Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.
- Forma~ Mode~ing and Prob~em-Oriented Research UNDERSTANDING THE DEMAND FOR ENERGY The past decade has demonstrated the need to improve understanding of U.S. energy demand. Demand projections have repeatedly proved inaccurate--usually on the high side--and there is little reason for confidence that today's projections will be any more accurate than those of the past. Yet accurate analysis of energy demand remains an important national priority. It is needed if policy makers are to find effective ways to avoid national crises when there are shocks in world oil markets. It is needed to set appropriate levels of investment in electric power generation so that the country can avoid the eco- nomic drain of unnecessary investment and the threat of widespread brownouts and blackouts. And it is needed to aid in decisions about how serious a public effort is needed in the areas of energy production and conservation. But in spite of this clear need, it has so far proved impossible to differentiate accurate from inaccurate pro- jections on the basis of their internal characteristics. For example, projections from detailed formal models of the energy system have often been as far wrong as the intuitive judgments of experts or forecasts from simpler models (e.g., Ascher, 1978; McNees, 1979; Zarnowitz, 1979). There are at least two inherent problems in trying to analyze energy demand for policy purposes. One problem is that major changes in the energy environment seem likely in the near future, changes similar in magnitude to those produced by the 1973 oil embargo and the 1979 revolution in Iran. It is unreasonable to expect energy analyses to predict such events because they stem from political and social forces outside the energy system:
2 they are, in a fundamental sense, not energy events. Because these events impinge on energy systems from out- side, even models based on accurate theory about the energy system will be wrong much of the time. The second inherent problem is that good energy models are few and far between. The dynamics of energy demand are far from well understood, even barring political cata- clysms. Economic theory describes some general depen- dencies of energy demand on income, output, and the costs of factors of production, but it leaves great latitude for the behavior of energy users, both individually and in the aggregate. When energy prices rise, for example, other factors of production (labor or capital, for example) tend to be substituted, but the degree of substitution and its path over time can only be determined empirically. This kind of empirical work is hard to do without a solid foundation of theory, method, and data. Direct observa- tion is difficult and yields only imprecise estimates because the behavior occurs in a complex and rapidly changing environment. The difficulties of modeling demand are general, with energy a representative and important case. Simple eco- nomic models, even those that track past experience well in the aggregate, do not completely describe the bases of consumers' responses. Such models treat consumers in general--and energy users in particular--as rational eco- nomic actors operating in an environment roughly describ- able as a market. But people often fail to take actions that would offer them clear economic benefits (e.g., Office of Technology Assessment, 1982; Solar Energy Research Institute, 1981; Stobaugh and Yergin, 1979). Furthermore, their investments in energy efficiency can be affected by a range of variables that are rarely esti- mated in energy demand models. Several of these variables are outlined below; they are discussed in more detail in the report of the Committee on Behavioral and Social Aspects of Energy Consumption and Production (Stern and Aronson, 1984). Preferences Consumer preferences cannot always be reduced to a common metric. People sometimes invest in energy efficiency to save money, sometimes to gain such intangi- ble consumer benefits as freedom from drafts in houses. Ways of incorporating such preferences in broad policy analyses have not yet been developed. Furthermore, claims of a rise of values of voluntary simplicity (Elgin, 1981) or spirituality (Yankelovich, 1982) suggest that past
3 relationships between increased income or gross national product (GNP) and energy demand may soon become weaker. Not enough is known to evaluate such hypotheses. Institutional Limits on Choice Energy demand is affected . by such institutional factors as regulatory constraints on energy prices; the split incentives that operate when the owner and the user of energy-using equipment and buildings are not the same (e.g., rental housing); and the power of manufacturers, standard-setting groups, lenders, and professional organizations to limit the range of investment alternatives available to energy users. Information is scanty, however, on how much influence each of these factors has and on the way each might respond to changes in policy or in the energy environment. As a result, policy analyses are often weak when they address institutional factors. Attitudes, Values, and Beliefs A number of survey and experimental studies show that various values, attitudes, and beliefs mediate the effects of structural, demo- graphic, and economic variables on energy-using behavior. Sometimes they appear to affect energy use more strongly than price does (e.g., Heberlein and Warriner, 1982). Stern and Oskamp (1984) have summarized this literature and offer a theoretical account that describes the joint effect on demand of economic, structural, demographic, and psychological variables. This framework may prove ana- lytically useful; it certainly underlines the extent of missing knowledge about energy demand. As yet, little is known about the conditions under which personal commit- ment, normative beliefs, or other attitudes or values make an important difference in energy demand. And even the rather detailed model of Stern and Oskamp does not address such potentially important variables as expectations of future prices and availability of fuels, beliefs about uncertainty of the same, willingness to seek and process information, and the existence of government or utility conservation programs. Communication There is considerable evidence that consumers' energy choices can be influenced in important ways by communication through informal social networks and by levels of trust in available information (for a review, see Stern and Aronson, 1984:Chapter 4). These social influences have many policy implications, but
4 little thought has been given to incorporating them in general analyses of energy demand. Convenience Economic theory recognizes the importance of the cost of gathering information and of uncertainty, and there is evidence that efforts to reduce such cost and uncertainty affect the choices of energy users. For example, some residential energy conservation programs seem to be successful partly because they simplify deci- sions and protect people from incompetent or unscrupulous contractors (Stern, Black, and Elworth, 1981). But the convenience factor has not been incorporated in energy models. Symbolic Meanings Energy users probably respond to government programs for conservation as a function of what conservation has come to mean to them in a public policy context (Stern and Aronson, 1984). Thus, President Carter's famous energy speech in his sweater symbolized an identity between conservation and being uncomfortably cold and reinforced this connection in people's minds. The effects of such symbolism, however real, are rarely considered explicitly in energy analyses. The Psychology of Shortage Some commentators say that public awareness about energy was changed in 1979 in a way that did not happen in 1973, even though in both periods there were rapid price increases for oil and the threat of serious shortage. Some hypothesize that it took a second "shock" to convince people that the energy environ- ment had changed. There is as yet no accepted way of evaluating the existence or quantifying the importance of this psychology of shortage. To the extent that factors such as those noted above determine behavior, the principles that govern the behavior of energy users do not lead to economically rational action as this concept is realized in existing models of energy demand. Thus, the most readily Whether all the relevant nonmonetary variables are measurable in principle and whether the behavior that flows from them can all be construed as rational under some noncircular definition of utility are two issues we do not address. It is clear, however, that some non- monetary variables not now addressed in economic models
s available analytical framework is almost certaintly incom- plete and may even be inaccurate or misleading. Formal models based on the assumption of rational choice may not capture the appropriate variables for analyzing energy users' responses to changing conditions and may, as a result, be quite wrong about the level of short-run response to changes in the energy environment and about the ultimate penetration of energy-efficient technology. Furthermore, an exclusive focus on the variables included in energy models implies neglect of other factors, some of which may be controllable by policy. A more complete framework for analyzing energy demand, suitable for general application and specified in quanti- tative form, is not now available, but the concepts and findings of behavioral research show promise of improving this situation. Behavioral research on energy users and their environment, including small-scale controlled ex- periments, can probably improve understanding of energy use in ways that will be useful for policy-related demand analyses, including formal energy modeling, and that may have policy applications. For example, laboratory experi- ments with different energy-efficiency labels for appli- ances can help determine the features of energy informa- tion that make the most difference in appliance purchase decisions. Such knowledge would be useful both for model- ing those decisions and for designing effective appliance labels. ABOUT THIS BOOK This book examines the state of energy demand analysis and develops recommendations for improving it. A recurring theme is the limitations of formal modeling--the dominant form of demand analysis. Despite the value of modeling, serious gaps in theory and empirical knowledge have led existing models to systematically overlook important fac- tors affecting energy demand, such as those noted above. We conclude that both energy demand analysis and the policy choices it supports can benefit from improved data do affect energy demand. It is also clear that the number of nonmonetary variables plausibly affecting demand readily outstrips the ability to measure them. AS a result, an analytical framework that can shorten the list of plausible influences has obvious value.
6 collection and from a more balanced reliance on a variety of analytic methods. These efforts will sometimes produce results that are easily incorporated into demand models by ~ - changing the parameters assigned to the variables in models, altering the functional forms of equations, or adding variables. Sometimes, however, new findings will not be so easily modeled, and more thorough changes in analytic tools will prove justified. This book distinguishes the formal modeling approach to energy demand analysis from what we call a problem- oriented approach. While the former employs mathematical techniques to build comprehensive descriptions of a com- plex energy system, the latter uses a wider range of research methods--including surveys, controlled experi- ments, exploratory analysis of existing data, and evalua- . , . ~ _ _ , ~ _ , talon research--to examine narrower analytic and policy questions. The rest of this chapter explores the strengths and limitations of each approach. Chapters 2 through 4 consider three issues the panel finds particularly important or interesting: the effects of prices on energy demand; the effects of financial incentives on investments in residential energy effici- ency; and the effects of information on energy demand. In these chapters, we pay particular attention to qualita- tive factors that have not been given careful considera- tion in formal models but that behavioral research has shown may have major effects on demand. For example, we explore the possibility that consumers respond to changes in price and not only to price levels; we consider the impact of qualitative differences among types of incen- tives, such as tax credits, rebates, and loan subsidies; and we examine the role of word-of-mouth communication in the adoption of energy-efficient technologies. We present the available evidence on these and other topics, evaluate the likely importance of some factors, and discuss methods that can be used to assess the roles such factors can play and to address their roles more fully in demand analyses. Chapter 5 is a case study of a particular type of behavior that affects energy demand and is in turn affected by prices, incentives, and information, as dis- cussed in Chapters 2 through 4. Taking appliance choice as an instance of such behavior, Chapter 5 identifies gaps in knowledge that impede understanding of such choice and suggests data collection efforts that could help close these gaps.
7 Chapter 6 presents our conclusions and recommendations about the appropriate roles of formal models, problem- oriented research, and data collection in energy demand analysis. It concludes by discussing ways different methods of analysis can complement each other. FORMAL MODELS OF ENERGY DEMAND Formal policy models are analytic tools that are built to provide quantitative projections of energy demand and to provide answers to questions about how demand might respond to alternative political and economic events or to policy choices. Formal models are often used to answer such questions as: "What will be the price of natural gas after full decontrol in the United States?n That will be the effect on the world price of oil of 4 percent annual GNP growth in the United States over the next two year so Much of energy policy analysis is now conducted via formal models. Two advantages of formal models over informal judgments for answering such questions are that they can ensure that all parts of the energy economy are included in the analysis and that their construction requires that the assumptions on which they are based are in quantitative form. Disagreements with a thorough and explicit analysis can be directed to particular elements of it rather than to some global expert judgment. Although models vary greatly in their levels of detail and in the amount of information they require to function, most energy models are computerized so that considerable detail can be included while it is still possible to respond quickly to forecasting or policy questions. Formal models are the favored analytical method among many policy makers and policy analysts for reasons only loosely related to their adequacy as tools for policy analysis or forecasting.2 Models are valued because 2 It is possible to distinguish the use of formal models for hypothesis-testing through analysis of relationships in existing data from their use for forecasting and policy analysis. Generally, forecasting and policy models tend to strive for completeness in the variables covered and to make projections into the future, while hypothesis- testing models tend to be less inclusive, more tightly derived from theory, and less concerned about projections.
8 they can respond to a range of policy questions and because their responses are quantitative. Furthermore, once created, models offer quick and relatively low-cost responses. These attributes are important to policy makers who must make many complex choices under severe time constraints. Models offer an added attraction for some policy makers: the mystique of computer analysis and economic theory gives an authoritative sense to the output from models that is quite unrelated to the adequacy of the models being used. For all these reasons, formal models are likely to remain the dominant tools for policy analysis in the area of energy demand. Types of Models Three frequently used approaches to building formal energy demand models are econometrics, system dynamics, and engineering process modeling. Econometric Models Econometric demand forecasting models typically include regression equations that estimate demand variables from other factors such as price, income, output, or noneconomic factors.3 When presented in "reduced form," each equation expresses a demand variable as a function of other variables. In principle, the family of regression equations is fitted to a data set based on observation of demand at various values of other variables included in the equations. Thus, the regression equations take into account the structure of correlations among the variables measured. In fact, however, judgment can be as important as measurement: the forms of the mathematical functions used to describe relationships among variables are generally selected by judgment or convention; the parameters in the equations are often adjusted judgmentally when the initial model results seem unreasonable; and even the data used to estimate regres- sion equations are often imputed, presumed, or guessed rather than measured directly. For example, the rate of capital investment in energy efficiency over time is often The discussion of models in this section primarily con- cerns models as used for projection and policy analysis. 3 The discussion here is focused primarily on models that are fitted to aggregate data.
9 estimated a priori, using an assumption that it responds to price changes with a time lag described by a particular mathematical function. For another example, data at the state level are sometimes constructed by allocating values measured at the national level according to some plausible formula. A detailed discussion of the variety of econo- metric energy demand models and the theoretical and empir- ical issues attending them can be found in Bohi (1981). Econometric models are limited in that they can only include variables for which data exist or for which data can reasonably be estimated; these may not be the most critical variables. The limitation is also an advantage, however, in that it imposes the discipline of tying models to data. Projections based on econometric models are problematic because they assume that the pattern of correlations in available data reflects structural relationships that will extend reliably into the future. There is little empiri- cal or theoretical justification for such an assumption in most cases, and the weaker the data base for the model, the weaker the assumption. Thus, projections from econo- metric models tend to enshrine as economic structure any systematic covariation, even if it was introduced by biased measurement or estimation; the greater the number of values and relationships that are estimated, presumed, or imputed, the more likely it is that the projections will contain systematic bias. This is a particularly important problem in energy forecasting because extrapo- lations ignore the possibility of major shocks to the system, such as those experienced in 1973 and 1979. System Dynamics Models System dynamics models consist of sets of simultaneous partial differential equations each of which, in effect, states an assumption about the response of one variable to the values of other variables in the immediately previous time period. The collection of equations defines relationships of energy demand to a set of determining factors and, with the addition of an initial set of values, can be used to generate values indefinitely into the future. In most cases, in which the causal relationships are not well understood, the modelers' judgments can determine the models' results. Compared with econometric models, in which causal rela- tionships can also be postulated by the modeler, system dynamics modelers are less likely to check their postu- lated functional forms and parameters against data. The reliance of econometricians on time-series data acts as a
10 rein on their judgment; this sort of safeguard is not an intrinsic feature of system dynamics models. Both econometric models and system dynamics models are typically checked by beginning with knowledge about some point in the past and demonstrating that by selecting the proper variables and coefficients, the models can generate values of energy demand close to those that have been observed from that point to the present. Once the vari- ables are chosen, coefficients for them are chosen to make the model fit the data. In selecting the variables, how- ever, econometric models are constrained to choose among those variables for which data exist; systems dynamics models can postulate new variables. Thus, system dynamics allows modelers great freedom to construct theory in the process of building the model. A system dynamics modeler can describe a variety of possible causal patterns and can generate richly textured accounts of the future. But because relationships can be postulated without observa- tion, system dynamics models can easily build more and more detailed pictures from less and less well-documented knowledge .4 What seems crucial for validating system dynamics models is empirical support for the causal relationships postulated in the modeling process. When this cannot be provided by analysis of existing data sets, it could be provided by experimental research to test the propositions embodied in the models' equations. This sort of approach, however, has not been a part of the research program of system dynamics modeling. Engineering Process Models Engineering process models rely on detailed data on present energy demand and focus on the determination of demand by technological choice. They account for present energy use by building up from energy-using technologies, aggregating according to cate- 4 A shortage of empirical support has been a problem with the system dynamics models presently available for use in the Department of Energy's Office of Policy, Planning, and Analysis (PPA). These include FOSSIL2 (Energy and Environmental Analysis, 1980), DEMAND '81 (Backus, 1981), OECD1 (Salama, Greene, and Krantzman, 1980), and HID (Marshall, 1983). The PPA recognizes the inadequacy of these models as a basis for demand analysis and has begun an effort--including the present work--to improve the situation.
11 gories such as appliance type, building type, end use (e.g., space heating, materials processing), and fuel type. To make projections, engineering models rely on assumptions from economic and demographic projections about such factors as household formation, building stock replacement, and the rate of improvement in the energy efficiency Of available technologies. The chief strength of engineering process models in comparison with other types of models is that they can describe the technological trade-offs in more detail. In practice, however, there are difficulties with such models. For example, models built from laboratory tests of the performance of technology are often in error as predictors of actual performance; predictions about the effects of building insulation may be the most extreme example. This problem can be addressed by conducting field studies of the technologies and validating the models on the resulting data. engineering calculations of the cost of more energy- efficient technologies may bear little resemblance to the price of those technologies in the market. Good data on manufacturers' decisions about pricing simply do not exist. Another problem is that Engineering models, like other types of models, also rely greatly on judgment. For example, they generally lack a theoretical framework for modeling choices among technologies. AS a result, their estimates of the rate of penetration of new technology are ultimately judgmen- tal. The obvious solution to this problem involves con- ducting separate studies of purchases to provide estimates of penetration. The above descriptions of three model types are ideal- ized and incomplete. Other classes of models exist (see, e.g., Greenberger, Crenson, and Crissey, 1976, for a dis- cussion), and hybrid models have recently been developed that combine the features of different types of models in _ an attempt to gain some of the advantages of each. For example, econometric models of appliance choice are now incorporating engineering data (the residential end-use energy planning system model, Goett and McFadden, 1982). The following comments generally apply, however, to for- mal models as an analytical approach. Problems with Formal Demand Models Serious criticisms have been increasingly raised, both by modelers and others, about formal demand models and their
12 use in policy analysis. We can only give a brief account of the criticisms here; for more detailed analyses, see, for example, Ascher (1978); Brewer (1983); Freedman (1981); Freedman, Rothenberg, and Butch (1983a); Greenberger, Crenson, and Crissey (1976). Some technical criticisms are serious: models have been found to be poorly documented; "overfitted, n meaning that they achieve close correspondence to past experience by making large numbers of unsupported and sometimes unreasonable assumptions; insufficiently explicit in their treatment of uncertainty; and inadequately tested and reviewed (see, e.g., Brewer, 1983; Freedman, Rothenberg, and Butch, 1983a). Some of the criticisms point to the lack of empirical basis for models' estimates of the ultimate magnitude and the rate of change of energy use in response to signifi- cant stimuli. Key parameter values are often postulated, and relationships are assumed to follow particular mathe- matical forms with little or no empirical support for the choices. Such procedures inspire little confidence. Unfortunately, little knowledge exists for supplying the coefficients needed to estimate the ultimate magnitude of change (steady-state response) or the rate of change (dynamic response) in the terms of existing models. Models have also been criticized for a poor track record as predictors. This criticism has been made of large-scale economic models generally (e.g., Ascher, 1978; McNees, 1979; Zarnowitz, 1979) and of energy models as special cases. The recent evidence suggests the predic- tive success varies widely from one energy model to another. Some of this evidence suggests that some models have failed in predicting energy prices and not in des- cribing the relationship of price to demand: if actual (realized) energy prices are substituted for the prices assumed in those models, they more correcty forecast energy demand up to the present (McFadden, 1983). Some of the criticisms point to the political implica- tions of large-scale energy modeling. Brewer (1983) argues that models tend to have a conservative bias because they extrapolate from past experience. He also argues that as policy analysis comes to rely on more and more technically complex models, the public tends to be closed out of policy debates. Still other criticisms emphasize the consequences of taking energy models seriously despite their flaws. Because of insufficient knowledge, virtually every model contains some judgmental elements that were added to make
13 the model conform better to the prevailing qualitative understanding of most energy modelers. These judgments are buried in models that appear superficially to be well grounded in science, and they interact with other elements of the models in unknown ways. When the judgmental ele- ments are not obvious, the users of models--especially complex models--can easily confuse the model and reality. They come to believe they have knowledge about the energy system when they have knowledge only of the model. And in the saying quoted by Freedman, Rothenberg, and Butch (1983b): "It ain't what you don't know that gets you into trouble, it's what you think you know that ain't so. n When policy analysts equate a model with reality, they begin to define issues in the terms most central to the model and to ignore the variables the model ignores. Among these ignored variables are many important behav- ioral factors for which quantitative data do not exist and that are not prominent in economic theory, for example: incompleteness of information, mistrust, the symbolic meanings of action, personal attitudes, social values, political conflict, and organizational routine. The phenomena that result from such influences are, of course, encompassed by formal demand models, but they appear under other labels and so may be misconceived in important ways. The Place of Behavioral Variables in Formal Demand Models In energy demand models, behavioral variables are usually subsumed under other broad concepts that are vague with respect to their behavioral basis. The concept of own- price elasticity, for example, describes the fact that, other things being equal, the quantity demanded of some- thing is inversely related to its price. But the concept says nothing about how the information embodied in price enters a consumer's awareness or about how awareness of price affects action. Thus, the concept of elasticity bypasses the behavioral phenomena that underlie the response to price; the behavioral questions are begged. Because the concept of elasticity is behaviorally atheo- retical, elasticity estimates cannot predict whether unanticipated conditions, such as news of an impending major oil shortage, will change the ways consumers respond to price signals. A focus on price elasticity also ignores factors that can mediate the effects of price,
14 such as the quality or trustworthiness mation or the ways that information is Because behavioral research focuses on of available infor- communicated. the underlying make forecasts processes, it may help demand analysts about new conditions. In this section, we discuss two concepts central to many formal demand models--the dis- count rate and the dynamics of response--to illustrate the ways behavioral concepts relate to the concepts used in formal models. The discount rate is a key concept in many demand models. Formal models estimate the ultimate penetration of energy-efficient technologies by making the assumption that energy users will, in the long run, make the economi- cally rational choice of the alternative with the greatest net present value in terms of investment cost plus subse- quent operating cost. What is economically rational for a consumer is a function of that consumer's discount rate, that is, the magnitude of the consumer's preference for present over future value. Thus, the concept of discount rate implies using a particular mathematical formula to account for the fact that the real dollar value of invest- ment in energy efficiency is generally less than the real dollar value Undiscounted) of the expected return. The magnitude of the discount rate is often simply postulated in formal models, but some empirical methods have been suggested for estimating discount rates. The methods usually used involve analysis of data on consumer choices among alternatives that vary in energy efficiency and investment cost (e.g., for appliance purchases, Hausman, 1979). From those choices, an implicit discount rate is calculated: it is defined as the discount rate that would make actual purchase behavior economically rational in terms of net (i.e., discounted) present value. The use of discount rates in models presents a major practical problem in that data are often inadequate for confidently estimating implicit discount rates. Sometimes markets do not offer a wide choice of energy efficiencies or do not present a trade-off between energy efficiency and initial cost, and so it is difficult, even in prin- ciple, to collect the relevant data. In other cases, as with some household appliances and home insulation, the needed information could be obtained through additional efforts at data collection and analysis; however, great effort might be required because implicit discount rates probably vary with characteristics of the investment and the consumer (e.g., for different income levels, Hausman, 1979; McFadden and Dubin, 1982) in ways that cannot
15 readily be predicted from the concept of time discounting alone. But the analyses are necessary if formal modeling is to be based on reliable estimates of discount rates. The use of discount rates can also breed a broader conceptual problem. The concept of time-discounting assumes that consumers make decisions based on the value they place upon time and that the relevant decision rule can be represented by a number that is constant for par- ticular types of consumers, or at least for individual consumers making particular types of purchases. Because time-discounting is implicitly a psychological process of the consumer, it is easy to interpret empirically calcu- lated implicit discount rates as representing an attribute of energy users.s But when implicit discount rates are calculated from data on purchases, the resulting number reflects a collection of variables--not only the degree of preference for present value--that may affect the level of investment. The calculated discount rate is affected, for example, by the extent to which information is imperfect, mistrusted, or ignored; by energy users' persistence in old habits; by the fact that consumers with limited capital do not always purchase what they would if they had more capital; and by various other factors that might change the rate of adoption of energy-efficient technology. To subsume all these influences under a single index and to call it the discount rate may be to misconceive the phenomena. More important for forecasting and policy analysis, such theoretical shorthand may lead analysts to think of some features of energy users' behavior as stable preference (part of the discount rate) when they may in fact be changed by economic or institutional forces or by 50f course, the argument can be made that consumers act as if they were making decisions about the value of time without going through any such psychological process. In this view, the discount rate is simply a mathematical shorthand for projecting the difference between the real dollar value of investment and the expected future return. This interpretation eliminates any psychological rationale for the functional form of net-present-value equations. It also lacks any theoretical structure that might offer insight into the variables that determine discount rates or that might suggest policies for changing the rate of acceptance of new technology.
16 policy. When important variables that influence behavior are treated as constants in analysis, erroneous forecasts may result. Analysis in terms of discount rates may also lead analysts to overlook potentially effective policies because the effects would be on something that has been implicitly defined as a stable attribute of energy consumers. Another central concept in many demand models is that of response dynamics. Investments in energy efficiency in response to new conditions occur slowly over time. Formal models commonly estimate the rate of investment a priori, using an assumption that the rate follows price changes or other stimuli with a time lag described by a particular mathematical function. A typical example is the original Oak Ridge National Laboratory model of resi- dential energy use (Hirst and Carney, 1978), which used a simple algorithm to estimate the extent and pace at which manufacturers and households improve the energy efficiency of equipment and structures in response to changes in fuel price. Neither theory nor data were available to validate the assumptions embodied in the algorithm, and little work has been done since then to correct the deficiency. There is some evidence, however, that the most typical algorithms do not give the best estimate of the dynamics of response to changing energy prices (Hill, 1983). Thus, formal models lack an empirical basis for estimating the dynamics of price response. Another aspect of the problem is the way time-lag specifications are used in econometric models. Lag coefficients quantify a phenomenon without identifying any of its multiple causes. 6 The slow response to new conditions is sometimes explained in terms of budget con- straints or the costs of replacing capital stock before the end of its useful life, but there are many other pos- sible explanations. To cite two examples, full informa- 6Modelers sometimes estimate dynamics from assumptions about decision processes rather than from assumptions of correspondence with a mathematical equation. Since this procedure is based on implicit or explicit theory, it may have testable implications about the causes of dynamics. But because the data needed to support the assumptions do not exist, postulating decision processes is not much of an improvement over postulating the functional form of a descriptive equation
17 tion spreads slowly, and new information sources take time to become credible. No doubt, many other behavioral and institutional factors also act to retard theiadoption of energy-efficient technology. AS with the discount rate, it would be a mistake to interpret a lag coefficient in terms of only one of the factors that affect it. Thus, for forecasting or answering analytical and policy ques- tions, a model of energy demand needs more than a lag coefficient that accurately describes past events. The coefficient should be based on an understanding of the ways various environmental factors, including behavioral ones, affect rates of response. 7 Formal demand models tend to address behavior through concepts that cover the outcomes of behavioral phenomena but offer little or no insight into the determinants of the phenomena. This approach is useful for forecasting only as long as the relationships among the important behavioral variables are relatively stable--but there is little reason for assuming such stability given the unprecedented changes that have been occurring in the energy environment. For policy analysis, the approach has the limitation that it directs attention away from some behavioral variables that may offer useful levers for policy. There are two possible ways to resolve these problems, and they are not mutually exclusive. One way is to direct resources to gathering the data needed to test key assump- tions and provide empirical grounding for the parameters used in policy models. Many data are needed; we discuss their specifics in the following chapters and offer recom- mendations in Chapter 6. The other way is to direct resources to other approaches to policy~analysis that focus on understanding the behavioral phenomena of inter- est in policy contexts. This problem-oriented way of doing energy demand analysis is discussed in the next section. 7 Models relying on lag coefficients can be effective in making projections under some conditions. But without understanding of the factors responsible for the dynamics of response, it is impossible to know what changes in the environment would invalidate the projections or what policy alternatives might alter the dynamics from past patterns.
18 PROBLEMORIENTED ANALYSES OF ENERGY POLICY ISSUES In contrast to formal policy models, problem-oriented analyses are by definition unsystematic, being addressed to specific questions rather than to the accurate descrip- tion of an entire social or economic system. An issue or policy option is raised, and information is gathered and analyzed by the best available methods to clarify the issue or the policy implications. This section describes the major research methods used; we discuss their uses in more detail in later chapters. Methods of Problem-Oriented Analysis Surveys National general-purpose surveys of energy users can col- lect data simultaneously on a range of variables relevant to energy use. Such surveys can allow for multivariate analysis of the relationships among economic, demographic, climatic, engineering, and social-psychological variables as they affect energy demand. They can include variables not presently addressed in models and can estimate their importance. Such surveys would use a nationally represen- tative sample of energy users. If repeated, surveys can gather time-series data essential for empirically estimat- ing formal demand models, particularly econometric models.8 National surveys are not, however, useful for all types of policy analysis. General surveys have little value, for example, for analyzing policies or programs that have not yet been tried or that few people have experienced, such as new appliance labels, utility load management programs, or comprehensive home retrofit pro- grams, because few people in a general population sample will have participated in such programs. This problem can be addressed by conducting more specialized surveys on samples drawn from program participants and comparison groups. 8Researchers distinguish among cross-sectional survey designs, which collect data only once; longitudinal designs, which collect data on the same variables over time; and panel designs, which collect data on the same variables from the same set of respondents over time.
19 Specialized surveys--surveys aimed at particular issues or types of energy users--can help answer important ana- lytic questions that cannot readily be addressed by models or in general national surveys. Thus, a national survey focused on solar energy can intensively probe for the factors determining intentions to invest in residential solar energy technologies (Farhar-Pilgrim and Unseld, 1982), or a utility company can survey its customers to assess the reasons for their response or nonresponse to a conservation program (see Berry, 1981). Specialized sur- veys are valuable because they can look closely at vari- ables, including many of those not included explicitly in formal policy models, that may affect important consumer actions. For example, a survey in Michigan estimated the magnitude of consumers' misunderstandings of household energy use (Kempton, Harris, Keith, and Weihl, 1982). But surveys suffer from some generic limitations. Surveys may have poorly worded questions, may induce respondents to give socially acceptable rather than accu- rate responses, or may fail to predict behavior because the respondents themselves cannot predict what they will do. Unreliability increases when surveys are used to assess responses to hypothetical situations, such as an experimental electricity rate structure; people cannot reliably predict their behavior in situations they have not experienced. A further problem with using survey results to project energy demand is that even when self- reports are accurate, inferences from them may not be. A major reason for this lack of correspondence is that energy-saving actions, such as investment in home insula- tion, do not always save the amount of energy that would be expected from technical estimates. This "failure" may be because energy use data are unreliable or because tech- nical estimates make unrealistic assumptions about how well the improvements are installed; it is also possible that people use some of the energy saved by technical improvements to increase their comfort (Hirst and Goeltz, 1983; Hirst, Hu, Taylor, Thayer, and Groeneman, 1983; Hirst, White, and Goeltz, 1983a). Thus, the best way to gauge the likely response to a new information or incen- tive program is to rely on actual observation. Specialized surveys often have problems attributable to their low budgets. Their sample sizes are small, the research staffs are often inexperienced in the finer points of survey design, and their findings are often poorly documented. There has frequently been a wide gap between intent and execution in specialized energy surveys.
20 Analysis of Existing Data The Energy Information Administration has several data sets that have only been partly analyzed, and utility companies have some of the best existing data on residential and commercial energy use. Utility data are particularly useful for analyzing responses to price because in some utility service areas, rapid price changes have recently occurred for one fuel supplied but not another, and within the regions, electricity prices have changed rapidly for some utilities but not for others.9 Analyses of past experience may fail for lack of crit- ical information, or they may be misleading for making projections because of differences between past and future situations. There have been problems getting access to existing data at the individual level because of concern about privacy. And there are also obviously limitations on the kinds and quality of the data available. Better data exist for analyzing energy use in the residential sector than in the commercial or industrial sectors; aggregate data are generally more available than disag- gregated data; energy-use data are better than data on equipment stocks; and data on attitudinal factors is par- ticularly adequate. In addition, some data, such as on demographic variables and local weather conditions, are often not available in a merged form with disaggregate data on energy use. Natural Experiments Rapid changes in the prices of fuels over the past decade have made it possible to study the price effects empiri- cally, and the fact that price changes were not uniform across fuels or localities provides useful comparison con- ditions. The recent sudden decrease in the inflation rate constitutes a natural experiment on the determinants of consumers' price expectations. Such natural experiments 9Existing data can be analyzed in various ways. One of these involves using the same econometric techniques that are sometimes used for building detailed formal models. Thus, small-scale, problem-oriented modeling can be part of an analytic alternative to system-wide formal demand modeling.
21 could provide much of the data needed for demand analysis with relatively little additional effort, if the data are collected regularly. Methods for evaluating the results of such natural experiments have been developed over the past decades (e.g., Campbell and Stanley, 1966; Cook and Campbell, 1979), but have rarely been used to learn from natural experiments on energy demand. Of course, a natu- ral experiment is riskier to interpret than a controlled experiment because it is harder to identify the causes of the phenomena observed. Controlled Experiments The experimental approach has been generally neglected in energy policy analysis. The best-known exception was the time-of-use pricing experiments conducted during the 1970s, some of which involved random assignment of house- holds to experimental electricity rates. Experimentation was the method of choice in those studies because there was no empirical basis for estimating the effect of prices based on time of use and because the experimental rates were so far from most energy users' past experience that self-reported intentions could not be relied upon. The same rationale suggests that experiments could provide the most valid answers to questions about the design of energy conservation programs, particularly for assessing the effects of interventions that are nonfinancial n charac- ter and for which, as a result, existing models are par- ticularly inadequate. Even laboratory experiments are sometimes appropriate for policy analysis. AS part of energy information efforts, it is necessary to choose what information to provide, how to design the layout of appliance labels, and so forth. For many of these choices, it would be useful to experiment with alternatives in a laboratory setting to see which alternatives are eye-catching, understand- able, meaningful, and considered useful by people like those for whom the information is intended. The greatest advantage of controlled experiments over other research techniques, of course, is that they can control for large numbers of extraneous variables that may covary over time with the variables under study and make the interpretation of nonexperimental data difficult. For this reason, experimentation is sometimes useful even when models, analyses of existing data, or survey research can produce empirically meaningful results. ThuS, small-scale
22 field experiments on the effects of rebates as an incen- tive for energy conservation (Geller, Winett, and Everett, 1982) can provide a check on the elasticity estimates derived from analysis of time-series data. Like other research methods, experiments have their problems as a policy tool. Some researchers, unfamiliar with practical policy concerns, have experimented with unrealistic treatments and consequently produced imprac- tical recommendations (see Stern and Oskamp, 1984). Experimental studies often meet practical opposition from program managers who are eager to get on with their pro- grams and who feel they know enough to act without await- ing the results of formal research. Experiments also face political opposition as unethical or unnecessary: if the policy is a good one, it should be made available to all, not just a small experimental group (for a discussion of such issues, see Mosteller and Mosteller, 1979). More- over, if experimental subjects believe an experiment to be temporary rather than a permanent change in policy, their belief may affect their behavior. Evaluation Research Past and present energy programs are a great untapped source of information about energy demand. Information programs run or mandated by government--such as Project Conserve, the Residential Conservation Service, the Energy Extension Service, and the energy-efficiency labeling of cars and appliances--all constitute experiments with information, but they are almost always uncontrolled, and few have been adequately evaluated. Incentive programs-- such as the federal and state tax credits for conservation and solar energy--have also been inadequately studied. And thousands of local energy programs, public and private (see Center for Renewable Resources, 1980) have been started, but rarely with an evaluation component (Stern and Aronson, 1984:Chapter 7). Evaluations of such pro- grams can provide knowledge that is unlikely to result from other research methods. In particular, evaluation studies can uncover variables in the implementation of conservation programs that would not be anticipated by before-the-fact analyses. Evaluation research can use any of the methodologies outlined above. The firmest conclusions, of course, can be drawn if energy programs are treated as experiments from the start. A variety of quasi-experimental research
23 designs that retain many of the advantages of experimen- tation can also be used. The call for experimental con- trol does not, however, imply that energy programs should be rigidly specified in advance for scientific purposes; that approach would limit the programs' abilities to adapt to their environments and could threaten their success. Rather, we are emphasizing the advantages of experimenta- tion that accrue from procedures of careful measurement and of random assignment to treatment conditions, which control for a variety of extraneous variables. Evaluation research usually has serious inference prob- lems. Because the call for evaluation usually comes after a program is in place, experimental design is impossible, and researchers sometimes attribute an effect to a program that may have been due to self-selection of program par- ticipants. Furthermore, when a study begins after a program, some of the essential data may not be available, and preexisting conditions cannot be reliably assessed. The Strengths and Weaknesses of Problem-Oriented Research Individually, the above research methods are limited in their usefulness for analyzing energy demand; together, they constitute a valuable set of tools. They provide policy analysts with several ways to answer particular policy questions. For example, to assess response to a tax credit for investments in energy efficiency, several approaches could be combined: surveys of willingness to invest given the proposed tax credit; an evaluation study of the effects of a similar incentive that has been tried in another state or country; small-scale field experiments in which the tax credit or a similar incentive was actu- ally offered and actual investment by participants was assessed. Analysis based on several methods is less prone to errors that arise from reliance on a single method. Problem-oriented studies can also be used to address issues that arise in constructing policy models. When data are not available for estimating discount rates, surveys to assess consumer choices among hypothetical technologies may, despite the limitations of such surveys, be the best available source of estimates. To estimate price elasticity over a range of prices that has never been experienced (e.g., premium prices for peak-period electricity), small-scale experiments are advisable and have sometimes been conducted.
24 Problem-oriented research is also useful for answering behavioral questions that are not explicitly addressed in models. If a policy maker needs to know how to design an energy-efficiency label for an appliance so that consumers understand it, or how much difference a well-designed label can make in consumer response, or whether consumers are likely to believe the information the government requires in a label, formal models are not equipped to offer answers. But laboratory experiments can help answer the first question, a field experiment can help answer the second, and surveys and field experiments can help answer the third. Compared with modeling, problem-oriented research offers certain advantages. It can often address questions more directly than models can, and it can investigate a number of questions that cannot readily be represented in models or for which representation is possible but no empirical data exist. Before modelers respond to a new policy question with an empirically unsubstantiated ad hoc adjustment to an existing model, a problem-oriented study can provide the data needed to address the question empirically. In addition, some methods of problem- oriented research, particularly experimental ones, provide more convincing evidence than modeling data can offer. Experiments, and even well-conceived quasi-experiments, are relatively free of the problems of spurious correla- tion that haunt econometric models and of the unsubstan- tiated causal assumptions that leave system dynamics models open to serious question. The problem-oriented approach has conspicuous weak- nesses. It is not by itself well suited to forecasting energy demand. It can easily overlook the possibility of complex or counterintuitive interactions of a proposed policy with other events in the economy or energy system. Because of its responsiveness to changing policy concerns, the approach may not benefit from systematic accumulation of knowledge. And if problem-oriented studies proceed from a lack of basic understanding of the behavioral variables policy is intended to influence or of the prac- ticalities of policy implementation, they will tend to frame questions in unproductive ways and yield impractical advice. In addition, each research method used for problem- oriented analysis has its own particular weaknesses, as noted above. An extensive literature on social science methodology details the strengths and weaknesses of the various research designs that can be used for analyses of
25 questions relating to energy demand (the classic brief statement is by Campbell and Stanley, 1966; see also Cook and Campbell, 1979). As with formal policy models, many of the weaknesses of individual methods of problem-oriented research are surmountable. Replication and criticism are two obvious methods of quality control. In addition, each research method provides a way to cross-validate results obtained by other methods. Since findings are best established if they are proved robust to choices of methodology, problem- oriented energy research may make its strongest contribu- tion as a method of quality control and validation. Although problem-oriented research cannot offer the broad view that energy models attempt to provide, it can be an invaluable part of the effort to understand the U.S. energy system. It offers ways to validate assumptions, to estimate the parameters of models, to see if variables that have not yet been considered may be important, and to explore in detail the behavioral phenomena that under- lie such broad concepts as price elasticity, time lags, and implicit discount rate. Problem-oriented research on energy demand continues to occur in response to the needs of policy makers and the interests of researchers. But the insights from this research have not been systemati- cally incorporated in formal analyses of the energy sys- tem, and the informational needs of formal modelers have not yet served as a significant impetus for problem- oriented research. CONCLUSIONS There are many critical gaps in knowledge about energy demand. Given the present state of knowledge, it is clear that most analyses are based on at least some erroneous assumptions and ignore at least some important variables. Formal demand models, in particular, are unreliable guides to policy analysis. Too often, there is no evidence to confirm or revise the assumptions in such models, to decide if the important variables have been correctly identified, to support parameter estimates, and to evalu- ate the importance of variables omitted from the models. Not only is-convincing evidence lacking to justify the assumptions, but data essential for obtaining the evidence are also lacking. This state of knowledge is an important reason that demand models are the subject of so much debate.
26 Problem-oriented studies have great potential value for filling gaps in demand analysis, but--even within their limits as analytic tools--they have been insufficiently used. Problem-oriented research can produce significant new knowledge, but we believe this knowledge will not always be readily translatable into the language of existing models. Rather, we expect that improved knowl- edge will bring about changes in existing models and that such changes will lead to improved analysis of energy demand. This report is a beginning effort to show how different analytic methods can be used in a complementary fashion to improve understanding of energy demand.
