In "The Levels of Selection" [Brandon 1984], Robert Brandon provides a suggestive but ultimately unsuccessful attempt to use the probabilistic notion of screening off in providing a schema for dealing with an aspect of the "units of selection" question in the philosophy of biology. I characterize that failure, and suggest a revision and expansion of Brandon's account which addresses its key shortcoming.
In moving away from simplistic reductionism, the philosophy of the sciences has run up against a recurring and general set of questions about the relationships between explanations of complex systems at different levels of organization. In discussions of genic, organismic, and group selection in evolutionary biology; individualism and holism in the social sciences (see, e.g., [Sober 1984]); and connectionist and cognitivist models in computational psychology (see, e.g., [Clark 1991], we must confront the general problem of picking out the "preferred" level of organization for the explanation of particular effects.
In his paper "The Levels of Selection" [Brandon 1984], Robert Brandon rightly sees the units of selection controversy as one particular case of this general problem. But his attempt to use the probabilistic notion of screening off in providing a kind of schema for answering an aspect of the "units of selection" question in the philosophy of biology is, although suggestive, ultimately unsuccessful. In what follows, I'll say why I think his account fails, and suggest how we might avoid its problems. Screening off is a fairly intuitive idea: A screens off B from C iff the assumption of A makes B probabilistically irrelevant with respect to C, but not vice versa. That is: A screens off B from C iff P(C|A&B) = P(C|A) but not = P(C|B)
Or to put it in a more intuitive but slightly less precise way: A screens off B from C just in case all the correlation between B and C is in virtue of their correlations with A.
It has been widely noted (see, e.g., [Salmon 1984]) that the screening off relation captures some aspects of the direction and proximity of causation: First, common causes screen off common effects from one another; e.g., atmospheric pressure screens off barometer readings from inclement weather. Fix the pressure, and any variation in barometer reading will not further correlate with a varying likelihood of stormy weather. And second, more proximal causes in a chain screen off more distal ones from their common effects---e.g., gas pedal pressure and throttle position are both correlated with engine speed; but fix the throttle position, and variations in gas pedal pressure no longer will correlate with changes in engine speed.
Brandon's suggestion is that screening off should be taken as a mark of the level of organization that is directly causally responsible for particular effects of a complex system. Roughly, entities or properties at the level of organization directly responsible for some particular effects should screen off entities or properties at other levels of organization from those effects. That is to say, stabilities and variations of the properties L1 at the "preferred" level of organization screen off the connection of properties L2 at each other level of organization of that structure or event from effect E---i.e., fixing L1 leaves L2 and E probabilistically independent; symbolically, that
Intuitively, again, it's to say that the correlation between L2 and E is in virtue of their links to L1.
The key illustration given of this is the case of phenotypes screening off genotypes from the reproductive success of the individual organism: Fix the phenotypical properties of the organism, and variations in the genotype no longer correlate with reproductive success. In this way, "... screening off provides us with the means for answering the question `At what level does the causal machinery of organismic selection really act?'" [No correlation alone will suffice to show causation, of course. But screening off here is not offered as a complete criterion or definition of the causal relation we're after, but as a notable mark or indicator.] [Brandon 1984, p.135]
But there is a clear limitation on this use of screening off. It might well be only accidentally correlated with the question of preferred levels of organization in the genotype/phenotype case. We can, after all, simply subsume the genotype/phenotype example under the general case of more proximal causes screening off more distal ones. Phenotypical properties are the more proximal causes of reproductive success; genotypical properties are relevant only as the more distal causes of those phenotypical properties.
Phenotypes should then screen off genotypes from reproductive success simply in virtue of being more proximal causes of those effects. The fact that they do should not then be taken by itself as providing any additional support for Brandon's view of levels. screening off might then facilitate picking the causally relevant level of organization only in cases where this in fact happens to coincide with greater proximity to the effect. The data offered here is just what we'd expect simply from the fact that phenotypical properties are temporally more proximal causes of reproductive success.
In the standard cases of causal proximity discussed earlier, it's clear how to think of holding one event or set of properties fixed while allowing another to vary: Since the two antecedent events or states are spatially and temporally distinct, we use this to isolate them. Fix the local and momentary properties of an event, and we can counterfactually vary the possible causal chains that might have led to or from it---e.g., fix the local properties of the bullet entering the body, and the effects of bodily damage are not further correlated with whether the bullet came from pulling the trigger or dropping the gun.
The more common and interesting case of sorting levels occurs when neither level is more proximal to the effect in question than the other, but where they are in fact token identical; i.e., where the properties at the two levels in question are properties of the very same physical structure---unlike the case of genetic and phenotypical properties of organisms. In cases of token identity, you can't take advantage of the fact that features at one level are more proximal to the effect in order to use screening off to pick between levels. What is there to say about cases where the relationship between the properties at different levels is one of token identity, and not of distinct points along a temporally ordered causal chain?
I think we can still make clear sense of the idea of fixing the properties of an event at one level while letting those at another level vary by taking advantage of the multiple realizability and context-dependence of higher-level properties. By using these tools as well, we can expand the use of screening off to pick between levels even when the properties are token identical. [See chapter 3 of [McClamrock 1993] for a more detailed account of this.]
Multiple realizability allows local lower-level differences without higher-level differences: the same higher-level property can be instantiated by various lower-level configurations and materials. Conversely, context-dependence allows higher-level differences without local lower-level differences: the same local lower-level structure can instantiate various higher-level properties, depending on its context.
We can vary (counterfactually, that is---just as in the earlier cases) the lower-level implementation of the fixed higher-level properties by using the higher-level properties' multiple realizability. Fix some higher-level property of a structure---say, the throttle's properties as an air flow controller. Then whatever additional lower-level properties that throttle may have (e.g., being a butterfly valve or a sideslip valve, composed of aluminum or plastic, and so on) turn out to be independent of the effects of the throttle on engine speed. Those properties can then be seen as screening off lower-level implementational properties in much the same way as in the standard cases: Fix the properties more directly responsible for the effect, and then note the independence of the effect from the various ways in which those properties might be produced---produced now not by different causal histories, but by different lower-level implementations.
Conversely, to fix lower-level properties while allowing higher-level properties to vary, we can use context-dependence. [Note that if the higher-level properties of components of complex systems were not context-dependent, fixing lower-level properties while varying the higher-level ones would be impossible---ruled out by supervenience.] Fix the local physical properties of a structure, and its higher level properties may still vary, depending on the surrounding context. So, if we specify the local physical structure of the choke in a carburetor, and then allow the surrounding context to vary, then whether it's still a choke or not will depend (among other things, of course) on its position vis-a-vis the fuel jets: If it's above the jets, it will control the richness of the fuel/air mixture, and will thus be a choke; but if it's below the jets in the new context, it will modulate the total amount of fuel/air mix going to the intake manifold, and will thus become a throttle. Once the point is clear, examples abound (see [McClamrock 1991]): Fix the chemical structure of a DNA molecule, and its genetic properties can still vary, depending on where in the overall genome it occurs, or on the particular nature of the coding mechanisms present; fix the machine-level properties of a given CPU operation in a computer, and its functional role in the higher-level program can still vary; and so on.
So multiple realizability allows lower-level properties to be altered within the constraint of fixed higher-level properties. And conversely, context-dependence allows for the possibility of lower-level properties of structures screening off their higher-level properties from some class of effects, and it does so in much the way that multiple realizability allowed in the possibility of higher-level properties of systems screening off the lower-level ones.
Questions about "preferred levels" recur across the sciences. How are we to pick the appropriate level to explain a class of phenomena? What are we to make of apparent conflicts between explanations at different levels? Is there a way in which apparent conflicts might be transformed from verbal disputes into real, at least partially empirical questions---questions with solutions that avoid both the triviality of simplistic reductionism ("only the little individuals matter") and the vapidity of complete ecumenicalism ("any level---individualistic or holistic---is as good as another")? How is it that views advocating the importance of higher-level accounts avoid implying that (as Elliot Sober puts it) "you must add some sort of occult social fluid to individuals and their interactions to get social groups"? [Sober84, p.185]
Brandon's suggestion of screening off as a mark of the level of
organization directly responsible for particular effects of
a system is promising one. Bringing out a shortcoming in his
presentation of it, and offering an expansion of the account which
addresses that shortcoming and perhaps widens the applicability of
the analysis should make some contribution to the realization of
that promise.
[Thanks to Daniel Gilman, Stuart Glennan,
Greg Mikkelson, David Walton, and William Wimsatt for their
comments on an earlier draft.]