Imagine you're in a neighborhood where everyone paints their house the same color — not because of an HOA rule, but because each homeowner looks at the houses around them before choosing paint. Over time, the neighborhood converges on a color that no single homeowner chose independently. Now replace "paint color" with "how much debt a company carries," and "neighborhood" with "companies selling similar products." That's the core mechanism here: firms don't just inherit their own past leverage — they inherit the leverage choices of their product-market neighbors. The committed claim: network-weighted lagged leverage from product-market peers accounts for roughly 45% of the combined own-plus-network persistence in book leverage, and ignoring this network channel makes individual firms look more sluggish than they actually are. This isn't the first paper to say peers matter for capital structure — that fight has been running since Leary and Roberts (2014) — but it is the first to combine time-varying text-based product-market networks with a heterogeneous dynamic panel estimator that allows firm-specific coefficients to vary with network position. The architecture is a spatial dynamic partial-adjustment model with interactive fixed effects, estimated via defactored instrumental variables. The "spatial" part means the estimator treats the product-market network like a geographic adjacency matrix, except the weights change every year based on textual similarity of firms' product descriptions (the Hoberg-Phillips network). The "defactored" IV approach handles the nightmare econometric problem of correlated unobservables in network data — if two firms share unobserved shocks (say, a tech bust), naive estimation would attribute that correlation to network influence when it's really common exposure. The ladder comparison is against the standard partial-adjustment model without network terms, which is the workhorse in the capital structure literature. When network dependence is omitted, own-firm persistence looks higher and the implied speed of adjustment looks lower. The paper doesn't claim to beat a competing network model on a public leaderboard — this is economics, not ML — but it does demonstrate that static networks (holding the graph fixed) produce a substantially larger network coefficient, meaning researchers who freeze the network overstate peer influence. The distinction between time-varying and static networks is the sharpest empirical contribution. Integrity is reasonable for an empirical economics paper. The instrumental-variable strategy is the main line of defense against endogeneity, and the authors explicitly test robustness to network specification. There are no pre-registered hypotheses (standard in this field), but the Hoberg-Phillips network data is publicly available and the econometric framework is transparent enough for replication. The key vulnerability: text-based product similarity might capture demand-side shocks rather than true strategic imitation, and no instrument can fully rule that out. The firm-specific estimates are the quiet star of the paper. They show that own-leverage persistence and the sensitivity to standard determinants (profitability, tangibility, size) cluster systematically by product-market position. This means the heterogeneity isn't noise — firms in similar product spaces don't just carry similar debt levels, they respond to the same fundamentals in similar ways. That's a structural finding that goes beyond "peers matter." The milestone question is whether this network channel is large enough to matter for macro-financial stability or just for academic capital-structure models. If network leverage transmission is real at the magnitudes estimated here, then a shock to one firm's financing constraints propagates through the product-market graph in ways that standard firm-level risk models completely miss. The practical next step is connecting this to default cascades or credit-cycle amplification — the paper stops at capital structure and doesn't go there.
