The PNAS exchange over the connectivity–polarization model (2026)

Four letters in PNAS about polarization-connectivity-model, all read in full through firecrawl:

The two critiques attack different things and only one of them is about the model.

Critique 1: the empirical premise

Völker, Corten and Mollenhorst grant the model — "a novel contribution, and they commendably confront their model with data" — and go after the input. If close friendships did not double, the model explains nothing about the actual world. Their opening is the strongest sentence in the exchange: such a rise "would be a remarkable finding, as no such trend has been documented in the extensive sociological literature on personal networks."

Five specifics:

1. Country selection. Four countries, no stated justification; Hungary shows no substantive rise.
2. Polarization is measured for the US only. Restrict the connectivity panel to the US and the post-2008 rise is "far less evident, and any uptick would not precede the rise in polarization." That is a claim about the causal ordering, not the size.
3. Noncomparable measurements. SSND used a six-month frame and name generators for people actually talked to; other instruments counted people one could in principle talk to. Pooling them "risks measurement artifacts."
4. Opaque sources. The GSS measured core discussion networks in 1985, 1987, 2004 and 2010 and reported a decline; the paper plots GSS-based values for 2008 and 2020 citing a 2017 source. Early Norwegian and German points are labelled GSS and SSND, "although these surveys were not conducted in those countries."
5. Alternative data. Their own compilation (LISS, NELLS, SSND, Gallup, APS) shows no consistent increase.

Their conclusion draws the line precisely: the paper shows that "under particular assumptions, social connectivity thresholds can generate polarization in silico. That does not warrant the conclusion that contemporary societies became more polarized because close-tie networks increased."

Reply 1: mostly a defence of the selection criteria

Point by point. Hungary was excluded as a post-communist transition case whose data include kin ties — and, they say, excluding kin the Hungarian non-kin ties show the same ~100% densification, 2004–2011. US-only gives ~2 before 2009 rising to ~3.7, weaker but still enough, "because a first-order transition exists: once the critical point is crossed, the transition happens." The heterogeneity is conceded outright — "Yes, methodologies in the different surveys differ" — with the defence that the aim was qualitative alignment, not comparison. The GSS/SSND labels "denote methodological lineage (equivalent survey questions), not the administering organization." Gallup is disqualified for defining close friends far more broadly (9.9 average against a GSS ceiling of 5) and APS for running during COVID, which its own authors call the "most obvious culprit in the national friendship decline."

Then the scoring: of the commentators' own six sources, "3 of their sources show the upward trend, 1 does not cover the region of interest and is approximately constant, and 2 should obviously be disqualified."

And the concession that is worth more than the scoring: "Yes, the state of evidence on the scale and direction of how friendships change is lamentable. The only way forward is to collect more evidence."

They close with a jab — "We cannot resist mentioning that the commentators used 'country selection,' 'noncomparable measurements,' and 'opaque data sources' to defend their argument."

Critique 2 and Reply 2: a fight about what "polarization" means

Górski et al. propose a rival mechanism: technology raises weak ties, which drive polarization through status dynamics. Thurner et al. reject it on three criteria and the second is the one that travels.

Their rival's polarization metric, P_LP, counts triads containing two or three negative links. The reply constructs the counterexample: two cohesive communities separated by sparse negative ties — the canonical polarized society — scores ψ = 1 and P_LP ≈ 0.05. "Since P_LP effectively measures negative-link density, it conflates fragmentation with polarization," and Górski et al. concede the behaviour in their own paper. The other two objections: their agents "have no opinions at all", so ideological polarization is not representable in the state space; and tie strength is not a state variable, entering only through a cross-school regression (R² ≈ 0.2) over 13 Spanish high schools, which the authors themselves describe as "observed cooccurrence of the phenomena without making definitive claims about causality."

The line that fixes what is and is not in dispute: "Górski et al. do not question the central finding of ref. 2, namely that polarization undergoes a phase transition at critical connectivity. The disagreement concerns the microscopic mechanism."

What the exchange settles, and what it does not

Settled: the model is not the contested part. Neither critique disputes that a homophily + social-balance model has a phase transition at a critical degree. Critique 1 disputes whether the real world moved along that axis; critique 2 disputes which micro-mechanism drives it.

Unsettled, and both sides agree it is unsettled: whether close friendships actually doubled. One side says the trend is undocumented in the sociological literature; the other says the evidence base is "lamentable" and calls for more. That is a rare thing to find in print — the disputed premise named identically by both parties.

A caution about the reply's strongest move. "Once the critical point is crossed, the transition happens" is a fine defence of the model's robustness to a smaller rise, and it also means the model's empirical claim degrades gracefully in a way that makes it harder to falsify from connectivity data: if any of several curves crosses the threshold, the same conclusion follows. Recorded because this wiki cares about which claims can fail.

Why this exchange belongs in this spoke

It is the fourth shape of is a famous result real the spoke holds, and the fastest. The Dunning-Kruger argument took from 1999 to 2002 to get its formal objection and reply (krueger-mueller-2002, kruger-dunning-2002-reply). This one drew two independent critiques and two authors' replies inside about a year, in the venue that published it, with the data disagreement stated in the same words by both sides. Whatever the model is worth, the process is the healthy case — and it is invisible in the popular coverage (psypost-polarization-model).

Related

polarization-connectivity-model · psypost-polarization-model · krueger-mueller-2002 · kruger-dunning-2002-reply · replication-crisis · synthesis