1

4 systems has been replaced by privatized retirement planning in which individuals choose from among a set of privately managed funds. 3 In this paper, we study a particular retail choice program to measure the size of consumer inertia and to better understand the underlying mechanisms behind the inertia. The recent opening of the Texas residential electricity market provides an excellent setting to investigate retail choice, using a detailed micro panel dataset on household s choices. We measure the size of various sources of consumer inertia and evaluate the welfare and distributional consequences of retail deregulation in the presence of search frictions. Beginning in 2002, residential electricity customers in many regions of Texas were allowed to choose their retail provider. Initially, all customers were assigned to the incumbent utility but could switch to one of several competitive retailers that competed on price. The retail choice program has been under the oversight of the Public Utility Commission of Texas, which actively sought to increase the transparency of the choice process, through pervasive advertisement and consumer education programs throughout the state. The PUC also administers the website a centralized source of information that is easily accessible to households, and informs them on the choices they have over retail electricity providers. Consumers can use this website to view electric plan choices and then to switch their provider. An initial look at patterns in the aggregate data suggest strong evidence of consumer inertia. The incumbent electricity providers have held on to their market share leadership during the first four years of retail choice. In Figure 1, we display the evolution of market shares of the incumbent vs. the competitive providers in an electricity service area of Texas. Although the incumbent has lost market share (from nearly 100% to around 60% over the four years), it is still by far the largest player in the market. This market share graph stands in stark contrast, however, to the rates being charged by the competing retailers. We see from Figure 2 that at least one competitive retailer has undercut the incumbent in nearly every month of the deregulation experiment. Indeed, as we will show in more detail in Section below, it appears that a large number of households have resisted switching away from the incumbent to one of the competitive retailers, even though they would have enjoyed substantial monetary savings from doing so. 3 See Duarte and Hastings (2011). 2

5 These patterns provide motivation for an in-depth study of the seemingly inertial behavior of residential electricity consumers. behavior: We model three mechanisms that could account for this inertial Search costs/inattention: Households may not choose to search for alternative providers. Even though is a click away, the household may not be aware of it, or the household may exhibit a status quo bias so that it does not consider the offerings of alternative providers. Incumbent Brand Advantage / Product differentiation: Even if the household is aware of all available options, consumers may believe that service during power outage events or other dimensions of power quality could differ across retailers. Such beliefs, even if technically incorrect, may be a source of vertical product differentiation. Alternatively, consumers may believe that characteristics of customer support such as ease of paying bills vary across providers. Switching costs: Even if the household is aware of all available options, switching to another provider may be costly either due to explicit monetary costs of switching (early cancellation fees), or due to the hassle of making the switch. The contribution of this paper is two-fold. First, we estimate the relative magnitudes of these sources of inertia because optimal policy to enhance the consumer surplus benefits of retail choice depends upon the source of inertia. For example, suppose that the primary reason that consumers continue to purchase from the incumbent rather than from a lower priced provider is that they believe that power quality depends upon the provider chosen. (Technically, the delivery of power via local powerlines and metering is not a function of retail provider, but consumers may not be aware of this). In this case, information campaigns to inform consumers that it s all the same power may induce households to choose lower priced providers. Moreover, if the brand effect naturally diminishes with time, then any brand advantage may be only a transition cost to retail competition. On the other hand, if search costs are substantial, then regulatory nudges for consumers to become informed about available retail options may reduce consumer inertia. However, if switching costs are substantial, then policymakers may need to expect there to be permanent sources of inertia when opening retail electricity markets to competition. 3

6 Second, we develop an empirical methodology to empirically quantify each of the three possible sources of inertia. Although the extant literature includes rich analyses of individual sources of inertia, we are not aware of any existing methodology that allows for all three sources of inertia. Such a methodology is valuable in its own right because these classes of inertia are likely prevalent in a variety of settings where a default option is replaced with consumer choice, including schooling, health, and utilities. Our econometric model of household choice in Section 4 nests all three sources of consumer inertia within a two-stage discrete choice framework. In the first stage of the model, the household decides whether to consider the available choices (the relevant time interval is a month). With some probability, no such consideration is made, thus the household stays with its current provider. Conditional on making the choice to consider all available alternatives (which are easily accessible through powertochoose.com), the household then makes a decision between the alternatives according to a standard discrete choice model. In this second stage we allow non-price differentiation factors to enter the household s decision, thus capturing potential brand advantage by the incumbent. The second-stage choice model also allows for switching costs by allowing the utility from the current provider to be different from the utilities of other providers. In Section 4.3, we also provide transparent conditions for the identification of model parameters from sample moments. We show that the first-stage search or decision probabilities are separately identified from the second-stage choice probabilities. Moreover, by exploiting the presence of some households who institutionally do not face search or switching costs, we are able to identify the switching costs that may affect the second-stage choice probabilities of incumbent households. Our results indicate that all three economic forces are at play in rationalizing the marketshare dynamics in this market. The incumbent enjoys an economically very significant brand effect consumers value the incumbent s brand at nearly $80 per month. While this effect appears to diminish over time, it still explains a large portion of the variation in market shares. The search and switching cost components of decision-making also play important roles. We estimate that consumers of the incumbent only search for other retail options in about 5% of months, or approximately once every 18 months. While the percentage of households who actively search in a given month is not large, the search activity shows intuitive seasonal patterns: consumers are most likely to search in summer months, during which electricity bills tend to be high. As far as switching costs, we find 4

7 evidence of costs of switching away from competitive retailers but no robust evidence of switching costs for incumbent customers. In order to evaluate the implications for policy, we perform a counterfactual experiment of a policy intervention to reduce two of the sources of consumer inertia. In this counterfactual, we change two factors of the market: (1) increase the probability that a household searches in a given month, and (2) reduce the relative size of the perceived brand effect. We have in mind a policy intervention that sends a flyer in the monthly bill that tells consumers they can go to to view options and that their power quality will be the same under any provider. We view this as a fairly low cost policy intervention that combines a nudge with an information treatment. Our policy experiment suggests that consumer surplus could increase by around $100-$200 per year per household. Our paper relates to several other sets of ongoing research. This paper is related to research that has studied the consequences of programs that extend consumer choice rather than impose a default option. Previous work has studied settings such as school choice (Hastings, Kane, and Staiger, 2009), retirement savings behavior (Madrian and Shea, 2001), long-distance telecom (Hausman and Sidak, 2004), and health (Handel, 2009; Abaluck and Gruber, forthcoming; Kling et al., 2009). In such settings, program designers have made concerted efforts to decrease information costs, yet consumer inattentiveness may impact decisionmaking. In contrast to some of the settings studied in these other studies, households choosing their electricity provider require much less mental math, as we describe below, which makes our findings of consumer inertia even more striking. Also, our paper is related to work studying the retail choice behavior in utilities (see, for example, Waddams, Guilietti, and Waterson, 2005 and Miravete, 2003). A variety of countries have offered retail choice in utilities such as natural gas and electricity, and these jurisdictions have led to a mixed record on the number of consumers who switch away from the incumbent. 4 2 Retail Electricity Choice in Texas As in many states, residential electricity customers in Texas historically were served by vertically integrated electricity providers at regulated tariffs. The state was divided into separate service territories, each with a vertically integrated firm. Beginning in 2002, residential electricity customers in most 4 For an analysis of the merits of retail choice in electricity, see Joskow, For a review of experiences in residential electric choice and a representation of consumer search in such markets, see Brennan,

8 parts of Texas were allowed to choose their retail provider. Starting on January 1, 2002, all customers were assigned to a retail firm that was affiliated with the old incumbent utility. However, in any subsequent month, a customer had the option to switch to any other retail electricity provider at no cost. In addition, the former utility was split into a lines company and a retailer. The operations of all electricity transmission, local powerlines, and meters is now operated by a regulated firm (with a different name) that is separate from the new (incumbent) retail firm. As a result, the quality of power service (e.g. outages) is technically independent of the retailer chosen by a household the household-retailer relationship is purely financial with no impact on the technical operations of the grid. (It is possible that consumers were not aware of these technical operations, and we discuss that below when interpreting results). The incumbent regulated retailer, referred to as the affiliated retail electric provider (AREP), was required to reduce residential rates by 6 percent from previous levels, and this regulated rate was called the price-to-beat. This rate was the only tariff allowed to be offered by the AREP. 5 The AREP was allowed to request an adjustment to the price-to-beat up to two times a year; the size of the adjustment was not chosen by the AREP but rather was determined by a formula indexed to the price of fuel (i.e. natural gas). Other competitive retail electricity providers (CREPs) can procure power in the wholesale market and market that power to residential customers. Prices by CREPs were not regulated. Some of the largest competitive retail firms were AREPs from other service territories. In 2002, most service territories had between three to five CREPs, and by the end of our sample in 2006 the choice set expanded beyond ten. Because of relatively low natural gas prices during the first few years of retail competition, the price-to-beat was higher than competitive prices for retail power. This so-called headroom was thought to provide new retailers with sufficient margins to encourage entry. As a result, CREPs were able to price more than one cent per kwh below the price-to-beat (the median monthly electricity consumption in our sample is 968 kwh). As we discuss below, this created the potential for mean savings of about $12 per month for switching away from the incumbent AREPs. 5 Starting in 2005, the AREPs could offer alternative rates to incumbent customers but were still required to offer the PTB tariff. Few AREPs offered rates other than the PTB during our sample period. In January 2007 after our sample period ends, the PTB was lifted and the AREPs could charge existing customers any rate. 6

9 Households had multiple sources of information on potential electric providers. The most salient source of information was a well-publicized website established by the state s public utility commission Consumers could enter their zip code into the web interface and view a list of providers. Households could follow links on this website to switch to a new provider in a process that was online and relatively fast. In addition to the website, retailers advertised through a variety of media including radio and billboards. The initial switch away from the incumbent to competitive providers did not impose dollar costs on households. In addition, a household switching away from the incumbent did not forfeit the right to switch back to the incumbent s price-to-beat later. The only explicit cost of switching were early termination fees assessed to households that signed 12-month fixed term contracts but chose to switch to another provider before the contract expired. 6 By regulation, the incumbent s rate was month-to-month so incumbent customers never faced these fees. Regardless of the retailer used, a household would still receive a single monthly bill that included charges for all electricity services (energy, transmission, distribution, metering, and billing). Thus, the primary non-monetary switching costs are those associated with paying a different provider. 3 Data We study the retail choice behavior of all residential electricity customers in the service territories covered by one of the formerly vertical integrated regulated utilities TNMP. This former utility has the appealing feature that it was formed by several mergers over the years and therefore has customers spread throughout the state. In particular, the areas of the state that we study include both urban and rural areas. We have monthly data on each of the approximately 192,000 residential meters in TNMP territory from January 2002 until April For each meter, we have the residence s address which we match to the Census block group to measure demographic characteristics of residents and some characteristics of housing stocks. In addition, for each meter we have the monthly electricity consumption 7 and the monthly choice of retail electric provider. We focus on the six retailers who had at least 1% of the market at some point in time from 6 Such fees are waived if the customer moves out of the residence. 7 Meter reads do not occur exactly on the last day of the month, so we perform an interpolation to estimate the consumption occurring during each calendar month. 7

10 January 2002-April These six retailers include the previous incumbent utility, two retailers that were the affiliated retailers in other service territories, and three retailers who were entrants into the marketplace and had no affiliation with incumbents in other service territories. Our data on monthly choice of retailer is reported for the meter/address. We do not know the names of the customers on the bills, so we can only measure when the residence is served by a different retailer. To allow for the possibility of turnover in the residents, we assume that the residents do not change unless there is a disconnect of service. If electric service is disconnected for more than 30 days, we assume that new residents occupy the residence and call that event a move-in. Otherwise, we assume the same residents are making decisions for the residence. Thus, for any change in occupancy that involves no disconnect or a disconnect shorter than 30 days, we cannot account for changes in the agent making choices. 13% of meters experienced at least one move-in at some point from Any new meter installed after 2002, e.g. a new house, had no default choice. Thus, households were required to explicitly choose a retail provider. About 13% of meters existing at the end of our sample period appeared after We will exploit this institutional feature in our empirical strategy. We are able to calculate each household s bill with high precision. For each household, we have monthly choice of retailer and electricity consumption. The Public Utility Commission of Texas collected monthly information on the rate plans offered by each retailer in the different service territories. In many cases, the retailer offered only a single rate plan, so we can precisely measure the monthly bill. In fact, only one rate plan was offered by four of the six retailers that we model, including the incumbent. However, a complication is presented by the fact that two retailers offer a menu of rate plans, and we have no information on which plan is chosen by a given household. In these two cases, we consulted an industry analyst and assume the households chose the plan judged to be most popular by the industry analyst. 9 Most of the retailers used multi-part tariffs. For the incumbent, the tariff is a fixed fee of $5.17 and then an increasing block tariff with the second block beginning at 400 kwh per month. 10 Three 8 We exclude one competitive retailer with 1.1% market share for which no price data are available. 9 For example, the affiliated retailer from another territory offered a month-to-month plan for the entire sample period. However, that retailer also offered a renewable plan for much of the sample period and a rate for a 12-month fixed rate contract for about the second half of the sample period. (The average rate for the 12-month contract was about 0.2 to 0.3 cents/kwh cheaper than the month-to-month plan.) An industry analyst told us that few customers were likely to have purchased the renewable plan, so we assume those customers chose the 12-month contract. 10 The second block has an additional fuel charge added from June to October. 8

11 of the other retailers had similar tariff structures fixed fees around $5 and increasing block tariffs beginning in the middle of the sample period. One retailer had a relatively higher fixed fee of $8.70 followed by a single block tariff. And the last retailer had a linear tariff no fixed fee and a single block tariff. 11 Because the rate plan is a very important determinant of consumer choices, we need to choose an appropriate measure of price that is likely to drive a household s decision process. Two options are available the marginal price and the average price. The marginal price is likely to be the same for all households with any given retailer because even those retailers with increasing block tariffs have the highest block begin at a low usage level (400kWh/month). However, the average price differs by (expected) consumption because all but one retailer employ non-linear tariffs. Although theoretical work on non-linear tariffs usually assumes that consumers respond to marginal prices, we believe that average price is more appropriate in this institutional setting. The standard assumption that consumers respond to marginal prices assumes that households have full information on the multi-part tariff function, how that tariff function varies over a year, and knowledge of expected consumption. This assumption is not likely to be reasonable for the Texas residential electric market for several reasons. First, powertochoose.com, the website created by the PUCT that lists all available retailers, saliently displays only the average price for customers consuming 1000kWh/month. (Other details that are displayed saliently are the renewable energy content of the power, the term of the contract and whether the contract has a cancellation fee.) Consumers who want to gain more detailed information about the tariff may also click to download the Facts Label that are required to contain specific parameters of a retailer s service. The rate information on these Facts Labels is the average price for customers consuming either 500, 1000, or 2000 kwh. It is, in fact, impossible to recover the shape of some of the nonlinear tariffs using this information! Consumers would have to contact individual retailers to get details on the nonlinear tariffs. A second reason to use average price is that recent empirical evidence suggests that residential electricity consumers may respond to average price more than marginal price (Ito, 2010). Because the average monthly consumption is closest to 1000kWh/month (the average consumption was We should note that about 6% of customers received discounts as part of a low-income program. However, eligibility for this program was independent of retailer. We do not have data on which customers qualified for this program, so we are forced to assume that these customers pay the standard tariff. 9

12 kwh/month and median consumption was 968 kwh/month), we use the retailer s average price at this consumption level as the price that determines retail choice. Figure 2 shows the evolution of the average price at 1000kWh by retailer in the TNMP service territory. Rates ranged from about 8.5 to 14 cents per kwh from Rates were generally rising over the sample period with much of this driven by rises in the price of natural gas, a primary determinant of wholesale electricity prices in Texas. In particular, rates jumped several cents in late 2005 following the natural gas price spikes caused by Hurricanes Katrina and Rita. The average rate of the incumbent (i.e. the price-to-beat rate) was systematically higher than one or more other retailers throughout most of the sample period. In fact, by the Fall of 2002, the first year of retail choice, at least one competitive retailer offered an average rate at least one-half cent cheaper than the incumbent in every month except one month in late Moreover, in many months in the middle of the sample, a competitive retailer s average rate was over one cent cheaper than the incumbent s price-to-beat. Also, it is worth noting that the green product is priced at or above the price-to-beat throughout the sample period. 3.1 Summary Statistics We begin by providing a basic description of observed switching behavior and the potential dollar magnitude of expenditure differences of purchasing from alternative providers. These summary statistics of the raw data yield many patterns that are consistent with results that arise from our model s estimates in Section 5, thus providing support for our modeling assumptions Switching: Timing and Frequency The incumbent maintained a large market share despite charging higher rates than competitive retailers. Figure 1 shows the market shares of the six largest retailers over the first four years of retail choice in TNMP. The incumbent exhibited a gradual erosion of market share throughout the sample, but still maintained over a 60% share by April Two other retailers had over a 10% share while the remaining three large retailers had less than a 5% share. The total number of switches per month was relatively low in the first year of retail choice but then rose in the following three years. Figure 3 displays the total number of switches from one provider to another each month in our sample. There appears to be a seasonality in switching behavior with 10

13 a peak in the summer months. The peak month for switching was July in 2002, June in 2003, and August in 2004 and The summer is also the season with the highest monthly bills due to the electric cooling. Figure 4 displays the 20th, 50th and 80th quantile of bills each month. Summer electric bills can be twice as high as winter bills. This seasonality in expenditures may create greater saliency of electricity choice and is certainly consistent with increased switching in the summer. Our model below allows for seasonality in the decision to consider alternative retailers. Also, we measure the frequency with which a given household switches retailers. Figure 5 displays the number of times a meter is observed to be served by a new provider from January April Recall that we do not have information on the occupant of each house, so we cannot always measure a change in tenancy. However, we can exclude new meters and move-ins to exclude some changes in tenancy and proxy for households with a consistent occupant throughout the sample. Figure 5 shows that approximately two-thirds of households never switch providers (which is consistent with the incumbent s market share at the end of the sample). Among those households that do switch, most switch only once (20% of meters) or twice (9% of meters) in the first four years of retail choice. Finally, we measure the retail choices of households that had no default option and were required to make an explicit choice. Institutionally, all residences built after 2002 and all customers moving into an existing residence after 2002 had to make a choice, else they would have no power. 12 As we discuss above, we proxy for these households using meters identified as new movers and move-ins. Among the three sources of search frictions identified in the introduction, decision costs about the possibility of retail choice is not likely to be present for these customers. Figure 6 displays the market share each month for new meter and move-in customers making their initial choice of retailer. Even among these customers, the incumbent has a very high market share. This suggests that neither search/decision costs nor switching costs are likely the sole determinants of consumer inertia; rather there is a sizeable incumbent brand advantage Potential Savings If there were no sources of consumer inertia (i.e. no brand advantage, no search costs, and no switching costs), then consumers would view electricity as a homogenous product and switch to the lowest- 12 Exceptions could include apartments where the building owner chooses the retailer. 11

14 priced provider. However, as we describe above, we observe a large market share by the higher priced incumbent, and we view this as evidence of some form of inertia. In order to document a magnitude of this phenomenon, we measure the savings to households of buying the same amount of power from an alternative provider. To do so, we calculate the bills for each household in months it purchased from the incumbent and the counterfactual bill if the household had purchased the same amount of power from other retailers. This exercise should not be seen as a welfare analysis but rather as an assessment of the magnitude of potential expenditure reductions from switching. Consider two extremes of the frequency with which a household switches. First, consider a scenario in which a consumer switches only once during the four year sample and does so in the first month of retail choice (January 2002). We calculate the monthly savings if each household had switched to one of the two large competitive retailers, which in this case are AREPs from other service territories. The mean monthly savings of purchasing from one of the larger retailers is $7.65 and the mean savings purchasing from the other is $9.92. At the other extreme, consider a scenario in which a consumer switches to the lowest price retailer each month. For households that continued to purchase from the incumbent at the price-tobeat, the mean savings per month of switching to the lowest price retailer is $12.41 and the median savings is $7.56. Figure 7 plots the distribution of monthly savings if the household had consumed the same amount of electricity but purchased from the lowest priced retailer. The savings is over $16 per month for about one quarter of the household-months and over $29 dollars for about one-tenth of the household-months Demographic Differences in Response to Retail Choice Finally, we provide some initial evidence that retail choice disproportionately benefits specific demographic groups. To do so, we calculate metrics of the fraction of potential savings that were realized by switching, as compared to a benchmark of purchasing from the incumbent at the price-to-beat for the entire sample period. We assume consumption remains constant and calculate the monthly bill size under three scenarios: 1) staying with the incumbent, 2) purchasing from the low price retailer each 13 We also assess the savings realized by those households that switched rather than purchase from the incumbent at the price-to-beat, using similar assumptions as above. For those months in which households purchased from any retailer other than the incumbent, the average bill would have been $8.79 higher per month if the same consumption were purchased from the incumbent. 12

15 month, and 3) the actual retail choices. Our upper bound measure of electricity expenditures is the bill size if the household had purchased from the incumbent for the entire sample period. We view this baseline as a proxy for electricity expenditures in a regime without retail choice. 14 Our lower bound of expenditures is the monthly bill size if the household had purchased from the lowest price retailer each month. Finally, we calculate the actual monthly bill under the observed retail choice by the household and compare it to these bounds. 15 Percent achieved is the percent of possible gains realized ((actual bill - incumbent bill) / (optimal bill - incumbent bill)). The mean Percent achieved across all household-months is 11.0%. This relatively low figure should not be surprising because nearly 60% of households purchase from the incumbent at the end of the sample period. We characterize correlations between gains realized and demographic characteristics by regressing Percent achieved on demographic characteristics of the household s neighborhood. Note that we do not have demographic data on the occupants of each household; rather we have characterics of the household s Census block group. Accordingly, we interpret these regressions to show correlations between realized gains of retail choice and demographics of the neighborhood rather than demographics of individuals. 16 Table 1 shows the results. Column (1) shows that households with a higher Percent achieved tend to be in neighborhoods with a higher educated population, a lower poverty rate, and a greater fraction of households in an urbanized area. In addition, a higher Percent achieved is realized in neighborhoods with ceteris paribus fewer senior citizens, more blacks, fewer Hispanics, and fewer houses with electric heating. In column (2) we add household-specific data on average monthly electricity consumption and find that higher usage households realize a greater Percent achieved. For just over 40% of the households in our sample, we are able to obtain individual-level char- 14 As we describe above, the incumbent s tariff is regulated and any adjustments to the price-to-beat is indexed to the fuel cost, so it is plausible that this tariff is a reasonable approximation of the tariff in a counterfactual regulated regime. 15 Our upper bound is not a literal upper bound for several reasons. First, the green provider charged rates higher than the incumbent for parts of the sample period, so this baseline ignores preferences for renewable energy sources and other forms of product differentiation. Also, there were some periods in our sample when one of the competitive retailers charged a price above the price-to-beat. Nevertheless, we use the incumbent to create the upper bound because it represents a default outcome if no active choices are made by existing residents and it is plausibly the counterfactual price in a regime with no retail choice. 16 Borenstein (2009) documents the heterogeneity within Census block groups and the shortcomings of using such metrics for distributional analyses in some settings. 13

16 acteristics of the house. We matched data on the address of each meter to information from the online real estate database Zillow. We should note that the houses that we could match to Zillow data oversample urban areas, so one should not view these matches as a random sample. Columns (3) and (4) of Table 1 show the results. We find that homes with higher value and higher electricity usage realize a higher Percent achieved. To the extent that house value is a proxy of occupant wealth, this suggests that wealth is positively associated with Percent achieved. The regressions also include Census block group demographics and much of the inference from columns (1) and (2) persist higher percentage gains are realized in neighborhoods with higher education levels, lower poverty, fewer seniors, and more blacks. Also, these results indicate that after controlling for house value, neighborhoods in this sample with more electric heating realize higher percentage gains. One should be cautious in interpreting the above correlations as demand-side behavior. Clearly, neighborhood demographic characteristics are correlated with other characteristics that may drive supply-side effects such as advertising. Nevertheless, the correlations suggest that certain demographic groups benefit more from retail choice, and this is important for understanding the distributional consequences of the policy. We should heavily emphasize that these calculations do not constitute a formal welfare analysis. The calculations do not allow for heterogeneous preferences for product characteristics such as brand effects, the quality of customer relations, whether the product is green, or the term structure of the retail contract. 17 Nor do these calculations incorporate the cost of search. Nevertheless, these simple calculations provide evidence on the potential magnitude of the consumer surplus effects of retail choice. 4 Model We model the household-level choice of electricity provider as a two stage process that occurs each month. Each month, in stage 1, the household has a current provider and decides whether to consider alternative providers with some probability. We refer to this stage as the Decision to Choose Stage. In stage 2, the households who choose not to consider alternative providers in stage 1 will maintain their 17 We have heard anecdotal evidence that consumers believed that the quality or reliability of power was determined by their retail provider, as it may have been during the deregulation of long-distance telecom. However, the delivery of electricity is entirely determined by a separate transmission and distribution utility, so loss/reconnection of power is independent of the choice of retailer. 14

17 current provider for the month. However, households who choose to consider alternative providers in stage 1 will then choose the provider that maximizes utility among the retailers in the market. These households may choose a different provider or may continue with their current provider. We refer to this second stage as the Choice Stage. We allow for heterogeneity across households and across time at both the Decision to Choose Stage and the Choice Stage, as we describe below. Some households in our sample are identified as moving into the residence in month t. 18 In that month, the household has no current provider and is modeled as searching with probability 1. For the remaining households, the search probability is estimated as we describe below. One empirical complication is that we do not observe the outcome of the Decision Stage. We only observe households who change providers, i.e. those who decide to consider alternative providers and then choose a different provider. From the analyst s viewpoint, households who do not switch are both those who do not consider alternatives and those who do but choose their current provider. We describe a model below that allows us to separately identify parameters of the Decision Stage and the Choice Stage. We need to make an assumption about consumers expectations of future retail electricity prices. Although we have no direct evidence on consumer beliefs in retail electricity, there is evidence from another energy commodity gasoline that consumer beliefs are consistent with a no-change forecast (Anderson, Kellogg, and Sallee [2011]). Thus, we assume consumers expect that current prices will remain constant, which also allows us to reduce the household-level choice to a static model. 4.1 Stage 1: Decision to Choose In this stage, each month a household decides whether to choose from the set of providers in the choice set or not. We model the probability of considering a (possibly new) provider to vary by several characteristics of the household. First, the probability varies by the household s current provider because a household s experience with its existing provider may induce it to consider alternative providers. Second, we allow for seasonality in the months of the year that customers actively decide upon their provider. This allows the model to attribute some of rise in switching during summer months observed in Figure 2 to result from more searching during the summer. And finally in some 18 These movers include both meters with service disconnected for at least 30 days (13% of meters) and new meters installed after the start of retail choice (also 13% of meters). 15

18 specifications, we allow the decision probability to vary in the household s demographic characteristics. We denote current providers by k and the chosen new provider by j (again, recall that k and j can be the same if the households searches and chooses its current provider). Month is indexed by t. We model the decision probability for a household that is currently a customer of provider k at time t, denoted λ k t, with a standard logistic S-curve : ew k t λ k t (γ) = 1 + e W t k where Wt k = r γ rzrt k and { Zrt} k is the set of observable characteristics used to estimate the decision probability. The variables include dummy variables for each existing retailer k, month of the year dummy variables to allow for seasonality in deciding to search, and in some specifications, demographic characteristics of the household s census block group. This specification envisions the decision to choose as a push rather than a pull effect. Households are driven to consider alternative providers in response to events such as a large summer bill from their current provider rather than being induced by another provider; this specification is consistent with anecdotal evidence from industry analysts. 4.2 Stage 2: Choice of Retailer In this stage, each household who enters the choice stage in month t chooses the provider from the choice set that yields highest utility. For households who enter the choice stage, the indirect utility for each household i of choosing retailer j in period t is: U (k) ijt = V (k) ijt (θ) + ε ijt (1) where V (k) ijt (θ) is a parameterized utility term, and ε ijt is a random utility shock that is i.i.d. across consumers, providers, and time. V (k) ijt (θ) is further specified as: We assume ε ijt to be a Type I Extreme Value random variable. V (k) ijt (θ) = s θ s X (k) ijt,s (2) Here, X (k) ijt,s is the s th characteristic of retailer j at time t. The product characteristics include the price, an indicator for whether the retailer is the incumbent, and the incumbent indicator interacted 16

19 with a time trend. This specification allows for non-price product differentiation by the incumbent. Specifically, the variables comprising X (k) ijt are the following: 1. p j,t is the tariff of retailer j in month t for 1000 KWH usage per month for the following 12 months (as reported on the website powertochoose.com ). This is observable by households and the researcher, as we describe in the data section. 2. d Incumbent is an indicator variable for the incumbent, allowing for an incumbent brand effect 3. d Incumbent MONT HCOUNT ER is the incumbent indicator interacted with the number of months since the market began, allowing for a time trend in the incumbent brand effect Because ε ijt is a Type I Extreme Value random variable, the probability that household i chooses retailer j in month t is given by the familiar logit probability: P ijt (θ) = exp (V ijt (θ)) k 1 exp (V ikt (θ)) (3) This probability is used in GMM estimation that is described below. 4.3 Simultaneously Estimating Decision and Choice Stages We simultaneously estimate both the decision to consider alternative providers and the choice stages. In order to do so, we exploit our data on observed switching behavior to derive a set of moment conditions. As noted above, one empirical challenge is that we do not directly observe the outcome of Stage 1. Rather, we observe switches to other retailers for those who decide to consider alternative retailers. It is possible that a household decides to consider alternative retailers in a given month but chooses to stay with the same retailer, in which case we do not directly measure the decision outcome. We address this complication with two strategies. First, there are some households that consider alternatives with probability equal to one, namely those households in the month when they move into a house and there is no default provider. For these households, we can estimate only the parameters of the choice stage (because search occurs with probability equal to one) and utilize moments generated by the stage 2 model of choice. 17

20 Second, we exploit the observed month to month aggregate switching from the old provider k to the new provider j to estimate the probability of search for different types of customers. This model of the flow of customers from one provider to another provides additional moments for our GMM estimation. Illustration of Empirical Strategy We illustrate the empirical strategy with a simple example to demonstrate how we separately identify the decision and the choice probabilities. For purposes of illustration, assume that we only observe two months of data the customer s retailer last month and this month (so we suppress the t subscript). In addition, assume that there are only 3 retailers. Let each household served the previous month by retailers 1, 2 and 3 decide to consider alternative providers with probability given by λ 1, λ 2, and λ 3 respectively. And let the probability of choosing retailer 1, 2, and 3 conditional upon entering the Choice Stage, be given by P 1, P 2, and 1 P 1 P 2, respectively. We want to estimate these five probabilities. Conceptually, we can create a matrix of counts of the number of customers switching from retailer k to retailer j during the month. The cells of this matrix provide us with statistical moments that we use to estimate the 5 parameters/probabilities. Consider all households who were served by retailer 1 in the previous month, and denote this number N (1). Some of these households will be observed to also use retailer 1 in the current month; such households are ones who did not consider alternative providers (occurring with probability 1 λ 1 ) and those who did consider alternatives but chose to remain with retailer 1 (occurring with probability = λ 1 P 1 ). Likewise consider households observed to use retailer 1 in the previous month and retailer 2 in the current month. These households are ones previously with retailer 1 who considered alternatives and chose retailer 2 (occurring with probability = λ 1 P 2 ). Likewise, we can characterize households previously with retailer 1 who considered alternatives and chose retailer 3 (occurring with probability = λ 1 (1 P 1 P 2 )). The expected number of customers who were initially with retailer 1 and continue to use retailer 1 is: N (1) [(1 λ 1 ) + λ 1 P 1 ]. The expected number of customers who were initially with retailer 1 and switched to one of the other retailers is: N (1) λ 1 P j for j = 2, 3. This provides 3 moments to match to sample moments on the number of customers flowing between retailer 1 the previous month and all 3 retailers in the current month. 18