****This chapter was corrected down to Causality section and corrected by Becky to there. After Becky, I added stuff at end CHAPTER 4-3 INTERPRETING INFORMATION Experience makes your life feel richer. Experience also can help you to understand, predict, or control the world around you. But to make use of your experience you must interpret the information you gain from experience. Deductively rearranging the information that experience provides can illuminate those data. Chapter 4-2 discussed those manipulations. Now we take up the matter of going beyond the information at hand -- combining the information with theory and other fact to obtain as much knowledge power as possible. There is not much that one can do with a single isolated piece of information. If you see Gertie standing next to the lamp post over there, about all you can do with that observation is to guess that the next time you look, Gertie will still be standing next to the lamp post, and not standing elsewhere. That is, in the absence of any other information, including the absence of any change that you know of (which means that the time span should not be very long) your best guess is that things will continue to be as they are now. The basis of useful information is an association between sets of observations on two dimensions -- the colder the climate, the smaller the windows in the houses; leg strength in a person generally increases until the age of 25 (or whatever), and then decreases; an increase in a country's money supply is accompanied by an inflation in prices; women have fewer auto accidents than do men, on average. Naked correlations such as these need to be bolstered by additional material in order to make them yield fuller knowledge. The various processes of interpreting the correlations within the information at hand -- explanation, prediction, causal attribution, and others -- partially overlap each other, but they also differ. Figure 4-3-1 diagrams the connections among these processes. Explanation and the determination of causality both take the raw information about the relationship between two sets of facts -- say, the movement of bathing suit sales, and the number of weddings -- and choose the additional interpretations to attach to that relationship and that particular set of facts. The aim is to extend the interpretation to new situations not included in the original information. For example, if you have just learned that your new fertilizer discovery will increase the growth of soybeans in the Midwest, should you conclude that the same benign effect will take place in the eastern states and in India? The concept of causality as a special source of difficulty. It is sometimes difficult to determine, and sometimes difficult to interpret. Prediction of the future is another sort of crucial extension of the information available from the past. Will the Orioles do as well in the second half of the season as in the first half? Below you will find a set of principles for forecasting accurately. The chapter also points out the main causes of error in prediction of economic and social phenomena. FIGURE 4-3-1 (from Research Methods) CAUSALITY Whether something "causes" something else can be both puzzling and important. Sometimes it seems easy to say what causes what, as when a baseball breaks a window. But sometimes the interpretation is not so simple, as when we ask why a golfer with a four-stroke lead six holes before the end of the Masters tournament finally lost, or whether cigarettes cause lung cancer. We are interested in causality when we want to decide how to act - whether to stop smoking, whether to play one quarterback rather than another, and whether to make Willie pay for the window which was broken. This action orientation is unlike explanation, which we seek in order to satisfy our curiosity and our general understanding of the world around us. Of course explanation and causal assessment overlap greatly, but here we focus on the difference between them. Let's postpone for a moment the question of what we mean when we use the word "causality", and note that there are two ways we can improve our understanding of whether one thing causes another: 1) get more and better facts, and 2) improve the interpretation of the facts we have. Getting more facts was discussed in Chapter 3-3 on science. Here I'll mention only that experimentation is an important method for producing useful information about causality, but by no means a perfect or conclusive method. Remember the gentlemen who got drunk on scotch and soda, whiskey and soda, and brandy and soda on successive evenings and concluded that soda causes hangover? We should also note that one can often think about the imputation of causality at many different levels, as is suggested by Figure 4-3-1 about auto accidents. Sometimes people talk past each other for this reason, as is often the case in discussions of crime. Figure 4-3-2 Now we will discuss the interpretation of the available facts with respect to the concept of causality. Especially since David Hume addressed himself to the subject first in 1739 and again in 1748, the concept of causality as used in scientific work has tantalized philosophers and has occasionally occasioned furious controversies. All that could be agreed upon has been lack of agreement. Hume's contribution was to assert this idea in crystal-clear fashion: All that can ever be known about events is what can be observed. And the most that can be observed is that there is a "constant conjunction" between events, a statistical correlation. Hume specifically denied that a causal relationship can be established by a priori logical analysis of the features of events or objects. Though Hume's analysis refuted much murky confusion, it is not fully satisfactory, because both in scientific work and in everyday life we are prepared to call "causal" some correlations but not others. A statement that the flight of birds overhead precedes rain seems to be a different sort of statement than is the statement that actuating the starter of the automobile precedes the starting of the engine; indeed, we behave very differently toward these two statements. And there seems to be a difference in meaning between the empirical relationship of prices on the Dutch stock exchange to the number of houses built in the U. S., and the empirical relationship of mortgage interest rates in the U. S. to the number of houses built in the U.S. This difference between statements that are only predictions and those that seem to have additional meaning causes people to continue wrestling with the concept of causality. To put it differently, Hume discovered the impossibility of a definition of causality in terms of physical properties. But Hume did not replace the material-property definition with a concept that fits the needs of the working scientist. And it is that nut that I sought to crack when I first began to think about the matter in the early 1960's. CAUSALITY AND EXPERIMENTATION Natural scientists often say that an experiment defines causality. Indeed, a positive experimental result is quite a good test of causality when experimentation is possible. If the stimulus is followed by response, and non-stimulus is followed by non-response, as in John Stuart Mill's canons, the stimulus- response relationship is commonly said to be causal. Experiments are repeatable, and hence the definition has high reliability. This explains why there is relatively little dispute in the experimental sciences about which relationships to call causal. A single experimental relationship is not, however, a perfectly valid indicator of causality. In the famous Hawthorne experiments, for example, variation in light intensity in the work room was followed by variation in the work performance. But it became obvious that it was other factors -- perhaps the attention of the experimenters, or perhaps associated changes in the rate of pay, an issue still the subject of controversy -- which caused the increase in work output. It is intuitively clear to experimental scientists that causality is shown better by a series of experiments that vary the parameters of the original experiment, than by a single experiment. One can therefore state the criteria of causality in experimental situations as follows: (l) Keeping all other conditions the same, vary the stimulus and observe the response. (2) If the variation in stimulus is followed by variation in the response, yielding a statistically significant relationship that is also strong enough to be of some importance, vary the conditions and repeat the experiment. (3) If the original relationship continues to appear even under different parametric conditions, call the relationship 'causal.' Two important points should be noted about these criteria of causality when experimentation is possible: (l) It is the actual operation of experimenting that defines the term;_t_h_e_ _e_x_p_e_r_i_m_e_n_t_ _m_u_s_t_ _a_c_t_u_a_l_l_y_ _b_e_ _c_a_r_r_i_e_d_ _o_u_t; the experiment is important as an act, and not as a model, in this context. This definition specifies that an actual experiment cannot be replaced by a hypothetical experiment. (2) Whether or not a relationship will be called "causal" is not an automatic and perfectly objective process; rather, it requires _j_u_d_g_m_e_n_t based on unspecifiable contextual knowledge, e.g., judgment about whether the _a_p_p_r_o_p_r_i_a_t_e conditions have been varied, whether _e_n_o_u_g_h conditions have been changed, and whether the observed relationship is sufficiently important or strong. CAUSALITY IN NON-EXPERIMENTAL CONTEXTS Now we move on to the much harder task, a set of criteria of causality for observed relationships that are not subject to experiment. One suggestion is what Wold called "the fictitious experiment," equivalent to this test: Judge whether the observed situation has the properties of a controlled experiment. If you so judge, call the observed relationship 'causal.' But this definition differs from the experimental working definition in that it does _n_o_t include the crucial operative phrase, i.e., "carry out the experiment...." Furthermore, it is clear that this definition has low reliability; that is, there is much room for disagreement among scientists about whether or not an observed situation does indeed have the properties of an experiment. One frequent suggestion has been to deny the label "causal" to any non-experimental observed relationship, to say "correlation does not prove causation." But there are several drawbacks to this suggestion: (l) The term "causal" is frequently used in scientific descriptions of non-experimental relationships, and therefore we need to discern its meaning when it is used. (2) The "pure" scientist may be able to withhold the appellation of causality (though it may well be a useful word in his vocabulary), but the decision-maker (or the scientist _q_u_a adviser for "decision-makers") certainly cannot duck the issue. The 1964 Surgeon General's Committee on Smoking and Health knew that many people would not decide to quit smoking who otherwise would if the Committee did not use the word "cause," and the progress of legislation might also depend upon whether they wrote "causal." Therefore they chose to use the word. (3) Our intuition tells us that there is an important difference among various observational relationships, a difference that corresponds to our usual sense of the word "causal." For example, there is a difference between a) the statement that when one clock's hour hand reaches l2, another clock strikes the hour, and b) the statement that when you remove the plug from the socket the electric clock ceases to run. For an economic example, we sense a difference between the observed association relating prices on the Dutch stock exchange to the number of houses built in the U.S., and the observed association relating mortgage interest rates in the U.S. to the number of houses built in the U.S. Similarly, in sociology there seems to be a difference between a statement that certain phases in the moon precede or accompany a rise in the murder rate, and the statement that a rise in the temperature precedes or accompanies a rise in the murder rate. Here is the working definition that I propose for the term "cause and effect relationship" in non-experimental situations: l. _S_t_r_e_n_g_t_h_ _o_f_ _C_o_r_r_e_l_a_t_i_o_n. The relationship must be strong enough to be interesting and/or useful. For example, one is not likely to say that wearing glasses "causes" (or "is a cause of") auto accidents if the observed correlation is .07, even if the sample is large enough to make the correlation statistically significant. (A correlation is measured by a number between -1.0 and +1.0, with zero indicating no correlation. In almost every discipline except perhaps educational research, a correlation of .07 is usually considered to be of no importance at all.) In other words, unimportant relationships are not likely to be labeled "causal." Of course this criterion by itself is not enough; that is the grain of truth in the expression "correlation does not prove causation." But nothing else "proves" causation, either; that is the larger truth. 2. _F_e_w_n_e_s_s_ _o_f_ _S_i_d_e_ _C_o_n_d_i_t_i_o_n_s. The relationship in question must not require too many "if's," "and's," and "but's." That is, the "side conditions" must be sufficiently few, and sufficiently observable, so that the relationship will apply under a wide enough range of conditions to be considered useful or interesting. For example, one might say that an increase in income "causes" an increase in the birth rate if this relationship were observed everywhere. But if the relationship is only found to hold true in developed countries, among educated persons, among the higher-income groups, among those who can be assumed to know about contraception, then one is less likely to say the relationship is causal -- even if the correlation is extremely high once the specified conditions have been met. 3. _N_o_n_-_S_p_u_r_i_o_u_s_n_e_s_s. For a relationship to be called "causal" there should be good reason to believe that even if the "control" variable is not the "real" cause (and it never is), some "more real" variables will change consistently with changes in the control variables. (Between two variables, v may be said to be the "more real" cause, and w a "spurious" cause, if v and w require the same side conditions except that v does not require a side condition on w.) This third criterion (non-spuriousness) is of particular importance to policymakers. The difference between it and the previous criterion concerning side-conditions is that a plenitude of very restrictive side- conditions may take the relationship out of the class of causal relationships even though the effects of the side-conditions are known. But the criterion of non-spuriousness concerns variables that are as yet _u_n_k_n_o_w_n and unevaluated, but which have a _p_o_s_s_i_b_l_e ability to upset the observed transformation. Examples of "spurious" relationships and hidden-third-factor causation are commonplace. For a single illustration, toy sales rise in December. One runs no danger in saying that December "causes" an increase in toy sales even though it is "really" Christmas that causes the increase, because Christmas and December (almost) always accompany each other. One's belief that the relationship is not spurious is increased if _m_a_n_y likely third-factor variables already have been investigated and none reduces the original relationship. This is a further demonstration that the test of whether an association should be called "causal" cannot be a logical test; there is no way that one can express in symbolic logic the fact that "many" other variables have been tried and have not changed the relationship in question. The more tightly a relationship is bound up with (that is, deduced from, compatible with, and logically connected into) a general framework of theory, the stronger is the relationship's claim to being called causal. For an economic example, the positive relationship of the interest rate to business investment, and the relationship of profits to investment, are more likely to be called "causal" than is the relationship of liquid assets to investment. This is because the first two statements can be deduced from neo-classical price theory whereas the third statement cannot. This element of theory in scientific thinking and in the explication of the concept of causality is perhaps the biggest difference between Hume's thinking and contemporary thinking. Hume focused on each relationship all by itself, and all by itself there is no more that one can say except that there is "constant conjunction". But the presence or absence of other statements of relationship that are either connected to the relationship in question by commonsensical logic, or even more strongly by an integrated body of theory, is very important in deciding whether it is sensible to think that the statement in question should be considered as only a predictive relationship, or whether one should go further and call it "causal". One likely reason for absence of this consideration in Hume's thinking is that in his time there was no branch of science, except perhaps physics, that possessed such an integrated body of theory. Economics lacked an integrated body of theory until Adam Smith came along to weld together the various fragmentary observations that already existed; William Letwin (1975) has persuasively argued that this was Smith's greatest achievement. Certainly there was at that time no philosophy of science that analyzed the importance of an integrated theoretical framework, as has been done for us by recent writers. Indeed, none of the social sciences other than economics yet has a well-developed body of deductive theory, and hence this criterion of causality is not weighed as heavily in those social sciences. Rather, the other social sciences seem to substitute a weaker and more general criterion -- whether the statement of the relationship is accompanied by other statements that seem to "explain" the "mechanism" by which the relationship operates. Consider, for example, the relationship between the phases of the moon and the suicide rate. The reason sociologists do not call it "causal" is because there are no auxiliary propositions that sensibly "explain" the relationship and describe an operative mechanism. On the other hand, the relationship between broken homes and juvenile delinquency is often referred to as causal, because a large body of psychological theory serves to explain why a child raised without one or another parent, or in the presence of parental strife, is not likely to "adjust" readily. The reader may remark on the absence from the definition of a time-direction concept. The reasons are twofold: First and most important, time-dating cannot help determine _w_h_e_t_h_e_r there is _a_n_y causal relationship between the two variables. Additionally, time-dating is itself a difficult and uncertain operation; often one cannot say which event preceded which. This is especially the case when human intentions and expectations of future events influence a person to take an action to affect another event; the effect event then precedes the causal events by way of the expectations. Of course one could argue a forward- moving chain of events, but this easily becomes ambiguous. Whether or not this is a _g_o_o_d definition must be considered from several points of view. First of all, it ought to fit common scientific _u_s_a_g_e. The definition given above actually evolved from an inductive study of statements of economic relationships, and it can best be understood in that concrete context. A second test of this definition is that it fit the reader's _i_n_t_u_i_t_i_o_n. One's intuition is closely related to one's experience with usage, of course, but the intuition has some life of its own, too. In other words, I hope that the reader agrees that the definition offered here really stands for what the concept means to the reader. A third test concerns the reliability of the definition. Clearly it is much less reliable than almost any other working definition in science; the need for contextual knowledge in judging causality assures that there will be many more cases about which judges disagree than for most other definitions. But the better question is whether this definition is _m_o_r_e reliable than other methods of classifying situations into "causal" and "non-causal." If this definition is a helpful improvement, then it may lead to others that are even better. One might argue that all of the criteria proposed above lack good definitions themselves, and hence listing them does not improve the situation. Perhaps so. But often one can make a better judgment when one breaks up an overall judgment into parts, e.g., one can usually make a better judgment about the height of a skyscraper if one estimates the number of floors, and then multiplies by a guessed-at height of floor, than if one has to guess the height of the building directly. Similarly if one at least examines a correlation coefficient, self-consciously thinks about the relationship to the body of theory, etc., one may arrive at a better judgment of causality than if one makes a judgment directly. SUMMARY Property definitions of causality are a dead end. Definitions referring to logical properties have failed, and must always fail. What is needed is a set of criteria of causality. This is the set of criteria I propose: A statement shall be called causal (a) if the relationship correlates highly enough to be useful and/or interesting; (b) if it does not require so many side-condition statements as to gut its generality and importance; (c) enough possible "third factor" variables must have been tried to give some assurance that the relationship is not spurious; (d) the relationship is deductively connected into a larger body of theory, or (less satisfactorily) is supported by a set of auxiliary propositions that "explain" the "mechanism" by which the relationship works. This definition is a checklist of criteria. Whether a given relationship meets the criteria sufficiently to be called "causal" is not automatic or perfectly objective, but rather requires judgment and substantive knowledge of the entire context. REFERENCES Barnett, Harold, and Chandler Morse, Scarcity and Growth (Baltimore: Johns Hopkins, 1963). Blalock, Hubert M., Jr., Causal Inferences in Non- experimental Research (Chapel Hill, U. of N. Carolina Press, 1964). Bridgman, P. W., The Logic of Modern Physics (New York: Macmillan, 1927). Burtt, Edwin A. (ed.), The English Philosophers From Bacon to Mill (New York: Modern Library, 1939). Einstein, Albert, Ideas and Opinions (New York: Bonanza Books, 1954). Ellis, Albert and Robert A. Harper, New Guide to Rational Living (Los Angeles: Wilshire (1961; 1975). Glymour, Clark, Richard Scheines, Peter Spirtes, and Kevin Kelly, with foreword by Herbert A. Simon, Discovering Causal Structure: Artificial Intelligence, Philosophy of Science, and Statistical Modeling ((San Diego: Academic Press, 1988). The reference is to an advertisement for this book. Hirschi, Travis, and Hanan C. Selvin, "False Criteria of Causality in Delinquency Research." Social Problems, 13 (Winter 1966). Hume, David, An Enquiry Concerning Human Understanding (Lasalle, Ill: Open Court, 1949). Kant, Immanuel, Critique of Pure Reason (New York: St. Martin's, 1781/1965). Lazarsfeld, Paul F., "Evidence and Inference in Social Research" Daedalus, 87 (1958). Reprinted in May Brodbeck (ed.) Readings in the Philosophy of the Social Sciences (New York: Free Press, 1968). Letwin, William, The Origins of Scientific Economics (Westport: Greenwood, 1975). Mausner, Judith S., and Anita K. Bahn, Epidemiology -- An Introductory Text (Philadelphia: W. B. Saunders, 1974) Orcutt, Guy H., "Toward Partial Re-direction of Econometrics." The Review of Economics and Statistics (August 1952), pp. 211-13. ________, "Actions, Consequences, and Causal Relations." The Review of Economics and Statistics (November 1952), pp. 305- 13. Pearl, Judea, "Embracing Causality in Default Reasoning", Artificial Intelligence, vol 35, 1988, 259-271. Planck, Max, Where is Science Going? (Woodbridge, Conn: Ox Bow Press, 1981). Russell, Bertrand, A History of Western Philosophy (New York: Simon and Schuster, 1945). Simon, Herbert A., Models of Man (New York, Wiley 1957). Simon, Julian L., "The Concept of Causality in Economics," Kyklos, Vol. 23, Fasc. 2, 1970, pp. 226-254. Waldrop, M. Mitchell, "Causality, Structure, and Common Sense", Science, September, 1987, pp. 1297-1299. Wold, Herman O. A., "The Approach of Model Building," in: Herman O. A. Wold (Ed.) Model Building in the Human Sciences (Monaco: Centre International d'Etude des Problemes Humains, 1966). FOOTNOTES 1 Albert Ellis and Robert A. Harper changed their language to E-prime when they revised their excellent self-help book, A New Guide to Rational Living (1961; 1975), and they assure us that it clarified their thinking greatly. My knowledge of E-prime comes from the introduction to their book. 2 Einstein's admiration for Hume is impressive and charming. "If one reads Hume's books, one is amazed that many and sometimes even highly esteemed phillsophers after him have been able to write so much obscure stuff and even find grateful readers for it. Hume has permanently influenced the development of the best of philosophers who came after him." (1954, p.21) 3 I do not suggest that mechanical systems of deciding whether or not a relationship should be called "causal" are impossible or impractical. With a sufficient number of specifications, carefully made, there is no reason why a computer program could not effectively sort into "causal" and "non-causal". For some discussion of this issue, see Pearl (1988). Page # thinking intrp43# 3-3-4d