THE QUANTITATIVE ANALYSIS OF SOCIAL PROBLEMS Edited by EDWARD R. TUFTE Department of Politics Princeton University ATTITUDES AND NON-ATTITUDES: CONTINUATION OF A DIALOGUE PHILIP E. CONVERSE Some years ago Carl Hovland (1959) undertook a systematic comparison of conclusions which had been reached concerning the modification of attitudes through communication within two broad research traditions: the experiment and the sample survey. Despite common substantive interests, these two traditions had remained rather insulated from one another, and as Hovland noted, their separate efforts had produced results easily taken as contradictory. The purpose of Hovland's essay was twofold. First, he wished to point out that these apparently divergent views about the ease of persuading to attitude change could be readily reconciled by proper understanding of the differences in the two types of research designs and the backgrounds of the investigators involved in each. To the best of our knowledge, members of both traditions felt that the reconciliation was well handled. Secondly, however, Hovland seemed concerned in a more general sense with a need to bridge the gap between experiment and survey, and in effect called for a more vigorous dialogue between the two traditions. This paper is written in an effort to continue the dialogue. It is not our intention here to retrace Hovland's steps in any detail, although a number of things to be said have relevance for his argument. Nor, for that matter, do we wish to rehearse the differential powers and shortcomings of the two methods, for we would subscribe with little amendment to most of the conclusions long since reached by parties interested in the subject. It seems that divergences in outlook are likely to arise between the two traditions if for no other reason than the fact that the experimentalist is able to study what can happen in situational configurations which he creates and controls, while survey analysts pay a more passive attention to what does happen in an actuarial sense, as a matter of relative empirical frequency. Thus, in the case Hovland discussed, it appears true that when audiences\ are exposed to certain persuasive communications on certain kinds of issues under certain further experimental conditions, considerable attitude change can be demonstrated. From an actuarial point of view, however, these conditions occur in nature infrequently: few people expose themselves to potentially contrary messages even though a torrent of such messages may be sent, and various other aspects of the experimental condition go unfulfilled as well. Revised version of a paper read at the Seventeenth International Congress of Psychology in Washington, D.G., August, 1963. Reproduced by permission of the author. 168 Perhaps the clearest function which sample survey results may fulfill in any dialogue, then, is to remind the experimentalist of the actuarial mainstream, pointing out sources of critical variation in "natural" attitude-change processes to which he may have become insensitive. It is in this spirit that our remarks are made. We will begin by a presentation of results from one analysis of sample-survey data on attitude change. We choose this analysis because it seems to have implications not only for some of the methodological practices common in experimental work on attitude change, but also for the way in which the attitude continuum is to be conceptualized. Carried a step or two further, the results suggest hypotheses which might be worthy of experimental test within the areas described by balance, congruity or dissonance theories of attitude change. Statistical Properties of Certain-"Naturalistic" Attitude Changes The data to be reported are drawn from a sequence of panel studies con-/ ducted on a national cross-section of the adult population of the United States over a four-year period from 1956 to 1960. The immediate occasions for the sur\ veys were the three national elections of 1956, 1958 and 1960. In principle, respondents were interviewed at five different points in time over this period: before and after the election in 1956; after the election in 1958; and once again, before and after the 1960 election. In practice, slightly less than 70 per cent of the original 1956 pool of respondents who were still alive and in possession of their faculties were successfully reinterviewed in the 1960 waves of the panel. However, analyses across a wide variety of social and attitudinal characteristics suggested that the 1960 survivors were a remarkably unbiased subset of the original pool of respondents. The panel aspects of the study sequence were fully utilized, in the sense that the interview schedules applied at the various points in time contained direct repetitions of a wide array of items bearing on attitudes and social situations.. Here we shall focus on some of the time-change aspects of the attitudinal data. The simplest property to assess from these attitude measurements has to do with the correlation of measurements of the same items over time. Empirically, the range in variation of such test-retest correlations on attitude measurements computed after a two-year interval was very wide, covering a space in the correlation continuum from roughly .10 to .80.1 Furthermore, the data made clear that these turnover correlations were quite stable for specific items over comparable periods of time: if the correlation of item A between tl and f2 turned out to be .4, for example, it was demonstrable that the same correlation computed for the same item between the interval t2 and t3 would rarely be much more than .03 or .04 from .40 as well. In short, then, varying items had very differential but very stable rates of attitude turnover associated with them over the two-year test-retest span. Cursory inspection of differences in turnover rates attached to varying items was sufficient in many cases to indicate why the turnover rate fell in the levels 170 experimental and quasi-experimental studies attitudes and non-attitudes 171 where it was found. Thus, for example, items which showed test-retest correlations of less than .20 typically had to do with the particular election as an object of estimation. Since three elections were covered, it is quite obvious that while the questionnaire items were identical, the objects of reference—the different elections—were in many respects different objects, so that almost complete turnover of opinion could hardly be taken as surprising. At the very high end of the turnover range, measures of generalized affect toward the major political parties approached a two-year test-retest correlation o"f .80. Here, of course, the objects are familiar ones evoking strong affect from many people, and are characterized by a reasonable amount of temporal stability in their attributes as perceived by the public. Included among the attitude items, however, was a battery of eight questions directed at the principal issues of public policy which were being debated during the period of the study. These included attitudinal items of a familiar Likert type on matters of civil rights, social welfare legislation, the relation of government to free enterprise, and problems of foreign policy such as aid to neutral countries and the like. We shall focus our attention on this set of attitudinal items for two reasons. First, they resemble closely the type of item to which the experimentalist often gravitates in his attitude studies. Secondly, the empirical properties of the items were quite striking. For while the meaning-substance of the items seemed to have changed very little during the period of the study, and while the items were chosen specifically to capture the cleavages in basic psychological commitments in public politics of the period, the test-retest correlations were all within the lower half of the empirical range, running roughly from .23 to .46 within both of the two-year time spans available for inspection. There are at least two possible reactions to such low coefficients. One is to imagine that public opinion on these items must have been in a high state of flux during this period, with a very considerable evolution of attitudes in the wake of changing national events. There were, however, several empirical problems which such an interpretation would encounter. In the first place, the marginal attitude distributions for the various time points were remarkably similar despite high rates of turnover within the tables.2 It would seem that if national events were exerting systematic forces on opinion in a manner which would produce meaningful evolution of public attitudes, the distributions of opinion should progress in one direction or another over time, rather than remain relatively stable, with almost all of the individual change in one direction being counterbalanced by an equal amount of individual attitude change in the opposing direction. Secondly, the items had been chosen to avoid superficial attitudes toward short-term events, in an attempt to plumb more basic and stable dispositions toward questions of public policy. Hence no very large measure of meaningful; change was to be expected. Indeed, examination of the differential rates of turnover within the battery of issue items suggested that the more basic and ideological i the issue dimension was, and the more remote its referents were from day-today change in national events, the higher the turnover of opinion. Thus, for example, the most unstable issue of all was one having to do with whether or not the federal government should "leave things like electric power and housing for private businessmen to handle." Somehow it seemed implausible that large proportions of the American population between 1956 and 1958, or between 1958 and 1960, had shifted their beliefs from support of creeping socialism to a defense of free enterprise, and that a correspondingly large proportion had moved in the opposite direction, forsaking free enterprise for advocacy of further federal incursions into the private sector. This being so, the more reasonable reaction to the low coefficients is simply that we were doing a very poor job of tapping the attitudinal dimensions at which we originally aimed, and that our results, viewed now as test-retest correlations in the strict sense, give witness to an incredible degree of measurement unreliability. Hence we should not talk of results at all until we go back and develop better measuring instruments. There is somewhat more internal evidence supporting this view than was the case for the evolution-of-opinion interpretation. One of these items of evidence is particularly relevant to our argument. Originally we had assumed that some of. these items, vis-a-vis some people in the population, would have very little meaning. Hence we had taken elaborate precautions to remove such people from any sense of obligation to respond to items which generated no affect for them. Thus, for example, the battery of issue items was prefaced by a statement which pointed out, among other things, that "different things are important to different people, so we don't expect everyone to have an opinion about all of these (things)." Furthermore, as each item in the battery was read, the respondent was explicitly asked whether or not he had an opinion on the matter and only if he said that he did was he further asked what his opinion was. These precautions may be fruitfully compared with many common attitude-measurement situations in which experimental instructions entreat the subject to express some kind of attitude other than the indifference point even though he may find it hard to do so. After all, "don't know's," equivocations and other forms of missing data are exasperating in analysis and are to be avoided at all cost. However, we felt that it was more important to deal with missing data than with measures laden with "noise." Our screening procedures were successful enough that variously acrossthe ' set of attitude items, anywhere from a handful to 35 per cent of the sample confessed that it had no genuine opinion on the matter under consideration. Ironically, however, the low test-retest correlations which we have cited for these items were computed for the subset of people who did lay claim to some opinion; the many no-opinion people were set aside. Furthermore, it was readily discovered that there was almost a perfect correlation between the issue items ordered according to proportions of people who said they had no opinion, and the ordering produced across the same items in terms of response stability. 172 experimental and quasi-experimental studies attitudes and non-attitudes 173 These pieces of evidence, of course, lead toward a strong suspicion of item unreliability, although it is rather distressing to learn that such unreliability remains after all of our precautions to avoid the measurement of non-existent attitudes. However, let us not be too hasty in our judgment, for the heart of the analysis which we wish to report turns out to throw an odd and unexpected light on the whole question of reliability. The most revealing statistical property of these attitude-change data emerges when we consider not simply the correlations between the same attitudes over two-year spans, but also the correlation for each attitude between- the initial and terminal interviews, a span of four years. For we discover that these /i-to-^ correlations tend to be just about the same magnitude as the t^to-t^ correlations, or the t2-Xo-t3 correlations. That is, surprising though it may be, one could predict 1960 attitudes on most of these issue items fully as well with a knowledge of individual attitudes in 1956 alone as one could with a knowledge of the more proximal 1958 responses. Furthermore, the tendency toward parity df the three correlations is clearest among the issue items with greatest turnover; iamong the more nearly stable items, the four-year correlation tends to be slightly .lower than the two-year correlations, a pattern which is of course much closer'to our intuitive expectations. At this point we will find it useful to shift some of our weight from the vocabulary of correlation to the vocabulary of Markov chains, a body of mathematical theory useful in "treating stochastic change processes. Within this vocabulary, the tables representing attitudes at different points in time can be converted into proportions by rows, and considered as empirical matrices of transition probabilities. The similarity of the pair of two-year tables for most of these issue items provides some presumptive evidence for the assumption that whatever else may be said, these matrices as reckoned in summary fashion over the population can usually be considered constant in their probabilities over time. Proceeding with this assumption, it can readily be demonstrated that the empirical parity of the three time-correlations could not occur if it were true that the issue responses of the total population were properly describable by a single constant matrix of transition probabilities through the two steps from t± to t3. For if such a condition held, then the four-year time-correlations would necessarily take values substantially below those of the two-year correlations, the differences ranging upward from .10 to .20 or more under common conditions. Hence it is clear that no single transition matrix constant over the two-year stages can account for the observed response behaviors. In other words, if we maintain the assumption of constant matrices over time, then we are forced to conclude that the time paths of response can only be treated as arising from some mixture of two or more transition matrices, and not one alone. Let us imagine what these matrices might'be. At first glance it is apparent that there is a whole range of matrices which, while mathematically possible, are entirely implausible in the empirical situation at hand, and which we can therefore rule out of serious consideration as describing the behavior of any member of the population. These are the subset of possible matrices which, if cast as second-order determinants, have negative values, or which in any form would generate negative time-correlations. In content terms, a genuine negative time-correlation would mean that respondents who had agreed with an item in 1956 must have gone out of their way to disagree with that item in 1958, and then remember to agree with it again in 1960. This would seem to be highly unlikely behavior, even if it were conceivable that respondents could remember specific responses they had made to an interview two years earlier; and since this seems quite inconceivable in itself, it seems fair to rule out such matrices as impossible. Once this constraint is added to the picture, however, we begin to have some strong leverage on our data. In particular, we can now say that if the total population were to be described in terms of a mixture of two transition matrices, then there are only two which in combination would generate our phenomenon j of three equal time-correlations. These two matrices are (1) the identity matrix, ; with probabilities of 1.0 in the major diagonals; and (2) what we might call a "random" matrix, of equiprobable responses. And quite naturally, the magnitude of the three equal time-correlations would be a simple function of the relative prevalence of these two matrices in the population. This model for response behavior we might call a "black-and-white" model, for it posits a very stringent division of the population into two sharply contrasting subsets. In content terms, one portion of the population would be perfectly stable in its responses over time, while the other portion would be given to response time-paths which in a strict statistical sense were random. On the face of it, this would seem to be an unlikely descriptive model. Furthermore, while most of the assumptions we have made about the nature of the data which have led us to this model seem quite palatable, our final assumption was that only two transition matrices were necessary to describe the behavior of the population. This constraint seems somewhat gratuitous: we would much sooner suppose that a continuous shading of different matrices were represented in the population. Happily, in this instance, the data permit a test of the goodness of fit of the simple black-and-white model. There are several statistical manipulations which would accomplish the same general end. We will describe one which is as simple to follow intuitively as any of them. Let us assume that the black-and-white model comprises an accurate and exhaustive account of the attitude responses generated over time. Then jt follows that any individuals who change from one side of an attitude scale (say, the "agree" side) to the other ("disagree") between t1 and t2 must of necessity belong to that portion of the population whose responses are random. They are a "pure" random group. However, they do not exhaust the total set of subjects following random response paths. For between tx and r2 certain subjects following random response paths would give two consecutive responses on the same side of the issue by chance alone. Fortunately, the proportion of such subjects in the cells . /mi experimental and quasi-experimental studies attitudes and non-attitudes 175 of stable /i-/2 attitudes is calculable, given the numbers of subjects who fall in the minor diagonals, or change cells. This does not mean, of course, that we can isolate the "random" subjects from those whose response paths are stable in a meaningful sense. But we can tell in considerable precision how "polluted" with such people the set of apparently stable people is. Therefore, we can define two sets of people on the basis of the pattern of tx and t2 responses. One is the pure random set. The other is a mixture, in known proportions, of the perfectly stable and the perfectly random respondents. If the black-and-white model is a proper description of response behavior in the population, then, we can predict with similar precision the nature of the relationship between l2 and t3 responses for each subset. The purely random people should show a t2-t3 correlation between responses of .00. The polluted subset should show a t2-t3 correlation greater than the total-population time-correlations, but falling well short of unity because of the remaining admixture of random respondents. Just how far short of unity this second correlation would fall can of course be readily calculated on the basis of knowledge of relative proportions of random and stable respondents. As an initial test of the model we chose data arising from the item for which the black-and-white model seemed on a priori grounds to be least inc'bnceivable. That is, we sought the item from our battery for which it seemed likely that genuine attitudes would be most deeply ingrained and hence immutable over time, and yet one which would be basically "ideological" enough that the issues posed might be truly beyond the ken of substantial numbers of people in a cross-section population. The item described above concerning the relative roles of government and private business matched these specifications in excellent fashion. It was not entirely coincidental that this was the item on which the largest proportion of respondents had indicated that they had no opinion, arid the item which had shown the highest response instability of any in the battery among those who did claim opinions. Respondents were then sorted into two groups in the manner described above, on the basis of the pattern of (1 and r2 responses. The predictions were that the pure random group would show a fa-r3 correlation of .00', and that the second group would show a correlation of .47, if the black-and-white model were indeed an appropriate description of the underlying response process. The results of the primary test showed t2-ts correlation values of .004 for the first group, and of .489 for the second group. In other words, the time data'generated by this issue item fit the predictions of the black-and-white model with remarkable precision.3 We then proceeded to test the same model for a number of the other issue items, although in varying degree they had face content and zero-order time-correlations which suggested in advance that they could be expected to depart from the model somewhat. The test showed that indeed they did, although the bifurcation of t2-t3 correlations between the two test groups remains quite extreme. Thus, for example, the key correlation for the putative ""random" subset departs from .00, but rarely rises above a figure of about +.09. Hence we might conclude that there remains a "near-fit" for the other items as well. This fact of "near-fit" may be conceptually more important than meets the eye at first glance. For there are widespread assumptions about processes of attitude measurement, as well as a few formal models, which presuppose some underlying continuum of latent response probabilities vis-a-vis any single attitude item. Thus, for example, in a heterogeneous population, one might expect that ideally one could isolate individuals for whom response probabilities towards taking one of two positions on a given item range continuously between .5 and 1.0. What is intriguing about the black-and-white model, along with the "real" data which fit it, is the demonstration of an absence of such continuity, with two maximally discontinuous classes (or three classes, if one distinguishes between the perfectly stable "pro" class and the equally stable "anti" class). Once data depart significantly from this simple model, the number and variety of models which could conceivably account for the data become large indeed, and the discriminatory power of our mathematical deductions evaporates accordingly. In these instances it is very easy to fall back on vaguer notions of latent-response probability continua. Perhaps this is appropriate. However, the fact that one set of these data fits the black-and-white model very well, and the other sets of conceptually comparable items only miss a fit with the model in modest degree, suggests that we should not abandon the black-and-white model completely in imagining the processes which underlie the responses to other items in the battery. In other words, it would seem likely that were the truth of the matter isolable, we would discover that a very large proportion of the responses to the other items in the battery could best be understood in terms of two sharply discontinuous classes of respondents, the stable and the random. What is new in these other items, and what leads further data to diverge somewhat from the black-and-white model, is the presence of some few people who are undergoing a meaningful evolution of attitudes on the issue in question. The crucial fact, from the point of view of our argument, is the strong likelihood that even the attitude items straying somewhat from the expectations of the black-and-white model are clouded by large numbers of purely random responses. ^ What psychological interpretation is to be placed on such random responses? It seems to us most simple to imagine that they came from people with no real attitudes on the matter in question, but who for some reason felt obliged to try a response to the item despite our generous and repeated invitation to disavow any opinion where none was felt. In this vein, it may be useful to analyze exactly what the objects were which we were asking our respondents to evaluate. For example, the key black-and-white item on government and private enterprise posed as an object of potential attitude not just the federal government or private business, but rather a type of relation between the two. Furthermore, the manifest content failed to make clear which of the two parties to this relation would feel helped or hurt by it. This means that respondents who may • ,\> cApeniiit-nrai ana quasi-experimental studies attitudes and non-attitudes 177 have had some prior feeling of generalized affect toward private business or toward the federal government (or both) could not respond stably, for lack of this further information which the question presupposes. The experimentalist may find this an incredible observation, for the information presupposed—that private business does not generally want further governmental expansions into its economic sector—seems almost a ubiquitous piece of the "common culture." However, the survey analyst rapidly comes to recognize that presupposition of any information about objects which lie beyond the daily ken of the subjects tested will miss the mark for substantial numbers of people in a heterogeneous population. Indeed, in the case of the attitude item fitting the black-and-white model, it can be calculated from the data that something less than 20 per cent; of the total sample fell into the category of real and stable attitudes on the item. The remaining 80 per cent represented confessions of "no opinion" or statistically random responses. Unfortunately, it was a minority within this 80 per cent which took advantage of our invitation not to bother fabricating an opinion. When attitudes are asked for in such a setting, people are remarkably obliging. Relevance of Findings for Experimental Studies At this point the experimentalist may well ask, "Of what interest is all this to me? I want to understand the implications of attitudes once people have psychological states worthy of the name. I am a student of attitudes, and not a student of non-attitudes." This is, of course, an impeccable position. It presumes, however, that what he typically studies are indeed attitudes, and not non-attitudes concealed with hastily-fabricated affective judgments, as was the case with a full plurality of our test population. A reading of the experimental literature over the years suggests to me, however, a remarkable insensitivity to this possibility. And when I keep this possibility in mind in reading any given study, I often end up with an interpretation of results that is quite oblique to the interpretation offered by the investigator. In other words, there seems to be food for thought here even for investigators who wish to limit their efforts td the study of genuinely-formed attitudes, but who do little to protect themselves against the measurement of non-attitudes by mistake. Therefore it is appropriate to explore the implications of these survey data several steps farther. The most obvious implication has to do with instrument reliability. Once we granted that the low-time correlations for these items were not likely to be accounted for by "true" attitude change, but rather should be seen asi test-retest coefficients of reliability, we were prepared to send the whole instrument back to the shop for repairs, since a reliability coefficient of .3 is disastrously low. Yet the fit with the black-and-white model suggests that where people actually had attitudes, the single item could scarcely be further perfected, for on a trial as stringent as a two-year test-retest, the reliability coefficient was indistinguishable from perfection. From this point of view, what needs repair is not the item but the population. Less facetiously, the moral is clear: where measurement reliability is at issue, the measurement of non-existent states is very unrewarding. And while the classical view of these matters took "reliability" to be a property (or number) attached to the measuring instrument, we could not have a more dramatic example of the fact that reliability in our field of inquiry is instead a joint prop- -erty of the instrument and the object being measured. Other aspects of psychological measurement may deserve review in this light. Speaking from personal experience, I would hypothesize that such a phenomenon as test fatigue is itself a direct consequence of pressures, felt by the subject to search for faint or non-existent bits of affect to fulfill the requirements of the attitude questionnaire. In those rare cases where an attitude item or battery dovetails nicely with thoughts or feelings I have experienced on my own with any strength or clarity before, even such an impersonal process as marking a questionnaire offers the reward of pleasant catharsis. Such pleasure seems somewhat infrequent, however, and the hunt-and-fabricate feeling is fully as familiar.* One outcome of such harassment is fatigue; another is a more or less conscious recourse to some response set touched off more by question form than question content. Underlying the overestimate of who has attitudes about what, perhaps, is the common view held by many social psychologists of the individual as a vibrant bundle of attitudes. Nothing we have said need call this view into question in the least respect: it is certainly an heuristic viewpoint and undoubtedly a faithful one as well. However, it is all too easy to assume from such a view that mere selection of a "familiar" object or controversy as a point of attitude measurement must evoke true attitudes in all or almost all of a test population. There is, of course, a very wide logical leap from the first of these propositions to the second. Possible objects of attitudes are infinite, and a person can be seen as a vibrant bundle of attitudes without any assurance that his attitudes extend to more than a very tiny subset of such objects. Phenomenological differences in information and attention almost ensure the contrary: it may well be difficult to find objects in most domains which will not be matters of non-attitude for many members of the test population. In sum, then, there is a very real sense in which attitudes take practice— practice which is genuine in the sense of having been powered by own psychic energy aside from the kind of transient situation created by the experimentalist or the survey interviewer. Where such practice has not occurred, the state to be measured is non-existent. The measurement of non-existent states gives maximally unreliable results. If the subject himself were helpful to the investigator in refusing to report very ad hoc feelings as "attitudes," then the problems would be greatly diminished. This does not occur, however. Hartley (1946) years ago collected a full set of ethnic attitudes toward groups that did not exist. We made a great effort to encourage holders of non-attitudes to bypass such items. Some accepted the invitation, but the majority did not. Whatever our intentions, the attitude questionnaire is approached as though it were an intelligence test, with the "don't know" and "can't decide" confessions of mental incapacity. 178 experimental and quasi-experimental studies attitudes and non-attitudes 179 It is true that on many grounds the survey analyst is more exposed to the dangers of studying non-attitudes than is the experimentalist. That is, despite great variety in experimental procedures, there are some rather typical aspects of attitude-change studies which intentionally or accidentally provide protection. One example of accidental protection is the use of college students as subjects. Non-attitudes on a wide range of matters which seem "common culture" to the investigator are an inevitable consequence of information impoverishment among the less well-educated strata of heterogeneous populations. While the professor is likely to be impressed that college sophomores are not very well-informed either, they remain, relative to the total population, a fairly alert group. Hence, if a specific attitude item were to show an 80 per cent non-attitude rate in a heterogeneous population, the nature of the population the experimentalist uses might well reduce the rate on the same item to something like 30-50 per cent, among college sophomores. More intentional steps for protection include such things as the multiple-item battery and the choice of attitude-objects which are "close to home" for all of the subjects, and hence far more likely to have become the object of genuine affect for any population member. Certainly one need not worry greatly about non-attitudes in the sense in which we use the term here when the object of evaluation is "mother" and the dimensions of evaluation so common (as in the more typical uses of the semantic differential) that they are part of anybody's common judgmental vocabulary. However, in attitude-change experiments perse there are pressures away from the havens of protection which either multiple-item batteries or the choice of very homely objects can afford. In the classic format of the "before" measurement, the persuasive message and the "after" measurement, the message itself must be of some limited content scope, and there are in turn only a limited number of attitude items which can be imagined within such a scope. Hence the item base of very many attitude-change measures is extremely limited, if indeed it exceeds one item. Similarly, the necessity of dealing with some "common" object of orientation to which the persuasive message can effectively pertain makes the use of objects which are phenotypically knit into the lives of test subjects ("mother," "my work," "my professor," etc.) rather awkward. Add to these difficulties the need to deal in objects which are controversial, along with an understandable desire to treat "socially significant" attitudes, and the common result is not only the use of a narrow item base, but attitude-objects of political, intellectual or social interest which tend to lie beyond what is very salient for many of the test subjects. It is, of course, in such areas that non-attitudes abound, even for college sophomores. If it can be granted that many attitude studies have measured an abundant number of non-attitudes on the supposition that they were attitudes, what difference has it really made? Do not the holders of non-attitudes, particularly with multiple-item measures, tend to gravitate toward the zero point of "indifference," where they belong? And even though they may abound, do they do anything more than add "noise" to the results, attenuating them rather than biasing them? First, let us take the question of the meaning of the attitude continuum,— for this is what is at stake when we say that non-attitudes should fall at the "zero-— point." It might be mentioned in passing that a frightening number of our ran- * dom respondents were capable of giving "strong agree" or "strong disagree" responses, probably under pressure to introduce some variety into their strengths as well as to give some attitude. With multiple items, however, it is indeed likely that inconsistencies (which will be frequent in the expression of non-attitudes) will drive the respondent toward the middle of any summary attitude measure. The question becomes, then, whether or not that is really where we want him. Now there are several ways of imputing meaning to zones on this basic attitude continuum. It is generally assumed that an extreme location has something to do with either an intense or at least a univocal attitude. One way of visualizing the matter is to reduce the molar concept of a generalized attitude toward a complex attitude-object into its molecular parts—the set of affective reactions which the individual holds toward all of the component properties of the object which he perceives. In such a reading, the proper location of the individual on the attitude continuum with respect to the molar object is some algebraic summation of ratio of component valences, weighted in one fashion or another. While we know remarkably little about what combining rules may pertain, we do know that we can expect some generalization of affect or "strain toward symmetry" (Newcomb, 1953) among these molecular valences, such that the molar attitude toward the object tends to be somewhat more univocal than the same affective components might be if the properties of reference resided in a scatter of dissociated objects. However, we also know that this is no more than a trend, and one which reality very often forestalls. People do maintain mixed I attitudes toward very many objects, especially those not lending themselves 1 to any easy dissociation. Indeed, the instance of the perfectly-mixed reaction occurs often enough that it has attracted a term of its own, which is ambivalence, an uncomfortable state but not a non-existent one. Where should ambivalent people" be located on the attitude continuum? Most certainly, it would seem, they deserve a position on the middle, at the zero-point. However, we now see that things are becoming somewhat crowded here, for we have already created something like Sherif's zone of indifference, and located our non-attitudes in the middle of this zone. In some instances, perhaps, the consequences of this overcrowding are few. But there are many instances of experimental treatment of attitude change in which it may matter a great deal. Take as a concrete example the implication of the Osgood congruity model (Osgood et al, 1955 and 1957) that attitudes near the zero point are moved greater distances by persuasive information than are attitudes initially located towards the extremes of the continuum, a matter which has received experimental confirmation in the typical format of the attitude change study. Perhaps such wider movement under the experimental conditions would be equally true of intense but ambivalent attitudes as well as of non-attitudes. Intuitively, however, we would expect that there would be rather drastic differences among holders of the two *u