UN Population Division, Department of Economic and Social Affairs,
with support from the UN Population Fund (UNFPA)
***************************************************************** This document is being made available by the Population Information Network (POPIN) Gopher of the United Nations Population Division, Department for Economic and Social Information and Policy Analysis, in collaboration with the European Association for Population Studies and the IUSSP. For further information please contact Professor G.C. Blangiardo, Local Organizer, EAPS Conference, Milan, University of Milan, Istituto di Statistica, V. Visconti di Modrone 21, Milan, Italy. ******************************************************************* EUROPEAN POPULATION CONFERENCE CONGRES EUROPEEN DE DEMOGRAPHE Milano, 4-8 settembre 1995 Plenary VI ®God has chosen to give the easy problems to the physicists¯ or why demographers need theory* by Guillaume Wunsch ®We ought not think of the social world as a system of phenomena in which we can expect to find a strong underlying order¯ 1. The need for theory The two quotations which open this paper, respectively drawn from C. A. Lave and J. G. March (1975) (two social scientists) and D. Little (1993) (a philosopher), reflect the bewildering nature of social phenomena and the difficulties demographers are confronted with, when they want to make sense of the phenomena they study. To be sure, demographers have developed sophisticated methods of analysis and benefited from the availability of large and diverse data sets pertaining to their field, but in how many instances can we say with confidence that we have explained and understood the demographic processes which are studied, that we are able to make reliable forecasts, and that we can recommend efficient policy measures? For example, suppose we want to determine if a woman's number of children influences her labor participation, or more generally find an answer to a ®Why?¯, ®What-is?¯, or ®What-if?¯ question concerning the occurrence of a particular event or relationship. For example, why is male mortality higher than female mortality in Europe? What is the impact of family allowances on French fertility? What would be the impact on internal migration if one decreased the price of public transport in Italy? Explanation is the search for suitable answers (research hypotheses) to these questions; plausible answers then have to be tested in order to be either confirmed or falsified by facts. To come back to our original problem, other determinants of labor participation might be the woman's age and other household income . This leads firstly to the following simple causal structure: Household income Number of children Labour participation Age of female This causal structure is then tested using, say, a logit regression model applied to survey data, and the logit coefficients are computed. It is seen that the presence of young children does have a significant impact on female's labor participation, controlling for household income and age of female. Have we however truly explained the relationship between the presence of children and labor participation, controlling for possible confounding factors? For example, instead of the simple causal system tested above, the relationship might be in the form of a causal chain such as Age Children Income Labor participation as one's number of children might be a determinant of other household income through e.g. family allowances. In this case, income should not be controlled as it is an intervening variable between presence of children and labor participation. Similarly, presence of children and income should not be controlled if one wishes to evaluate the impact of woman's age. More complex causal structures, taking account of other confounding variables (such as social category and education) and translated by a system of simultaneous equations, can also be proposed and the subsequent results might be quite different from those obtained in the former cases. This rather trivial example shows that explanation in demography does not only entail pointing out the variables or possible risk factors associated with the effect one is studying, through the use of a more or less sophisticated statistical technique. Epidemiological studies have discovered for example some 280 or more variables associated with coronary heart disease (CHD), including snoring, grey hair, wifely love, premature baldness, etc. (W. E. Stehbens, 1993). All these factors have been shown to be associated with CHD, but as one knows good correlations are not necessarily good explanations! Further data possibly derived from other types of studies (non-epidemiological studies, in this case) should confirm the statistical evidence, and especially a plausible explanation linking causes and effects should be provided, so that the phenomenon becomes intelligible or comprehensible. Statistical techniques are moreover not always the best tools for certain purposes, such as determining the reasons underlying a specific behavior; qualitative analyses could be preferred in those cases. A good explanation therefore requires uncovering the generative process, the causal history or mechanism, or finding the ®reasons¯ of the causes as R. Franck (1994a) would say, such that we can understand why a cause produces an effect. In the social sciences, this causal process is usually based on the actions of individual agents, taking account of the impact of groups, social structures, and institutions, and on the relationships between man and his context or environment. This is the ®causal mechanism¯ aspect of explanation. Other views consider that explanation should also lead to making a phenomenon expectable, to unifying different phenomena, or to resolve curiosity and reduce contingency taking into account the context in which the ®Why? or What-is/if? question¯ is asked (see D. M. Hausman, 1993a). I believe that, at the minimum, explanation should state the ®reasons¯ of the causes, as pointed out above; from this viewpoint, the other characteristics are important but not necessary. Finally, the causal system must be related to its environment or context. May we for example assume the closure of the causal system, such that no outside (implicit) variable has an impact on more than one explicit variable in the system? Furthermore, does a cause(s) and effect(s) approach necessarily lead to a true understanding of the relationships involved, such that our results form the basis of possible predictions and policy actions? These are some of the topics which will be examined in the present paper, taking population as an example. Finally, I have intentionally included many references to the non-demographic literature especially, to which the interested reader is referred for further information. 2. Theories, laws, and explanation Demographers do not usually proceed by asking for example what are all the causes, both proximate, intermediate, and ultimate, of mortality in contemporary Europe. This is the view of modern science, however. In the past, scientists have on the contrary searched (unsuccessfully) for a global interpretation of society in its total manifoldness, hoping to develop a synthesis of all social sciences … la Talcott Parsons. Max Weber, among others, has been a strong opponent of this aim. Indeed, a ®total¯ explanation of mortality is practically impossible, to say the least. Causes are so numerous and so interrelated that we would end up with a fuzzy explanation, a huge system of variables, and an untestable theory. It has indeed been conjectured (J. F. Traub and H. Wozniakowski, 1994) that there exist unanswerable scientific questions. Such would be the case of many problems (especially those which are discrete) with a large number of variables, as well as those concerning the behavior of chaotic systems (systems extremely sensitive to initial conditions). From this perspective, it would probably be unwise to develop very large models in a causal or a systems approach, as their solution would possibly be intractable due to their computational complexity. In most cases therefore, we are interested in knowing the impact of one (or a small number of) possible cause(s) on an effect (or response). If we want to see for example if there are differences in mortality according to income, we must first delimit the period of time and the area we are interested in, then define what we mean by mortality and by income, find suitable indicators of these two concepts, obtain relevant data and measure the variables, and check if the relation observed is statistically significant. Theories employed in making observations, i.e. the so-called measurement theories, should be distinguished from explanatory theories dealing with the possible causal mechanism producing these observations. It is the latter type of theories which is the subject of this paper. We must also control possible confounding variables, i.e. those which on the one hand have an impact on mortality and on the other hand are associated with income. If confounding variables are not controlled, the relationship between income and mortality cannot be adequately determined. Suppose there is a variable, such as education, which is associated with income and has an influence on mortality. Controlling for education could therefore significantly change the relationship between income and mortality. Furthermore, why is income associated with education? If one time-orders the variables, education could be a cause of income, higher educated people earning more money. Another reason would be that, on the contrary, income is a cause of education, as higher income enables one to take better advantage of the educational system. Or education and income could be the effects of a common cause, such as the social class of one's parents. All these possible reasons explain the association between education and income, and all are a priori equally valid interpretations of this relationship. In the first case, the causal model is: education income mortality One could also have the following model, where there is both a direct and an indirect path leading from education to mortality: income education mortality The second case is represented by the following causal chain: income education mortality Here too one can postulate a ®direct¯ effect of income on mortality (through possibly a better use of the health care system), and an indirect effect through education, leading to the causal graph: education income mortality Finally, in the third case, the causal structure would be: income social mortality class education As the relationship between cause and effect usually depends upon their role in the wider causal structure, it is necessary to specify the interrelations between possible causes and confounding variables, that is to develop a causal theory linking the variables together while specifying their order, causes preceding effects in time. This theory, among other things, helps us to determine which variables have to be controlled and those which should not be controlled. The theory is developed according to the prior knowledge, experience, and imagination of the scientist, taking account of the state of the art at that time. In this respect as in many others, scientific theories are hardly different from everyday inference or judgement under uncertainty, i.e. the psychology of causation or how people make judgements of probable cause; see e.g. H. J. Einhorn and R. M. Hogarth (1986). Theories can therefore change over time and space, and differ from one scientist to the other; as always, the proof of the pudding is in the eating, i.e. are the theories confirmed by the data on hand or not? In particular, a population theory  should not only take into account the state of knowledge of demographers at the present time, but it should also consider the state of the other sciences dealing with individuals and groups. For example, a theory of demographic behavior cannot ignore all the work done on human behavior in cognitive psychology, in decision theory, in artificial intelligence, and in the philosophy of mind. Demographic propositions on behavior should therefore not contradict findings in other fields; if they do, there should be very good reasons. To give another example, a population theory developed at the micro (individual) level should not neglect the impact of institutions and groups, at the macro level, on the actions of the human agent; demography must therefore integrate the results of sociological research or, at the very least, cannot ignore them. In this sense, demographic explanations should necessarily be multidisciplinary, but the same could be said for the other social sciences too, such as e.g. economics. Going back to our example, to complicate things further, the impact of income on mortality is not necessarily the same at all levels of education. In this case, we would also have to take into account the interaction between these two causes, interaction being the joint impact of the two causes (income and education) in addition to their individual action on the response variable (mortality). As we see, even with a very small number of variables, the possible causal structures are numerous. Theory enables the scientist to select a small sub-set of possible structures or ®possible worlds¯ (usually only one) from the many which can be proposed with the variables on hand. For an interesting discussion concerning the selection of one or a small handful of theories among the many possible alternatives, see C. Glymour et al. (1987). The model thus chosen is then submitted to empirical testing, in order to be either confirmed or falsified . If the plausible explanation is strong and if there are no good alternatives, it will be kept as the best theory under uncertainty till a better explanation is proposed. Finally, we must say something about the rest of the world, i.e. the environment of our explanatory system. Can we assume that there are no other implicit or hidden variables which could act as confounders (®strong¯ closure of the explanatory system), or can we consider that the latter affect at the most one explicit variable and are not interrelated (®weak¯ closure of the system)? If not, are we prepared to accept the possibility of hidden confounders which could play havoc with our results? For example, many statistical models actually assume the weak closure of the system, but this hypothesis can only be tested by the progress of science. Further research may indeed show that some hidden confounders were not taken into account in previous studies. This is the process of accumulation of knowledge, one of the main characteristics of science. If one has followed these guidelines, and if our hunches are empirically confirmed, one can at the most provide a good description of the possible relationship between income and mortality, controlling for various parasitic variables such as education. This exercise does not lead us however to understand why there exists such an association, because we have not yet provided an explanation of this relationship. Even if we know that there is a mortality gradient from the lower to the higher income classes, we have not uncovered why these differentials exist. We cannot furthermore generalize our results to other times and other places, nor predict the future, nor come up with effective policy measures. For example, one could suggest that to eliminate mortality differentials in the UK, it is sufficient to move everyone from the lower to the upper income classes, but this proposed policy is not very illuminating for the British health authorities! We must therefore find better answers to ®explanation-seeking-why-questions¯, as Carl Hempel (1965) has coined them, and this is the purpose of an explanatory theory. As I have stressed at the very beginning of this paper, theory-development in the social sciences is beset with difficulties. Firstly, in most cases we cannot have recourse to controlled experiments in order to check our hypotheses, for ethical or practical reasons. We cannot, for example, randomly distribute people among income classes and then see how and why they die! In addition, though randomization is usually considered an efficient method for controlling possible hidden confounders, this is however not absolutely true if there are interaction effects between explicit and implicit causes (see G. Wunsch 1994, section II). Secondly, mortality has numerous causes or determinants, and the interrelations between these causes are not well-known. Furthermore, many causes probably do not act in a deterministic way; probabilistic explanations are therefore required. There are also probably numerous risk factors specific to each individual; the theory we develop will therefore explain perhaps only a small part of the total variance. There are other problems too. Most concepts in the social sciences are abstract constructs, such as social class. These concepts must therefore be defined and operationalized in order to be measurable, and all demographers do not necessarily agree on a joint meaning and do not reach consensus on a common set of indicators or operational definitions. Finally, the social sciences deal with people and not only with things as in the natural sciences. Things such as stellar objects have no beliefs, desires, intentions, motives, reasons, and goals, contrary to human beings. Explaining human deeds therefore requires not only observing human actions (meaning here voluntary behavior) but also unravelling the reasons which cause these actions, i.e. understanding purportful (intentional) and purposeful (goal-oriented) behavior. This implies that it is not always sufficient to find the neurobiological, psychological, or social conditions which determine human actions, but that it can be necessary to discover too the purposes and reasons of the agent's actions, even though these reasons may grow up in many cases out of specific socio-cultural contexts . For example (J. D. Greenwood, 1988), an aggressive action can be a self-determined act of revenge, independent of all internal or external causes such as aggressive drives or environmental stimuli. On the other hand, as one's knowledge and value-system (determinants of many behaviors) are both acquired through observational learning and imitation, education in a broad sense is a causal factor explaining many human actions. I cannot therefore agree with the reductionist viewpoint of methodological individualism, as social contexts (inter- personal relations and institutions) affect individual behaviors. In particular, human beings do not act only on the basis of pure rationality; any theory of behavior neglecting the social context of human action is therefore incomplete and inadequate. Explanation therefore needs to determine the causes of the generation of the action, including the agent's reasons which have lead to this action as well as the social determinants of his action. Some of the things we do can be causally explained by the reasons we have for doing them (on this topic, see F. Dretske, 1988). However, as often in causation, the same causes may lead to quite different effects: people having similar reasons do not necessarily undertake similar actions, though their decisions are probably not arbitrary . Actions can therefore be predictable in various degrees, taking account among others of the pressure (constraints) of social institutions. This does not rule out the possibility that some actions might be truly undetermined, as the result of free agency, but we should nevertheless not give up too soon in our search for explanation. In fact, to some extent it is easier for us to understand human actions than plate tectonics for example, as we can in the first case draw upon our personal experience (see e.g. T. Abel, 1953). Many human actions are therefore expected or familiar, on the basis of our own behavior. This capacity to understand the behavior of someone else is, however, not per se a method of analysis; it helps in making assumptions or hunches, and in developing a plausible theory of human action, but it cannot be applied to verify a theory. Understanding through introspection or empathy is therefore no substitute for a good explanation, pace some sociologists who believe otherwise. Formally, a causal theory is a set of variables and of relations between these variables, developed in order to provide a possible answer to an explanation-seeking- why-question (Why?, What-if?, What-is?) concerning some particular domain in space and time, in our case a particular demographic phenomenon such as the fall in fertility in Western Europe during the latter half of the sixties. In addition, according to the practice of scientists (A. Hohenester et al., 1988), a theory should satisfy various criteria of simplicity and applicability. It should, for example, have relatively few parameters compared to the actual situation which it tries to explain, as a theory is a simplification or idealization of reality. A theory should be internally consistent or valid, e.g. not contain contradictory propositions. It should also be able to make new predictions. Finally, a theory should be clearly testable and externally valid, i.e. confirmed by the data, if we want to distinguish a scientific theory from a purely speculative one. In particular, as A. Piaser (1994) has stressed, a theory should ®mime¯ the process one is studying, i.e. translate into one's language the essential characteristics of the production mechanism. This language should be structurally homologous or similar to the process one is studying. Ideally, there should therefore be some sort of isomorphism or correspondence between the elements x in the real world and the elements x* of the language being used. For many philosophers, such as R. N. Giere (1994), mature scientific theories are formulated in the language of mathematics; such is the case with physics for example. Piaser criticizes this viewpoint, rightly I believe; mathematics is the language of physics not so much because physics is a ®mature¯ science but because mathematics corresponds so well to many physical phenomena. In this case, mathematics can be used not only to represent existing phenomena but also to predict new results. Moreover, what is a mature science? One which uses mathematics as its language? Is economics a mature science because its theories are often expressed mathematically? I believe the maturity of a science is not measured by the language it uses but by its capacity to explain its domain. Even in physics, mathematics cannot always be applied in such cases as many- variable, non-linear, stochastic problems, which are best expressed e.g. in statistical terms or by simulation. In the social sciences and in the humanities, mathematics is not necessarily the most adequate language available: try explaining in formulas the process of demographic transition in China, the fall of the Roman Empire, or the success of New Age music! The complexity of the social world is probably better convened in statistical terms, in graphs and graphics, and in ordinary language, than in the highly formalized language of mathematics. The latter would be useful nevertheless if the phenomena observed are simple and regular, rare cases in our field. This does not mean, I insist, that mathematics are useless in demographic theory, but that the quality of a theory in the population sciences is definitely not proportional to its mathematical content. Furthermore, this conclusion does not obviously condemn the use of mathematics in developing methods of demographic analysis, nor in demographic modelling; what would these fields be without mathematics! The use of ordinary language does not necessarily sanction however sloppy theorizing, fuzzy definitions, and faulty reasoning. Clear definitions, logic, and careful deductions (or inductions) are a prior quality of theories expressed in ordinary language, just as they are required as well in mathematical theories. In this respect however, a scientific theory expressed in a mathematical language is less prone to slipshod reasoning and to ill-defined concepts, due to the axiomatic nature of mathematics. Theories expressed in ordinary language require therefore an even greater effort in precision and in deductive reasoning. In particular, care should be taken not to take ordinary words for granted: concepts in demographic theories should therefore be clearly defined and made distinct, if necessary, from their ordinary meaning. For example, ®parity¯ in fertility theory does not mean ®equal status or pay¯, as defined in my Oxford Paperback Dictionary! In addition to interpreting an actual process in one's mind using an homologous language, another important characteristic of a theory is the property of generalization. Take the following case: Nancy had a traffic accident because she was dreaming of John and did not see the car coming from her right. This explanation is a reason of Nancy's car crash, but it is not a scientific theory ... except if day- dreaming is a common characteristic of female drivers. On the other hand, demographers have shown that driving a motorcycle is more dangerous than driving a car, that the young and the old have greater risks of accidents than the middle- age groups, that some months of the year, some days of the week, and some hours of the day are more accident-prone than others. Reasons have been given for these general findings, and though they do not obviously apply to all times and to all regions of the world, one can consider that demographers have provided a good general picture of traffic accidents and their causes. These relations are not deterministic, as some young drivers are more careful than their fathers for example, but they apply on average to the population of drivers. Do explanations therefore lead to laws (or lawlike statements) in the social sciences? All depends upon the definition of a ®law¯. If a law must be exceptionless and universally valid, as in physics, there are few cases of laws in the social sciences and in demography in particular, except for laws having a biological basis. For example, a woman's fertility is located between menarche and menopause; and the span of life of the human species is limited and equal to more or less 115 years. Otherwise, higher education does not always lead to lower fertility or infant mortality, and labor participation is not always reduced by the presence of young children. Even though these relationships are common, they are not true in each and every situation, and the strength and direction of the relation is not always the same. On the other hand, if by ®law¯ one means what should usually happen if our theory is correct, then social and population theories can indeed lead to laws valid for a certain place and time. Good explanations do not have to yield completely general laws. In this sense, belonging to an upper social class in the UK does usually lead to a better health status and a lower mortality. Wearing a safety belt in Italy is usually a sound behavior when driving a car. There are exceptions to these rules, but they nevertheless apply in most circumstances. What social laws formulate are typical actions, relations, or events, and not necessities. Social laws hold only ceteris paribus, and several social scientists have claimed that the ceteris paribus clause disqualifies laws in our domain, as all other things never remain equal in the real world . I believe this position is incorrect to some degree. In the long run, things obviously do change and the ®law¯ on safety-belts proposed above cannot be applied to the Roman Empire nor even to the contemporary world before the invention of that device! I have stressed however that explanations and laws in the social sciences are never completely general or universal; they are only valid for a certain time and place, i.e. for a specific niche in history. Moreover, laws do not have to be exceptionless. This is due to the fact that explanations and scientific laws are abstractions, idealizations, and generalizations of the real world; exceptions therefore can and do occur, but this does not necessarily disprove the explanation nor the law. A theoretical law states what would happen if various simplified conditions hold. If these conditions do not perfectly hold, i.e. if the ceteris paribus clause is not completely satisfied, laws can still cite factors, aspects, and tendencies of a complex situation, as H. Kincaid (1990) has stressed. To quote Kincaid more or less freely, here are six practices which lend credence to ceteris paribus laws, even though they nevertheless do not waive all criticism: 1) It can sometimes be shown that in some narrower range of cases, the ceteris paribus law is satisfied. 2) We can sometimes show that although other things do not remain equal, it makes little difference in practice. 3) We can sometimes explain away the failure of the ceteris paribus law, invoke the relevant reasons thereof, and give at least an approximate prediction of the combined effect of the counteracting factors and relevant laws. 4) Sometimes, we can provide inductive evidence for a ceteris paribus law, by showing that as conditions approach those required by the ceteris paribus clause, the law becomes more predictively accurate. 5) The law may still make some predictions which are borne out and which are unexplained by alternative hypotheses. 6) A ceteris paribus law may provide a single explanation for some set of diverse phenomena, even though we cannot fully specify the ceteris paribus clause or explain away all counteracting factors. It must be stressed that the ceteris paribus clause does not imply that all causes except one (the risk factor of interest) must remain equal. Once one has estimated for example the structural parameters of the causal model, it is possible to determine the impact of, say, Xk on Xn taking account of the changes in the other Xi. The clause means that the parameters themselves have to remain stable and not change significantly, due to a modification in context that would alter the causal structure by e.g. adding new variables or changing the relationships between existing variables. It is often assumed that a law is any true generalization which supports prediction. A law is however not a statistical regularity, as regularities are not explanatory by themselves. Indeed, regularities are neither necessary nor sufficient conditions for causation. Social regularities are only associations, for example a positive relation between income and health, and as such they may lead to wrong causal inferences. On the other hand, the observation of an empirical regularity is often a first step towards understanding the causal mechanism producing the association, that is discovering a plausible explanation, such as Durkheim's theory of suicide. To give another example, if health is positively correlated with income, this might be due to particular reasons such as a better access to the health care system in the absence of an adequate social security system. Regularities may therefore lead to explanations; if the latter are confirmed or not falsified by the data, explanations in turn may lead to lawlike statements. Laws derive from explanatory theories, and not vice versa! As D. Little (1993) has pointed out, once we have discovered and confirmed the causal mechanism involved, we no longer even need to use the empirical regularity in order to explain. In the above example, access to health care becomes the prime determinant of health and not income any more: a good social security system would be just as valid as a higher income. All regularities do not however lead to a satisfactory explanation (see E. Sober, 1987); for example, two time-series can be highly correlated without one being the cause of the other or both having a common cause. This is the case for example if both variables have increased over time, such as the increase during the past centuries in average temperature and the world population; no one I hope believes this is a true causal relationship, at least during the pre-industrial era! A causal theory usually leads to the type of proposition ®If X then Y because Z¯, where X represents causes, Y effects, and Z reasons. In the causal process, causes can be internal or external to the agent. For example, education of the mother is an external cause of her lower infant mortality, but the reasons thereof relate to the human agent and not to the external cause, as Caldwell's explanation of the relationship between infant mortality and education has exemplified so well. To give another example, drawn from R. Franck (1994b), if I pour water on the flowers in my garden, they will grow; but the reason of their growth is mainly contained in the flowers themselves and not in the water: try pouring water on pebbles instead and see the result! The same external cause can therefore produce different results, taking account of the internal causes particular to each agent. In other words, a good explanation implies not only determining the external causes of an agent's action, but also the agent's reasons leading to this voluntary behavior. The ®If...then¯ statement must be qualified however. Consider the following two examples: ®If fire then smoke¯ and ®If smoke then fire¯. Both of these statements are correct in English. As one sees in this case, common-sense reasoning in natural language may yield contradictory propositions. As J. Pearl (1988a) has stressed, it is therefore necessary to distinguish, from a causal viewpoint, between expectation-evoking (If fire then smoke) and explanation-evoking (If smoke then fire) propositions, the first being a causal rule and the second an evidential rule. In this paper, we deal mainly with causal rules or propositions leading from the cause to the effect (if ®cause¯ then ®effect¯), and not vice-versa. Evidential rules are however not devoid of explanation: they point from evidence to hypothesis or from effect to cause, such as in the case of symptoms Y suggesting disease X (or smoke suggesting fire). Explanation-evoking propositions are currently used in medical practice for example, but they cannot combine diagnosis with prediction. On the contrary, causal propositions predict the effect. For example, smoking causes lung cancer (among other diseases), and therefore one predicts for Nancy (who smokes) a higher probability of developing eventually a lung cancer than if she did not smoke. On the other hand, Joe's lung cancer suggests or is evidence for his addiction to smoking, but in fact his disease could be due to asbestos pollution in his factory (actually, Joe is a non-smoker). This is the essence of what is sometimes called ®causal asymmetry¯ (not to be confused with causal priority): two or more causes (smoking and asbestos) can have a common effect (lung cancer), while two or more consequences (lung cancer and bronchitis) can have a common cause (smoking). Common causes compete with each other for the effect, while common consequences support each other; for an impressive and thorough discussion of the logic of causal dependencies, see J. Pearl (1988b). In two recent articles, D. M. Hausman (1993a and 1993b) has persuasively shown why causes explain their effects while effects do not explain their causes, by establishing a link between causal asymmetry and explanatory asymmetry. Briefly, the argument goes as follows. Suppose X(3) is dependent upon X(1) and X(2). If the value of X(3) doubles, this has no implication for the values of the causal variables: all one can say is that at least one of them increased by enough to make X(3) double. To quote Hausman (1993a): ®One cannot make causes wiggle in any definite way by manipulating their effects¯. On the contrary, if X(1) is a cause of X(3) and X(4), and if one knows the relationship between these variables, a change in X(1) can predict the changes in the two effects, under the assumption of the weak closure of the system of variables. The same result is valid in the first example: if X(1) is independent of X(2), a change in X(1) can be used to predict the change in X(3), in a deterministic or in a stochastic situation. Note that in these examples, X(1) and X(2) are independent of each other, while X(3) and X(4) are not as they have a common cause X(1); this is the essence of causal asymmetry, as stated above. To sum up, why- questions are requests for causes and not for effects, as knowledge of causes is more valuable in many cases than knowledge of effects. As one knows, policy measures are therefore based on the manipulation of causes and not of effects! 3. Do demographers have theories? The answer to this question is obviously positive, since demography during the past decades has become less descriptive and more explicative. In fact, theories abound in demography, even to the extent that it is difficult to extract from the mass (or mess) of research hypotheses, more or less confirmed, some strong conclusions and some definite explanations. Demography has never had a grand explanatory paradigm, such as the postulate of rationality and the concept of utility in microeconomics; the postulate, in sociology, of the dependence of conduct on social forces outside the individual; structuralism in linguistics or in social anthropology; Gestaltism or behaviorism in psychology. Some demographers have however suggested, from an anthropological viewpoint, that population processes lead to the adjustment of population size to the capacities of the environment (see B. de Bruijn, 1993). A central paradigm in demography could therefore be based on the following postulate: every population adopts social and individual conducts leading to the eventual adaptation of its numbers to the environment's carrying capacity, the latter varying across time and space according to material, social, and individual constraints and needs. Fertility, mortality, and migration changes, brought about by individual and social conducts, would then be seen as strategies designed to adapt population numbers to the changing environment, taking account of constraints, needs, and values. The lack of a grand unified approach is not necessarily a disadvantage, as the break between Freudian and Jungian psychoanalysts shows for example. A general paradigm, such as Freudism, can become canonized and it can then stifle research. If Freud's work stands as a bible for psychoanalysts, how could this field progress beyond and possibly in contradiction to what Freud has written? To take another example, can one develop Marxism without Marx, if someone is still interested? Luckily, demography has always avoided canonizing its founding fathers, and contrary to sociology or to philosophy we are fortunate to have few doctoral theses devoted to the thought and writings of e.g. L. Henry, A. J. Coale, or J. C. Caldwell, even though these demographers, among others, have had a profound impact on population research. Their work has indeed become integrated in the mainstream of demographic thinking, and that is the best compliment we can pay to them. On the other hand, the fact that the population sciences lack a basic demographic paradigm implies that demographers have to cannibalize other fields of inquiry, in order to found their explanations. For example, as two recent general reviews of the field of mortality have shown  and to which the reader is referred, explanations have drawn upon findings in history, epidemiology, public health, nutrition, sociology, anthropology, economics, and political science. The spectrum is probably wider still, as findings from e.g. the psychology of health could also be considered in a micro approach. Though social facts, and demographic phenomena in particular, should be explained in a multidisciplinary approach, one can raise nevertheless the following related questions. These questions have an impact, among others, on the training of demographers: in addition to demographic modes of observation and data analysis, what indeed should be taught to burgeoning demographers? What is specifically demographic in the explanation of mortality levels and trends? Are all explanations in demography necessarily drawn from other fields? Do demographers therefore need to be at the same time specialists in these different fields, a hopeless task indeed? Is population thus to be studied only by multidisciplinary teams, and what would be the specific role of demographers in these teams? Is the specificity of demography only to gather together the results of various approaches to the explanation of population trends and levels? Would a demographer therefore be the conductor of a multidisciplinary orchestra in which there are no demographic instruments? If one looks at the demographic literature, too many papers still do not define the concepts used, nor explain why such and such indicators are taken instead of others. Very often too, the explanatory framework (if there is one) is a mere sketch, and frequently one does not learn the reasons for postulating relationships between some variables and not between others. Population journals have their part of responsibility here: if concepts are not defined, if the link between indicators and concepts are not clarified, if the conceptual framework is unsubstantiated, the paper should be rejected for publication. Actually, as experience shows, most population journals apply these selection criteria none too often. For example, O. Lopez Rios, A. Mompart, and G. Wunsch (1992) have related regional mortality differentials in Spain to the demand and supply of health care, controlling, among others, for the level of social development and of economic development in the Spanish provinces. Are these macro-concepts relevant for causal analysis, as they are so wide and fuzzy? Though the authors have given a definition of ®social development¯ and of ®economic development¯ in their paper, have they really pointed out all their dimensions and justified the battery of indicators they use in their LISREL model? Are all the relations postulated between the latent variables based on sound theoretical grounds and prior research findings? The answers to these questions are most probably negative. Too many studies are still confined to pointing out the most relevant determinants, controlling for possible confounding factors. We thus know that having a car or being a tenant is strongly related to mortality in the UK, that toilets outside the house are a significant determinant of mortality in Norway, etc. These factors are actually indicators of social class differentials; what we need instead is the knowledge of the causal mechanism (the intervening variables and paths) leading from social class to mortality, i.e. the reasons of the causes. To give another example, if education of the mother has a strong impact on the survival of her children, what are the reasons thereof, and have we confirmed or falsified the postulated mechanisms? In an excellent recent review of studies dealing with social class differentials in mortality, G. Davey Smith et al. (1994) have shown that the usual explanations of differential mortality by social class do not hold: the persistence of social class differences in mortality cannot be explained solely by artefacts, social selection, or behavioral factors. They stress the need for relating social selection and behavioral factors to life circumstances, and for taking into account the pattern of inter- relations between the various possible causes of social class differentials in mortality. For example, in the British Regional Heart Study, when social class differentials in ischemic heart disease (IHD) risk were considered, they were adjusted for smoking behavior which was said to largely account for them. In the same study, when the relationship between smoking and IHD risk was examined, socio-economic position was not taken into account. Using the Whitehall data, the authors have however demonstrated that a significant part of the association between smoking and mortality appears to be due to the relationship between smoking and socio- economic position. A life course analysis of the link between social class, morbidity, and mortality has therefore to be performed, using longitudinal data. Ideally, this study should also take into account familial background and household characteristics, as stressed in M. Salhi et al. (1995). In the field of family formation and fertility, the situation is probably not better, though I lack the expertise in this domain in order to be a good judge . Take however a recent example, drawn from an otherwise excellent paper by R. Lesthaeghe and G. Moors (1994). These authors postulate that one's system of values (SV) determines to some extent decisions on family formation (the ®selection¯ effect) and/or that the choice of one's family and domestic type (FT) reinforces or weakens some of one's values (the ®affirmation¯ effect). This leads to the temporal causal chain SV FT SV which, ideally, should be confirmed or falsified with longitudinal data. The authors test this theory with period data drawn from the 1990 European survey on values in four EU countries. Firstly, as the authors themselves acknowledge, cross- sectional data cannot discriminate between the two effects, as the data are not ordered in time. Secondly, as the survey was not conducted for the purpose of testing this particular theory, do the available indicators tap all the relevant attitudes and values? Along the same line of argument, the authors have named their latent variables (dimensions of conservatism) by exploratory principal components analysis applied to the available set of items, in an inductive approach. A sounder way would probably be to use a deductive approach, starting with the selection of possible relevant concepts, constructing scales for each attitude or value, and checking the validity of the items by confirmatory analysis. This approach would however require the collection of original data, and naturally be much more expensive than solely using secondary data already obtained for other purposes. Thirdly, and more importantly for the quality of a theory, the authors do not test the ®reasons¯ of the causes: what is the mechanism through which SV influences FT, or vice versa? Finally, though Lesthaeghe and Moors take account too, as independent variables, of the level of formal education, of the type of religious upbringing, and of current socio-economic status, can the authors assume the closure of their system? There might be e.g. a common cause (socialisation itself?) influencing both SV and FT, which should be controlled in order to derive the true relationship between the two sets of variables. Moreover, what are the relationships between the independent variables themselves? One can assume, for example, that formal education and religious upbringing influence one's current attitudes towards conservatism, and that the level of formal education has an impact on one's socio-economic position. The residuals approach, adopted by the authors for dealing with multicollinearity between independent variables in a logistic regression, does not adequately take account of a possible order among causes. As Lesthaeghe and Moors are specifically concerned with explanatory theory, these issues cannot I believe be lightly dismissed, however interesting their results. 4. Discussion and conclusions Demographers, like many other social scientists, often complain about the lack of theory in their discipline. In fact, as I have previously pointed out, demography has developed numerous conceptual frameworks and theories explaining fertility, mortality, and migration levels and trends. We are indeed confronted in each field with a mosaic of theories from which no general Gestalt emerges. This is partly due, as we have seen, to the absence of a central paradigm of explanation in demography. It is also due to the fact that many theories are so local in scope, so time and space dependent, that cumulation of knowledge is difficult to achieve. Having shown for example that the recent decrease in Syldavian mortality is due to less smoking and alcohol intake, and that the fall in Bordurian mortality results from better access to the health care system, how do we put these two explanations together? There are probably many Humpty-Dumpties similar to this example in demography. A more troublesome problem related to the diversity of theories is the level of explanation. Some theories in the population sciences are developed at the macro (societal) level and others at the micro (individual) level; combining both types is once again difficult to achieve. Like all social sciences, demography deals with individuals in a social context, i.e. with people behaving in interaction with others and with institutions. As Emile Durkheim or Max Weber have already stressed a long time ago, individual actions are therefore to a large extent socially and culturally determined. Population theories thus have to take account of this multi- level determination: theories developed at one level only are bound to be inadequate. In particular, I doubt if demographic theories (and other social theories) can legitimately be grounded on the so-called rationality principle, in at least the traditional sense of utility-maximizing behavior, such as in the micro-economic theories of fertility (see B. de Bruijn, 1992). If this is the case, demographic theories are firmly rooted in the individual; society only intervenes as constraining individual decisions. Firstly, many people do not act rationally, as many studies in the decision and behavioral sciences have shown, due to e.g. the ambiguity of the context, a lack of logical competence, their unconscious, their emotions, or individual and social habits. In fact, as W. Berkson (1989) has suggested, the social sciences can perhaps dispense with rational behavior: what we are probably much more interested in is predictable action, whether it is governed by rationality or not. For example, young and drunk late night irresponsible male drivers are not acting rationally, but they are acting predictably as mortality statistics have unfortunately shown. Rationality is in the mind of the scientist, who tries to make sense of the data, and not necessarily in the minds of the actors themselves! Secondly, and more importantly perhaps, the rationality principle focuses the theory, as I have said, at the micro level. The latter cannot show, I believe, how we move from individual decisions to social conducts, from personal characteristics to social structures, or from lower level to higher level organisation. Collectivities do not act like individuals, and their ®behavior¯ cannot be considered solely as the aggregation of the actions of their members. This possibly implies moving from the individual to interpersonal networks and systems, as in the study of the behavior of small groups such as families. Much still needs to be done however in the study of large groups, with which demographers are usually concerned. To conclude, good explanations of the causes and consequences of population trends require theories taking account of both the micro and the macro levels, of individuals and society. Multi-level theories cannot therefore be based on a rationality principle which would be firmly grounded at the individual level only. Individuals are members of a family and a household, of a neighbourhood, of a workplace, of a community, of a social class, etc. Like Russian dolls, individuals are therefore imbricated in a succession of wider more or less overlapping entities, which transcend to some extent the individual decision-making process. Demographic theories thus have to tell us how and why society influences individual behaviors concerning birth, death, and migration, but also how and why individual behaviors are themselves organised into higher-level social conducts. In other words, population theories must integrate individual behavior and social structure. Demography is, after all, a Social Science; it should therefore be both social and science! 5. References Abel T. (1953), The Operation Called Verstehen, in H. Feigl, M. Brodbeck (eds.), Readings in the Philosophy of Science, Appleton- Century-Crofts, New York, 677-687. Berkson W. (1989), Testability in the Social Sciences, Philosophy of the Social Sciences, 19, 157-171. Burch T. K. (1994), Icons, Strawmen and Lack of Precision: Reflexions on Current Demographic Theorizing about Fertility Decline, Paper presented at the 1994 Meeting of the Population Association of America, Miami. Davey Smith G., D. Blane, M. Bartley (1994), Explanations for Socio-economic Differentials in Mortality: Evidence from Britain and Elsewhere, European Journal of Public Health, 4(2), 131-144. de Bruijn B. (1992), The Concept of Rationality in Social Sciences, Pdod-Paper n. 9, Population Research Centre, University of Groningen. de Bruijn B. (1993), Interdisciplinary Backgrounds of Fertility Theory, Pdod- Paper n. 16, Population Research Centre, University of Groningen. Dretske F. (1988), Explaining Behavior. Reasons in a World of Causes, Mit Press, Cambridge, Mass. Einhorn H. J., R. M. Hogarth (1986), Judging Probable Cause, Psychological Bulletin, 99(1), 3-19. Ekstr”m M. (1992), Causal Explanation of Social Action, Acta Sociologica, 35, 107-122. Franck R. (1994a), Introduction g‚n‚rale, in R. Franck (sous la dir.), Faut-il chercher aux causes une raison?, Librairie Philosophique J. Vrin, Paris, 24-40. Franck R. (1994b), Deux approches inattendues de la causalit‚: Aristote et les Sto‹ciens, in R. Franck (sous la dir.), Faut-il chercher aux causes une raison?, Librairie Philosophique J. Vrin, Paris, 166-182. Giere R. N. (1994), The Cognitive Structure of Scientific Theories, Philosophy of Science, 61, 276-296. Glymour C., R. Scheines, P. Spirtes, K. Kelly (1987), Discovering Causal Structure, Academic Press, Orlando. Greenwood J. D. (1988), Agency, Causality, and Meaning, Journal for the Theory of Social Behaviour, 18(1), 95-115. Hausman D. M. (1993a), Why Don't Effects Explain Their Causes?, Synthese, 94, 227-244. Hausman D. M. (1993b), Linking Causal and Explanatory Asymmetry, Philosophy of Science, 60, 435-451. Hempel C. G. (1965), Aspects of Scientific Explanation, The Free Press, New York. Hohenester A., L. Mathelitsch, M. J. Moravcsik (1988), The Usage of ®Theory¯ and ®Model¯ in Scientific Conceptualization, Scientometrics, 14 (5-6), 411-420. Kincaid H. (1990), Defending Laws in the Social Sciences, Philosophy of the Social Sciences, 20(1), 56-83. Lave C. A., J. G. March (1975), An Introduction to Models in the Social Sciences, Harper and Row, New York. Lesthaeghe R., G. Moors (1994), Expliquer la diversit‚ des formes familiales et domestiques: th‚ories ‚conomiques ou dimensions culturelles?, Population, 49(6), 1503-1526. Little D. (1993), On the Scope and Limits of Generalizations in the Social Sciences, Synthese, 97, 183-207. Lopez Rios O., A. Mompart, G. Wunsch (1992), SystŠme de soins et mortalit‚ r‚gionale: une analyse causale, European Journal of Population, 8, 363-379. Loriaux M. (1994), Des causes aux systŠmes: la causalit‚ en question, in R. Franck (sous la dir.), Faut-il chercher aux causes une raison?, Librairie Philosophique J. Vrin, Paris, 41-86. Pearl J. (1988a), Embracing Causality in Default Reasoning, Artificial Intelligence, 35, 259-271. Pearl J. (1988b), Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference, Morgan Kaufmann Publishers, San Mateo. Piaser A. (1994), Entre causalit‚ et processus. L'explication en sociologie historique, in R. Franck (sous la dir.), Faut-il chercher aux causes une raison?, Librairie Philosophique J. Vrin, Paris, 381-419. Salhi M., G. Caselli, J. Duchˆne, V. Egidi, A. Santini, E. ThiltgŠs, G. Wunsch (1995), Assessing Mortality Differentials Using Life Histories: a Method and Applications, in A. Lopez, G. Caselli, T. Valkonen (eds.), Premature Adult Mortality in Developed Countries: from Description to Explanation, Clarendon Press, Oxford. Sober E. (1987), Explanation and Causation, British Journal for the Philosophy of Science, 38, 243-257. Stehbens W. E. (1993), The Quality of Epidemiological Data of Coronary Heart Disease and the Lipid Hypothesis of Atherogenesis, Journal of Clinical Epidemiology, 46(12), 1359-1364. Tabutin D. (to be published), Transitions et theories de la mortalit‚, in H. G‚rard, V. Pich‚ (sous la dir.), Sociologie de la population. ThiltgŠs E., J. Duchˆne, G. Wunsch (1995), Causal Theories and Models in the Study of Mortality, in A. Lopez, G. Caselli, T. Valkonen (eds.), Premature Adult Mortality in Developed Countries: from Description to Explanation, Clarendon Press, Oxford. Traub J. F., H. Wozniakowski (1994), Breaking Intractability, Scientific American, Jan., 90B-93. Wunsch G. (1994), L'analyse causale en d‚mographie, in R. Franck (sous la dir.), Faut-il chercher aux causes une raison?, Librairie Philosophique J. Vrin, Paris, 24-40. _______________________________ * The comments and suggestions of Robert Franck and Hubert G‚rard, among others, are gratefully acknowledged. 1. For a recent discussion of the possible determinants of female labour supply, see e.g. K. Henkens et al. (1993). 2. I make no distinction, in this text, between population and demographic theories, and use both as synonyms. 3. Proving one's theory is however a very difficult task! See G. Wunsch (1994), section VII. 4. This point has been stressed by Max Weber; see M. Ekstr”m (1992). 5. For an exponent of this view, see M. Loriaux (1994). 6. E. ThiltgŠs et al. (1995); D. Tabutin (to be published). It seems that migration theories are also based mostly on non- demographic explanations, economics being the prime source in this case. 7. For an interesting overview, see B. de Bruijn (1993); T. K. Burch (1994) also provides a stimulating discussion of various issues related to theories of fertility decline.