What is This?

Data aggregation is the process of gathering raw data and summarizing it to make meaning. This is a necessary process in research and inherently introduces researcher bias. The decisions that a researcher makes to decide how to summarize their data lead to the inclusion or exclusion of groups. Researchers need to be intentional in their summary choices, particularly when it comes to grouping people to minimize the bias introduced. This is a form of aggregation bias. 

For example, a researcher is seeking to understand the relationship between household income and the average number of years of education within the household. They gather data from four cities and plot this data to find the line of best fit. In so doing they find a strongly positive correlation between average years of education and household income. They conclude that more years of education is strongly positively correlated with household income.  However, they have aggregated all four cities into one analysis instead of looking at them separately. Grouping by city reveals that in two of the four cities, there is actually a negative correlation between average education and household income.  

Often, due to small sample sizes, underrepresented and minority groups are aggregated into larger groups for analyses. It is important in QuantCrit and equity-based research to disaggregate groups based on race/gender and other identifiers that have been systemically and systematically oppressed in the societies we study. 

The Fine Balance of Data Aggregation

While aggregating is necessary to achieve large enough sample sizes that regression outputs and interpretations are reliable, aggregation can also hide important voices that would signal change is needed. 

Recommendations:

1. Know the limits of the tools you are using. If the package or estimator you are using to analyze your dataset require a minimum sample size, be aware of that minimum and use it to maximal effect. For instance, our LASSO database has over 15,000 completed student assessments for the Force Concept Inventory. We could look at how all 15,000 of those students are doing but we would be missing important data from different groups of interest. White men dominate our dataset, as is the case for STEM in general. So, by looking at the entire set of 15,000 cases, we would be seeing the average student’s performance – a white man. However, our tools require a minimum sample size of 200 cases. We identified each group that would meet that minimum threshold to find ten race/gender groups we could study. This not only allowed us to more accurately represent the data, but we could then identify educational debts owed to marginalized groups. We still fully understood the performance of white males, but we gained a deeper understanding of all the groups we studied. 

2. Be careful about your analysis! If you do not disaggregate your data, you cannot say that your findings are generalizable to all groups – only to the dominant group in your sample. There were many groups we did not study because their population in the dataset did not meet the sample size cutoff. We have been cautious about not making any claims about those groups or being so vague in our findings as to inadvertently include them.