Gender Bias in Remote Pair Programming among Software Engineering Students: The twincode Exploratory Study

Abstract

Context. Pair programming has been found to increase student interest in Computer Science, particularly so for women, and would therefore appear to be a way to help remedy the under–representation of women in the field. However, one reason for this under– representation is the unwelcoming climate created by gender stereotypes applied to engineers in general, and to software engineers in particular, assuming that men perform better than their women peers. If this same bias is present in pair programming, it could work against the goal of improving gender balance in computing. Objective. In a remote setting in which students cannot directly observe the gender of their peers, we aim to explore whether Software Engineering students behave differently when the perceived gender of their remote pair programming partners changes, searching for differences in (i) the perceived productivity compared to solo programming; (ii) the partner’s perceived technical competency compared to their own; (iii) the partner’s perceived skill level; (iv) the interaction behavior, such as the frequency of source code additions, deletions, validations, etc.; and (v) the type and relative frequencies of dialog messages used for collaborative behavior in a chat window. Although there are some studies on pair programming performance and gender pair combination, to the best of our knowledge there are no studies on the impact of gender stereotypes and bias within the pairs themselves. Method. We have developed an online platform (twincode) that randomly classifies students into gender–balanced groups, arranges them in pairs for remote pair programming (sharing an editor window and a chat window), and can selectively deceive one or both partners regarding the gender of the other via the use of a clearly gendered avatar. Several behaviors are automatically measured during the pair programming process, together with two questionnaires and a semantic tagging of the pairs’ conversations. We will perform a series of experiments to identify the effect, if any, of possible gender bias in remote pair programming interactions. Students in the control group will have no information about their partner’s gender; students in the treatment group will receive such information but will be selectively deceived about their partner’s true gender. To analyze the data, apart from checking reliability of questionnaire data using Cronbach’s alpha and Kaiser criterion, for each response variable we will (i) compare control and experimental groups for the score distance between two in–pair tasks; then, using the data from the experimental group only, we will (ii) compare scores using the partner’s perceived gender as a within–subjects variable; and (iii) analyze the interaction between the partner’s perceived gender (within–subjects) and the subject’s gender (between–subjects). For the (i) and (ii) analyses we will use t–tests, whereas for the (iii) analyses we will use mixed–model ANOVAs.