Moffitt 2005 — Remarks on the Analysis of Causal Relationships in Population Research

Summary

The full-length, more technical companion to Moffitt 2003 — Causal Analysis in Population Research, published in Demography 42(1): 91–108, February 2005. Moffitt presents the complete structural equation system for causal identification (outcome and selection equations with heterogeneous $\beta_i$ ), formalizes the area fixed-effects estimator as a two-equation instrumental variables (IV) system in first differences, names and distinguishes the population-segment fixed-effects design from difference-in-differences (DiD), develops two distinct types of extrapolation failure, and applies the framework to three illustrative causal questions — teenage childbearing, economic growth and internal migration, and race and educational attainment. The conclusion explicitly counsels methodological pessimism and synthesis across approaches.

Key Claims

Every causal estimate requires a minimum of one identifying assumption; that minimum assumption is untestable. The burden of justification rests on a priori argument, theory, or outside evidence — "there is no mechanical algorithm for producing a set of 'assumption free' facts or causal estimates" (Heckman 2000).
The structural model is $Y = \alpha + \beta_i T + \gamma X + \varepsilon$ (outcome) and $T = \delta + \theta X + \phi Z + \upsilon$ (selection). Ordinary least squares (OLS) is biased when T is correlated with $\varepsilon$ (unobservable confounders) or with $\beta_i$ (selection on individual gains from treatment). IV with a valid Z recovers the average $\beta_i$ for "switchers" — those who changed T in response to Z (equivalent to Angrist, Imbens, and Rubin's [AIR] "compliers").
The area fixed-effects (FE) model is the preferred extension of ecological IV to panel/repeated cross-section data: $\Delta Y = \alpha + \beta_i \Delta T + \gamma \Delta X + \varepsilon$ and $\Delta T = \delta + \theta \Delta X + \phi \Delta Z + \upsilon$ , where $\Delta$ denotes first differences and $\Delta Z$ is a change in the area-level environmental variable. $\Delta Z$ instruments for $\Delta T$ , removing area fixed effects while allowing individual changes in T to remain endogenous. Distinct from individual fixed-effects models, which assume differencing alone eliminates bias.
The individual fixed-effects model for panel data "has seen decreasing support among economists": simply differencing does not explain why individual changes in T occur, and those changes may well be endogenous.
The population-segment fixed-effects design (Moffitt's term for demographic DiD) uses nationwide policies affecting demographic groups differently as the implicit instrument. Mathematically equivalent to area FE but the parallel-trends assumption is "much more suspect" — demographic groups are harder to argue as comparable absent the policy.
Two distinct types of extrapolation failure: (1) Mechanism specificity — each Z estimates only the effect of T induced by that particular cause; whether postponement-via-abortion-policy has the same effect as postponement-via-labor-market-shock is an empirical question the model cannot answer; "the" effect of T is ill-defined without specifying the mechanism. (2) Range restriction — Z induces variation in T only across a particular range; extrapolation to the full population requires additional assumptions; instruments inducing larger variation are preferred but often have weaker internal validity claims.
Reduced-form estimation (effect of Z on Y without identifying T) requires weaker assumptions (Z need only be exogenous, not excluded from Y) but loses mechanism knowledge and may falsely attribute effects to the intended T when a different mechanism is operative.
Three illustrations: (a) Teenage childbearing → child outcomes: multiple mechanisms (education, earnings, maturity, contraception, mate supply) make "the" effect of postponement ill-posed; candidate Z — contraceptive availability, abortion policy, miscarriage — each faces mechanism and validity concerns; (b) Economic growth → internal migration: aggregate complexity + reverse causality + small N make the causal effect essentially unknowable; (c) Race → educational attainment: well-defined as a gedanken (thought) experiment but unanswerable with confidence given myriad uncontrolled unobservables.
Synthesis: weight evidence across studies with different mixes of internal and external validity; convergence across methods is the strongest available evidence; formal theory and ethnographic evidence merit positive weight.

Concepts Introduced or Extended

Instrumental Variables (area fixed-effects formalization, mechanism specificity extrapolation type, population-segment FE)
Natural Experiments (four instrument types with full objections; internal/external tradeoff developed)
Difference-in-Differences (area FE vs. individual FE distinction; population-segment FE distinguished)
Local Average Treatment Effect (switchers terminology; range restriction extrapolation)
Rubin Causal Model (historical attribution: Neyman 1923/1935, Heckman 1978, Roy 1951, Haavelmo 1943/1944)

Entities Mentioned

Quotes

"There is no mechanical algorithm for producing a set of 'assumption free' facts or causal estimates based on those facts." (Heckman 2000, cited by Moffitt)

"Modesty of claims for truth is the clear lesson from this review."

"There is a danger in maximizing internal validity at the expense of external validity. To do so would lead to a field consisting only of narrowly defined exercises without generalizability and to a collection of miscellaneous facts that do not add up to any general knowledge."

"The individual fixed-effects model for panel data has seen decreasing support among economists. Simply assuming that the change in an individual's T from one time to another is exogenous leaves unspecified why individual changes in T occur."

My Take

Moffitt (2005) is the more useful of the two companion papers for understanding his framework, because the structural equations make the identification logic precise and the area FE formalization clarifies why individual fixed effects are insufficient. The two-type extrapolation distinction — mechanism specificity and range restriction — is the conceptually richest contribution and underappreciated: it explains why local average treatment effect (LATE) estimates from examiner IV (Maestas et al. 2013; French and Song 2014) cannot be mechanically generalized to either the full applicant pool or to a policy that moved allowance rates by a different amount. The paper's pessimism about individual FE models is a direct challenge to a large empirical literature in demography, and the population-segment FE critique (weaker parallel-trends assumption) is a sharp formulation of what DiD requires. The teenage childbearing illustration remains the clearest pedagogical demonstration that "the effect of T" is incoherent without specifying the mechanism.