A Paradox of Experimentation and Other Anadromous Thoughts

Was Mozart a prodigy because he was a master later?

Posted Jan 21, 2012

The law of causality, I believe, like much that passes muster among philosophers, is a relic of a bygone age, surviving, like the monarchy, only because it is erroneously supposed to do no harm.
~ Bertrand Russell

A year ago, I argued sharply against the hypothesis of retrocausation, or the idea that later events can influence earlier events. The contexts for those posts (e.g., this one) was Daryl Bem's (2011) research, which reversed the sequence of events in the protocols of classic social-cognition experiments and still yielded statistically significant results. For example, it is well known that people come to like stimuli (e.g., images of Chinese ideographs) once they have become familiar with them through repeated exposure. Bem replicated this effect but also found that participants already liked those stimuli to which they were repeatedly exposed later. To disable forward causation, Bem used a random-event generator that assigned participants to either the repeated-exposure condition or to a no-exposure control condition. Bem's crucial claim was that the event generator was truly random, although its output was correlated with participants' responses to the stimuli that were shown before and after. To Bem, the only true causal path went from the responses to the later stimuli to the responses of the early stimuli, bypassing the random-event generator. This causal model seems incoherent to me because it selectively ascribes causal force to some correlations and not to others.

The Bem affair will either expire with a whimper or it will be remembered as a scientific break-through that changed the way we view the Universe. Science is conservative and demands extraordinary evidence for extraordinary claims. Provocative results must be independently replicated (witness the current struggle over the finding that some particles can travel faster than light). I recently came across a piece of work that looks similar in design to Bem's, and which has not been published as it awaits independent replication itself. This work was conducted by Michael Franklin and Jonathan Schooler (FS) at the University of California at Santa Barbara. On the one hand, the need for further replication is understandable; on the other hand, this work is already a conceptual replication of Bem, and, setting time aside, Bem's research is a preplication of Franklin's work.

In FS's experiment, participants first did a shape identification task. There were four different, though somewhat similar-looking irregular shapes (see image). In Phase 1 of the experiment, these shapes appeared on the screen one at a time and in unpredictable order. The participant had to press a key (A, B, C, D) to identify the shape. Using accuracy and latency measures, a performance score was then calculated and keyed to the outcome of the spin of a roulette wheel, such that RED = shape A and BLACK = shape B. After the roulette wheel was spun, the participants were assigned to gain more practice with either shape A or with shape B in Phase 2 of the experiment (see image).
Over 4 replications of their experiment, involving a total of 579 bets (on RED or BLACK), FS found a total weighted hit rate of 59%, which is handily significant when compared to the baseline of 50%, and, as an effect size, it is about 3 times the size of the effect Bem found in his research.

So what's going on?

According to FS,"because the ball landed on RED, subjects were assigned to practice with Shape A in phase 2. It is this future practice with Shape A, in phase 2, that retrocausally influences performance in phase 1. Therefore, by measuring the subject's performance in phase 1 we can predict the future practice shape (Shape A). The successful prediction of the roulette outcome is thus a consequence of the shape being mapped onto a specific roulette outcome (RED)" (p. 6 of their report).

Let's review. There are three events: Performance in Phase 1, the outcome of the "random" spin, and practice/performance in Phase 2. The latter two we can collapse into one because by design there is a perfect mapping of RED on A and BLACK on B. We now observe a correlation between Phase 1 (performing better on A or B) and the subsequent events (RED/A vs. BLACK/B). Assuming that all marginal probabilities are .5, the correlation is .72, which is huge.

If true, what are the implications of this finding?

One possibility is that retrocausality will be added to ordinary, progressive causality. According to this view, time flows both forward and backward, and therefore, causality can be conceived of as moving forward and backward. I don't like this idea because I cannot imagine a flow in one direction being both true and co-existent with its own opposite.

Another possibility is that there is no flow at all. Time, Kant notwithstanding, is an illusion. Why we might have this illusion is another question; let's just say we have it. Then, the idea of precedence, i.e., the demand that a cause precede the effect is moot. As the idea of precedence is essential to all ordinary conceptions of causality, the term retrocausality is a misnomer. What we would have is contingency, not causality. In an essay on determinism, Bertrand Russell (1913) noted that if it is true that early events entail late events, then it is also true that late events entail early events. To a principled determinist, the notion of causality dissolves (but also see Russell, 1948, for some revisions to this view).

Which brings me to a paradox of experimentation.

We conventionally assume that we conduct experiments by manipulating independent variables and observing the dependent variables. The independent variables are under our control. The causal chain of events begins with our decision to act on one part of our environment. Our action thereby becomes an intervention. For example, we water a plant with the intention to make it grow. Usually, that works and we conclude that water causes growth. With full bi-directional determinism, however, we say that water is contingent on growth just as much as growth is contingent on water. It follows that our decision to use water and not nothing is contingent on the act of providing water just as much as the act of providing water is contingent on our decision to use water and not nothing. More disturbingly, our decision to use water is also contingent on events that preceded it (e.g., our study of plants and our attitude toward them), which implies that past events that led to our intention to do the experiment are contingent on our present intention just as much as that intention is contingent on the past events.

With full determinism, we must surrender the idea that we can intervene in nature (because we are nature), and hence the idea that we can design experiments that are independent of everything else that is going on. From the ordinary textbook perspective, designing an experiment is an act of human agency. We are, presumably, free to run an experiment or not, to design it in one way and not the other. From the perspective of bi-directional determinism, however, the way we design an experiment and the results it will yield if we run it are as contingent on the findings just as the findings are contingent on the experiment.

The irony is that determinists like myself, who argue against the notion of free will, still cling to the idea of free human agency and the privileges of forward causality when we design experiments whose objective it is to reveal deterministic relationships. Perhaps we should get over it and accept the pure, unbiased notion of determinism, according to which the Logos or Dao of the Universe has no direction, only a web of relationships.

Those who wish to participate in an online version of the Franklin and Schooler experiment may go to www.michaelsfranklin.com, where they can also find contact information. A copy of the report on the experiment can be found here. In order to keep track of where subjects are coming from, after the experiment is initiated there is a screen where you can enter an experiment code -- Please enter the code 'PsychToday'.'

Now, an afterthought comes to mind. What might be the effect of having read this essay or the FS report before taking the experiment? Participants now know the hypothesis and they might act strategically. Suppose half of them deliberately perform better with shape A than with shape B, e.g., by slowing down their responses on the latter. This looks like an intervention. Will this nullify the contingency of their own performance in the here and now on their performance in the here and later? But then again, we just abandoned our belief in intervention.

Sometimes I don't know what the Universe wants.

Bem, D. J. (2011). Feeling the future: Experimental evidence for anomalous retroactive influences on cognition and affect. Journal of Personality and Social Psychology, 100, 407-425.

Russell, B. (1913). On the notion of cause. Proceedings of the Aristotelian Society, 13, 1-26.

Russell, B. (1948). Human Knowledge. New York: Simon and Schuster.