ETD

• centrality
• growth
• international trade network
• null models
• panel regression

A growing literature finds that a country’s centrality in the International Trade Network has explanatory power for per capita income beyond aggregate openness. Standard inference asks whether the centrality coefficient differs from zero. This thesis asks a stricter question: whether the observed coefficient is also distinguishable from the distribution of coefficients obtained on random networks that preserve selected local features of the observed ITN. If the coefficient is reproduced by such networks, centrality is largely consistent with information already captured by the extensive and intensive scale of trade. If it is not, the coefficient contains variation that is not reproduced by the selected local-margin constraints and is therefore consistent with a role for higher-order bilateral structure: who trades with whom, how reciprocal flows are and how strongly countries are embedded in regional or global value chains.
The empirical analysis uses BACI bilateral trade flows over 2000–2019 and a balanced
panel of 189 countries. Sixteen specifications are estimated across direction, network type, outcome and the inclusion of a lagged dependent variable. Each specification is re-estimated M = 100 times on samples from a suitable maximum-entropy null model: DCM (Squartini, Fagiolo and Garlaschelli, 2011a) for the binary network and CReMA (Parisi et al., 2020) for the weighted network.
The results show a clear binary/weighted asymmetry. Binary coefficients are approximately reproduced by DCM. Weighted coefficients are not reproduced by CReMA: in every level specification, the observed coefficient lies above the entire null distribution.
The topological diagnostic is consistent with this pattern. The weighted null reproduces the binary topology and the strength sequence, but misses weighted clustering, the empirical distribution of bilateral weights, weighted reciprocity and the upper tail of Katz-Bonacich centrality. Weighted reciprocity remains roughly twice as large in the data as in the adjacency-preserving null model throughout the sample. These gaps indicate that the weighted centrality-income association is not reducible to local margins under the constraints considered here. The contribution is methodological, by proposing a null-model benchmark for network regressions; empirical, by documenting a binary/weighted asymmetry in ITN centrality-income regressions; and conceptual, by clarifying how centrality coefficients in growth empirics should be interpreted once they are benchmarked against null models with explicitly chosen constraints.