Contribution
It separates two ways industry can bias science. The familiar one corrupts individuals through conflicts of interest. The one this paper names, industrial selection, changes which methods prosper in a community while leaving every individual’s judgment untouched, and it shows the second in a model built so that funding cannot affect any agent’s beliefs.
Key insight
A funder does not need to change anyone’s mind. Where researchers honestly disagree about how to measure success, funding the ones whose measure favours your product makes them more productive and more likely to train the next generation, so the community drifts toward their method.
Assumptions
- Agents are “(myopically) rational Bayesian agents” whose beliefs and choices are “unquestionably unaffected” by funding (p. 1012).
- Methodological bias: each agent’s measured success rates are drawn around the true rates with variance σ², standing for different metrics of “success” (p. 1013).
- Productivity varies (50–100 trials a round, uniform) and sets influence and training.
- Turnover: with probability 0.02 a round one agent leaves; the newcomer takes another agent’s method with probability proportional to that agent’s productivity (a Moran process) (p. 1013, n. 5).
- Industry funds anyone whose measured success for the drug exceeds its true rate by more than T, adding F trials a round (p. 1015). No treatment is in fact better than the drug.
Key results
- Diversity alone lowers reliability. Read from Figure 1, with no industry the share of agents on the superior act falls from about 95% at σ² = 0 to about 64% at σ² = 0.06.
- Industry lowers it further: at σ² = 0.04 about 64% with F = 20 and about 46% with F = 100. Without diversity there is no effect: “If there is no methodological diversity to begin with, then industrial selection does not occur” (p. 1015).
- Two channels: the funded agent influences peers more now and trains more newcomers later (p. 1015).
- Independent funding (Figure 2): with no NSF funding about 56%; a meritocratic NSF, which counts industry-funded trials, about 48–49%; a selective NSF, which does not, about 77–81%.
- The case: in the antiarrhythmic drug story, funded researchers’ views did not change with funding, and industry “simply chose not to fund” those who held unfavourable views (p. 1011).
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Industry funding can bias a community without corrupting any individual | moderate | an existence result in a simulation, plus one historical case |
| C2 | Methodological diversity, merit-based influence and selective funding are sufficient for the effect | moderate | the model, at a few parameter settings |
| C3 | Without methodological diversity the effect does not occur | moderate | the model |
| C4 | A merit-based independent funder worsens it, and one that discounts industry-funded work counteracts it | moderate | Figure 2, one parameter setting |
| C5 | The results hold across dense and sparse networks | weak | asserted in footnote 4, no data shown |
| C6 | Conflicts of interest were not the primary cause of the antiarrhythmic disaster | weak | the authors’ reading of a historical account (Moore 1995) |
Method
Agent-based simulation of 20-agent communities in Zollman’s (2010) bandit model: each round agents perform the action they think better as many times as their productivity, update by Bayes’s rule on their own and neighbours’ results, and turn over. Figure 1 uses pA = .5, pB = .45, T = 0.03 and F = 0, 20, 100 across values of σ²; Figure 2 uses T = 0.04, F = 20, an NSF productivity threshold of 75, and a σ² value missing from the caption. The number of runs, run length or convergence criterion, the network used for the figures, and the units of NSF funding are not reported.
Concepts
- industrial selection — bias produced because “it was because of the views researchers antecedently held that industry contracted them in the first place” (p. 1012), not because funding changed them.
- methodological bias — an agent’s systematic deviation in measured success, from its choice of metric.
- meritocratic and selective funding policies — an independent funder counting, or not counting, industry-funded trials toward its productivity threshold.
Connections
It builds on Zollman’s model (LIT-735, read in NOTE-564) and cites the authors’ own 2015 paper (LIT-751, unread). Its turnover dynamic is borrowed from evolutionary biology, as the authors note.
- The Zollman effect (THEORY-143). There, sparse communication protects a community’s reliability when evidence is scarce. Here the qualitative result is said to hold on “both dense and sparse communication networks” (n. 4), so network structure offers no defence against industrial selection. That rests on a footnote, with no data.
- Transient diversity (THEORY-148). That account treats diversity in what members pursue as what protects a learning group. Here diversity in how members measure success is what a funder exploits, and it lowers reliability even without a funder. The two are different kinds of diversity, of belief and of method. The contrast is the record’s; the paper does not discuss it.
- Longino (THEORY-151). In the case the paper describes, venues for criticism existed (“conferences with dissenters regularly occurred”) and the FDA oversaw the process, yet the community was steered. Funding gave some researchers more influence through productivity, which bears on Longino’s tempered equality of intellectual authority. The paper does not cite Longino; the reading is the record’s.
Bearing on the record
Source of THEORY-146. For the owner’s essay:
- §8.2 (Longino, Ostrom, epistemic networks). It shows that an epistemic network meeting the usual integrity safeguards can still be steered from outside, and that the remedy is at the community level: how an independent funder counts merit.
- §6.4 (alien interests). The record reads industrial selection as an outside interest reshaping a community’s practices without capturing any member, a case of the essay’s institutional niche construction. The paper does not use those terms.
Limitations
One historical case, read from a secondary history, and a small simulation at a handful of parameter values, with the run count and the network unreported and one figure caption incomplete. The robustness claim for network structure is a footnote. The model’s agents never change their method, so the paper does not show what happens when researchers adapt to funders.
Open questions
Whether the effect survives agents who revise their methods in light of evidence; how large methodological diversity is in real fields; whether a selective funding policy is feasible, which the authors concede “might face opposition”.