Reflection in Linear-in-Means Network Models

Why Peer Effects Create a Reflection Problem

Suppose an analyst observes that people connected in a network have similar outcomes. One explanation is a peer effect: a person’s outcome responds to the outcomes or characteristics of their peers. But peers may respond to one another at the same time. If a person’s outcome affects a peer’s outcome and that peer’s outcome feeds back, it is not enough to regress one person’s outcome on the observed outcomes of their peers. The peer outcomes are jointly determined with the person’s own outcome.

This simultaneous determination is the reflection problem. It is distinct from the algebraic question of whether a system of simultaneous equations has a unique solution. A unique equilibrium can exist even when the data do not contain enough independent variation to distinguish the peer-outcome effect from the effect of peer characteristics or shared group conditions.

Model and Assumptions

Consider \(n\) individuals. Let \(y_i\) be individual \(i\)’s outcome, \(x_i\) a vector of observed characteristics, and \(w_{ij}\) the weight individual \(i\) assigns to individual \(j\). Collect the weights in the known \(n \times n\) matrix \(W\). For the basic model, assume \(w_{ii}=0\) and that each row of \(W\) sums to one, so \((Wy)_i=\sum_j w_{ij}y_j\) and \((Wx)_i=\sum_j w_{ij}x_j\) are weighted peer means. These conventions describe the model here; other applications may use different network weights.

Let \(D\) be a matrix of group indicators, and let \(\lambda\) contain the corresponding group effects. The linear-in-means model is

Here, \(X\) contains individuals’ own observed characteristics, \(\delta\) their coefficients, \(\beta\) the endogenous peer-effect coefficient, and \(\gamma\) the contextual-effect coefficient. The vector \(\varepsilon\) contains individual-specific unobserved factors. A shared group effect can represent, for example, a common environment affecting members of a classroom or neighborhood. If it is omitted and correlated with included variables, the corresponding estimates may be confounded.

The term \(Wy\) is endogenous in the structural equation because \(y\) is determined jointly across the network. Estimating its coefficient requires more than specifying the equation: the relevant variation must distinguish \(Wy\) from other included or unobserved influences. For the algebraic equilibrium calculations below, assume \(W\) is fixed and known and \(I-\beta W\) is nonsingular. For an interpretation in which feedback converges by repeated propagation, a common sufficient condition is that the spectral radius of \(\beta W\) is less than one. That stability condition is not, by itself, an identification condition.

Worked Example: Writing and Solving a Two-Person Network Model

Why the Peer Outcome Is Endogenous

Rearrange the model to isolate the simultaneous system:

When \(I-\beta W\) is nonsingular, the reduced form is

This expression shows that individual shocks can propagate through the network. In general, a peer outcome \((Wy)_i\) therefore contains a component affected by the shock \(\varepsilon_i\) in individual \(i\)’s own structural equation. That association between a regressor and the structural disturbance is why ordinary least squares on the structural equation does not generally identify \(\beta\). Assumptions about observed characteristics, shared effects, and available identifying variation are also needed before a particular estimation strategy is justified.

Reflection and the Limits of Group Variation

The reflection problem is especially clear in a group where every member is equally connected to every other member. Suppose the group has \(m\) individuals, each has weight \(1/(m-1)\) on every other member, and there are no self-links. Write \(\bar y_g\) for the group outcome mean and \(\bar x_g\) for the group mean of a scalar characteristic. For any member \(i\),

These identities show that a peer mean is not an independently varying quantity in this network. Once the group mean and individual value are given, the peer mean is determined. With group effects in the model, between-group differences in group means cannot automatically be interpreted as peer effects: those differences may also reflect the shared group conditions.

Solving Is Not Identifying

A model has an equilibrium solution when its simultaneous equations can be inverted under the stated conditions. Identification is a different question: whether distinct values of the structural parameters imply distinguishable distributions of the observed data under the assumptions being used. A nonsingular \(I-\beta W\) gives a unique reduced form for a given parameter vector, but does not ensure that \(\beta\) and \(\gamma\) can be distinguished from one another or from group effects.

In particular, if peer outcomes and peer characteristics have no independent variation after accounting for own characteristics and shared group conditions, the data cannot isolate their separate effects without additional identifying restrictions or variation. In a complete group, the identities above expose this limitation directly. In other networks, the exact collinearity may disappear, but simultaneity remains: the existence of network variation alone does not establish that the endogenous peer outcome is identified.

Common Mistake

A common mistake is to reason, “The network matrix is known and the simultaneous equations have a unique solution, so the peer-effect coefficient can be estimated.” This confuses existence of an equilibrium with identification of its parameters. Invertibility establishes that outcomes are determined for specified parameters; it does not supply independent variation that separates the endogenous peer-outcome effect from contextual effects or shared conditions. The correction is to state the equilibrium condition and then examine identification on its own, including whether the relevant regressors can be distinguished under the model’s assumptions.