The role of mechanisms,integration,and unification in science and psychology

Mechanisms, integration, and unification

Wesley Salmon, who is one of the greatest philosophers of science in the world, provides an informative discussion of the difference between explanatory mechanisms versus unifying theories (Salmon, 1998). My favorite of Salmon’s examples includes a boy holding a helium-filled balloon at the end of a string, sitting in a jet that is taking off. Does the balloon move forward or backward? Counter to most people’s common sense, it moves forward and Salmon suggests two explanations, both of which are correct, but one of which is more mechanistic whereas the other is more unifying. The mechanistic explanation is that the rear of the jet pushes on the nearby air molecules, which push on more of them, and so on, until the balloon is forced forward. The unifying explanation invokes Einstein’s general theory of relativity, and specifically the equivalence principle. In brief, the equivalence principle states that acceleration in a particular direction is equivalent to a gravitational force in the opposite direction. Thus, when the jet accelerates forward, this is equivalent to introducing a gravitational force towards the rear. The less dense helium-filled balloon is displaced by air molecules and moves in the opposite direction of the gravitational force induced by the jet’s acceleration, which is forward.

Salmon provides several examples such as the foregoing one to demonstrate the distinction between mechanisms versus unifying theory. Salmon does not argue that either mechanisms or unifying theories are bad; both are nice to have. Rather he simply makes the distinction and points out the particular value of unifying theories. At this point, I will take it as given that unifying theories are important in successful sciences. Does psychology have unifying theories?

To answer this question, it is necessary to make a distinction that, to my knowledge has never been made. Specifically, I wish to distinguish between integration versus unification. To explain the distinction, I will relate the essence of a conversation I had several years ago with a famous social psychologist (name omitted to protect the guilty). I had mentioned what I considered to be a lack of unifying theories in psychology (and social psychology too). The famous social psychologist agreed with me but was not concerned about it. He felt that with all of the impressive findings that social psychologists are accumulating with modern social cognition methodologies, someone would integrate them in a Psychological Review article and then we would be on our way. In the conversation, it was clear that what the famous social psychologist meant by integrate was putting the findings into the same review and making a classification scheme to categorize them. More generally, he thought that what I meant by unify is what he meant by integrate. Given, then, that unify equals integrate and the strong likelihood that someone would eventually integrate social psychology findings in an article, the famous social psychologist arguably was being rational in not joining me in my concern of a lack of unifying theories.

But alas, I disagree that unify equals integrate. It seems to me that integrating is rather easy and that psychologists do it every time they review a body of literature. In contrast, to unify, it is necessary to come up with a deeper insight that causes researchers to perceive the “facts” and mechanisms in a different way (Whewell, 1840). The great scientists of the past did much more than classify findings or mechanisms; they had profound insights that changed our ways of thinking and made us perceive the existing findings and mechanisms as special cases of more basic principles (Trafimow & Rice, 2009b; Whewell, 1840).

The perceptive reader might have noticed that I included Mendeleev’s periodic table as an example of unification in chemistry. As this periodic table certainly classifies the elements and I just seemed to argue against classification, am I undermined by my own example?

Au contraire! Mendeleev’s periodic table was the result of a deep insight and was much more important than mere classification. Consider that the custom in Mendeleev’s time was for chemists to try to organize elements by atomic weights. In contrast, Mendeleev’s insight was to use valence—the tendency for elements to combine with other elements—as the most important principle. (Today we know that this has to do with the fit between the number of electrons and the number needed to complete electron shells.) The consequence was that some of the elements seemed “out of order” as there were cases of lighter elements being where heavier elements should be and vice versa. Worse yet, there were holes in the table and Mendeleev argued that these holes represented elements that had not yet been discovered. Mendeleev used the valence principle to predict what characteristics these elements would be found to have when discovered. The big shots in chemistry at the time thought this was all laughable until the hypothesized elements were discovered and found to have precisely the characteristics that Mendeleev predicted. So we see that Mendeleev did much more than classify. He came up with a deep insight (that valence was the crucial underlying issue), upon which he based the classification, and also made shocking new predictions that eventually were confirmed.

Is unification possible in psychology?

The importance of unification in the successful sciences suggests some possible choices about unification with respect to psychology. One choice is to assume that unification is not possible in psychology. In that case, we must either resign ourselves to psychology never being able to reach the heights that other sciences have achieved or else make the implausible argument that psychology can reach these heights even without unifying theories. The philosopher of science who wishes to traverse this latter road has the unenviable task of explaining how psychology, or any science, can make great achievements without unifying theory. Although I am open to the possibility of such an explanation, it seems unlikely that anyone will succeed in generating it in a plausible manner. Therefore, at least for now, let us take it as given that without unification, sciences are extremely limited in what they can achieve. It follows that if unification is impossible in psychology, then psychology is extremely limited in what it can achieve.

I resist the conclusion that psychology is extremely limited in what it can achieve. This is admittedly, at least in part, because of emotion. But if I am to take my own resistance seriously, I am forced to draw two conclusions. The less problematic conclusion is that it would be desirable for psychologists to discover unifying theories. The more problematic conclusion is that psychologists can discover unifying theories. Obviously, if it is impossible for psychologists to discover unifying theories, then it is pointless for me to recommend that they do it regardless of the desirability of the accomplishment. Thus, I have to make the case that unifying theory in psychology is possible. My acknowledgment that most of psychology is nowhere near the precision of the hard sciences contributes to increasing the difficulty of making this case. I also acknowledge that there has been much debate about unification in psychology, particularly in the pages of Theory & Psychology (e.g., Goertzen, 2008; Henriques, 2008), and no clear winner has emerged. That the debate has been inconclusive further contributes to the difficulty of making my case.

The easiest way to argue that something is possible is to point out that it has been done. Unfortunately, although there have been some attempts at unifying theories in psychology, these attempts have not been successful. Theories in psychology that have seemed to have a unifying character have been disconfirmed by data or have been deemed by psychology researchers to not be falsifiable in principle (but see Trafimow, 2009 for a contrary argument) or to have provided a moving target. A special case of the moving target problem includes cognitive theories with many parameters that can be adjusted to account for practically any conceivable data (see Roberts & Pashler, 2000 for a powerful argument). Finally, as I pointed out in a previous section, the many attempts at integration in psychology do precisely that; they integrate but do not unify. Because of the problems attempts at unifying theories in psychology have encountered thus far, none of them are widely accepted and so none of them can be used to provide the argument I need that unifying theories in psychology are possible.

Given that I cannot take the easy way out by stating that unifying theories in psychology already have been achieved, is there another possibility? Perhaps if I could argue that there is a theory that has gone part of the way, it might be more plausible to believe that we could go the rest of the way with sufficient effort or luck. Toward that end, consider a recent theory by Trafimow & Rice (2008, 2009a) termed potential performance theory (PPT) that already has been applied to disparate areas of psychology. PPT was developed originally in the context of morality but it quickly became apparent that it was more generally applicable. There is insufficient space to describe PPT in detail but it is reasonably easy to describe the assumptions and explanatory power across various domains in psychology.

PPT is a highly mathematical theory that depends on the basic notion of randomness. It has been known at least since 1904 that over the long haul, randomness decreases effect sizes (e.g., Spearman, 1904). But although we normally think of effect sizes as being across groups of people randomly assigned to different conditions of an independent variable or a large group of people who provide correlation coefficients, the notion that randomness decreases effect sizes also can be applied to individual persons. Thus, for any person who performs less than perfectly, there are two classes of underlying causes; these are systematic and random ones. Note the absence of any typical “psychology” words such as attitudes, intelligence, and so on. There is only systematic stuff and random stuff and everything else derives from that distinction.

It is important not to be fooled by words such as “perform” into believing that the theory only applies to areas such as free throw shooting, recall of words, and so on. In fact, practically any behavior can be considered to be “performance.” For example, Trafimow and Rice (2008) pointed out that agreeing on where to eat dinner can be considered to be performance in that agreement is better than disagreement. Although PPT is applied easily to task performance when there are objectively correct and incorrect answers, it also can be applied to task performance when the answers are completely subjective. In fact, in the original article, Trafimow and Rice (2008) applied PPT to understanding the random and systematic processes underlying relative (not absolute) moral decisions.

Consider an area of research that has nothing to do with morality or even social psychology. For decades, researchers interested in the interaction between humans and automation have performed studies as follows (see Rice, Trafimow, & Hunt, 2010 for a review). The researcher seats participants in front of a computer. The participant performs a task, such as indicating whether there is a particular object on the screen, and the computer provides advice at whatever level of probability the researcher has selected. If the automated advice has been selected at a sufficient level of probability, the human’s best strategy is to simply agree with the automated advice each time. For example, when the automated advice is 95% accurate, the combination of a human and the automated aid performs less well than the automated aid by itself. Thus far, this is not surprising. What is surprising is that even when participants are told about the accuracy of the automated aid and are told explicitly that their best strategy is to always agree with the automated aid, they nevertheless fail to do so. So what is going on; are people just plain stupid?

PPT suggests an answer. Perhaps people are not stupid and understand the researcher’s instructions in their entirety. But human behavior is, at least in part, random, and randomness pushes performance in the direction of the chance level; randomness generally reduces performance. To test this, Rice et al. (2010) performed a typical automated aid experiment but used PPT to find the “potential scores” which indicate how participants would have performed had there been no randomness in their behavior. As a surprise to researchers in the area, potential scores met or exceeded the performance of the automated aid by itself, regardless of the level at which the experimenters had set it. Thus, PPT provided the solution to the puzzle. Participants in psychology experiments are not stupid or incapable of understanding the best strategy. Rather, they understand the proper strategy quite well but, being human, they are plagued by randomness. When the randomness is taken into account by PPT, it is clear that participants generally meet or exceed the performance of the automated aid.

PPT also has been applied to meta-attributions pertaining to negative behaviors (Trafimow, Hunt, Rice, & Geels, 2011). It has long been known that some kinds of negative behaviors cause stronger trait attributions than do others. For example, dishonest behaviors cause observers to strongly attribute dishonesty as a personality trait of the person who performs them but unfriendly behaviors do not cause observers to strongly attribute unfriendliness as a personality trait of the person who performs them. But what about meta-attributions where the goal is not to describe the traits of the person who performs a dishonest or unfriendly behavior but, rather, is to guess the attributions other observers would make. Using PPT, Trafimow et al. (2011) were able to show that meta-attributions are actually more accurate when dishonest than when unfriendly behaviors are performed. More important, and contrary to what we saw earlier with automated aids where randomness was responsible for the phenomenon of interest, in the present case a systematic process was responsible for the effect.

Possibly the most spectacular PPT success thus far is in its application to signal detection theory (Trafimow, MacDonald, & Rice, 2012). According to signal detection theory, there is a distribution of random noise in the percepts resulting from either the presence or absence of the signal. When the signal detection task is reasonably difficult, these two distributions overlap. Consequently, the participant sets a criterion level along this perceptual dimension that provides the dividing line for responding that the signal is either present or absent. Given random noise and a criterion level, it becomes clear that there are two potential reasons for imperfect responding: (a) the random noise can cause people to have less than perfect sensitivity to the stimuli and (b) the placement of the criterion level might be biased. Signal detection theory has dominated cognitive and perceptual psychology for over half a century.

Suppose that it were possible to magically remove all random noise. In that case, as long as people use a reasonable decision rule (e.g., always responding that the stimulus is absent is not reasonable), people should perform perfectly. In more formal terms, as long as the decision rule is between the peaks of the two distributions, removing randomness should cause perfect responding. Trafimow, MacDonald, and Rice (2012) used PPT to obtain potential scores that estimate how participants in a signal detection theory paradigm would have performed in the absence of random noise. They only used participants whose decision criteria were between the peaks of the two distributions, thereby removing bias as a relevant factor. If signal detection theory is correct, controlling for randomness should cause potential scores to be near 100%. The data strongly contradicted this conclusion, thereby demonstrating that there is a systematic (non-random) element in the perceptual process that is not accounted for by signal detection theory.

Conclusion

I commenced with the assumption that unification is important in successful sciences. I then distinguished unification from both mechanisms and integration. Mechanisms are good to have but they do not substitute for unification. Also, integration falls well short of unification. Psychologists should unify but, at best, they integrate. Finally, though admitting that we do not have any successful unifying theories in psychology, it seemed to me that we at least have some theories that unify in a small way that give hope that more is possible. PPT is an example of such a theory. By briefly describing some PPT successes in very different psychological contexts, I wished to provide some hope that unification is possible in psychology. My argument is not that psychologists should do PPT research. Rather, it is that if PPT can achieve a small degree of unification, perhaps with some creative insight, psychologists can discover a better theory that can unify in a big way. This is my version of the Holy Grail.