Reductionist research strategies and their biases


16


PhilGoetz

I read an extract of (Wimsatt 1980) [1] which includes a list of common biases in reductionist research. I suppose most of us are reductionists most of the time, so these may be worth looking at.

This is not an attack on reductionism! If you think reductionism is too sacred for such treatment, you've got a bigger problem than anything on this list.

Here's Wimsatt's list, with some additions from the parts of his 2007 book Re-engineering Philosophy for Limited Beings that I can see on Google books. His lists often lack specific examples, so I came up with my own examples and inserted them in [brackets].

  1. Conceptualization
    1. Descriptive Localization: Describing a relational property as if it were monadic, or a lower-order relational property.
      • Fitness treated as a property of phenotype (or even of genes) rather than as a property of phenotype and environment.
      • [This may be equivalent to assuming that you can apply linearization to remove variables from a function. You often do this to analyze the stability of equilibriums. So often it's a useful assumption.]
    2. Meaning Reductionism: Assuming lower-level redescriptions change meanings of scientific terms, while [going back to?] higher-level redescriptions [does] not.
      • Philosophers (who view themselves as concerned with meaning relations) are inclined to a reductionist bias.
      • [What he might mean: Modernist theories of meaning begin by studying sentences in isolation, as logic propositions. Take a classic example, "Bachelors are unmarried men," represented as BAUM = forall(X, B(X)=>U(X),M(X)). They show that there are instances #bob where B(#bob) holds, yet affirming U(#bob) or M(#bob) would seem peculiar, and then say therefore not(true(BAUM)), then generalize to say that true(forall(X, P(X)=>Q(X))) is not meaningful in general. But they don't consider that we may use the English word "truth" differently when talking about propositions with quantified or unbound variables (BAUM) than when talking about propositions with only bound variables ("Justin Bieber is a bachelor"). But I don't think this is the main problem with these modernist analyses; compare the similar "birds can fly" / "*penguins can fly" argument.]
    3. Interface Determinism: Assuming that all that counts in analyzing the nature and behavior of a system is what comes or goes across the system-environment interface.
      1. black-box behaviorism: all that matters about a system is how it responds to given inputs
        • [Systems with hysteresis cannot be made sense of with a timeless black-box analysis.]
      2. black-world perspectivalism: all that matters about the environment is what comes in across the system boundaries and how it responds to system inputs.
        • [At first, this seemed true to me. But if the environment behaves predictably, and the system studied relies on this predictable behavior, any analysis that doesn't model the environment will try futily to find mechanisms inside the system that produce the needed information. For instance, if you were studying circadian rhythms, and your model of the environment specified the sky's brightness at any particular moment, but didn't model brightness as a 24-hour cycle, you would run into serious difficulty trying to explain the organism's cyclic behavior entirely in terms of internal components.]
    4. Entificational anchoring: Assume that all descriptions and processes are to be referred to entities at a given level, which is particularly robust, salient, or whatever. This is the ontological equivalent of assuming that there is a single cause for a phenomenon, or single level at which causation can act.
      • Thus the tendency to regard individual organisms as primary [in selection, presumably].
      • Cf. methodological individualism for rational decision theorists and other social scientists. [This is a particularly important point to inject into the FAI/CEV discussion of "human values". The values encoded into the behavior patterns of individual humans by individual selection, the values encoded by kin selection, the goals they develop through interaction with the environment (which are probably not distinguishable, on later inspection of the brain, from "final goals"), the values they hold consciously, and the socially-condoned values of human groups, are all different, and encoded in a variety of representations and levels of abstractions, and often oppose each other. A rational agent is by definition rational only within one representation and with one set of non-contradictory goals. I haven't seen discussion in the FAI literature of this problem.]
      • Similarly for genes for some reductionist neo-Darwinians. [Not sure if anybody actually holds such a position.]
  2. Model Building and Theory Construction
    1. Modeling Localization: Look for an intrasystemic mechanism to explain a systemic property, rather than an intersystemic one. Structural properties are regarded as more important than functional ones, and mechanisms as more important than context.
      • [I don't know what he means by "functional".]
      • [the example above of trying to model circadian rhythms without modeling environmental cycles would also be an example of this bias]
      • [Chomsky positing that children must have a built-in universal grammar because he didn't do the math]
      • [See all of behavior-based robotics and everything written by Rodney Brooks for objections to this bias in artificial intelligence.]
    2. Simplification: Simplify environment before simplifying system. This strategy often legislates higher-level systems out of existence or leaves no way of describing systemic phenomena appropriately.
    3. Generalization: When starting out to improve a simple model of system environment, focus on generalizing or elaborating the internal structure at the cost of ignoring generalizations of elaborations of the external structure.
    4. Corollary: If the model doesn't work, it must be because of simplifications in description of internal structure, not because of simplified descriptions of extrenal structure.
  3. Observation and Experimental Design
    1. Focused Observation: Reductionists will tend not to monitor environmental variables, and thus will often tend not to record data necessary to detect interactional or larger-scale patterns.
      • [Nearly every drug toxicity study ever, for failing to sample each subject's gut microbiome, which is a primary determinant of how ingested drugs are broken down]
    2. Environmental Control: Reductionists will tend to keep environmental variables constant, and will thus often miss dependencies of system variables on them. ("Ceterus paribus" is viewed as a qualifier on environmental variables.)
      • [Mouse experiments often use sedentary mice in HEPA-filtered cages environments fed ad-libitum. Interventions that extend lifespan in such experiments, such as rapamycin or caloric restriction, may work less well in other environments.]
    3. Locality of Testing: Make sure that a theory works out only locally (or only in the laboratory) rather than testing it in appropriate natural environments, or doing appropriate robustness analyses to suggest what are important environmental variables and/or parameter ranges.
    4. Abstractive Reification: Observe or model only those things that are common to all cases; don't record individuating circumstances.
      • Raff (1996) [3] notes that evolutionary geneticists focus on intraspecific variability, while developmental geneticists focus only on genes that are invariant within the species. This produces problems both of methodology and of focus when trying to relate micro-evolution and macro-evolution or evolution and development.
      • Cognitive developmental psychologists tend to look only for invariant features in cognition, or major dysfunctions, rather than populational variation.
    5. Articulation-of-Parts (AP) Coherence (Kauffman/Taylor/Schank): Assuming that studies done with parts studied under different conditions are valid when put together to give an explanation of the whole.
      • [There's a classic case in cell biology of a chemical that has opposite effects on cells in vitro and in vivo, though I can't recall now what it is.]
    6. Behavioral Regularity (Schank/Wimsatt): The search for systems whose behavior is relatively regular and controllable will result in selection of systems that may be uncharacteristically stable because they are insensitive to environmental variations.
      • Schank: Regular 4-day cyclers among Sprague-Dawley rats are insensitive to conspecific pheromones. [This is probably a reference to this article. I think Wimsatt's point is that biologists chose to study ovulation cycles using a particular strain of rat because it had regular cycles, and it seems that that particular strain of rat had regular cycles because of an inbred genetic deficit in its regulation of ovulation cycles.]
      • [The initial resistance to chaos theory and nonlinear systems theory was due to linear analysis having done a very good job for centuries on problems that were studied because linear analysis worked on them.]
  4. Functional Localization Fallacies
    1. Deficit Reification: Assuming that the function of a part is to produce whatever the system fails to do when that part is absent, or produced when it is activated or stimulated.
      • spark plugs as "sputter suppressors"
    2. Assuming 1-1 Mappings Between Parts and Functions:
      • Stopping the search for functions of a part after finding one; e.g., hemoglobin also functions in NO+ transport
      • Ignored division of labor when a part's necessity is shown through deletion studies, thus missing the roles of other parts
      • [The NCBI stores data on all bacterial genes in a format that assumes each gene has exactly one function [4]]
    3. Ignoring interventive effects and damage due to experimental manipulations
      • in neurophysiological studies
      • marking specimens in mark-recapture studies may affect their fitness
    4. Mistaking lower-level functions for higher-level ends, or misidentifying the system that is benefited:
      • [I think his examples here are mistaken]
      • eliminative reductionists who want to deny the existence of large domains of cognitive function [long discussion of this here, which I recommend not reading; suffice it to say that I think the ERs being complained of are not misidentifying the system benefited, but just using language a little sloppily]
    5. Imposition of incorrect set of functional categories.
      • Common in philosophy of psychology when it neglects ethology, ecology, and evolutionary biology.

Wimsatt (1980) then says:

There are at least two possible corrective measures... The first is robustness analysis--a term and procedure first suggested by Richard Levins (1966) [2]. The second, which I will call "multilevel reductionist analysis," involves using these heuristics simultaneously at more than one level of organization--a procedure that allows discovery of errors and their correction....

It should be clear that these heuristics are mutually supporting, not only in their effective use in structuring and in solving problems, but also in reinforcing, in multiplying, and, above all, in hiding the effects of their respective biases.... Whatever can be said for theories or paradigms as self-confirming entities, as much and perhaps more can be said similarly for [sets of] heuristics.



[1]. William Wimsatt (1980). Reductionist research strategies and their biases in the units of selection controversy. In T. Nickles, ed., Scientific Discovery: Case Studies, Dordrecht: Reidel, p. 213-259.

[2]. R. Levins (1966). The strategy of model building in population biology. American Scientist, 54:421-431.

[3]. Rudolf Raff (1996). The Shape of Life: Genes, Development, and the Evolution of Animal Form. Chicago: U of Chicago Press.

[4]. They let you use multiple GO tags, and put multiple names within a protein's name field if separated by slashes, but these are not adequate solutions.