Does the Global "Brain vs. Brawn" Hierarchy Dissolve Over a Career?

International co-authorship in science has a well-documented division of labor. Researchers from scientifically advanced countries tend to hold the last-author slot, which usually signals the person who conceived and supervised the work. Researchers from developing countries tend to hold the first-author slot, which usually signals the person who did the hands-on work. What we do not know is whether this is a stage or a trap. One possibility is that a young researcher from a developing country does the operational work early and moves into leadership as their reputation grows. The other is that researchers from these countries stay in support roles at every career age. The evidence so far pools papers across years and compares co-authors within the same paper. That design cannot tell a pipeline from a ceiling, because it never follows the same person over time.

This proposal follows individual researchers across their publication histories. For each researcher it estimates a starting point (how likely they are to lead a paper early in their career) and a rate of change (how quickly that likelihood rises). The tool is a mixed-effects model, a regression that gives every researcher their own intercept and slope while still estimating population averages. The design also has to handle a problem that pooled analyses ignore. Researchers who stop publishing internationally, or leave science, are probably not a random subset. If the researchers who would have stayed in support roles are the ones who exit, the surviving senior researchers will make the ceiling look like a pipeline.

Core Claim

The claim is that a longitudinal model of authorship roles can separate three things: the gap in who leads at the start of a career, the gap in how fast leadership is gained, and the spread of individual slopes within each country tier. The model has to include selective exit and cross-country migration, and it has to keep a within-paper comparison so that paper quality does not confound the result. Two outcomes would each be informative. A tier gap in intercepts with converging slopes means the hierarchy is a pipeline with an entry delay. Parallel or diverging slopes mean the hierarchy is a structural ceiling. A large spread of slopes within a tier would mean the pipeline exists but is open only to a minority. That last quantity, the fraction of researchers whose estimated trajectory crosses into regular leadership within a fixed horizon, cannot be recovered from the pooled paper-level design. It is also the policy-relevant one.

Relation to Prior Work

The motivating result is Persistent Hierarchy in Contemporary International Collaboration. Using paper-level fixed effects on 17.3 million Web of Science papers from 2008 to 2020, it shows that researchers from Developing and Lagging countries have 37% and 46% lower odds of holding strategic roles. It validates authorship position against contributor-role statements in about 4,200 PLOS papers. We keep its within-paper comparison and its four-tier country classification. What we add is the individual researcher as a unit observed repeatedly over time, which that design conditions away.

Two other scientometric studies are adjacent but model different things. Collaboration Diversity and Scientific Impact computes cross-sectional probability ratios for multi-institution teams reaching top-cited status. The Effects of Remote Working on Scientific Collaboration and Impact tracks population-level distance and impact trends in arXiv and OpenAlex. Neither models individual careers or authorship roles. Do Researchers Benefit Career-wise from Involvement in International Policy Guideline Development? does follow individuals, using embedding-based matching for 101 WHO guideline authors. Its exposure, however, is a single binary event, whereas international collaboration is continuous and recurring.

The method comes from longitudinal epidemiology. Characterising the genetic architecture of changes in adiposity during adulthood using electronic health records separates baseline BMI from BMI slope with mixed models and finds that the two have quite different genetic architectures: genetic correlation 45.2%, heritability of about 25% for baseline versus 1.98% for change. The same paper warns that people who appear often in health records are systematically different, which is the analog of our frequent-publisher problem. Family Poverty Affects the Rate of Human Infant Brain Growth shows a similar pattern in which groups start at comparable levels and diverge through different velocities. That is exactly the kind of result a pooled snapshot would misread. Finally, Longitudinal Citation Prediction using Temporal Graph Neural Networks shows that author history strongly predicts outcomes. It is built for prediction accuracy rather than for decomposing variance, so it constrains our choice of tool rather than competing on our question.

Research Program

The data source is OpenAlex, the open bibliographic database already used by Venturini et al., restricted to papers with authors in at least two countries. Each researcher's origin tier is set by the country of their first affiliation, using the Miao et al. tiers. Current affiliation country enters as a time-varying covariate. Without this separation, a Lagging-country researcher who moves to an Advanced-country institution would look like a "pipeline" success when the change is really migration.

Career age is measured as years since first publication. The outcome for researcher $i$ on paper $j$ at career age $t$ is whether they hold the last-author slot. We restrict to fields where author order is not alphabetical, and we flag alphabetical-order papers automatically and drop them. The primary model is a cross-classified logistic mixed model:

$$\operatorname{logit} P(y_{ij}=1) = \beta_0 + \beta_1 t + \beta_2 g_i + \beta_3 g_i t + x_{ij}^\top \gamma + b_{0i} + b_{1i} t + c_j$$

The symbols are defined as follows. The indicators $g_i$ encode origin tier. The vector $x_{ij}$ holds time-varying covariates: current country, team size, number of Advanced-country co-authors, and field-by-year terms. The pair $(b_{0i}, b_{1i})$ is researcher $i$'s random intercept and random slope. The term $c_j$ is a paper random intercept that plays the role of Miao et al.'s paper fixed effect. The pipeline-versus-ceiling question is carried by $\beta_3$. How open the pipeline is depends on the variance of $b_{1i}$ within each tier.

To handle selective exit, we fit a shared-parameter joint model. Each researcher's hazard of ceasing international publication for five or more years depends on the same $(b_{0i}, b_{1i})$:

$$h_i(t) = h_0(t)\exp(\alpha_0 b_{0i} + \alpha_1 b_{1i} + z_i^\top \delta)$$

Here $h_0$ is a baseline hazard, $z_i$ holds tier and field, and $\alpha_1 \neq 0$ means that people with flatter leadership trajectories exit at different rates. This assumes that exit depends on the latent trajectory rather than on unobserved shocks. We treat that as a named assumption and probe it with a sensitivity analysis. For scale, we fit the model on stratified samples of researchers within field-by-tier cells, with a Bayesian implementation using variational initialization. Because full estimation over millions of researchers is expensive, we check stability across independent subsamples instead.

Baselines

The naive reference point is pooled logistic regression of last-authorship on tier, career age, and their interaction, with standard errors clustered by author. It uses the same papers and covariates. Any conclusion that changes between this model and ours shows what the longitudinal structure contributes. Agreement would show that the simple model was adequate for the mean trend.

The strongest existing method is the Miao et al. paper-fixed-effects logit, which reported 37% and 46% lower odds of strategic roles for Developing and Lagging researchers. We extend it with tier-by-career-age-bin interactions. We first reproduce its tier odds ratios on our OpenAlex sample as a data sanity check. Close agreement is needed before any comparison can be trusted. We then run it on the identical analysis sample with the same covariates. The joint model beats it if either of two things happens. First, the corrected tier-by-age slope differs from the age-binned fixed-effects estimate by more than their combined uncertainty, and a simulation with planted selective exit shows that the fixed-effects estimate is the biased one. Second, the joint model delivers the within-tier slope distribution, which the fixed-effects model cannot estimate at all.

A third baseline is an author-fixed-effects linear probability model with tier-by-career-age interactions. It identifies differential mean slopes without assuming that random effects are independent of covariates.

Two ablations isolate the key ingredients. The first removes the dropout submodel, leaving a mixed model with random slopes only, and tests whether the exit correction changes the answer. The second removes the current-country covariate and tests how much apparent upward mobility is really migration. All models share the same sample, covariates, and outcome definition. Each gets a comparable specification search, fixed in advance: the same set of field controls and the same career-age spline options.

Why not just…?

Why not just add career-age interactions to the Miao et al. paper-fixed-effects model? This is the most serious alternative and it is included as a baseline. Its weakness is that each career-age bin contains different people. Senior Lagging-tier researchers still publishing internationally are survivors, and entry cohorts differ, so a cross-sectional age gradient mixes within-person change with changes in who is in the sample. The planted-exit simulation described below quantifies how large that bias can get at realistic exit rates. If, on real data, the age-binned fixed-effects slope matches the joint-model slope, we will report that the simpler design suffices for the mean. The slope-dispersion result would still stand.

Why not just use author fixed effects with tier-by-career-age interactions? For the average differential slope this is a legitimate competitor. It is also more robust to correlation between individual ability and covariates, which is why it is a baseline. It still conditions on researchers observed at multiple ages, so it shares the survivor problem. It also yields no distribution of individual slopes. Where its mean slope and ours disagree, a Hausman-style comparison indicates whether the random-effects assumptions are doing damage. In that case we would report the fixed-effects mean and keep the mixed model only for dispersion.

Why not just model time to first last-authorship with survival analysis? It is simple and answers part of the question. However, a single last-authored paper does not show that a researcher now leads regularly, and time-to-first-event treats one "brain" paper followed by permanent return to support roles as a success. We fit the survival model anyway. The two approaches should agree on ordering, with Lagging-tier researchers taking longer. If our model's crossing fraction and the survival model's cumulative incidence diverge sharply, that divergence measures how often leadership fails to persist.

Why not just predict roles with a temporal graph neural network? The GCN plus LSTM model of Holm et al. reaches roughly 0.75 to 0.82 mean absolute error on log citations, but our question is about an interpretable contrast and a variance component rather than predictive error. A predictor would not say whether the tier gap lives in intercepts or slopes, so we do not treat it as a competitor.

Proposed Experiments and Decisive Checks

The first check is recovery on synthetic careers. We simulate author histories with known tier-specific intercepts, slopes, slope variances, and an exit hazard that depends on the slope, at publication rates and irregular spacing resampled from OpenAlex. Each estimator must recover $\beta_3$ and the slope variance. The joint model passes if its bias stays within its own uncertainty while the pooled and age-binned fixed-effects estimates drift in the predicted direction.

The second check is replication. We reproduce the direction and approximate size of the Miao et al. tier gap on our sample and confirm that authorship position tracks contributor roles in the PLOS subset they used.

The headline analysis reports the corrected tier-by-career-age slope. It also reports the within-tier slope variance and the implied fraction of researchers who reach a 50% leadership probability within fifteen career years, broken out by tier and field. These estimates are shown with and without the exit and migration terms.

The decisive falsification test concerns the Lagging tier. The pipeline reading is refuted for this tier if its corrected slope is not steeper than the Advanced-tier slope and its crossing fraction stays near zero after stripping out migrants.

As a secondary test, we check whether the post-2020 rise in remote cross-border collaboration reported by Venturini et al. shifted slopes for career cohorts entering after 2020. Because the post-2020 follow-up is short, this result is exploratory and does not count toward the main conclusion.