Why raw mean differences by sex are not enough

A simple comparison of average salaries for men and women confounds sex with career structure. Years of experience, academic rank, disciplinary field, and administrative appointments are all associated with pay, and their distributions often differ by sex. An unadjusted mean gap therefore mixes genuine sex-associated differences with differences in who holds which roles and how long they have been in the system.

This project treats those factors as control variables (potential confounders) and asks whether sex-based gaps remain in adjusted comparisons. The statistical framing is multivariable: hold experience, rank, field, and administrative roles in view, then examine residual associations with sex. The Shiny app is the delivery layer; the substance is confounding, adjustment, and careful interpretation of observational salary data.

Three analysis questions

Compensation is not a single outcome. The work separates three related but distinct questions so that entry-level gaps, promotion dynamics, and later-career salary are not collapsed into one average:

• Starting salary: Do men and women differ in pay when they first enter the faculty, after accounting for field and related entry conditions?
• Promotion raises: Do salary increases tied to promotion from Associate to Full Professor differ by sex, once other factors are controlled?
• Overall salary: In the most recent observed salary, do sex differences remain after multivariable adjustment for experience, rank, field, and administrative roles?

Separating these questions matters because a gap that appears in career-total salary may reflect composition of ranks and tenure, differences in starting offers, differences in raise trajectories, or some combination. Adjusted models help localize where associations persist.

Confounding, composition, and selection

Beyond individual controls, composition of the faculty pool can drive crude averages: if women are underrepresented in high-paying fields or senior ranks, raw means will diverge even when within-group pay is similar. Longitudinal records from 1976-1995 also raise selection issues: who remains in the panel, who is promoted, and who takes on administrative roles are themselves outcomes that can differ by sex and by era. Adjusted comparisons reduce some of this confounding but do not automatically fix selection into the observed sample.

Association vs causation, and limits of inference

The data are observational and longitudinal, not a randomized experiment. Estimated coefficients and adjusted gaps speak to association after controlling for measured covariates, not to causal claims that sex alone “causes” a pay difference, or that equalizing one covariate would close the gap. Unmeasured factors (negotiation norms, leave patterns, department-level practices) can still bias adjusted comparisons.

Generalizability is likewise bounded: results describe this faculty population and period (1976-1995), not necessarily today’s academia or other institutions. The app supports exploratory and adjusted descriptive analysis; responsible use keeps those design limits visible alongside any numerical gap.

Explore the work

Use the live Shiny app to inspect patterns interactively, and see the GitHub repository for analysis code and project materials.