Who's Learning? Using Demographics in EDM Research

Algorithmic Bias in Education

“…Bias of this nature can also arise even when demographics are not explicitly encoded in models, for instance when a proxy for a demographic variable is unintentionally included as a predictor. This form of bias is surprisingly common in education -in a recent survey of the role of demographics in the Educational Data Mining (EDM) community, Paquette et al (2020) found that roughly half of papers including demographics in analyses used at least one demographic attribute as a predictive feature within the model, without incorporating demographics into model testing or validation.…”

“…Evaluation bias occurs when the test sets used to evaluate a model do not represent the eventual population where the model will be applied. As reviewed in Paquette et al (2020), many models in educational data mining are developed on non-representative populations, and many papers do not even report what populations the models were tested on, making detection of evaluation bias quite difficult. Finally, deployment bias involves a model being used in inappropriate ways --being designed for one purpose and then used for a different purpose, such as using a model designed to identify student disengagement for formative purposes to assign participation grades.…”

“…Ultimately, when possible, the best path to addressing both representational and measurement bias is to collect better data --data that includes sufficient proportions of all groups of interest, and where key variables are not themselves biased. This step is recognized as essential among learning analytics practitioners (Holstein et al, 2019) but is less emphasized among researchers thus far (see, for instance, Paquette et al, 2020, which reports that many papers involving educational algorithms do not even report aggregate demographic information). Holstein and colleagues (2019) argue that, rather than conducting research on less biased algorithms, one of the most important steps for enhancing fairness in the use of algorithms in education would be for researchers to find ways to support practitioners in "collecting and curating" higher-quality data sets.…”

“…It is impossible to evaluate if a model is fair towards a specific group of learners if we do not know which learners are in that group. These efforts should work to collect traditional census categories, absent in many data sets used in our field (Paquette et al, 2020), but this is a minimum --as our review has shown, many other categories are associated with algorithmic bias. It is too early to list out all categories that should be included in an effort of this nature --there are too many unknown unknowns, and more research is needed.…”

Should College Dropout Prediction Models Include Protected Attributes?

“…On one hand, the observed historical gaps capture systematic inequities in the educational environment of different student groups, which may well apply to future students from the same groups and therefore contribute to similar gaps. In this sense, explicitly using demographic data can result in more accurate predictions and improve the efficiency of downstream interventions and actions based on those algorithmic decisions [33]. From an ethics and equity perspective, however, including demographic variables may lead to discriminatory results as predictive models would systematically assign differential predicted values across student groups according to the records of their historical counterparts.…”