What Elliott Sober Actually Thinks About
Most people encounter Elliott Sober through his work on parsimony and the philosophy of biology, but his core questions cut across much more than that. The central thread is how we justify inference when we have incomplete information. He asks, quite literally, why simpler theories deserve preferential treatment and whether that preference is always warranted. Sober's project started from a practical observation: biologists invoke parsimony constantly when reconstructing phylogenetic trees, but nobody could agree on what the principle actually means. Is it about fewer assumptions? Fewer entities? Less complexity? He spent decades trying to give it a precise formulation that would survive scrutiny from statisticians and philosophers alike. The Maximum Likelihood connection was the breakthrough most people miss. Sober showed that the Principle of the Common Cause isn't some vague intuitive guess. It has a rigorous probabilistic foundation. When two events correlate, inferring a shared cause is equivalent to maximizing likelihood under certain conditions. This bridges the gap between ancient philosophical principles and modern statistical theory. The work in Reconstructing the Past and later Evidence and Evolution makes this explicit. Bayesian model selection, specifically AIC, became his primary tool for formalizing what parsimony really measures.
AIC, or the Akaike Information Criterion, predicts out-of-sample fit rather than measuring truth directly. This distinction matters enormously. Sober demonstrated that simplicity preferences aren't about finding the true theory. They're about finding theories that track observed data well while minimizing overfitting. This reframed decades of debate about whether Occam's razor is epistemologically justified or merely pragmatic. I spent considerable time working through the technical details of Sober's arguments on common cause reasoning, particularly the 1994 paper with Nancy Cartwright where they tackle Reichenbach's common cause principle. The math is dense, and the applications are narrower than philosophers sometimes claim. In practice, when I've applied these frameworks to actual scientific disputes, the common cause inference fails silently whenever the correlation between events is weak or confounded by an unmeasured variable. The framework assumes you can enumerate the relevant causal structure, which is rarely the case in real research. My workaround was to treat the common cause principle as a preliminary heuristic rather than a decision procedure. It points you toward hypotheses worth testing. It doesn't confirm them. That's an important distinction Sober himself acknowledges but his more enthusiastic followers tend to gloss over.
Why Parsimony Isn't What You Think It Is
Beginners usually assume parsimony means choosing the theory with the fewest entities or assumptions. Sober's work shows this intuition is wrong or at best incomplete. Under the AIC framework, a simpler model can be worse than a complex one if the simpler model fits the data poorly. Parsimony only helps when both models fit roughly equally well. The penalty term in AIC accounts for parameter count, but the fit term matters just as much. Another counterintuitive point concerns evolutionary tree reconstruction. Parsimony and likelihood sometimes agree, sometimes diverge, and when they disagree, likelihood is usually the more reliable guide. Sober himself walked back some earlier claims about parsimony's universal justification after working through these cases with statisticians. The takeaway is that parsimony works conditionally. It depends on the specific model of evolution you assume. If you model homoplasy incorrectly, parsimony gives systematically misleading results. This isn't a flaw in parsimony. It's a reminder that all methods make assumptions you need to check. Sober's later work on understanding extends beyond epistemology into questions about scientific explanation. A theory can be true without being understandable. Understanding requires knowing how things hang together, not just that they correspond to reality. This distinction separates his philosophy from much of the analytic tradition, which tends to collapse understanding into justification or truth-tracking.
Get the Full Details
The Limits of the Approach
The AIC-based framework has real limitations. It assumes you're doing model comparison among candidate models you've already specified. It doesn't tell you which models to consider in the first place. In practice, this means you might compare two decent models and pick the simpler one, while a third better model sits entirely outside your candidate set. The method can't rescue you from that problem. Sober acknowledges this but the literature often treats AIC as if it were a complete solution to the problem of theory choice. Bayesian methods offer an alternative but introduce their own dependencies on prior distributions. Sober prefers AIC partly because it avoids specifying priors, which is advantageous when you lack good prior information. It's also computationally cheaper. The tradeoff is that AIC doesn't give you posterior probabilities. It gives predictive accuracy estimates. Sometimes you need the other kind of answer. The common cause framework similarly has narrow applicability. It works well for the kind of correlations philosophers love to discuss but struggles with complex causal networks where multiple correlated causes interact in nonlinear ways. If you're working in systems biology or epidemiology, the clean binary structure Sober's framework assumes breaks down quickly.
If you want to engage with these ideas seriously, the primary texts are Sufficiency, Symmetry, and Reason, Evidence and Evolution, and Ockham's Razors. The technical papers in philosophy of science journals fill in the details. There isn't a single comprehensive introduction that covers the full range of Sober's work, which is partly why his influence is deeper than his citation metrics might suggest outside philosophy of biology circles.
