Reaching the Upper-Middle Class in 1994

Personal income is a key variable in economic and social research, and its predictors can help guide policy. This analysis looks at what factors most influenced reaching the upper-middle class in 1994, which I define as earning $50,000 or more that year. Framing it as a threshold has three advantages: it focuses on a specific, policy-relevant cutoff; logistic regression is less sensitive to outliers than modeling raw income directly; and it sidesteps the heavily skewed distribution of income itself.

The data is the well-known "Adult" dataset from the UCI Machine Learning Repository, extracted from the 1994 US Census. After pre-processing — re-grouping some categorical variables to avoid low cell counts — I considered ten candidate predictors: sex, age, race, work-class, occupation, marital status, US birth, years of education, and hours worked per week. I tested education both as a category and as a continuous years-of-schooling count; the continuous version fit better and is the one used in the final model. About 24% of the 32,561 respondents made $50,000 or more.

I fit a logistic regression predicting the odds of crossing $50,000 from nine variables at once, holding the others constant, then tested the null hypothesis that each coefficient was zero. Every variable in the findings table below was statistically significant at the 5% level.

By far the largest effect was marital status: married respondents had odds of crossing $50,000 that were almost ten times as high as unmarried respondents. Because that's an odds ratio rather than a probability ratio, it doesn't mean married people were literally "10 times more likely" to earn $50,000+ — but it's still the strongest effect in the model by a wide margin. Years of education came next, with each additional year raising the odds by about a third. The chart below shows the two compounding: predicted probability of crossing $50,000 climbs much faster with education for a married profile than an unmarried one, holding everything else at a typical 1994 value.

A classification tree trained on the same variables, splitting mainly on marital status and education, correctly classified 82% of respondents — a modest improvement over the roughly 76% you'd get by simply predicting that everyone earned under $50,000, since only about 24% of the sample crossed the threshold. McFadden's pseudo-R² for the logistic model was 0.35, generally considered a good fit by that measure.

Two methodological limitations are worth flagging: grouping some categorical variables may have obscured more detailed patterns, and standard logistic regression doesn't test interaction effects — for example, whether education's effect on income differs by work-class. More fundamentally, this is an observational census sample, not an experiment, so these are associations rather than proven causal effects. Marriage and income most likely influence each other in both directions, and both correlate with unmeasured factors like age and household structure — so I'd treat education, not marital status, as the more actionable and less confounded finding here.