Colleen Carey looks at the effects of recessions on income inequality and finds some surprising results
Picture a group of twenty-five year olds. Most of them have finished their education and are just embarking on their work lives. They will earn varying amounts of money, due mostly to differences in education. But there won’t yet be any millionaires (Mark Zuckerberg excepted), any entrepreneurs, or any titans of industry. Compared to an older group, the variance of their incomes will be small. It turns out that if you follow this group as they age, the variance of their incomes will, in general, grow. As the titans arise, or the middle managers climb, incomes spread out.
Consider the twenty-five year olds born in 1905, who started working in 1930, rode the roller coaster of the Depression and the post-war boom, and retired (all at the same time, for the sake of argument) at age sixty in 1965. Compared to 1930, how similar did they look in 1940? In 1950? Now think of the twenty-five year olds who started working in 1960, enjoyed almost a decade of economic expansion, and then the stagflating 70s, to retire in 1995. Which group ended up with more income inequality? When did incomes spread out the most? Why?
A basic research question emerges: does income inequality within a cohort grow more quickly in good macroeconomic conditions or in bad? Do recessions exact a double price, both lowering average incomes and increasing income risk? While previous research has sometimes suggested yes, a new paper (by me and Stephen Shore, both of Johns Hopkins University) disagrees. For anyone interested in economic inequality, it’s helpful to know the economic conditions that foster it. What can we expect in the wake of the Great Recession?
This figure illustrates the growth in cross-sectional variance within a cohort over its work life. Defining a cohort as male household heads (employed or unemployed) aged 25 to 59 and born in the same five-year period, we calculate the cross-sectional variance of income for cohorts born between 1891 and 1980. The general upward trend is apparent. But the figure also illustrates an obstacle to empirical analysis. Between 1960 and 1970, the variance grows slowly (or even declines), while it grows quickly between 1970 and 1990. There were also many more months of NBER recession (gray bars) in this latter period. We may be tempted to conclude that NBER recessions lead to higher income volatility. But perhaps Lyndon Johnson’s Great Society programs reduced income volatility, and the well-documented increase in wage inequality, thought to be due to structural changes, could be causing the steep increases between 1970 and 1990. How does the Baby Boom, and changes in the distribution of educational achievement, affect these cross-sectional variances? In any period in the data, an alternative story that does not rely on macroeconomic factors can rationalize our observations. Depending on the chosen age, year, and cohort controls, the years of data included, and the treatment of standard errors, we can generate a positive, negative, or null correlation between the growth in cross-sectional variance and macroeconomic conditions.
Our paper tries a new strategy –we use state-level variation in macroeconomic conditions. Such variation is substantial; some states experience booms when the rest of the country is in recession (Wyoming, presumably because of its energy sector), others experience the national boom-bust cycle more intensely (Michigan), and any state may lead or lag national conditions. We redefine a cohort as male household heads born in the same five year period in the same state (excluding Alaska and Hawaii due to data availability issues). The state setting affords us many more observations than the national, and we can therefore control for simultaneous national-level policies such as the New Deal or the Great Society.
We first considered the association between Gross State Product (GSP) and a state-cohort’s cross-sectional variance. We use the sum of GSP over the cohort’s work life and the number of years when GSP was negative (a “state recession”). We found, to our surprise, that income risk appears procyclical – that incomes disperse less quickly when state economic conditions have been bad.
Involuntary job loss seems the most plausible reason why income risk would rise in a recession. If bad state economic conditions do not cause job loss in the same way as national, then our result is fully consistent with countercyclical income risk. We needed to look more closely at employment. State employment figures are only available from 1992, but the Bureau of Labor Statistics supplies national employment in each of 15 “supersectors” (e.g., Durable Goods Manufacturing, Leisure and Hospitality) for each year from 1939. Each state differs in its exposure to these “supersectors”. For each year, we impute the share of the state’s workers in each industry (using the Census’s industrial classification of each worker), and multiply it by the national change in employment in that industry. Just as with GSP, we find that the growth in cross-sectional variance is lower when the state’s expected employment change is less favorable.
We are not sure exactly what mechanism produces procyclical income risk. But we find that our effect is driven by economic conditions ten to twenty years ago. Then when economic conditions are good, incomes will disperse more quickly in the medium-to-long term (we do not find conclusive evidence on the short-term effect.) Other research has suggested that recessions change individual’s behavior – they hold less wealth in stocks and are less entrepreneurial. And of course, small-business credit can evaporate in recessions. We might be picking up the income-flattening effect of lower economic risk-taking.
More About the Methods
Controls: In national analyses, we observed seven cohorts in each of seven Census years (1950 to 2000 plus the 2005 ACS) for 49 observations, whereas using state-level data we now have (49*48=)2,352 observations of a birth-state cohort’s cross-sectional variance.
We can now flexibly control for all national-level variation as well as much state-specific variation. In particular:
- Dummies for each age in each year control for anything common to the national-level cohort
- Dummies for each cohort in each birth state mean we are comparing growth rates (not levels) of income dispersion
- Allowing each state its own linear age profile means incomes disperse at a baseline rate in each state
Given these controls, results are identified by comparing the rate at which cross-sectional variance grows over time within a state-cohort during better-than-average or worse-than-average economic conditions for that state.
Robustness: The paper contains a large number of robustness checks. We tested different adjustments for interstate migration. We dropped years of Census data, years of economic data, geographical regions, and birth-year cohorts. The results hold across racial and educational categories. In addition to the household head’s total income (the definition used heretofore), we also test the cross-sectional variance of wage income (which excludes business, asset, and public assistance income) and total household income. Above, “cross-sectional variance” means the variance of excess log income, or the variance of the residuals from a regression predicting log income separately for each year using year of birth, age X race, and age X education. The results hold for raw log income (instead of excess), various ways of handling nonpositive incomes, and using the split between the 90th and 10th percentiles (instead of variance).