Can Network Structure Deconfound Its Own Effect? Testing the Inverted-U of Subgroup Structure on Innovation

A well-known finding in innovation research is that firm collaboration networks work best when split into a moderate number of tight subgroups. Too few subgroups and partners never build trust. Too many and the subgroups become silos. This "inverted-U" curve is estimated by regressing patent output on a structural measure of the network, using lagged predictors and firm controls. But firms choose their partners. If unobserved capability, shared location, or executive friendships drive both whom a firm clusters with and how much it patents, the curve could come from that sorting rather than from structure. Existing tools for removing hidden confounding on graphs were built for a different question, and they break in a specific way when the network structure is itself the thing whose effect we want to measure.

This note proposes a way to estimate the dose-response curve of subgroup structure while correcting for hidden firm traits. The correction learns about those traits from data that reflects them but does not determine the exposure: a firm's earlier solo patent text, its prior-window ties, and its governance links. The work also builds a semi-synthetic benchmark on real patent co-application graphs. In it, the true curve is planted as either flat or inverted-U, so we can measure how often current practice "discovers" an inverted-U that is not there.

Core Claim

The closest existing method is the embedding deconfounder of Using Embeddings for Causal Estimation of Peer Influence in Social Networks. That paper learns a low-dimensional summary of each node's position (a node embedding) from the adjacency matrix and uses it as a stand-in for unobserved homophilous traits. On a semi-synthetic Pokec benchmark with true effect 1.0, it recovers 1.17, where an unadjusted model gives 7.40 and a stochastic block model gives 5.28.

In that setting the treatment is a peer's outcome, not the graph. In the innovation setting, the exposure is a deterministic function of the graph. Examples are the ratio of network size to cohesive subgroups used in How Does Tie Strength Dispersion within Inter-Organizational Networks Affect Agricultural Technological Innovation? Evidence from China, and the rolling standard deviation of tie strengths ("divisive faultlines") used in Divisive Faultlines and Knowledge Search in Technological Innovation Network: An Empirical Study of Global Biopharmaceutical Firms.

An embedding rich enough to proxy the confounder from the same-window graph will also nearly encode the exposure. Adjustment then leaves almost no exposure variation, which is a failure of overlap. Weakening the embedding restores overlap but brings the bias back.

Our claim is that this tension is resolved by channel separation. The proxy for latent traits should come from information that shares the confounder but is not a function of the exposure-defining ties:

  • pre-window solo patent text,
  • earlier collaboration windows,
  • governance layers such as board interlocks.

We also claim a second, more mechanical problem. In this literature the network and the outcome are built from the same patents, so firm patenting volume links them by construction. A credible test must therefore measure innovation on patents that do not enter the network.

Written out, let $S_i$ be firm $i$'s structural exposure computed from its collaboration ties in window $[t-5, t-1]$. Let $Y_i$ be innovation in year $t$, $C_i$ the unobserved traits, and $W_i$ the separated proxy. The target is the curve $\mu(s) = E[Y_i(s)]$. We estimate it by adjusting for a learned summary $\lambda(W_i)$ rather than $\lambda(A_i)$, under the requirement that $S_i$ retains variation given $\lambda(W_i)$. That requirement can be checked directly by how well $\lambda(W_i)$ predicts $S_i$.

Research Program

The work has three parts.

Benchmark. We build a semi-synthetic benchmark that keeps real firms, real proxies, and realistic graph sparsity, but plants the truth.

  • Latent traits. We take the firm set and pre-window filings of a patent co-application cohort. We draw a latent trait $C_i$ that is partly predictable from each firm's pre-window solo-patent text and partly pure noise. The noise share is a dial controlling how much confounding the proxy cannot see.
  • Ties. We regenerate weighted collaboration ties from a latent-space model in which firms with similar $C_i$ form stronger, more repeated ties. This produces faultlines and subgroups through homophily alone.
  • Outcome. We set $Y_i = h(S_i) + \gamma C_i + \varepsilon_i$, with $h$ either flat, linear, or inverted-U.
  • Dials. Varying $\gamma$, the noise share, and the shape of $h$ lets us map when each estimator invents or erases curvature.

Estimator. The estimator learns $\lambda(W)$ from three proxy channels:

  • a text encoder over the firm's earlier solo filings;
  • a graph embedding of the firm's prior-window ties, trained with the joint reconstruct-and-predict loss of the Veitch et al. paper;
  • governance links where available.

It then fits a cross-fitted, doubly robust continuous-dose estimator of $\mu(s)$. This combines a flexible outcome model and a model of the exposure's conditional density, each trained on held-out folds. Curvature is tested with a formal inverted-U test on the estimated curve, with confidence bands from a firm-level block bootstrap. We use the bootstrap because, as the file notes, embedding deconfounders lack closed-form asymptotic variance.

An overlap diagnostic reports how much of the variance in $S$ is explained by $\lambda(W)$. It is reported for every estimate. If it exceeds a preset threshold, the estimate is declared unidentified rather than reported.

Real-data re-analysis. We then re-estimate the published relationships, choosing outcomes that avoid the mechanical coupling:

  • forward citations to the firm's solo patents;
  • distinct IPC classes entered through solo filings, the breadth measure of the biopharma faultlines paper.

The Harvard USPTO Patent Dataset contains rejected and abandoned applications alongside granted ones. Where assignee information permits, it lets us rebuild networks from applications rather than grants and check whether the curve depends on conditioning on grant success.

Baselines

Current practice (the naive reference). This is the specification used across the innovation-network papers: pooled regression of patent output on the structural measure and its square, predictors lagged $t-5$ to $t-1$, firm size, age, R&D, and density controls. It is the reference because it produced the claim under test.

  • The agricultural paper reports an inverted-U for subgroup structure.
  • The biopharma paper (509 firms, 1,798 observations) reports a positive linear effect on search depth and an inverted-U on breadth.
  • The file gives no exact coefficients for either: no reported result found on Lacuna for the subgroup-structure quadratic coefficients.
  • The nearest reported figure in this literature is the structural-hole coefficient of 1.844 in A STUDY ON THE INFLUENCE OF KNOWLEDGE NETWORK STRUCTURE EMBEDDING ON ENTERPRISE INNOVATION PERFORMANCE, which we re-estimate under the same pipeline.

We run this baseline exactly as published on identical samples.

Firm and year fixed effects with lags. This is the strongest classical competitor. It removes every time-invariant firm trait without any learned proxy. It receives the same samples and covariates.

Same-graph embedding deconfounding. This is the Veitch et al. estimator applied naively, with the embedding learned from the exposure-window graph. It is the strongest existing method for latent homophily, with 1.17 against a true 1.0 on Pokec in that paper. It gets the same embedding dimension grid, the same training budget, and the same dose-response head as our method. The only difference is the input channel.

Flexible outcome model under interference. This is CauGramer from Causal Graph Transformer for Treatment Effect Estimation Under Unknown Interference. It aggregates multi-hop neighbor features with cross-attention and reports 28% to 80% lower effect error than baselines on BlogCatalog and Flickr. We use it as the outcome model inside the doubly robust estimator to check whether spillovers across subgroups change the curve.

Ablations. Two ablations isolate the key ingredients. The first swaps the proxy channel, comparing same-window graph, prior-window graph only, text only, and all channels. The second swaps the outcome between network-coupled patent counts and solo-patent outcomes.

What counts as beating the baselines. On the benchmark, our method must do both of the following:

  • keep the false-inverted-U rate near the nominal test level when $h$ is flat and confounding is strong, where we expect the naive regression and same-graph embedding to fail;
  • recover a planted inverted-U's turning point with lower curve error than fixed effects when the confounder drifts over time.

Why not just…?

Why not just add firm fixed effects? This is the real threat. If homophilous traits are stable within firms, fixed effects remove them and our machinery adds nothing. The benchmark settles it: we include settings where $C_i$ drifts across windows, such as a strategy shift that changes both partners and output. In the real data, the decisive comparison is whether fixed effects alone already flatten the inverted-U. If fixed effects and our estimator agree, we will report that the simpler fix suffices. That would still be a substantive correction to a literature that mostly uses pooled lagged regressions.

Why not just stratify by technology class and firm size to check for Simpson's paradox? Beyond Simpson's Paradox: A Cascade of Confounders in AI Agent Pull-Request Co-Authorship shows how far stratification can go. Its pooled −26.0 point gap in pull-request merge rates turned into per-agent effects that largely vanished within repositories; for example, Devin went from +33.5 to +1.6 points, p=0.73. We adopt this as a required robustness check across IPC sections and size strata. However, stratification only handles observed groupings. It cannot address sorting on tacit capability, which the agricultural paper itself lists as omitted from patent data.

Why not just use the embedding of the collaboration graph, as Veitch et al. do? In this setting the exposure is computed from that same graph. An embedding good enough to capture homophily can largely reconstruct faultlines and subgroup counts, leaving no variation to estimate a curve from. The proxy-channel ablation and the overlap diagnostic test this directly. If the same-graph embedding retains overlap and matches our estimator, channel separation is unnecessary and we will say so.

Why not just propensity-score match on observed firm covariates? Matching across continuous doses works under full observability, as the non-bipartite matching study in Understanding the Effects of the Neighbourhood Built Environment on Public Health with Open Data shows. That paper also notes its dependence on all confounders being observed. We run matching on observed covariates as a further baseline in the benchmark. We expect it to behave like the naive regression whenever $C_i$ is unobserved.

Proposed Experiments and Decisive Checks

The first decisive check is the benchmark's false-discovery map. Across confounding strengths, with a flat planted $h$, we measure how often each estimator's inverted-U test rejects. If the published specification rejects far above nominal, the literature's curve is not evidence of an optimal subgroup structure by itself.

The second is the real-data re-analysis. It uses the biopharma Derwent cohort, the Chinese electronics panel behind the structural-hole result, and HUPD application-based networks. We ask whether the inverted-U in breadth and innovation survives three separate changes, each reported on its own:

  • separated-proxy adjustment;
  • solo-patent outcomes;
  • application-based rather than grant-based networks.

The third check is internal. We report the overlap diagnostic and the block-bootstrap bands for every estimate. An estimate where the proxies nearly determine the exposure is reported as unidentified, not as a null effect.

A surviving curve with a stable turning point would give managers a defensible target for subgroup structure. A collapsing curve would show that a widely repeated design rule reflects who partners with whom.