Scientific, Medical, and Social Worlds › Economic and Social Simulation

Economic and social worlds often contain agents who respond to expectations, incentives, and one another. Adaptive feedback is not exclusive to these domains: robots, markets, and controlled physical systems can all form feedback loops. What matters for a particular social model is which behavioral responses and institutions its state and transition law represent.

An economic state may include stocks, flows, institutional rules, and beliefs about future outcomes. Expectations can influence contracts, prices, and investment, while outcomes can revise expectations. An expectation of inflation does not mechanically guarantee inflation: the result depends on behavior, constraints, shocks, and policy responses. Publishing or acting on a forecast can also change incentives. A model should specify these response channels rather than treat reflexivity as a universal explanation.

The Lucas critique concerns using historically estimated behavioral relationships to evaluate a changed policy regime. If decision rules respond to the regime, coefficients that fit the old regime need not remain invariant under the new one. We can write the dependence schematically as $\theta(\rho)$, with $\rho$ denoting a policy regime, rather than treating $\theta$ as policy-independent. This is a warning about invariance assumptions, not a theorem that all econometric evaluation is impossible or every policy change invalidates every parameter.

For a world model, the corresponding test is whether an action changes only a represented input or also the unmodeled response rules. Learned and hand-written models both need an argument for the mechanisms they hold fixed. This connects to Part XII, where deployment assumptions and feedback are considered further.

Several modeling approaches represent these issues differently:

• Dynamic stochastic general equilibrium models specify dynamic choices, constraints, shocks, and equilibrium conditions. Families vary in heterogeneity, financial frictions, and behavioral assumptions; they do not all consist of identical fully rational agents. Internally coherent policy calculations still depend on whether the specified responses and parameters describe the intended regime.
• Agent-based models explicitly simulate interacting agents and can represent heterogeneous decision rules and networks. That flexibility does not establish realistic behavior or parameter identification. Different micro-level mechanisms can fit similar aggregates, so validation should include the behavior and interactions needed for the query.
• Structural microsimulation and tax-benefit calculations can apply a specified rule to individual records. Knowing a tax formula does not eliminate uncertainty about inputs, eligibility, take-up, implementation, or behavioral response. A static arithmetic calculation and a dynamic behavioral forecast should be labeled separately.
• Financial network models represent institutions and exposures, with losses or payment shortfalls propagating according to a selected rule. Cascades can occur in some networks and shock regimes; there is no single critical-connectivity statement valid for every clearing or default model. Exposure data and assumptions about recovery, liquidity, and response constrain the interpretation.
• Language-model social simulations generate responses or agent actions from prompts and model outputs. Their applicability depends on whether those generated distributions reproduce the relevant human behavior under the intended conditions.

Language-model outputs depend on learned parameters, prompting, decoding, and post-training; they are not literally samples from the empirical distribution of training text. Nor does plausible language establish a representative human sample. Argyle and colleagues report agreement for particular GPT-3-conditioned survey comparisons, whereas Santurkar and colleagues find substantial opinion misalignment in their evaluated models and demographic groups. These are scoped empirical results, not universal validation or universal invalidation.

A study should examine response distributions by group and question, not only mean opinion. Potential failure modes include prompt sensitivity, excess agreement, exaggerated stereotypes, and insufficient within-group variation. Whether one occurs must be measured for the actual model and setup. A correct mean with inaccurate variance can misrepresent tails or minority responses; an inaccurate simulator may nevertheless be useful for exploratory hypotheses if it is not presented as population evidence.

In a changing multi-agent system, another agent's adaptation can alter the transition law seen by one agent. Equilibrium analysis and dynamical analysis answer different questions. For a game with specified payoff functions $u_i$ and strategy sets, a Nash equilibrium $\pi^*$ means no agent improves its payoff by unilateral deviation while others keep their equilibrium strategies:

$$u_i(\pi^*) \geq u_i(\pi_i, \pi^*_{-i})$$

for every agent $i$ and deviation $\pi_i$, where:

• $u_i$: the payoff (utility) function for agent $i$
• $\pi^*$: the joint policy profile at equilibrium
• $\pi^*_{-i}$: the policy profile of all agents other than $i$
• $\pi_i$: a deviation policy for agent $i$

The Nash condition alone supplies no convergence rate, stability guarantee, or path to equilibrium. Those require a specified adjustment process. A market or agent model can have an equilibrium while its proposed learning dynamics fail to reach it. Conversely, a simulation of transients does not establish that its endpoint satisfies an equilibrium definition. General-equilibrium conditions are a different formal construction and should not be substituted for the Nash inequality.

Social experiments can randomize individuals or clusters, though some aggregate interventions cannot feasibly be assigned that way and interference may complicate interpretation. Quasi-experimental methods offer alternative identification routes under explicit assumptions:

• Instrumental variables require a relevant instrument and justified independence and exclusion conditions; additional assumptions determine which effect is identified. Exclusion alone is insufficient.
• Difference-in-differences uses a specified untreated-trend assumption together with conditions about anticipation, comparison groups, and interference. Similar pretrends can inform a diagnostic but do not prove the post-treatment counterfactual trend.
• Regression discontinuity identifies a local contrast under appropriate continuity or assignment assumptions at a cutoff, with attention to manipulation and treatment definition. It does not automatically identify effects far from that cutoff.
• Synthetic control requires an appropriate donor comparison and assumptions supporting the constructed untreated path, including attention to spillovers and post-period stability.

These designs do not universally outrank structural models. A design and a structural model can support one another, and both need a clearly defined estimand. If a query is not identified by the available evidence and assumptions, a precise simulation result should not conceal that limitation.