10 Is Research Based on Causal Scientific Principles Possible in the Social Sciences? I.

The use of mathematical and statistical techniques is becoming more and more widespread in the social sciences. It is becoming all the more important, therefore, to demonstrate by a detailed description of these techniques that there are reasons to doubt their applicability in the field of the social sciences.

The aim of this paper is to offer such a demonstration with specific reference to the techniques which allow us to take a given set of data and determine the values of the constants by means of which a variable, interpreted as a dependent variable, can be brought into a law-governed relationship with other variables. It is irrelevant whether this relationship is linear or nonlinear, whether there is one or more than one independent variables, whether—as in time-series analyses—the dependent variables themselves also function (time-shifted) as independent variables, and whether the relationship is recursive or nonrecursive, deterministic or statistical. The critique applies to all techniques, from simple linear regression to the comparatively complex procedure of time-series analysis, insofar as such techniques are used to determine the value of constants (including those with values which vary according to some consistent pattern). It hardly needs pointing out that the use of such techniques is on the increase. In the field of economics, econometrics is steadily establishing its position as the home of these techniques, ¹ despite criticism from the advocates of pure economics. ² In sociology, too, the systematic introduction of econometric techniques is being seen increasingly as a universal panacea, a trend fostered above all by the works of Blalock. ³

II.

To illustrate the following argument let us assume that the values of the constants b₁ and b₂ in the multiple regression equation

Y = a + b₁X₁ + b₂X₂ + e

have been determined on the basis of a given set of data. Y—the dependent variable—is interpreted as a linear function of the independent variables X₁ and X₂ and a magnitude of error e with a mean value of 0.

The b-constants in this equation can be interpreted in either of two ways. They can be given an innocuous, but completely unusual, interpretation, or they can be given their normal interpretation. This second interpretation, however, is no longer innocuous. It involves assumptions which, as will be shown, are inappropriate in the social sciences.

According to the first innocuous interpretation, the partial regression coefficients represent nothing more than a verifiable statement of how best to predict the Y values on the basis of the X₁ and X₂ values (assuming both linearity and that one is dealing with additive effects). The constants are historico-mathematical facts. They have no significance beyond the historical data with respect to which their values were determined.

There can be no objection to this interpretation. It has the consequence, however, that setting up an equation of the kind given above becomes downgraded in importance. One would not be establishing a theoretical principle but merely providing a description of the facts, and what is more, a description of a kind that can be generated at will for any set of data simply by varying one’s assumptions about the types of functions and by enlisting the aid of a computer.

It is improbable that anyone has ever performed a regression analysis intending to achieve only what is implied by this innocuous interpretation. The act of setting up the above equation is normally interpreted as formulating a general hypothesis which can be falsified by new data and which asserts that the relationship between Y, X₁ and X₂, determined by the constants in the equation is universally valid.

III.

The implicit assumptions demanded by this interpretation can be reconstructed by considering the following situation. Using fresh data, an attempt has been made to reproduce the results obtained by analysis of the initial set of data and formalized in the above equation using constants with precisely determined values. Let us assume that the outcome of this attempt is that the multiple linear regression analysis performed for both the Y and the X₁ and X₂ variables of this second set of data produces b constants which diverge significantly from those obtained for the first set of data. According to the innocuous interpretation of such equations, this result would not have any particular consequences. With the first set of data one has established a certain historico-mathematical fact, and with the second another fact. The two are different, and that is all there is to it. According to the normal interpretation, however, the failure to reproduce the results falsifies the hypothesis.

The prerequisite for being able to say “falsify” is the “constancy principle,” the conviction that observable phenomena are in principle determined by causes which are constant and are time-invariant in the way in which they operate, and that in principle contingency plays no part in the way in which causes operate. Only if the constancy principle is assumed to be valid does it follow from any failure to reproduce a result that there is anything wrong with the original hypothesis.

Obviously, the constancy principle is not simply based on experience. As has been known since Hume, there is no observable link connecting events, and even if such a link existed, experience could not show whether it was time-invariant or not. The principle cannot be disproved by experience either, for once it is accepted, any event which appears to disprove it (such as a failure to duplicate a result) can be interpreted from the outset as if experience shows here merely that one particular variable is not the cause of another variable requiring explanation (otherwise the result would have been successfully duplicated). No conclusion can be drawn as to whether any other variable might actually be found which turns out to be time-invariant in the way it operates with respect to the dependent variable in which we are interested. To the extent that experience cannot exclude this possibility, the validity of the constancy principle cannot be disproved.

Although neither derived from nor disprovable by experience, the constancy principle is nevertheless a necessary condition for there being experiences which can be regarded as either confirming or falsifying each other (in contrast to the isolated and unconnected experiences connected with historical facts). ⁴ The failure to duplicate results could be interpreted as falsifying the original hypothesis in accordance with the normal interpretation of the regression equation, and one might consequently feel prompted to explain the differing values of the b constants by asserting that in one sample one or more factors F were implicitly involved in causing Y which were not present in the other sample or did not operate in the same way. Finally, one might feel the need to explain these factors F hypothetically and to incorporate them in the initial hypothesis which assumed only the systematic operation of X₁ and X₂, thus replacing one hypothesis by a new one. But all this is only possible to the extent that one has already assumed the validity of the constancy principle for Y and all the factors causing Y.

IV.

We have asserted that there is a sphere of objective reality which cannot be regarded as determined by laws, and that, therefore, no equation describing its behavior (such as the regression equation given above) can be formulated which can be given a normal interpretation.

Since the validity of the constancy principle cannot be doubted on the evidence of external, sensory experience, it can only be on logical grounds that the principle can correctly be regarded as inapplicable in any particular sphere.

The constancy principle is an operational schema, a method. One does not experience and learn that there are causes which always operate in the same way, rather one establishes that phenomena have particular causes by following a particular type of investigative procedure, by refusing on principle to allow any exceptions (instances of inconstancy), and by being prepared to deal with them by producing a new hypothesis each time one is required. The world by itself is not sufficient to establish the constancy principle. It requires the existence of an active, perceiving subject. For his part, this active subject—the prerequisite for a world determined by causes having constant effects—cannot, for logical reasons, assume the validity of the constancy principle with respect to the states of his knowledge (and to the intentional actions which might draw on that knowledge). In order to guarantee the unconditional validity of the constancy principle, the subject must himself be able to learn. He must start from the assumption that he can assimilate falsifying experiences and replace old hypotheses with new ones. If on the other hand one were to view the state of one’s knowledge as caused, and if (absurdly) one were to treat anything not yet known as being predictable in principle, one would deprive oneself of all possibility of using one’s ability to learn, that is, to form new, previously unknown hypotheses, as a way of maintaining the law-governed nature of that sphere of reality, which is not constituted by knowledge or actions drawing on that knowledge.

The result of this logical analysis of the constancy principle as the principle of a nonempirically based operational schema underlying causal investigation is that the principle can only be valid in that objective sphere which is not constituted by one’s own knowledge or actions manifesting that knowledge. (In this sphere, the question of whether there are law-governed constants on the basis of which it becomes possible to make ex ante predictions is positively determined independently of experience, and empirical factors play a role only in the question of which concrete variables are causally linked to which concrete effect variables and which are not.) In the sphere of knowledge and action, on the other hand, it cannot be valid. (In this sphere, the question of whether or not there are constants is itself empirical in nature and can only be decided for a given variable on the basis of past experience, that is decided ex post.)

V.

For anyone who is capable of learning, his or her knowledge and actions cannot logically be regarded as determined by a complex of causes operating in a constant way (whether statistically or deterministically). There can only be constants in relation to the causes of events where one is dealing with a world of nonlearning objects, or more correctly, where one conceives of an objective sphere of reality as a world of nonlearning objects. One cannot, however, think of oneself as nonlearning. Not only is an intellect functioning in accordance with the constancy principle necessarily a learning intellect (we learn about how objects conceived of as non-learning behave), but the statement “I can learn” also proves to hold true in other respects. It is in principle not falsifiable, for in order to falsify it one would need to be able to learn. And from another point of view, one cannot justifiably argue against the statement since, qua argument, there must be possible replies to it, and as the validity of an argument (as opposed to that of a stimulus) would be independent of the nature of the reply, such possible replies must be regarded as contingent reactions, and therefore it must be possible to learn.

No scientific advance can ever alter the fact that one must regard one’s knowledge and actions as uncaused. One might hold this conception of “freedom” to be an illusion, and from the point of view of a “scientist” with cognitive powers substantially superior to any human, that is, from the point of view of God, such a description may well be correct. However, we are not God, and even if freedom is illusory from His standpoint, for us human beings it is a necessary illusion. ⁵ We cannot predict in advance the future states of our knowledge and the actions manifesting that knowledge on the basis of previous states; we can only reconstruct them after the event. ⁶

VI.

Let us return to the regression equation given at the beginning of the discussion:

Y = a + b₁X₁ + b₂X₂ + e

and bring the argument full circle. Let the dependent variable Y in this equation be any intentional action (an action which attempts by some means to achieve a goal preferred over a starting point and other alternatives, and which in various ways manifests knowledge).

By setting up this equation or by determining the values of the constants for a particular set of data, if we accept the normal interpretation, we are making the following assertion:

There exists a complex of causes operating in a constant fashion which causes Y, and it is possible on the basis of our knowledge of this complex and of the way in which it operates (its function type) to predict the occurrence or nonoccurrence of the intentional act Y (conceived of as a dichotomic 0/1 variable). On the basis of experiences in connection with a particular set of data, the causal relationships explaining Y are described by a provisional hypothesis as in the above equation (with the values of the constants determined with respect to magnitude). New experiences may mean that these concrete assumptions about the causal variable and function types have to be revised. The equation may be replaced by others incorporating different assumptions. Some examples might be:

(1) Y = a + b₁X₁ + b₂X₂ + b₃X₃ + e

(2) Y = a + b₁Z + b₁X₁ + e

(3) Y = a + b₁ log X₁ + b₂X₂ + e

(4) Y = a + b₁X₁ + b₂X₂ + b₃X₁X₂+ e

In equation (1) it is assumed that X₃ is a causal variable as well as X₂. Equation (2) assumes that a variable Z (correlated with X₁) is the cause of Y and not X₁ as originally assumed. Equation (3) no longer assumes a linear relation in respect to the effect of X₁. Lastly, equation (4) assumes an interactive as well as an additive effect with regard to the two variables X₁ and X₂. No matter which equation is substituted for the original one, however, nor whether the original one is repeatedly found to be valid, it remains the case that Y can be predicted, however much one may argue about the precise details of the equation.

Our previous discussion demonstrated that this assertion is untenable. Y, qua intentional act, cannot in principle be predicted. This conclusion follows from the argument which we may summarize here:

(1) I and—as possible opponents in an argument—other people are able to learn. ⁷ (This statement cannot be challenged without implicitly admitting that it is correct. Above all, it must be assumed by anyone undertaking research into causes. To this extent, proposition (1) is valid a priori.)

(2) If it is possible to learn, one cannot know at any given time what one will know at any later time and how one will act on the basis of this knowledge. (If one did know at any given time what one will come to know at some later time, it would be impossible ever to learn anything—but see proposition (1) on this point.)

(3) The assertion that it is possible to predict the future state of one’s own and/or another’s knowledge and the corresponding actions manifesting that knowledge (i.e., to find the variables which can be interpreted as the causes) involves a contradiction. If the subject of a given state of knowledge or of an intentional act can learn, then there are no causes for this; however, if there are causes, then the subject cannot learn—but again see proposition (1).

The putative causes of Y qua intentional act and the putative constants by means of which Y and these causes are brought into a relationship with one another are in truth nothing more significant than variables which have been found in contingent, covariant relationships with Y at particular points in time. It is also a purely contingent historico-mathematical fact (but not a confirmation of a hypothesis!) if these covariant relationships were reproduced exactly or even merely approximately with new data. It is in principle only coincidence that people in the same situation defined by the same set of variables act in the same way (bring the same knowledge to bear). For if one is able to learn it is obviously impossible to predict whether a person will actually learn or not from one point in time to the next. It can only be ascertained post festum, as an already established fact. Ultimately, any change in these kinds of covariant relationships must be seen as a contingent fact (and not as a falsification of a hypothesis!). For if one can learn, then not only is it impossible to predict whether one will actually learn in any particular situation, but it is equally impossible to predict what, if anything at all, one will learn.