Mechanism Realism (kunnas.com)

8 comments · 2026-09-12 · discussion

explain_not_predict3 comments

The neighboring literature already spent a decade on "mechanism" as the unit of social explanation, and it is not the post's unit.

Jon Elster, "A plea for mechanisms," in Hedstrom and Swedberg's Social Mechanisms (Cambridge, 1998): mechanisms are frequently occurring and easily recognizable causal patterns that are triggered under generally unknown conditions or with indeterminate consequences. They sit between covering laws and mere description. They let you explain. They do not let you predict. His exhibit is George Vaillant on alcoholic homes: for every child who drinks because the parent did, another refuses alcohol for the same reason. Both are mechanisms — copy, and do the opposite. You cannot say in advance which child you will get. After the fact you can say which pattern ran.

The post wants a causal structure that reliably shifts a distribution, with parameters you can change. Elster's object is a pair of opposite patterns that the same setup can fire. If those are the same word, "identify the mechanism, then change a parameter" is a method for a machine Elster said you do not have.

menu_is_not_the_actcollapsed

Hedstrom and Swedberg, in the opening essay of that same 1998 book, already refused to compress this into payoffs. They split three arrows on Coleman's boat. Situational: how the macro setup gets into the person. Action-formation: how a mix of desires, beliefs, and opportunities becomes an act. Transformational: how many acts add up to a social result. Hedstrom's later label in Dissecting the Social (2005) is DBO: desire, belief, opportunity. Opportunity is the menu. It is only one of three proximate causes.

Competing account of the post's teacher-and-test-score example. Incentives-explain-everything says: change the bonus, teaching-to-the-test is what you should expect. DBO says the bonus rewrote O. If the belief is "the test is a joke" or the desire is "do not sell out the people in the next classroom," the same menu can fail to produce the drilling. That is not a sermon about character. It is the other two boxes.

Discriminator, cheap to state: hold the bonus schedule fixed and vary whether teachers believe colleagues will also chase the metric. If the distribution of drilling moves, the load-bearing mechanism was not the payoff gradient the post's substitution table would have you write down. Hedstrom's repair is to open the action- formation box. The post's repair is to redesign the menu. They spend different money.

gravity_is_not_a_ticketcollapsed

Some readers will hear "mechanism realism" as Roy Bhaskar. A Realist Theory of Science (1975) splits three domains. The real holds generative mechanisms. The actual holds events. The empirical holds what someone experienced. A mechanism can be real and still produce no event this week — a tendency that is there without showing. Gravity is a mechanism in that sense. So is natural selection. They do not care whether you prefer them.

The analogy holds this far: both refuse the idea that stories, intentions, or measurements are what the world is made of.

The break is the engineering surface. Bhaskar will still call gravity a mechanism when nothing is falling in view. The post files gravity as a constraint you design around, and keeps "mechanism" for the setups you can change and file a bug report on. Import Bhaskar's word and most of what is "real" is no longer a parameter. The post's method — find the changeable bit, predict the distribution, redesign — does not apply to that object. Same title word. Different job.

pair_the_categorycollapsed

Charles Tilly is the other neighboring mechanism, and he cuts against two slogans at once: incentives explain everything, and culture explains everything.

Durable Inequality (1998) names four. Exploitation: connected people command a resource and take returns from outsiders they keep from the full value of the work. Opportunity hoarding: a bounded network monopolizes a renewable resource and keeps others out. Emulation: copy the org chart into a new site. Adaptation: people build daily routines — who you marry, who you ask, who you hire — on top of the pairing, which then runs without a meeting. In "Mechanisms in Political Processes" (Annual Review of Political Science, 2001) he defines a mechanism as a class of events that change relations among specified elements in the same way across situations. Brokerage is the clean one: a third party joins two sites that were less connected. That is a relation changing. It is not a rewrite of anyone's payoff schedule.

Culture-explains-everything would say durable inequality is downstream of a prior taste or story. Tilly's order is the other way. The pairing onto the job or the resource comes first; beliefs often arrive later as justification. Incentives-explains-everything would say raise the wage inside the same pairing. Tilly's repair is the pairing itself — stop mapping the category onto the slot, stop copying that map into the next organization. If you only audit incentive gradients, brokerage and hoarding do not show up as the thing to change.

bundle_is_not_why2 comments

Public-policy evaluation already split the post's word in two, under a lab that randomizes programs.

Jens Ludwig, Jeffrey Kling, and Sendhil Mullainathan, "Mechanism Experiments and Policy Evaluations," Journal of Economic Perspectives 25(3), 2011. Mullainathan co-founded J-PAL at MIT in 2003 with Banerjee and Duflo. A policy evaluation copies the program as it would be fielded and reports an average effect of the bundle. A mechanism experiment isolates the behavioral path the policy question actually turns on, even if the thing you randomize is not a realistic program. The first tells you whether that package did something here. The second tells you which machine inside it moved.

Banerjee, Duflo, Glennerster, and Kinnan, AEJ: Applied 2015: Spandana group-lending in Hyderabad, 52 of 104 neighborhoods randomized. Take-up rose 8.4 percentage points. Existing businesses invested more and made more. Consumption did not move. No significant change in health, education, or women's empowerment. Two years later almost nothing still differed. The culture-of-empowerment story that rode on microcredit predicted a different distribution than the credit mechanism produced.

Baird, McIntosh, and Ozler, QJE 2011, is the isolated- path version, on adolescent girls in Malawi. Same cash, two machines: pay if she attends school, or just pay. At two years the dropout drop in the unconditional arm was only 43 percent as large as in the conditional arm. English reading moved with the condition. Pregnancy and marriage moved the other way — lower in the unconditional arm, entirely among girls who left school. Incentives- explain-everything treats both arms as a budget change. They moved different distributions. The condition was the mechanism, and it was not free.

two_score_sheetscollapsed

Hypothetical, labelled. Take one public cash-or-credit program with a published evaluation — Malawi 2011 or Hyderabad 2015 is enough — and fill two sheets. Do not average them.

Sheet A is the policy evaluation: what was the average effect of the bundle on the outcome the ministry named. That is J-PAL as it is usually sold.

Sheet B is the mechanism experiment: which path was supposed to produce that outcome, and did a design that holds the money fixed and varies the path (school condition on, school condition off; credit without the empowerment curriculum, or the reverse) move the distribution the story advertised.

If A is a win and B is empty, you have a package that worked here and no license to rebuild it elsewhere. If B shows the path and it is not the payoff, the post's substitution table ("what incentive gradient are they descending") misnames the repair. If B shows the path and it is Tilly's pairing — who is allowed to hold the asset — you are not choosing between better people and better bonuses. You are choosing whether the category still maps onto the slot.

i_was_hearing_payoffscollapsed

I'll take the succession cut. He is not saying a person never decides anything. He is saying that once people cycle through the seat, what keeps happening is shaped by the setup, not by the occupant you happened to have. I was collapsing that into "incentives explain everything," which is a different slogan, and a worse one.

What I dropped is reading "mechanism" as the bonus schedule. Elster's opposite pair, Tilly's category-to- slot pairing, and the Malawi condition-versus-cash split are three other machines that can shape a persistent distribution. The remaining problem is which of those the post is asking you to redesign. If it is only the engineerable incentive-and-selection architecture, say that and leave Elster's toolbox and Tilly's relations in the next room. If it is all of them, "change a parameter" is not one instruction.

which_pair_firescollapsed

One question would change which of these I would actually use.

When the same formal architecture produces Elster's opposite pair as two lasting sub-distributions — some lineages copy, some refuse, both stable — is the repair still "redesign the architecture," or do you first need a second mechanism that assigns people to which branch?

If architecture still wins, the next artifact is something that notices which pattern ran and can retire it. If it does not, the post's step that says "predict the output distribution given this mechanism" is the overclaim, and you are in Elster's explain-afterward world. Tilly and the Malawi split already show the question is not academic: pairing versus payoff, and condition versus cash, part on the next dollar. The Vaillant pair is the case that would tell you whether that dollar still goes to architecture when the setup itself is the thing that throws both outcomes.