Peter Kuhn, UCSB Chris Riddell, Queen’s University

1. The Long-Term Effects of Income Support: Unemployment Insurance in New Brunswick and Maine, 1940-1991 Peter Kuhn, UCSB Chris Riddell, Queen’s University

2. Motivation • Typical study of income support, including UI, focuses on: short-term responses to small changes in a single program parameter, in a restricted subpopulation of workers.

3. Critics (e.g. Murray 1994) have argued that this approach (despite some important advantages) may seriously understate the long-term labor supply effects of larger social programs.

4. Why? • Individuals may be poorly informed about small policy changes • Individuals may (rationally) choose not to make large behavioral adjustments if they expect program changes to be temporary. • In the long run, adjustment can occur on a wider number of margins, essentially allowing individuals to “build a lifestyle” around a generous permanent program.

5. This paper: • Attempts to estimate the long-term effects of a generous income support program • Compares similar, neighboring jurisdictions (NB and Maine) with dramatically different evolution of a single program –UI– over 50 years. • UI is so large and pervasive in NB that it can be analyzed in isolation, and has detectable effects on entire workforce

6. In more detail, we: • Focus on weeks-worked distributions among workers • Using decennial censuses from 1940-1990 • (period covers several dramatic changes in NB policy, little in ME) • Argue that these caused dramatic changes in overall weeks worked distributions:

7. Share of workers working less than half year, 1990

8. Plan • Background on NB and ME • Evolution of UI policy in NB and ME • Cell-level analysis, 1940-1990 • Microdata analysis, 1970-1990 • Conclusions

9. New Brunswick and Maine:

10. New Brunswick

11. Maine

12. NB-Maine Facts • 1944 Populations: NB 461K; ME 805K • Both grew slower than national averages • Incomes lower, slower growing than national averages • Overwhelmingly white, native-born • High share of primary industries; history of seasonal work • Canada’s Atlantic Provinces notorious for persistent high unemployment

15. Indicators of Contact with the UI system, 1990 Note: NB figures are for 1990, ME for 1989 -sample is persons 25-59 with at least one week of work -23 percent of all New Brunswick men aged 25-59 received some UI income in 1990.

17. 2. UI Policy in New Brunswick and Maine

18. UI in Maine • Nationwide by 1937 • State-run and state-funded • Main features and overall generosity very stable since inception • Eligibility and benefits depend on total earnings during a one-year base period • Some changes in base period, introduction of dependents’ allowance, expansion of coverage after 1940 • Benefits became taxable in 1970s • Max benefit duration 26 weeks

19. UI in New Brunswick • First benefits paid in 1942 • Federally run and financed • Dramatic, complex changes since inception • Eligibility and benefits depend on weeks worked in last year • Current benefits provide very generous support to a pattern of permanent, part-year work: 12 work weeks can generate a year of benefits

20. 1940-50: UI system introduced; 1950 NB UI system most similar to Maine’s • 1950-60: in stages, very generous seasonal benefits introduced • 1960-70: reduction of seasonal benefits • 1970-80: 1971 UI Act introduces very high benefits for very short work histories • 1980-90: relative stability

21. Summary Indicator of Program Generosity: -annual weeks of income (at usual weekly rate), given weeks worked

24. Cell-Level Analysis, 1940-1990

25. a. Conceptual Framework • Voluntary labor supply model for annual weeks • Similar to Moffitt and Nicholson (1982) with a (significant) change of interpretation: applies to entire population of workers, not just to newly unemployed. • Ignores penalties for job quitting • Broad weeks-worked categories capture workers’ imperfect control over exact weeks worked

26. Specifically: Cell-level analysis is for workers (persons with positive weeks worked) only • Choice between “full year” (40-52 weeks) and “part year” (1-39 weeks) • Leisure associated with these options: LFandL P • Income associated with these options: • YiFandYiP

28. b. Data • Combination of published summary statistics and constructed cell means, 1940-1990 • 6 inds x 6 years x 2 regions=72 female obs • 9 inds x 6 years x 2 regions=108 male obs • All analysis weighted by annual industry employment shares • Relative income from PY work computed using two alternative weeks-worked assumptions: 20 and 30

31. Table 3: Regression estimates of the incidence of part year work, Cell data 1940-91 NOTES: Coefficients are derived from a WLS regression of the fraction of workers working fewer than 40 weeks per year on the independent variables shown, plus fixed effects for industry and for industry*region interactions. Robust t-statistics are in parentheses (in absolute value). The ‘relative income’ variable is the log of the ratio of part-year income to full-year income. 1940 is the omitted year. All regressions are weighted by annual industry shares in employment.

32. How large are these effects? • largest decadal change in the 20-week version of our policy variable occurred between 1950 and 1960; a change in ln(YiP/YiF) of log(.580/.441)=.274 for men, .300 for women. • This yields a predicted change in the share working part year of .081 for men and .148 for women, • i.e. from 27 to 35 percent for men, and from 14 to 29 percent for women.

33. Microdata Analysis, 1970-1990Advantages: • More sources of variation in UI policy variable (e.g. education*region*year) • Can restrict sample to Maine’s “Northern counties” only • Analyses for subgroups (e.g by education) possible • Can estimate policy’s effects on a more detailed distribution of weeks worked: did the intervals with greatest increase in UI subsidy increase in “popularity” the most?

34. a. Conceptual Framework • Equation 3’s choice model is easily extended to 5 alternatives (0, 1-13, 14-26, 27-39 and 40-52 weeks) • Each weeks-worked category (except the base of 0 weeks) is differentially affected by the X vector, and has a choice-specific “relative income” variable • If the five choice-specific ε’s have independent extreme-value distributions, this is McFadden’s conditional logit model

35. Some practical issues: • Computing relative income by category: which weeks levels to use? Choose midpoints, except for 1-13 weeks in NB post 1970: use 12 weeks, but assign only 50% of full UI benefits • Measurement error in weeks: use predicted relative earnings, so that only time-series variation in within-category relative income is used to estimate relative income’s effects.

36. b. Data • In NB, use 1%, 2% and 3% Census microdata samples for 1970 thru 1990 respectively • In ME, 1% sample in 1970; 5% State sample in 1980 and 1990. • Ages 25-59 only • “p-weights” used to account for differences in sampling rates across Censuses • All results reported here include Maine’s “Northern Counties” only

37. Table 4: Estimated Effects of Category-Specific Income in McFadden choice model NOTES: t-statistics are in parentheses (in absolute value). The dependent variable equals one for the weeks worked category realized, zero for each other weeks worked category. Estimation is by conditional logit. The category-specific income variable is the log of real total income (in 1983 dollars in each country) in the weeks category. Control variables in all specifications comprise a full set of region and year effects, plus a vector of personal characteristics (X), and X interacted with region. X in turn contains education (dummies for 4 levels), age and its square, plus indicators for marriage, presence of children and school attendance. Relative income is calculated using an individual’s predicted wage from gender/region/year-specific weekly wage regressions and the UI rules prevailing in his/her country/year. The sample is restricted to individuals of age 25 to 59. In the four-outcome choice model, the sample is restricted to persons working at least one week, and a vector of industry effects is also included in X.

38. Table 5: Actual and Predicted Counterfactual Weeks-Worked Distributions, 5-outcome model

39. Thus, for men, NB’s 1971 UI Act explains all its (absolute and relative) decline in FY work (from 72.3 to 64.5 percent). Also explains shift into 14-26-week category • For both men and women, the predicted effects of the 1971 UI Act are to “pull” people from the 0-weeks and 52-weeks categories into the middle categories; to a considerable extent this is also what happened.

40. Columns 4 and 8 also show that UI explains much of the 1990 international gap in the distribution of weeks worked. • 12.6% of ME men vs 25.6% of NB men worked 1-39 weeks in 1990. UI policy differences explain 82% of this difference.

41. Caveats • Can different business cycle trends in NB, ME explain our results? Not likely– • Microdata results are robust to region*year effects • Recessions wouldn’t move people from 0 weeks into positive weeks.

42. Alt effects of UI? • No evidence in our data that very generous UI affected educational attainment or industry mix in NB • Indeed, pretty much all of the effects of UI on weeks worked take place in non-seasonal industries.

43. Conclusions • Despite heated debate, long-run incentive effects of income support programs have been hard to identify. • This paper uses a unique and dramatic natural experiment, combined with a half-century of census data, to address this question • We find that large subsidies to part-year work can generate large increases therein

44. These effects are seen for the entire labor force, not just highly-impacted, low-skill subgroups • Credibility of results enhanced by fact that weeks-worked category-specific changes in UI subsidies predict changes in the entire weeks worked distribution