Gender inequality in academic promotion trajectories in Japan

Abstract

Despite growing awareness of the lack of gender equity in academia, women’s disadvantages in many aspects, including academic promotion, remain a persistent issue. Using career-trajectory data for 74,344 researchers from Japan’s researchmap platform, this study examines gender inequality in promotion trajectories across two key transitions: PhD to associate professor, and associate professor to full professor. Specifically, we examine how gender gaps in promotion timeline vary across cohorts, disciplines, and career stages, and further assess their associations with career-history and institutional factors. The results show pronounced heterogeneity over time and across fields. Gender disparities in promotion timeline are more significant in early-career transitions than in later transitions, although gaps have narrowed in more recent cohorts. Survival-based robustness checks that account for right-censoring confirm that Stage 1 remains the primary locus of gender inequality, but also indicate that completed-transition comparisons can overstate female advantage in promotion timeline in recent cohorts, especially in STEM. By foregrounding promotion timeline as a core dimension of career stratification, this study discusses how institutional structures and environments differentially shape academic career advancement by gender, providing evidence relevant to Japan and comparable academic systems.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-54562-5.

Introduction

Gender inequality remains a significant structural problem in academic careers around the world¹. Although the proportion of women attaining senior positions has increased across many disciplines in recent years², they still face significant inequalities in the promotion process³. In academia, promotion not only confers institutional rewards—such as job security, compensation, and formal authority—but also carries symbolic significance, including prestige, professional recognition, and representation within one’s field^4,5. Importantly, inequalities in promotion are reflected not only in the underrepresentation of women at specific academic ranks but also in disparities within the promotion process itself—most notably in promotion timelines and the non-linear trajectories toward advancement. From this perspective, the promotion timeline provides a process-sensitive indicator of career inequality, revealing how gendered differences in advancement emerge, accumulate, and potentially reverse across career stages.

Despite widespread interest in documenting and explaining gender disparities among academic faculty⁶, assessing the scale and variability of these gaps at the system level remains challenging because academic careers are shaped by substantial heterogeneity across disciplines, career stages, and PhD cohorts^7–9. Existing studies have often focused on specific disciplines, institutional contexts, or single promotion milestones, making it difficult to compare findings across settings and to characterize how gendered promotion timelines evolve across multiple career transitions^1,7,10,11. As a result, large-scale, nationwide longitudinal evidence on gendered differences in academic promotion timelines is still relatively scarce^7,12. This gap has important implications: without stage-specific and cohort-sensitive analyses of promotion timelines, it is difficult to identify when gender disparities emerge, whether they accumulate across successive transitions, and where institutional intervention may be most needed.

Recent scholarship across higher education systems further suggests that gender inequality in academic careers cannot be understood fully through static rank distributions alone. Qualitative and mixed-methods studies show that women’s career progression is shaped by gendered service burdens, subtle bias in evaluation, male-dominated networks, and exclusionary academic cultures^13–15. Research in STEM likewise emphasizes that early-career inequalities are often systemic, cumulative, and embedded in organizational environments rather than reducible to individual choice or performance alone¹⁶. At the same time, recent comparative evidence from Europe suggests that gender inequality may involve a two-level pattern: women may be less likely to advance overall, yet among those who do advance, promotion timeline can be shorter than for men¹⁷. Together, these studies indicate that gendered promotion processes should be examined not only in terms of whether advancement occurs, but also when it occurs, and for whom. However, evidence that jointly compares stage-specific promotion timeline patterns across cohorts, disciplines, and historical periods remains limited, especially outside Euro-American contexts.

Against this backdrop, Japan provides a particularly informative case for studying gendered academic career progression. Japan has long exhibited persistent gender inequality in academia, and OECD indicators show that it has one of the lowest shares of women among academic staff across OECD countries¹⁸. At the same time, Japan offers a valuable empirical setting because researcher careers can be traced at scale through researchmap, a widely used national researcher platform containing detailed profile information on education, appointments, and academic activities.

Another reason for focusing on Japan is that its academic ranks are relatively consistent over time, which makes career progression easier to compare across cohorts. In this study, we examine two transitions: Stage 1, from PhD completion to first appointment as Associate Professor, and Stage 2, from Associate Professor to Full Professor. We use Associate Professor as the endpoint of Stage 1 because it is a clearer and more comparable marker than lower-ranked positions. Positions below Associate Professor, including postdoctoral appointments, assistant-professor-type posts, and koshi (lecturer) roles, have varied more across institutions and historical periods in title, status, and role^19,20. Associate Professor therefore provides a clearer basis for comparison across cohorts.

Using researchmap, we identified 78,208 researchers with verifiable PhD completion years and reconstructed time-stamped promotion trajectories from standardized appointment records; the discipline-mapped analytical dataset used for the main figures contains 74,344 researchers.

By analyzing these trajectories across disciplines and cohorts, this study moves beyond static representation gaps to examine how gender inequality is structured across career stages, historical periods, and observed promotion samples. Additional details on gender inference, title standardization, and analytical sample construction are provided in the Methods and Supplementary Information.

Results

Stage-specific gender gaps in promotion timeline

We begin with a descriptive comparison of promotion timelines across two critical stages in Japanese academia: Stage 1 and Stage 2. Figure 1a illustrates the gender composition of each stage, highlighting the selective nature of the analytical samples. The analytical samples are first defined by record completeness of each transition endpoint: Stage 1 includes researchers with an observed PhD completion year and a subsequent associate-professor year, whereas Stage 2 requires an observed associate-professor year followed by a full-professor year. Among the total cohort of 74,344 PhD holders in the discipline-mapped analytical dataset, women represent 17.87% (). Relative to this baseline, women account for 18.03% of the Stage 1 sample after applying completeness criteria (3752 women vs. 17,056 men), but their share decreases to 15.85% in Stage 2 (1703 women vs. 9,044 men). For duration-based descriptive analyses, we then restrict the observed-transition samples to strictly positive promotion durations (), thereby excluding zero- and negative-duration records. Under this definition, the corresponding female shares in the final positive-duration analytical samples are 16.94% for Stage 1 (3184 women vs. 15,610 men) and 15.73% for Stage 2 (1638 women vs. 8,778 men). This pattern is consistent with differential selection into senior-rank transitions, which may reflect a combination of attrition, delayed promotion, and unobserved later-career moves.

Fig. 1.

Multi-dimensional patterns of gendered academic promotion in Japan. (a) Sample composition across career stages showing the selective representation of female scholars. (b) Violin plots illustrating the distribution of promotion timelines (years). (c) Disciplinary variations in median promotion timeline, highlighting the contrast between STEM and Humanities.

Generally, women who reach these promotion milestones exhibit shorter promotion timelines than men in both stages. Using the positive-duration analytical samples, we summarize gender differences with median gaps () and mean gaps (), and later assess whether these gaps vary systematically across cohorts and disciplines. As shown in Fig. 1b, in Stage 1, the median promotion timeline is 7 years for women and 8 years for men, yielding a median gap of year. The mean gap is smaller: women average 7.83 years and men average 8.77 years, giving years. This pattern persists in Stage 2: the median duration is 8 years for women and 8 years for men, with years, and the corresponding means are 8.22 and 8.60 years, respectively, yielding years. Taken together, these descriptive gaps are consistent with selective processes in the promotion pipeline, whereby the subset of women who reach the next rank tends to progress relatively quickly.

However, this descriptive temporal pattern is not uniform, but exhibits structured heterogeneity across disciplines (Fig. 1b, c). In the positive-duration discipline-mapped sample, STEM women and men are at near parity in Stage 1 by the median, whereas in Stage 2 women show shorter durations than men. In the Humanities, women and men are also close to parity in both stages by the median, with only modest mean differences. Interdisciplinary fields likewise show near-median parity in Stage 1 but somewhat shorter Stage 2 durations for women. This divergence indicates that the mechanisms of gendered promotion timelines are deeply intertwined with disciplinary cultures and their respective “leaky pipelines.” This structured heterogeneity motivates our discipline-stratified models with a gender-by-cohort interaction, which test whether the direction and magnitude of the gender gap change across historical PhD cohorts within each field. Next, we assess cohort dynamics in these gender gaps and their disciplinary variation.

Cohort dynamics of gender gaps

To trace the historical evolution of gender gaps in academic promotion, we adopt a cohort-based design. Our analyses focus on researchers who earned their PhDs between 1970 and 2024, a window selected to balance temporal coverage with the availability and comparability of career-history information in the underlying records. The full API pull contains earlier entries (dating back to 1900), but these cohorts are sparse and less consistently documented; the rationale for the analysis window and the corresponding sample-construction steps are reported in the Supplementary Information (see Supplementary Table S1). By grouping researchers into five-year PhD cohorts, we facilitate comparisons of career trajectories across historical periods. We treat PhD conferral as the baseline of the academic career pipeline, as it marks the point at which individuals become at risk of subsequent promotions. For each cohort, we assess gender disparities along three dimensions: (i) cohort composition at successive pipeline points, (ii) cohort-level realization of observed promotion transitions relative to the PhD cohort size, and (iii) the time-to-promotion among those with observed completed transitions.

We first examine cohort dynamics in the academic promotion timeline. As shown in Fig. 2a, the gender gap in the transition through Stage 1 displays pronounced cohort variation. In earlier PhD cohorts, women typically required substantially more time than men to attain associate professorship, reflecting a clear historical disadvantage in promotion timelines. However, across later cohorts, the magnitude of this gap generally declines and eventually reverses. Specifically, from the mid-1990s onward, the median gap narrows and in several cohorts turns negative (favoring women). However, the precision of these cohort-specific differences varies by cohort and field, particularly for recent cohorts with limited exposure time and smaller numbers of observed transitions. For cohorts completing their PhDs after 2010, the observed timeline differences are considerably smaller. However, as a large share of individuals in these recent cohorts has not yet completed the promotion, these estimates are subject to substantial uncertainty; gender differences in these latest years are better characterized as indistinguishable rather than clearly eliminated. Overall, for Stage 1, the gender gap documented in previous sections is largely driven by cohorts that entered their academic careers in earlier historical periods.

Fig. 2.

Cohort dynamics of gender differences in academic promotion (1970–2024). (a, b) Median promotion timelines (years) for Stage 1 (PhD to Associate Professor, a) and Stage 2 (Associate Professor to Full Professor, b). Lines represent median values for men (blue) and women (red). (c)Female representation funnel across cohorts, showing the share of women among overall PhD holders (black), at Stage 1 entry (green), and at Stage 2 entry (magenta). Here, “Stage 1 entry” refers to the observable Stage 1 transition sample, defined as researchers with a verifiable PhD year and an observed Associate Professor (AP) start year; it therefore reflects observed inclusion in the Stage 1 transition sample rather than the full set of researchers at risk immediately after the PhD. (d) Promotion realization rates by stage and gender. For both stages, rates are normalized by the number of PhD holders in each gender-by-cohort cell. Solid lines indicate Stage 1 transitions; dashed lines indicate Stage 2 transitions. Thus, the Stage 2 series represents the share of each PhD cohort eventually observed to complete the AP→FP transition, rather than the conditional completion rate among Stage 2 entrants. The marked decline in the 2020–2024 cohort reflects truncated observation windows rather than a contraction of opportunities. All estimates are aggregated into 5-year PhD conferral cohorts.

By contrast, gender differences in the promotion timeline for Stage 2 do not exhibit a consistent cohort-based pattern. As shown in Fig. 2b, median gender differences in the Stage 2 transition are generally small and fluctuate across cohorts without a clear directional trend. Relative to Stage 1, Stage 2 gender gaps display greater instability and show no systematic tendency toward intergenerational convergence or divergence. This suggests that gender differences in later-career promotion timelines do not primarily evolve along cohort lines, but instead reflect heterogeneity within the stage itself rather than cohort-driven structural change.

These timeline dynamics occur against a backdrop of evolving institutional representation. Figure 2c illustrates the “funnel pattern” of female representation across successive points of the promotion pipeline. We find that while the proportion of women among PhD holders has increased steadily since 1970, it remains substantially lower than that of men throughout the entire period. When the sample is restricted to individuals included in the observable Stage 1 transition sample, the female share increases modestly relative to the overall doctoral cohort but remains clearly underrepresented. This pattern suggests that the composition of observed promotion samples differs from that of the full PhD cohort, although the figure should not be interpreted as a direct measure of selection at the immediate post-PhD risk set. Figure 2c further shows that female representation declines again among those observed in the Stage 2 transition sample, indicating that relatively few women appear in the observable pathway to full professorship after earlier stages of academic filtering. Although representation at both entry points increases across cohorts, the trend is non-linear and marked by period-specific fluctuations, indicating that improvements occur unevenly across generations.

A more complete picture of these dynamics emerges when promotion rates are considered alongside the timeline. As shown in Fig. 2d, the proportions of researchers who successfully achieve promotion vary across cohorts and differ between the two stages. Importantly, both stage-specific rates are normalized by the initial PhD cohort size in each cohort-by-gender cell. Accordingly, the Stage 2 line represents the cumulative realization from PhD conferral to Full Professor status, rather than a conditional rate restricted to those who had already attained an Associate Professor position. Notably, the sharp decline in observed promotion rates for the 2020–2024 cohort reflects a truncated observation window: many individuals in these recent cohorts have not yet reached the typical time-to-promotion threshold. This decline should therefore be interpreted as a consequence of insufficient exposure time (right-censoring) rather than evidence of a structural contraction in promotion opportunities.

Disciplinary heterogeneity and cohort-by-field interactions

We estimated interaction models separately for each stage and discipline to derive cohort-specific female-to-male differences in promotion duration (measured in years). Researchers were grouped into five-year PhD cohorts, with the 1970–1974 group serving as the reference category in each stage-by-discipline model. All estimates are reported with HC1-robust 95% confidence intervals to account for potential heteroskedasticity.

Regression results indicate that gender differences in promotion timelines are most pronounced in the transition from PhD to Associate Professor (Stage 1), with substantial heterogeneity across fields and cohorts (Fig. 3). In Humanities, female effects are positive in several earlier cohorts (especially around the 1980s) and generally contract toward near-parity in later cohorts. In STEM, early cohorts show large positive female effects (on the order of about 5–6 years in the 1970s bins), followed by a steady decline across cohorts; in several post-2010 bins, the completed-transition OLS estimates become negative.

Fig. 3.

Cohort-specific marginal female effects on promotion timelines across disciplines and career stages. Points show estimated marginal effects of being female on promotion timeline (Female−Male, in years) from unweighted OLS Model 2 with female-by-cohort interactions, estimated separately by discipline (Humanities, STEM, Interdisciplinary) and stage, using positive completed-transition durations. Error bars indicate HC1-robust 95% confidence intervals; the horizontal dashed line marks no gender difference (0 years). Hatched shading marks the 2015–2019 cohorts as limited-exposure bins: with an observation cutoff at 2024-12-31, potential follow-up is approximately 9 years for 2015 PhD cohorts and 5 years for 2019 PhD cohorts. This window is substantially shorter than typical promotion-completion durations, so cohort-specific duration estimates are more susceptible to incomplete transitions and selection into completed samples.

In comparison, gender differences in the transition to Stage 2 exhibit substantially less structure across cohorts and fields (Fig. 3). In Humanities and STEM, many cohort-specific estimates cluster around zero with no sustained directional trend, although some early bins show larger point estimates and wider uncertainty because of smaller completed-transition samples. For the 2015–2019 cohorts in particular, estimates are based on shorter effective follow-up (approximately 5–9 years to the 2024 cutoff), which is shorter than the typical duration needed to complete promotion transitions; accordingly, cohort-specific estimates are more vulnerable to incomplete transitions and selection into completed samples. Interdisciplinary estimates are comparatively imprecise, with several cohort bins showing very wide confidence intervals.

Synthesizing results across both stages indicates that gender differences in Japanese academic promotion do not follow a single, uniform pattern. Instead, they reflect a structured interaction between cohort timing and disciplinary context that is most pronounced in Stage 1, where sizeable early-cohort female disadvantages contract substantially over successive cohorts. By contrast, Stage 2 effects are generally smaller and less systematically patterned across cohorts and fields. These results align with the cohort-level descriptive patterns in Section 2 while showing that the magnitude and direction of gender gaps depend strongly on both field context and cohort timing. We additionally assessed the robustness of these cohort-specific female effects to stage-specific productivity adjustment by re-estimating the Fig. 3 OLS specification with controls for papers and projects accumulated within each promotion stage. The resulting estimates show partial attenuation, especially in Stage 1 STEM, but do not alter the broader conclusion that gender differences in promotion timelines are concentrated primarily in Stage 1, whereas Stage 2 effects remain comparatively weak and show no clear systematic pattern. Full variable definitions, model specifications, and sensitivity analyses are reported in Supplementary Note 4.

We further assessed the cohort patterns using Cox proportional hazards models that treat incomplete transitions as right-censored rather than excluding them from the risk set (Supplementary Note 4). The survival-based estimates corroborate the main substantive conclusion that gender inequality is concentrated in Stage 1, but they are more conservative than the OLS estimates reported above. In particular, Stage 1 female hazards remain lower than male hazards in multiple STEM and Humanities cohorts, indicating slower progression to associate professorship once censoring is incorporated. This attenuation is especially consequential for STEM: the later-cohort female speed advantage suggested by completed-transition OLS estimates is not reproduced as a stable pattern in the Cox specification. By contrast, Stage 2 hazard differences remain comparatively weak and irregular across cohorts and fields, reinforcing the interpretation that early-career promotion is the primary stage at which gendered timing differences accumulate.

Discussion

Despite recent efforts within Japanese academia to improve women’s representation, earlier studies have continued to document persistent gender inequality²¹. Using a large-scale longitudinal dataset, this study builds on that work by showing how promotion timelines differ across career stages, cohorts, and disciplines. A key contribution of this study is to show that patterns seen in pooled descriptive results can look quite different once cohort-specific and discipline-specific variation is taken into account, and that these patterns become more cautious once right-censoring is explicitly addressed. This process-focused interpretation is also in line with recent qualitative and mixed-method research showing that gender inequality in academia is shaped not only by rank composition, but also by organizational context, evaluation practices, and informal academic structures^13–16.

First, our results demonstrate that gendered promotion differences are primarily a Stage 1 phenomenon. While the pooled descriptive analysis suggests a median 1-year speed advantage for women, this aggregate figure masks a dramatic historical reversal and should be interpreted cautiously because it is based on completed transitions only. Our cohort-stratified models show substantial female promotion delays in multiple early Stage 1 cohorts, especially in STEM, with a weaker but still visible pattern in Humanities. The apparent female advantage in pooled data is therefore not a general feature of the promotion system. Once censoring is incorporated through Cox models, Stage 1 gender gaps remain clearly present, but the later-cohort reversals become more muted, especially in STEM. In contrast, Stage 2 transitions exhibit much smaller and less systematically patterned gender differences across cohorts. This does not necessarily imply that Stage 2 is fully gender-neutral; rather, it is consistent with two non-exclusive interpretations: barriers are strongest before first senior-rank entry, and later-stage observed samples may already reflect selective retention.

Second, we show that the evolution of these gender gaps is deeply contingent on disciplinary context. In the completed-transition OLS models, STEM shows particularly large female disadvantages in earlier Stage 1 cohorts, while Humanities also exhibits positive female effects in several earlier cohorts; both fields move closer to parity in later cohorts. However, the survival-based robustness checks indicate that this convergence should not be interpreted as a stable pattern of shorter promotion timelines among women in STEM. Rather, censoring-adjusted estimates suggest that later STEM cohorts move closer to parity than the OLS results imply, but do not provide consistent evidence that women systematically outpace men once all researchers still at risk of promotion are retained in the analysis. This cohort pattern is broadly consistent with gradual institutional change in Japanese academia, though our design cannot attribute it to any specific reform^21,22. The relative instability of Stage 2 gaps across disciplines further suggests that once the initial barrier to senior-rank entry is overcome, subsequent promotion timeline differences are smaller and less structured. The visible shift between earlier cohorts and those beginning around the mid-1990s should be interpreted in the context of broader changes in Japan’s higher education and gender-equality environment. In particular, this period overlaps with the continued expansion of higher education, the increasing participation of women in advanced education, and the gradual institutionalization of gender-equality policy in Japan^23–25. These developments provide plausible historical context for the cohort pattern observed here, although the present design does not identify any single policy or reform as its direct cause. At the same time, these broader changes do not remove the need for caution when interpreting observed later-stage samples.

These patterns also need to be interpreted in light of selection into observability and survivor bias in observed transitions. We use “survivor bias” in its standard selection sense: completed-transition samples overrepresent those who remain in the system long enough to be observed at later stages²⁶. The “funnel” patterns identified in our descriptive analysis show that observable promotion samples are shaped by non-random selection: a substantial share of female PhD holders do not appear in the observed transition to associate professor, and many more recent careers remain administratively incomplete. Accordingly, descriptive or completed-transition estimates can give an overly optimistic impression of female progression speed in later cohorts. We treat this as an interpretive framework rather than a mechanism directly identified by our data. Current findings suggest that women observed in later-stage samples may not fully represent the broader pool of potential candidates for promotion, but rather a subgroup shaped by selective retention. In supplementary analyses, we partially examined this possibility by incorporating stage-specific publication and project output. However, the available data still cannot directly capture broader evaluation processes, organizational context, or informal career mechanisms, and therefore remain insufficient to fully distinguish among these possible explanations.

This study has several limitations. First, as an observational analysis of large-scale career trajectories, it does not allow for causal inference underlying gender differences in promotion timelines. Although we model promotion stages separately and examine heterogeneity across cohorts and disciplines, the core models reported in the main text do not directly incorporate the full range of mechanisms that may shape promotion outcomes. Supplementary robustness checks add selected mobility, career-path, and stage-specific productivity controls, but important factors such as family responsibilities, internal evaluation practices, informal academic networks, service burdens, and other organizational processes remain unobserved. Second, the career-trajectory data used in this study are current as of December 2024. Estimates for more recent PhD cohorts are less precise because many individuals have not yet completed later-career transitions, and these results may evolve as additional observations accumulate. Finally, gender was inferred using large language models applied to names and related information. While this approach enables coverage at scale, some classification error is unavoidable; future research could improve accuracy through stronger validation procedures or, where feasible, multi-source data linkage.

Overall, our results suggest that the temporal organization of academic careers in Japan has undergone substantial cohort change, with early-career gender gaps in promotion timelines contracting markedly over time, while later-stage promotion timeline differences remain comparatively small and less systematically patterned. At the same time, the persistent underrepresentation of women at successive stages of the academic pipeline, together with the more conservative censoring-adjusted estimates and the attenuation of some cohort-specific female effects after productivity adjustment, indicates that improvements in observed promotion timelines among those who advance do not by themselves resolve broader inequality in representation. We therefore encourage future research and policy attention to look beyond isolated promotion outcomes and consider the full set of formative processes shaping academic careers, thereby contributing to a more equitable, transparent, and inclusive academic promotion system.

Methods

Data collection

The data used in this study were acquired from researchmap, a national researcher information platform operated with support from the Japan Science and Technology Agency (JST). The platform systematically documents the academic careers and research outputs of scholars affiliated with Japanese universities and research institutions²⁷. For each registered researcher, researchmap provides a publicly accessible profile page that consolidates personal background information, degree history, research fields, academic titles, and scholarly outputs.

Using the researchmap Web API²⁸, we programmatically collected information from individual profile pages. As of July 1, 2025, the dataset comprised records for 259,908 researchers across a wide range of disciplinary fields. From these profiles, we extracted researchers’ names, primary research fields, degree types and award years, educational histories, and academic career records, including the start year of each recorded career stage. We identified 78,208 researchers with verifiable PhD completion years in the full dataset; subsequent analyses further restricted this set to the observation window and inclusion criteria described below.

To ensure a robust and consistent benchmark for analyzing promotion timelines, we restricted our analytic window to the period between 1970 and 2024. This temporal boundary was necessary to mitigate measurement biases; records prior to 1970 exhibit substantial and non-random missingness in critical career markers, such as doctoral conferral and initial appointment dates. Including these early-career records would risk systematically inflating estimated promotion durations. Against this backdrop, our cohort was refined to 74,617 researchers who obtained their PhDs within the specified window. For discipline-stratified analysis, we further mapped these individuals into aggregated categories, resulting in 74,344 researchers with valid field information. After excluding the residual “Other” category to maintain the granularity of disciplinary comparisons, the final cohort comprised 73,995 researchers.

The analysis of career progression was conducted by identifying stage-specific transitions based on the observability of appointments. In the discipline-mapped analytical dataset used for the main figures, we initially identified 20,808 researchers with observable Stage 1 transitions and 10,747 with observable Stage 2 transitions. For duration-based analyses, we further restricted these samples to strictly positive promotion durations (), yielding final analytical sets of 18,794 researchers for Stage 1 and 10,416 researchers for Stage 2. This sequential filtering ensures that our promotion timeline estimates remain stable and minimizes the noise inherent in large-scale observational career data.

Career indicator construction

The identification and validation of doctoral degrees, associate professorships, and professorships followed a unified multi-stage data-processing pipeline, the overall structure of which is illustrated in Fig. 4. This pipeline takes as input degree information and raw academic career texts provided by researchmap and integrates LLM–based structured extraction with rule-based consistency checks, semantic similarity aggregation, and systematic manual verification. Together, these procedures enable the identification and standardization of key career stages across researchers’ academic trajectories.

Fig. 4.

Overview of data processing and sample construction.

Career-text preprocessing and title extraction: For academic career records, raw curriculum vitae texts were first subjected to standardized preprocessing, including the normalization of punctuation and the removal of extraneous whitespace. We then employed GPT-5 Nano, using a fixed prompt template, to extract academic title from career descriptions in a structured format. The model was instructed to return only titles explicitly stated in the text (e.g., associate professor, professor), while deliberately ignoring preceding modifiers related to departments, disciplines, or institutions. All model outputs were returned in a strictly defined JSON structure to ensure determinism and reproducibility in subsequent processing steps.

Hard alignment constraint and semantic aggregation:To minimize noise introduced by automated extraction, a career record was retained for analysis only if the extracted title could be strictly aligned with the original appointment text or corresponding metadata. This alignment check was implemented as a hard constraint: any model output failing to meet this criterion was discarded outright and did not enter either manual correction or probabilistic weighting procedures. Titles that passed this verification step were subsequently encoded as vector representations and aggregated based on semantic similarity to consolidate cross-linguistic and synonymous expressions. All aggregated records underwent full manual review to confirm their truly associate professorship and professorship titles.

Exclusion of non-equivalent titles and final title records:To prevent honorary, temporary, or otherwise non-equivalent titles from being misclassified as formal career stages, we further traced all preliminarily identified associate professor and professor records back to their original curriculum vitae texts. Records containing markers such as honorary, acting, temporary, or assistant were systematically excluded. Following this cleaning and standardization process, the final dataset comprised 51,157 researchers associated with 51,169 associate professorship records and 44,836 researchers associated with 44,857 professorship records.

Operational definition of promotion timeline:Career-stage timeline was defined using the following operational assumptions. For any researchers, the year in which a given academic title (associate professor or professor) was first recorded was treated as the promotion year to that title. The corresponding timeline was assumed to span from the first to the last recorded year (inclusive), even if intermediate years were not explicitly documented.

Construction of stage-specific productivity measures: For supplementary productivity-adjusted robustness checks, stage-specific productivity measures were constructed from differences in cumulative output counts recorded between successive career milestones. For Stage 1, productivity was defined as the additional number of papers and projects accumulated between PhD conferral and first appointment to Associate Professor; for Stage 2, it was defined similarly as the additional number of papers and projects accumulated between first appointment to Associate Professor and first appointment to Full Professor. Because these indicators were highly right-skewed and included many zero values, they were incorporated into the robustness models in log-transformed form, łog (1+x). An additional Stage 2 sensitivity analysis excluded cases in which the PhD year postdated the Full Professor year. Full definitions and results are provided in Supplementary Note 4.

Doctoral degree identification and manual verification:Doctoral degree identification constituted a parallel branch of the pipeline. Degree titles recorded in researchmap were first encoded as vector representations and matched against prototypical semantic representations of doctoral degrees to identify candidate PhD records. All records classified algorithmically as doctoral degrees were subsequently subjected to manual verification to ensure accuracy. Through this process, we confirmed 151,446 valid doctoral degree records corresponding to 147,051 researchers.

Assignment of PhD completion year:Doctoral degree completion years were determined using two complementary approaches. First, when a degree awarded year was explicitly reported in a doctoral degree record, that year was directly adopted; in cases where multiple doctoral degrees were recorded for the same individual, the earliest completion year was used. Second, for researchers lacking an explicit degree awarded year, we matched doctoral degree–granting institution codes (provided by the researchmap Web API) to corresponding institution codes in educational history records and assigned the latest recorded end year under the matched institution as the doctoral completion year. Using these procedures, we identified doctoral completion years for 78,208 researchers, of whom 74,617 obtained their doctoral degrees between 1970 and 2024 and constitute the final analytical sample used in this study.

Parsing research field labels and initial mapping:Researchers’ disciplinary affiliations were derived from the research field labels reported in their researchmap profiles. Each label was first parsed by retaining its leading segment, defined as the substring preceding the delimiter “|”, which captures the primary disciplinary designation used by the platform. These parsed labels were then mapped to three broad disciplinary domains—Humanities, STEM, and Interdisciplinary—using a predefined lookup table reflecting their conventional disciplinary orientation.

Handling ambiguous cases and residual categories:To account for disciplinary ambiguity, profiles containing multiple field segments were classified as interdisciplinary, while the leading segment was retained as a reference main discipline. For single-segment labels, assignment followed the lookup table, with an additional rule applied to improve conceptual consistency: labels initially mapped to the sciences were reclassified as engineering if the label name contained the character denoting “engineering,” reflecting their applied orientation. Labels that could not be mapped using these rules were classified as other, and missing field information was treated as unknown.

Rationale for aggregation strategy:This aggregation strategy was adopted to support stable cohort-by-discipline comparisons. Collapsing fine-grained field labels into three broad domains reduces dimensionality and avoids sparse cells that would otherwise compromise estimation stability in interaction models. Explicitly identifying interdisciplinary cases further prevents the introduction of subjective or non-reproducible assignment rules, while retaining interpretability in cross-cohort analyses. Additional implementation details are provided in the Supplementary Information(see Supplementary Table S3).

Ethics statement

This study analysed publicly accessible information about researchers’ career histories obtained from the researchmap Web API and associated public profile pages. No experiments were conducted on human participants, and no interaction or intervention with individuals took place. All analyses were based on information that was voluntarily disclosed on public webpages. Therefore, institutional ethical approval and informed consent were not required for this work.

Gender inference and dataset evaluation

Researchers’ gender was inferred from personal names. Prior research has shown that given names are often strongly associated with gender categories in population-level data²⁹. Gender inference was implemented using Gemini 3.0 Flash Preview, with gender defined as the category most commonly associated with a name in a statistical sense rather than as an individual’s gender identity. No personally identifiable information (PII) is reported, and the manual audit was conducted exclusively to verify the aggregate-level reliability of the inference procedure.

To assess the reliability of name-based inference, we conducted a manual audit on a sample of cases by searching publicly available information on the open web at the individual level. Specifically, for each audited researcher, we attempted to corroborate the inferred gender using verifiable public cues linked to the researcher (e.g., profile photographs or explicit gender information on publicly accessible platforms). Because many researchers cannot be reliably located through public search or do not provide verifiable gender cues, manual verification was performed only for cases where gender could be determined with sufficient clarity; cases with insufficient information were treated as non-verifiable and excluded from the audit rather than being forced into a classification.

As an additional aggregate-level check of sample representativeness, we compared the annual gender-specific PhD counts in the processed researchmap cohort with the corresponding official statistics from e-Stat (the MEXT series on doctoral graduates). The two series show strong concordance for both men and women (see n Fig. S1), indicating that the processed cohort closely tracks the official gender distribution of doctoral graduates in Japan over time.

Because confidence scores produced by LLMs are not directly comparable across languages and writing systems³⁰, and because internal model probabilities do not correspond to posterior probabilities in real-world populations, we did not impose explicit confidence thresholds on model outputs. Instead, we reduced systematic uncertainty in gender inference by controlling the quality of name information at the input stage through a stratified approach³¹.

Specifically, given that Japanese names convey markedly different amounts of gender-related information across writing systems, we prioritized names written in kanji as inputs for gender inference. Prior linguistic and computational research shows that kanji names, which encode both semantic and graphical information, typically exhibit greater gender discriminability than names represented solely through phonetic or transliterated forms³². When kanji representations were unavailable, romanized names were used as an alternative input. Romanized names more closely resemble phonetic or transliteration-based representations and do not preserve the semantic and visual cues embedded in kanji; they may also amplify ambiguity arising from multiple possible readings. Accordingly, romanized names were treated as a secondary and less informative source in the analysis.

It is important to emphasize that the gender variable in this study is used exclusively for group-level statistical comparisons and trend analyses, rather than for making definitive gender classifications at the individual level. The primary analytical objective is therefore to ensure that potential inference errors do not generate systematic bias at the sample level, rather than to optimize predictive confidence for individual observations.

Author contributions

R.P. and Z.L. conceptualized the project. G.W. and Z.W. were responsible for the methodology. B.A. and F.L. collected and processed the data. R.P. performed the visualization. Z.L., M.Y and X.F supervised the study. R.P. and Z.L. wrote the main manuscript text. All authors reviewed and approved the final version of the manuscript.

Funding

This research was supported by the JST SPRING Program (Grant Number JPMJSP2114). The funder had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Data availability

The data underlying this study were derived from the publicly accessible researchmap Web API (operated by the Japan Science and Technology Agency). Due to third-party terms of service and the dynamic nature of the platform, the raw records are not redistributed. However, the complete reproducible pipeline and analysis-ready derived datasets (including sample filtering outputs, constructed career-trajectory indicators, and processed tables) are available via Harvard Dataverse at: https://doi.org/10.7910/DVN/ZJXZEC.

Declarations

Competing interests

The authors declare no competing interests.