Identification in Dynamic Dyadic Network Formation Models with Fixed Effects

This paper establishes set identification results for dynamic dyadic network formation models with fixed effects by combining inequality-based integration and signed-subgraph comparisons, ultimately deriving point identification under specific assumptions regarding error distributions and additive individual fixed effects.

The Gist

Imagine you are trying to figure out why people become friends.

You have a big photo album of a group of people, but instead of just one photo, you have a time-lapse video showing who is friends with whom over several years. You want to know:

1. Do people like others who are similar to them? (e.g., "Birds of a feather flock together"?)
2. Do people like their friends' friends? (e.g., "My friend's friend is my friend"?)
3. Is there something invisible about certain pairs of people that makes them stick together, regardless of what we can see? (e.g., "They just have good chemistry," or "They grew up in the same town.")

This paper is a detective's manual for solving this mystery. The authors, Wayne Yuan Gao and Yi Niu, are trying to separate the "visible" reasons for friendship from the "invisible" ones, even when the data is messy and complicated.

Here is the story of their solution, broken down into simple concepts and analogies.

The Big Problem: The "Ghost" in the Machine

In the past, economists had a hard time with this. They could see the "visible" reasons (like shared hobbies or mutual friends), but they couldn't easily separate them from the "invisible" reasons (the Fixed Effects).

Think of the Fixed Effect as a Ghost that haunts a specific pair of people.

• Maybe Alice and Bob have a ghost that makes them friends forever, even if they have nothing in common.
• Maybe Charlie and Dave have a ghost that makes them enemies, even if they are very similar.

If you just look at the data, you can't tell if Alice and Bob are friends because they both love jazz (visible) or because of their ghost (invisible). The paper asks: How do we catch the ghost so we can measure the jazz?

The Two Detective Strategies

The authors propose two main ways to solve this, which they call "Integrate-Out" and "Difference-Out."

Strategy 1: The "Time-Lapse" Trick (Integrate-Out)

Imagine you are watching a specific pair, Alice and Bob, over 10 years.

• Year 1: They are friends. They both love jazz.
• Year 2: They are still friends. Now, Alice hates jazz, but Bob still loves it.
• Year 3: They break up.

The authors say: "If we watch them long enough, the 'Ghost' (their unchanging chemistry) stays the same, but the 'Jazz' (their changing hobbies) changes."

By treating each pair like a short movie, they can mathematically "average out" the Ghost. They don't need to know exactly what the Ghost is; they just know it doesn't change. By looking at how the friendship changes relative to how the hobbies change, they can estimate the importance of the hobbies without needing to see the Ghost.

Strategy 2: The "Magic Cancellation" Trick (Difference-Out)

This is the more clever, algebraic trick. Imagine you have a group of four friends: Alice, Bob, Charlie, and Dave.

• Alice is friends with Bob.
• Charlie is friends with Dave.
• But Alice is not friends with Charlie.
• And Bob is not friends with Dave.

The authors realized that if you compare these relationships in a specific pattern (a "signed subgraph"), the "Ghosts" cancel each other out like a magic trick.

• Alice's ghost + Bob's ghost = Total Ghost for Pair 1.
• Charlie's ghost + Dave's ghost = Total Ghost for Pair 2.
• If you subtract the two scenarios, the ghosts disappear because they appear on both sides of the equation with opposite signs!

This allows them to look at the "Jazz" (the visible data) without the Ghost ever showing up.

The "Super-Tools" (Sharpening the Results)

The authors realized they could make their detective work even better if they had a few extra clues. They found three ways to sharpen their results:

1. The "Perfectly Random" Clue: If we assume the "shocks" (sudden changes in friendship) happen randomly and independently over time, the math becomes much cleaner. It's like assuming the weather changes randomly every day; you can predict the pattern much better.
2. The "Individual Ghost" Clue: Usually, the "Ghost" is a unique thing between two people. But what if the Ghost is actually just Alice's personal vibe + Bob's personal vibe? If the Ghost is just the sum of two individual personalities, the math gets even easier. It's like realizing the "chemistry" isn't a magical force between them, but just Alice being nice and Bob being nice.
3. The "Logit" Clue (The Golden Ticket): If we assume the randomness follows a specific, well-known mathematical shape (called a Logit distribution), and we combine it with the "Individual Ghost" idea, the paper unlocks a Super-Tool.

The Super-Tool: This allows them to use any combination of friends and time periods to solve the puzzle.

• In the past, you could only compare friends who were friends on the same day (like a snapshot).
• With this new tool, you can compare Alice's friendship with Bob in January against Charlie's friendship with Dave in July.
• This is like solving a puzzle not just with the pieces on the table today, but by using pieces from yesterday, tomorrow, and next week all at once.

Why Does This Matter?

Before this paper, if you wanted to study how friendships form over time, you had to make very strict, unrealistic assumptions (like "everyone is exactly the same" or "there are no hidden ghosts").

This paper says: "We can handle the ghosts. We can handle the changing hobbies. We can handle the complex web of friends."

They provide a flexible toolkit that allows researchers to:

• Measure how much "homophily" (liking similar people) really matters.
• Measure how much "transitivity" (friends of friends) really matters.
• Do all this without needing to know the exact secret sauce of every single pair of people.

The Bottom Line

Think of this paper as upgrading the GPS for Social Networks.

• Old GPS: "You are here, but we don't know why, and the map is blurry."
• New GPS (This Paper): "We can filter out the static (the ghosts), track your movement over time, and tell you exactly which roads (hobbies, shared friends) are leading you to your destination, even if the map is complicated."

The authors have built a bridge between real-world messiness (people are complicated and have hidden traits) and mathematical clarity (we can still measure what matters).

Technical Summary

1. Problem Statement

The paper addresses the econometric challenge of identifying parameters in dynamic dyadic network formation models when the data includes:

• Time-varying observed covariates (homophily based on observable characteristics).
• Lagged local network statistics (e.g., common friends, friends-of-friends, transitivity, indirect effects).
• Unobserved heterogeneity in the form of time-invariant dyadic fixed effects ($A_{ij}$).

The Core Tension:
In dynamic settings, observed link changes result from a mix of structural state dependence (lagged links), observed homophily, and unobserved time-invariant factors. Introducing fixed effects creates a "incidental parameters" problem. Standard identification strategies often fail because:

1. Nonlinearity: The link formation is a binary choice (0/1), making it a nonlinear panel model.
2. Endogeneity: Lagged network statistics (e.g., $D_{ij,t-1}$) are endogenous regressors.
3. Fixed Effects: Unobserved dyad-specific heterogeneity ($A_{ij}$) is correlated with covariates and cannot be simply differenced out without specific structural assumptions, especially when combined with lagged outcomes.

The paper asks: Can we separate structural parameters (homophily and network effects) from unobserved heterogeneity in a dynamic network panel without imposing strong parametric assumptions on the error distribution?

2. Model Setup

The authors propose a dynamic index model for the link indicator $D_{ijt} \in \{0, 1\}$ between nodes $i$ and $j$ at time $t$:
$$D_{ijt} = \mathbb{1}\{ Z'_{ijt}\alpha_0 + X'_{ij,t-1}\lambda_0 + A_{ij} - U_{ijt} \geq 0 \}$$
Where:

• $Z_{ijt}$: Time-varying observed dyadic covariates (e.g., distance in attributes).
• $X_{ij,t-1}$: Vector of lagged local network statistics (e.g., lagged common friends, lagged link status).
• $A_{ij}$: Time-invariant unobserved dyad fixed effect.
• $U_{ijt}$: Idiosyncratic time-varying shock.
• $\theta_0 = (\alpha_0, \lambda_0)$: Parameters of interest.

Key Feature: The network statistics $X$ are lagged. This avoids contemporaneous simultaneity issues (strategic interaction at time $t$) but introduces endogeneity via the lagged dependent variable.

3. Methodology: A Unified "Difference-Integrate" Framework

The authors develop a semiparametric identification strategy that treats the two main approaches to handling fixed effects as endpoints of a spectrum: Integrating out (treating dyads as short panels) and Differencing out (algebraic cancellation via signed subgraphs).

A. Route 1: Dyadic Panel Identification (Integrate-Out)

• Logic: Treat each dyad as a short panel ($T$ periods).
• Technique: Uses the "bounding-by-$c$" technique (from Gao and Wang, 2026).
• Mechanism: By assuming the error distribution is homogeneous over time (though potentially serially correlated), the authors construct bounds on the probability of link formation. They integrate out the fixed effect $A_{ij}$ with respect to an unknown distribution.
• Result: Generates conditional moment inequalities that bound the parameter space $\Theta$. This relies on the assumption that shocks are independent across dyads and have homogeneous marginals over time.

B. Route 2: Signed-Subgraph Identification (Difference-Out)

• Logic: Construct comparisons across time and nodes where fixed effects cancel algebraically.
• Technique: Uses dynamic signed-subgraph comparisons.
• Mechanism: By comparing edge-time cells $(i,j,t)$ such that the net count of each dyad is zero (balanced signed subgraphs), the term $A_{ij}$ cancels out exactly (e.g., $A_{ij}$ appears with a $+$ sign in one term and a $-$ sign in another).
• Result: This yields inequalities that do not require assumptions about the distribution of $A_{ij}$ or its correlation with covariates, only the exogeneity of the shocks. This works even under arbitrary serial correlation.

C. Unified Partial-Differencing Perspective

The paper synthesizes these into a general framework (Proposition 4).

• Partial Differencing: Some dyads are fully differenced out (balanced), while others are partially retained.
• Partial Integration: The retained dyads are integrated out via a latent CDF.
• Significance: This provides a taxonomy of identifying restrictions, showing that the "panel" and "subgraph" approaches are complementary, not redundant.

4. Key Contributions and Sharpening Results

The paper demonstrates how adding specific structural assumptions sharpens the identification results, moving from set identification to point identification.

Contribution 1: Known Marginal CDF + Serial Independence

• Assumption: Errors $U_{ijt}$ are serially independent within dyads and follow a known continuous distribution (e.g., Normal, though not Logit).
• Result: The composite error in differenced equations (e.g., $U_{ijt} - U_{ijs}$) has a known distribution (convolution of marginals).
• Impact: The moment inequalities become explicit bounds (sandwich inequalities) rather than abstract envelopes. This allows for "Max-Score" type identification at the median ($c=0$).

Contribution 2: Additive Node Effects

• Assumption: The dyad fixed effect is additive: $A_{ij} = \nu_i + \nu_j$.
• Result: This allows for weighted node-differencing. Instead of requiring dyad-level balance (which is restrictive), one can construct configurations where the sum of node weights is zero.
• Impact: This significantly enlarges the class of admissible identifying configurations (e.g., weighted stars, longer cycles) compared to the unrestricted dyad effect case. It sharpens identification even with an unknown error distribution.

Contribution 3: The Exact Conditional Logit Representation

• Assumption: Combines Additive Node Effects ($A_{ij} = \nu_i + \nu_j$) with i.i.d. Logit Shocks.
• Result: The authors derive an exact conditional logit likelihood for any completely node-balanced configuration of edge-time cells.
  $$\frac{P(Y^+_C = 1 | Z_C)}{P(Y^-_C = 1 | Z_C)} = \exp(\Delta_C W(\theta_0))$$
• Innovation: This generalizes Graham (2017)'s static "tetrad" logit.
  • Within-date tetrads: Standard 4-node comparisons at a single time $t$.
  • Intertemporal tetrads: Comparisons across different time periods ($t_1, t_2, t_3, t_4$).
  • Triadic cycles: Comparisons involving 3 nodes over 6 periods.
• Significance: This creates a much larger set of identifying restrictions. Point identification is achieved if the support of the covariate contrasts from all these configurations spans the parameter space. This is a weaker condition than requiring within-date tetrads alone to span the space (crucial for small networks or discrete lagged covariates).

5. Significance and Implications

1. Bridging Static and Dynamic: The paper successfully extends the identification logic of static network models (Graham, 2017) to dynamic settings with lagged endogenous regressors, overcoming the "stable neighborhood" limitations of previous dynamic work (Graham, 2016) when time-varying covariates are present.
2. Semiparametric Robustness: It provides identification results without assuming a specific error distribution (semiparametric), relying instead on moment inequalities and structural assumptions (additivity, independence).
3. Flexibility in Data Usage: By introducing "intertemporal tetrads" and "triadic cycles," the paper shows how to extract identification power from time-series variation in networks, which is vital when cross-sectional variation (number of nodes) is limited.
4. Unified Framework: The "difference-out / integrate-out" spectrum offers a new theoretical lens for analyzing dynamic discrete choice models with fixed effects, applicable beyond just network formation.

6. Conclusion

Gao and Niu establish that dynamic dyadic network models with fixed effects are identifiable under semiparametric conditions. By combining panel-style integration with dynamic signed-subgraph differencing, and further sharpening these results with additive node effects and logit specifications, they provide sufficient conditions for point identification. Their work expands the toolkit for econometricians studying network evolution, allowing for the estimation of homophily and transitivity effects in the presence of unobserved heterogeneity and time-varying covariates.