Pensions and Unemployment Dora L. Costa
Massachusetts Institute of Technology

For the old, unemployment is frequently a prelude to retirement. The 19th century was no exception. Margo (1993) finds that the probability of long-term unemployment rose with age and that the older a worker was when he became unemployed, the more likely he was to subsequently retire. This paper investigates the impact of the first large government transfer, the Union Army pension program, on unemployment rates circa 1900. In a simple search model (Lippman and McCall 1976a, 1976b), pensions should induce higher entry into and exit from unemployment. In fact, recent studies have found that unemployment insurance decreases the probability of leaving unemployment (Meyer 1990; Katz and Meyer 1988) and increases the probability of entry into unemployment since the availability of benefits will lead firms to reduce their labor force through layoff (Feldstein 1978; Topel 1983, 1985). If Union Army pensions increased the incidence of long-term unemployment, then, because older workers who were unemployed were more likely to retire, they were a causal factor behind early retirement.

The relationship between age and unemployment has changed since the beginning of the century. Census data shows that in 1900 the incidence of unemployment rose with age among both agricultural and non-agricultural workers. But, by 1980, unemployment no longer dramatically rose with age. Several factors may account for these differences in unemployment status by age. First, as retirement incomes have risen, withdrawal from the labor force has become an alternative to the continuation of a job search. Secondly, labor markets have changed in ways that benefit older workers who are already employed.

The difficulties faced by older workers at the end of the 19th century were widely recognized by contemporaries. Compared to younger men, older workers were more likely to experience some unemployment and when they were unemployed, they were unemployed for longer periods of time. Men in seasonal industries were especially likely to suffer unemployment.

Unemployment had become a great social concern by the end of the 19th century. Although some men were fully compensated for layoff risk, there were wide disparities in the degree of compensation. For many men wage premia covered only half of income lost because of layoffs (Fishback and Kantor 1992; Hatton and Williamson 1991). Contemporaries cited unemployment as the single most common cause of poverty (Dubofsky 1975: 22; Lauck and Sydenstricker 1917: 76). Unemployed men faced the danger of gradually drifting into the casual laboring class as their health and skill deteriorated. Unemployment of the head of the household frequently resulted in sending children to work (Lauck and Sydensticker 1917: 170-171; Goldin 1979, 1981).

This paper uses Union Army pension records to study the impact of the first large-scale government transfer - the pension program covering Union Army veterans, on unemployment rates among older workers in the nineteenth century. Union Army veterans were eligible for a pension that was on average equal to 55% of per capita income. Receipt of a pension did not depend upon current income or labor force status, but did depend upon health, as judged by a panel of three examining surgeons appointed by the Pension Bureau. Men who could trace their disabilities to their war-time service received a greater pension for the same disabilities than men who could not. Therefore, the effect of poor health and of pensions on unemployment rates can be distinguished.

The sample used in this research is restricted to men who were in the labor force and who reported months of unemployment in the year June 1, 1899 - May 31, 1900. The total number of observations is 379. Because the sample is restricted to men who were in the labor force, men who were omitted from the sample were on average collecting larger pensions and were in worse health. The restriction to men currently in the labor force implies that there are only two possible states, employment and unemployment. Retirement is not an option. Therefore, the findings are conditional on current labor force participation.

Unemployment in the veteran sample is compared with a random sample of white, non- institutionalized men drawn from the Public Use Sample of the 1900 Census. The random sample contains both veterans and non-veterans. Veterans cannot be identified in the random sample, but approximately 16.7% of all white men aged 50-64 were on the pension rolls and about 11.0% of all men aged 65-81. Compared to the general population, veterans were more likely to be unemployed and when they did experience unemployment, they experienced more of it. Differences between veterans and the general population persisted even when 1) the random sample was restricted to men either born in a Union state or, if foreign-born, who immigrated prior to the Civil War and when 2) the restricted random sample was adjusted to have the same age distribution as the sample of veterans. Greater periods of unemployment among veterans do not result from veterans' being in more seasonal jobs than the general population. I divided non- farmers into 24 occupational categories and looked at the mean months of unemployment for white males aged 25-49 in the Public Use Sample of the 1900 Census (Preston and Higgs). Within each occupational category, veterans experienced more months of unemployment.

Men who were receiving higher pensions and men who were rated by the examining surgeons as being in good health experienced less unemployment. When the sample was grouped by health status, men who were receiving larger pensions experienced more unemployment. However, these cross-tabulations are inconclusive. Men who could trace their disabilities to the war and hence were eligible for larger pension amounts were not chosen at random. Men with war-related disabilities were employed in different occupations and differed in terms of marital and head of household status.

In order to test whether the relation between months of unemployment and pension amount persisted when I controlled for other characteristics, I divided months of unemployment into three classes: 0 months, 1-5 months, and 6-12 months. I separated the unemployed into the long-term unemployed (6-12 months) and into the short-term unemployed (1-5 months), since the chance of short-term unemployment is more likely to depend upon seasonal factors and thus may be unaffected by pension income. I then used a multinomial logit regression. The use of a multinomial logit model allows me to distinguish between short-term and long-term unemployment without imposing too many assumptions on the data. I later estimate entry and exit hazards, but I can do so only under very strong assumptions. I use an unordered logit model rather than an ordered logit model because if long-term unemployment does indeed differ from short-term unemployment, then the slope parameters will not be equal. In fact, I rejected the hypothesis that the slope parameters were equal.

Both farmers and non-farmers are included in the sample. When the sample was restricted to non-farmers, the magnitude and the signs of the coefficients remained unchanged, but most of the variables were no longer significant.

I control for seasonality in the regression by using the mean months of unemployment within job classifications calculated from the public use sample of the 1900 Census (Preston and Higgs 1980) as one of my independent variables. A better proxy for seasonality would be the coefficient of variation around the mean of monthly employment by industry. Although monthly employment information is available in the 1900 census of manufacturing, I have many occupations, including that of farmer, for which I do not have information on monthly employment. Therefore, I prefer to use mean months of unemployment within a job classification.

The regression results indicate that 1) mean months of unemployment within an occupation significantly increases the probability of either short-term or long-term unemployment instead of no unemployment, but not of unemployment of 1-5 months instead of 6-12 months; 2) compared to farmers, laborers and artisans were more likely to be unemployed 1-5 months rather than 0 months, but not 6 months instead of 0 months or instead of 1-5 months; 3) compared to professionals and proprietors, artisans and laborers were more likely to be unemployed 6-12 months and 1-5 months instead of 0 months, but not 6-12 months instead of 1-5 months; 4) the larger the pension the more likely the veteran was to be unemployed 6-12 months instead of 0 or 1-5 months; and 5) pension size did not affect the probability of unemployment of 1-5 months rather than 0 months.

The significant impact of mean months of unemployment within an occupation on the probability of some unemployment instead of no unemployment, but not on the probability of unemployment of 1-5 months instead of 6-12, suggests that while seasonal factors are an important component of unemployment, seasonal factors alone cannot explain long-term unemployment. Health and pensions may help explain some of the long-term unemployment among Union Army veterans. A 25% to 183% increase in pensions increased the probability that the veteran would be unemployed 6-12 months instead of 0 months by 0.10 and increased the probability that the veteran would be unemployed 6-12 months rather than 1-5 months by 0.24. A 125% to 400% increase in pension amount increased the probability that a veteran would be unemployed 6-12 months rather than 0 months and 6-12 months rather than 1-5 months by 0.18 and 0.46, respectively. When pensions are reduced, it is mainly long-term unemployment that falls. Although the percentage of men experiencing no unemployment rises slightly, there is an increase in the percentage of men unemployed 1-5 months.

While pension size affects the probability of long-term unemployment, pension size does not significantly affect the chance of unemployment of 1-5 months rather than 0 months. One possible explanation is that while pensions decreased the probability of exiting unemployment, pensions did not increase the probability of entry into unemployment. This possibility is investigated in the next section.

Since information on the duration of an unemployment spell, or whether the veteran is currently unemployed, is unknown, the probabilities of entering and exiting unemployment can be estimated only under very strong assumptions. These assumptions are spelled out below. The discussion closely follows Margo (1990b; 1993).

The census enumerators reported the number of months of unemployment experienced in the year June 1, 1899 to May 31, 1900. Let u(0) be the unemployment rate on June 1, 1899, p the probability that a worker employed at the beginning of the year would experience some unemployment, and f the fraction of men who had experienced some unemployment in the year. Then, f=u(0)+(1 u(0))p.

Several assumptions are necessary in order to express u(0) and p in terms of probabilities of entering and leaving unemployment. These are 1) that the size of the labor force did not change during the year, 2) that employment and unemployment were the only possible labor force states, 3) that the labor market was in steady-state equilibrium, and 4) that the probabilities of moving from employment to unemployment and vice-versa did not depend upon time. The steady-state unemployment rate, u, and p can then be written as p = eb

u =u(0) = b


where b is the probability of entering unemployment, or the cumulative entry hazard, and d is the probability of leaving unemployment, or the cumulative exit hazard. The steady state unemployment rate can be estimated as the average months of unemployment divided by average labor force months. Following Margo (1990; 1993), I assume a full-time work year of 50 weeks or 11.5 months. The average monthly entry and exit hazards are b/12 and d/12.

Average monthly entry and exit hazards are calculated for men aged 50-64 in the veteran sample who were collecting a monthly pension of less than or equal to $8 and of more than $8. The sample is also subdivided into farmers and non-farmers and men in poor health and good health. Age-adjusted hazards are calculated. The sample was restricted to men aged 50-64 because entry and exit hazards can be calculated only under the assumption that retirement is not a possible labor force state. The estimated hazards suggest that pensions exerted a much larger impact on the probability of entering unemployment instead of on the probability of exiting unemployment.

The main findings are that pensions increased the probability of long-term unemployment, but not of short-term unemployment. Pensions appear to have mainly decreased exit hazards, but they also slightly increased entry hazards. The findings are consistent with a job search model. Union Army pensions were a pure income transfer and, therefore, led men to search longer. They may also have led men to try searching for better employment instead of continuing to work in their current jobs. If so, pensions increased the already high quit rates characteristic of 19th-century labor markets.

The findings permit the quantification of the impact of a program that affected a large proportion of the population. A reduction in pensions would have decreased long-term unemployment and increased the incidence both of no unemployment and of short-term unemployment. Because pensions increased the prevalence of long-term unemployment and because older workers who were unemployed were more likely to retire, the results imply that pensions led to an increase in early retirement.

The findings do not pertain only to the Union Army pension program. They also shed light on the secular rise in decade unemployment averages. From 1800 to 1960 there has been a slight increase in decade unemployment averages from 1-3% to 5%, partly driven by a decrease in self- employment (Lebergott 1964: 189). In 1970-1979, average unemployment was 6% and in 1980- 1990 the average was 7 percent. Shifts in industry type can explain most of the difference in the probability of entry into unemployment, but not the greater duration of unemployment (Margo 1990). Two possible explanation for the increased duration of unemployment are the prevalence of unemployment insurance and rising wealth, permitting longer search until the worker finds an attractive job. Lebergott (1964: 165) noted, "As real incomes have risen over the decades so has the ability of the unemployed to refuse the first job available (if it is below standard), to delay until he finds 'suitable' work. And as social institutions have changed (in particular, as unemployment compensation has been introduced, supplemental benefits pioneered, health service extended), this effect has been intensified." A pure income transfer, the Union Army pension, mainly decreased the probability of leaving unemployment. Thus, the findings suggest that the income effects arising from unemployment insurance and the secular rise in incomes are significant factors in explaining the rise in unemployment rates over the course of the century.