Multiple Regression allows engineers to add more variables to predict an outcome. Since Eurocentric datasets are heavily correlated, the algorithm may assume that these correlations represent the underlying truth. When it processes a different type of data, the intertwined variables can cause the model’s coefficients to fluctuate, leading it to output a highly inaccurate or suppressed estimate of the underlying reality.
Logistic Regression uses a Sigmoid curve to establish classification boundaries. Because historical training labels associate Eurocentric values almost exclusively with specific Western views, the algorithm can squeeze the probability score of an alternative profile down, potentially locking it out of the top classification entirely.
Multiple Regression & Logistic Regression
Multiple Regression can suffer from omitted variable bias when important localised factors are excluded, effectively shifting their influence into the error term. When these variables are omitted, their mathematical weight is absorbed into the error term (E), making the model structurally blind to factors that may be important to the actual outcome.
Logistic Regression relies on maximum likelihood estimation based on past patterns. It can fail when its classification curve is optimised around a historical majority that does not adequately represent the population being assessed.
Logistic Regression doesn’t predict a continuous number; it predicts a probability between 0 and 1 to classify data into categories. It uses a Sigmoid curve to establish a decision boundary. If the classification boundaries are heavily influenced by the traits of the historical majority, alternative data profiles may receive lower probability scores and be classified incorrectly.
Conclusion
Ordinary Least Squares (OLS) forces a straight trendline through data by minimising squared residuals. When a dataset contains a disproportionate concentration of Western data, those observations can exert substantial influence over the resulting trendline, causing a highly unique data point to appear as a large “error.” The regression model can consequently smooth over this point to fit the global trend.
Therefore, regression models can fall into the same biased traps when they are trained on skewed historical data. Regression models do not necessarily find the hidden truth.
They build a mathematically optimised representation of the patterns contained within the data.
Day 18/ 30 of the #AlRewardmaxxing Series.
Tomorrow in Part 19, we're breaking down why simply throwing more variables at a flawed model fails to fix the structural problem.
What's your take? Have you seen Al systems mask missing, localized context behind complex mathematical probability? Let's discuss below!