การถดถอยลำดับ

Many outcomes are categories with a meaningful order but without a defensible numeric spacing: satisfaction ratings, severity levels and educational stages are examples. A nominal classifier ignores this order. Ordinary regression on integer codes assumes numeric distances and can produce predictions between categories or beyond the scale. Ordinal regression preserves ranking while modeling probabilities for each ordered level. A common formulation assumes an unobserved continuous tendency, such as satisfaction, related linearly to predictors. Thresholds divide that latent scale into observed categories. In an ordered logit model, the cumulative probability of being at or below a category is linked to a threshold minus the predictor score through a logistic function. The thresholds are estimated in order, while a shared slope often represents how predictors shift the latent tendency. Ordered probit uses a normal cumulative distribution instead. For a hypothetical four-level rating, an increase in a favorable predictor may shift probability away from low ratings and toward high ratings. It does not necessarily increase every category probability: middle categories can gain or lose depending on where the case lies relative to thresholds. Coefficients therefore describe movement on a latent or cumulative-link scale, not a direct fixed increase in the probability of every better response. The proportional-odds assumption in the common ordered-logit model says predictor effects are shared across cumulative splits, such as low versus fair-or-higher and low-or-fair versus good-or-higher. This parsimonious assumption may not fit every predictor or dataset. Assess it, inspect predicted probabilities, and compare with alternatives when needed. Also preserve category ordering explicitly; software may sort labels in a way that does not reflect intended semantics. Evaluate on representative data and choose metrics that respect the order and the consequences of different mistakes. Ordinal regression cannot make ambiguous category definitions or inconsistent human ratings reliable by itself.