Linguistic Analysis of Letters of Recommendation for Physics Residency Program Applicants
Abstract
Purpose
Determine if linguistic differences are present in letters of recommendation (LOR) for applicants to medical physics residency programs (MPRPs).
Methods
Three MPRPs provided letters from applicants in 2019-2021. Letters were analyzed using the Linguistic Inquiry and Word Count (LIWC) software to determine percentage of words in LIWC categories and custom dictionaries for technical/research skills and personality. Letters were classified by applicant gender (AG), letter writer gender (LWG), and letter writer academic rank (LWR). For each LIWC category, linear mixed effects models with a random effect for applicant were used to assess differences in mean word use by AG, LWG, and LWR. Interaction terms between AG and LWG, and between AG and LWR were used to investigate effect modification.
Results
Analysis was performed on 714 letters for 173 applicants. There were more male letter writers (LWs) overall (598 (84%) male versus 113 (16%) female). There was no significant difference in the distribution of LWG by AG (p = 0.15). Several linguistic features showed significant differences by AG. Words associated with conflict (p=0.017) and friend (p=0.028) had higher mean usage for male applicants, while words related to wellness (p=0.002) and social referents (p=0.005) had higher mean usage for female applicants. A significant interaction between LWG and AG was observed for positive tone (p=0.006), with female LWs using more positive tone words for male compared with female applicants. Letters by full professors were more likely to have linguistic features of work (p=0.001), lifestyle (p=0.002, and tech (p=0.018), while assistant professors were more likely to have features of skill/knowledge (p=<0.001) and social behavior (p=0.002). There were no statistically significant interactions between LWR and AG.
Conclusion
Linguistic features showed significant differences between AG, LWG, and the LWR. Discussion of methods to mitigate linguistic bias, such as standardization or bias training, could improve LOR.