Women are steadily underrepresented as last authors, the byline slot that signals the lab head. In NLP, women were about 25.5% of last authors against roughly 30% of authors overall, and in cardiology COVID-19 papers they were 16.4% of senior authors. A common response is that the gap "goes away" once you control for productivity. That move was decisive for self-citation, where adjusting for prior publication count shrank a claimed >50% male excess to about 2%. This proposal argues that the same move is unsafe for senior authorship, and it builds the tool to say how unsafe.
The reason is a structural one. Early-career productivity is itself shaped by things bibliometric databases never record: lab budget, protected research time, leadership encouragement, and a partner at home. Survey evidence says these favor men. The same unrecorded factors also drive who later becomes a lab head. "Controlling for productivity" therefore compares a woman and a man with equal output who may have very different hidden support. That comparison can create or hide a gap in either direction. We propose to estimate the direct effect of perceived gender on reaching senior authorship, holding early productivity fixed, as a range rather than a single number. The range is indexed by how strong the hidden institutional advantage is allowed to be. We then anchor that strength against a piece of institutional capital that bylines do reveal: who a researcher's early mentor was.
Core Claim
The question "is there a senior-authorship gap beyond what productivity explains?" is a question about a direct effect through a mediator. A mediator is an intermediate variable on the path from cause to outcome, here early-career output. The question is threatened mainly by hidden confounding between that mediator and the outcome, not by confounding of gender itself. Existing sharp sensitivity tools explicitly assume this kind of confounding away. Sharp Bounds for Generalized Causal Sensitivity Analysis bounds mediated effects only under no unobserved mediator–outcome confounding. Bibliometric practice ignores it entirely. Our claim has three parts. First, by treating the mediator as the "exposure" in a second-stage sensitivity model, we can obtain sharp bounds on the controlled direct effect of gender on time-to-seniority. Second, the bound's strength parameter can be calibrated empirically from mentor identity extracted from bylines. Third, the resulting picture is substantively different from regression adjustment. We expect productivity-adjusted regressions to understate the direct gap, because men with a given output level had on average more hidden help reaching it.
Formally, let $A$ be name-inferred gender, $X$ pre-career covariates (entry year, subfield, country of first affiliation), $M$ early-career trajectory (first-author count, venue mix, and collaborator count in academic years 1–5), $U$ unrecorded institutional support, and $Y$ the indicator of reaching $t$ last-authored papers within 15 academic years. The target is the controlled direct effect $\mathrm{CDE}(m) = E[Y(a{=}\text{f}, m)] - E[Y(a{=}\text{m}, m)]$, the gap if both groups were held at the same early trajectory $m$. We take $A$ as unconfounded given $X$, since it is fixed before the career begins. We let $U$ confound $M$ and $Y$ within each gender, constrained by
$$\Gamma^{-1} \le \frac{p(m \mid a, x, u)}{p(m \mid a, x)} \le \Gamma ,$$
so $\Gamma = 1$ means early output carries no hidden-support signal and larger $\Gamma$ allows stronger hidden selection into productivity. This is the marginal-sensitivity form of Sharp Bounds for Generalized Causal Sensitivity Analysis, applied with $M$ in the role of the treatment, separately within each gender stratum. That reuse is what makes sharp bounds on $\mathrm{CDE}(m)$ available without new identification theory. The deliverable is the bound curve over $\Gamma$, together with the "breakdown" $\Gamma^\ast$ at which the bound first includes zero.
Positioning Against Prior Work
The motivating evidence comes from three descriptive studies that establish a prestige gradient but make no causal decomposition. Gender Gap in Natural Language Processing Research documents the last-author deficit in the ACL Anthology. The glass ceiling in NLP shows women take longer to reach $t$ last-authored papers, with a significant gap at $t=10$ ($p=0.03$). Gender Disparities in Cardiology-Related COVID-19 Publications reports the same slope in medicine.
Identifying Positive and Negative Factors That Affect the Promotion of Clinical Faculty supplies the substantive reason to fear mediator–outcome confounding. Its survey of 58 faculty found lack of leadership encouragement cited by 74.1% of women versus 38.7% of men. Men also benefited more from stay-at-home partners and protected time. These are exactly the variables that would raise both early output and later lab-head status.
The work we most directly challenge is Self-citation is the hallmark of productive authors, of any gender. It logistically regressed self-citation on gender while conditioning on prior publication count, byline position and venue. That is the right instinct for removing career-length artifacts. But it implicitly assumes publication count has no hidden common cause with the outcome. For self-citation this assumption may be harmless. For promotion to lab head it is the crux.
Exploring the Confounding Factors of Academic Career Success predicts IEEE/ACM Fellow election with Transformers over node2vec and graph features. It reports a "gender effort gap" in which female Fellows had out-published male ones. That finding is conditioned on the outcome, the people who were elected. The paper itself flags causal analysis as future work. We treat it as a target for reanalysis rather than as a method.
On the methods side, three papers supply machinery we reuse or contrast against. The sharp-bounds paper above supplies the bound computation. Auditing Fairness under Unobserved Confounding supplies the reporting convention of comparing the breakdown parameter to the confounding strength of an observed covariate. Its racial Paxlovid gap was robust to $\gamma = 1.12$, which exceeded the strength of diabetes. Causal Fairness under Unobserved Confounding bounds path-specific unfairness under the same sensitivity family. It focuses on training fair predictors and keeps confounding on the treatment side; we audit a population quantity and put the confounding on the mediator. Using Embeddings to Correct for Unobserved Confounding in Networks is the strongest point-identification competitor and is discussed under baselines. Finding and Listing Front-door Adjustment Sets constrains the design. Front-door identification needs a mediator that $U$ does not touch, and early output is the opposite of that, so we do not pursue it.
Research Program
The first component is a career panel built from the ACL Anthology. The file reports about 44,894 papers, and the cleaned subset of The glass ceiling in NLP covers 15,609 researchers. For each author we record entry year, early trajectory $M$, and time to $t \in \{1, 3, 5, 10\}$ last-authored papers. Attrition counts as not reaching seniority, so the outcome is defined for everyone and the total gap needs no survival assumption. A secondary, persistence-conditioned estimand gets its own sensitivity parameter and is clearly labeled as a selected population. Gender is name-inferred at a high-confidence threshold. Misclassification is handled by re-running the analysis under published-style error rates that differ by name origin. We exclude names the tools cannot classify and report how that exclusion shifts the sample by country, given the known failures on East Asian Romanized names.
The second component is calibration of $\Gamma$, which is where bylines give leverage that clinical audits lack. The last author on a researcher's earliest first-author papers is usually their mentor. That mentor's characteristics are observed: prior last-author count, citation level, network centrality, and gender. These are a measurable slice of $U$. Following the benchmarking logic of Auditing Fairness under Unobserved Confounding, we drop mentor features from $X$ and compute the implied $\Gamma$ they exert on $M$. That gives a concrete yardstick: "the hidden advantage would need to be $k$ times as strong as having a well-cited mentor." Mentor homophily, documented in Voices of Her, means mentor features may partly lie on the causal path from gender. We therefore report calibration both with and without mentor gender.
The third component is estimation. Nuisance models for $p(m \mid a, x)$ and $E[Y \mid a, m, x]$ are fit with cross-fitting. Bounds come from the sharp-bounds procedure for discrete and continuous treatments, applied stratum-by-stratum in $a$, with bootstrap intervals on the bound endpoints. The fourth component is replication in a second field. We run the identical pipeline on Author-ity PubMed records, the dataset behind the self-citation study. Running it there lets us show, on that paper's own data, how its adjustment strategy behaves once the outcome is senior authorship instead of self-citation.
Baselines
The naive reference point is the unadjusted gender difference in the probability of reaching $t$ last-authored papers within 15 years, stratified only by entry cohort. It is the total effect under our assumptions and is a legitimate quantity in its own right. We report it alongside everything else so readers can see how much of it each method attributes to the direct path. No reported result was found on Lacuna for this exact quantity on the ACL Anthology. The nearest number is the significant mentor gap at $t=10$ ($p=0.03$) in The glass ceiling in NLP.
The most influential competitor is productivity-adjusted logistic regression in the style of Self-citation is the hallmark of productive authors. It is the strongest baseline because it is what the field actually does, and because it has a famous success: shrinking a >50% self-citation gap to 1.9% (first authors) and −2.1% (last authors). We run it on the same panel with the same covariates $X$ and $M$, read off the gender coefficient, and convert it to a risk difference comparable to $\mathrm{CDE}(m)$.
The strongest point-identification method is the embedding approach of Using Embeddings to Correct for Unobserved Confounding in Networks. It learns semi-supervised co-authorship embeddings as stand-ins for $U$ and feeds them to a doubly robust weighting estimator. On Pokec it moved a severely confounded estimate from 4.51 to 2.03 against a truth of 1.0. We apply it with $M$ as the exposure, using embeddings of the year-1–5 co-authorship graph so it addresses the same mediator–outcome confounding we target. It gets the same cross-fitting folds and the same hyperparameter search budget as our nuisance models.
We also run off-the-shelf mediation bounds from Sharp Bounds for Generalized Causal Sensitivity Analysis, which place confounding only between gender and outcome. This isolates our key ingredient, moving the sensitivity model onto the mediator. A second ablation removes mentor-based calibration and reports bounds over a generic $\Gamma$ grid. That shows whether the calibration, not just the bounding, changes conclusions.
In semi-synthetic tests, "beating" a baseline means our intervals cover the true $\mathrm{CDE}$ at nominal rate across injected confounding strengths, while the regression and embedding estimates miss it. On real data, the contribution is a breakdown value $\Gamma^\ast$ expressed in mentor-benchmark units. It is informative if it falls clearly above or below the calibrated range, rather than straddling it.
Why not just…?
Why not just control for career age and publication count, as the self-citation study did? Because for this outcome that control is the problem. Publication count is caused by gender-differential hidden support, so conditioning on it compares unequally supported people as if they were equal. The semi-synthetic experiment below plants exactly this structure on the real ACL graph and measures the regression's bias as a function of injected strength. If the bias stays negligible at the calibrated $\Gamma$, the regression is vindicated and we will say so.
Why not just use co-authorship embeddings as a proxy for hidden support? This is the real threat. Institutional capital plausibly shows up in who you collaborate with, and Using Embeddings to Correct for Unobserved Confounding in Networks gives a consistent estimator when homophily captures $U$. Two cautions apply. Early collaboration networks are themselves downstream of gender, a form of mediator contamination. And even on Pokec, the method's estimate remained double the truth. We test it in two semi-synthetic regimes: one where $U$ drives collaboration choices and one where it drives only funding and time. If embeddings recover the truth in both, they become the preferred point estimate, and our bounds serve as a robustness envelope around it.
Why not just report the total gap and stop? The total gap needs no hidden-confounding correction and we will report it. But the policy argument in this literature is explicitly about the residual gap beyond productivity. That is a direct-effect claim, and it is currently being settled by unadjusted regressions.
Why not plug in a learned latent confounder, as PriMeD does for health records? Mitigating health disparities in EHR via deconfounder reported small gains (AUROC 0.7013 vs 0.6984). Its validity rests on an untestable assumption that the latent factor captures all confounding. Our design reports what follows if that assumption fails by a measured amount.
Proposed Experiments
The decisive check is semi-synthetic. We keep the real ACL Anthology authors, bylines, gender labels and covariates. We generate a hidden support variable $U$ that differs in mean by gender, with strength set to multiples of the calibrated mentor benchmark. We then regenerate $M$ and $Y$ from fitted models plus $U$, with a known true $\mathrm{CDE}$, including a null-direct-effect setting. We report bias and interval coverage for the naive gap, the regression, the embedding estimator, gender-side bounds and our mediator-side bounds. We expect regression bias to be signed, typically making the direct gap look smaller, and to grow with $U$'s strength. We expect our intervals to retain coverage up to the specified $\Gamma$. Failure of coverage at moderate $\Gamma$ would falsify the method.
On real ACL data, we report bound curves for $t \in \{1,3,5,10\}$ and the breakdown $\Gamma^\ast$ in mentor units, and we repeat the analysis on PubMed. Two findings would each settle something. A large $\Gamma^\ast$ means hidden support far beyond the mentor benchmark would be needed to erase the direct gap, so the gap is robust. A $\Gamma^\ast$ near 1 means existing "beyond productivity" claims are not identified from bylines.
A final check reanalyzes the IEEE/ACM Fellow "effort gap" from Exploring the Confounding Factors of Academic Career Success. We simulate what productivity difference would appear among elected Fellows under a zero direct effect but gender-differential $U$. That tests whether the reported gap could be produced purely by conditioning on election. The authors report F1 of 85.6% (ACM) and 84.2% (IEEE) for their predictor, but no causal estimate to compare against.
References
- Sharp Bounds for Generalized Causal Sensitivity Analysis
- Gender Gap in Natural Language Processing Research: Disparities in Authorship and Citations
- The glass ceiling in NLP
- Gender Disparities in Cardiology-Related COVID-19 Publications
- Identifying Positive and Negative Factors That Affect the Promotion of Clinical Faculty at the Wayne State University School of Medicine: Does Gender Matter?
- Self-citation is the hallmark of productive authors, of any gender
- Exploring the Confounding Factors of Academic Career Success: An Empirical Study with Deep Predictive Modeling
- Auditing Fairness under Unobserved Confounding
- Causal Fairness under Unobserved Confounding: A Neural Sensitivity Framework
- Using Embeddings to Correct for Unobserved Confounding in Networks
- Finding and Listing Front-door Adjustment Sets
- Voices of Her: Analyzing Gender Differences in the AI Publication World
- Mitigating Health Disparities in EHR via Deconfounder