Does Census Hiring Stimulate Jobs Growth?

In order to correctly apportion representatives, the US Constitution requires that “Enumeration shall be made... within every subsequent Term of ten Years” (US Constitution, Article I, Section 2). In keeping with this requirement, the United States has held a census every 10 years since 1790. As in other countries, the format has evolved steadily (Ruggles and Magnuson, 2020).

Detailed data on US decennial census employment, 1950–1970, are available in US Census Bureau publications, although interpolation is necessary. Procedural histories of the 1940 and 1980 censuses report total and/or peak employment but give no guidance on the time path of employment changes. In 1940, “[there] were approximately 120,000 enumerators employed” (Robert Jenkins, 1983, Procedural History of the 1940 Census of Population and Housing, Center for Demography and Ecology, University of Wisconsin, p. 25). The Bureau of Labor Statistics (BLS) published a separate series on temporary census employment for the 1990–2010 censuses. Sources are listed in Appendix A.

In order to use the data on training and employment of temporary employees in the 1950–1970 censuses, I normalize it to match the employment concept of the Current Employment Statistics (CES) survey. The CES intends to count all workers who were employed during the pay period that includes the 12th day of each month. In order to construct a census employment series that matches the CES, I interpolated day-by-day census employment from Census Bureau sources. Then, I matched the federal pay periods and used the highest employment day during the pay period containing the 12th of the month as monthly census employment. The census-year data are shown in Table 1. The resulting series yields month-over-month changes that are correlated with month-over-month changes in federal employment at 0.49 for 1948–1972. By comparison, changes in census employment almost perfectly match changes in federal employment (correlation = 0.97) from 1990 onward.

For 1950, the procedural history lists month-end enumerator employment during the key months and quarterly employment updates from 1948 to 1952. I used linear piecewise interpolation to fill in the daily numbers.

For 1960, a Census Bureau memorandum lists weekly (or better) enumerator employment from April 4 to July 11. I used piecewise interpolation during that period to create daily employment estimates. In order to code the buildup, I used training schedule data, which indicates that enumerators were trained en masse on March 27 and 28. A few thousand other employees were trained in advance. I added these other employees to the enumerators and patterned their drawdown after the duration of post-enumeration employment in 1950.

For 1970, I used the detailed training schedule to construct the buildup in employment. I found that trainees exceeded by 19% the peak employment number given in the Census Procedural History. In order to reconcile these figures, I included a daily attrition factor, calibrated to match peak employment. The calibrated rate amounted to 23% attrition per month. Attrition alone cannot match the speed of the post-peak drawdown in census employment, so I used the 1960 drawdown to guide estimates of the decline in employment and verify that the results are consistent with the office closure calendar given in the 1970 Procedural History.

The Procedural History of the 1980 census, which I did not use in the empirical estimation, noted that “460,000 persons were employed at one time or another in the census district offices, with 270,000 working at the peak of activities in April and May 1980” (1980 Census of Population and Housing: History, U.S. Bureau of the Census, pp. 1–14). Although this provides no evidence on the time path of employment, it is useful to note the size of the discrepancy between total and peak employment.

For 1990–2010, I made no adjustments to the BLS data. It is probable that some unrecorded temporary census employment occurred in 1989 and 1991, although none was published for those years by BLS. However, I made no attempt to estimate census employment in those years.

Comparing the extrapolated data to federal employment data, I found that significant discrepancies remain. The census creates a large and easily identifiable bump in federal employment. However, the biggest increase (or decline) in CES federal employment often occurs in a different month than it does in census employment in the series I constructed.

The 1960 census, shown in Table 2, exemplifies the discrepancy. Federal employment in the CES increases by 189,000 in March, but the Census Procedural History indicates that the enumerators were not trained until the 27th of March – well after the Federal pay period of record.

One possible explanation for the discrepancy is that census enumerators are not paid as regular federal staff. Thus, they may be counted using a different pay period. Another explanation is that the CES series may preserve anachronistic reporting of temporary Census Bureau employees. Until 2002, federal employment data were provided to the CES by the Office of Personnel Management and did not correspond to the usual CES definition. In 2002, the historical series of employment was rebuilt from employment records. However, for years prior to 1975, CES did not have the records with which to rebuild the series; instead, they relied on existing data and a benchmark.

In order to test for either of these possibilities, I constructed a second monthly series of census employment for 1948–1972, in which I used the entire month as the pay period. I find that this alternative census series better matches the CES federal employment series than my principal series. In Section 7.5, I showed that the results using the alternative series are not substantially different than the baseline results.

Canada has held a decennial census since 1871 and a quinquennial census since 1956. Table 3 gives a published estimate of employment for all but four censuses since 1871. The census shifted to “self-enumeration” in 1971, but the change did not lead to an immediate decrease in employment intensity. Unlike in the US, however, the temporary workforce has declined modestly in the 21st century.

Jane Lavigne of Statistics Canada provided monthly employment data for the Canadian censuses of 2006, 2011, and 2016 in personal communication (2012, 2016). These data are presented in Table 4 and show that the timing of census employment has shifted around among recent censuses. The data may not be perfectly comparable: the 2006 data are based on the week with the highest payroll within the month. For the record, those were the weeks of April 23, 2006; May 21, 2006; June 4, 2006; July 2, 2006; and August 6, 2006; the 2011 and 2016 data were transmitted as monthly counts of unique employees.

Unlike in the case of the US., there is not a single, clearly preferred survey of employment in Canada, nor is there a single series for federal employment. To analyze the potential impact of censuses, one must either mix and match data sources or restrict oneself to a shorter analytical period.

The most commonly used measure of broad employment is the Labour Force Survey (LFS), a household survey (which thus counts employed persons, not jobs). The LFS does not provide federal employment counts, but it has a long series.

The database CANSIM2 supplied data on the federal government workforce in Canada from 1981 to 2012, at which point the series was discontinued.

In addition, since 1991 Canada has conducted the Survey of Employment, Payrolls and Hours (SEPH). SEPH provides total and federal employment figures and measures of employees rather than persons employed. SEPH covers a smaller universe, counting 11% to 18% fewer jobs than the LFS count of people employed. SEPH counts 24% to 32% fewer federal jobs than CANSIM2. SEPH’s methodology changed in 2001, and Statistics Canada has published many SEPH series beginning in 2001.

Although SEPH is a census, not a sample, Statistics Canada wrote, “This survey is a census with a cross-sectional design. Data are collected for all units of the target population; therefore, no sampling is done ... Responding to this survey is mandatory”. It has flaws that may be responsible for its rare use in research. Google Scholar returns 25,100 results for the search Canada “Labour Force Survey” and only 313 for Canada “Survey of Employment, Payrolls and Hours” (Google Scholar web search, performed by the author, October 4, 2018.). Stephen Gordon wrote, “Like almost everyone else, I’ve been treating the SEPH as almost an afterthought ... But now that I’ve taken a closer look, I’m going to start making a bigger deal of the SEPH numbers from now on” (Gordon-SEPH). It appears to be noisier than LFS, the reverse of the situation for the analogous surveys in the US. From January 1991 to March 2012, the broadest employment series (seasonally adjusted) of the US CES payroll survey has monthly first-difference standard deviation of 0.18%, while that of the US Current Population Survey is somewhat higher, at 0.25%. In Canada, however, the broadest employment series of SEPH has a first-difference standard deviation of 0.33% and that of LFS has 0.20%. The patterns persist in non-seasonally adjusted data. Of course, higher variance does not imply a lower signal-noise ratio, but lower variance is associated with the better US survey and with more recent data; older data in all surveys have higher variance.

The choice of broad employment series may significantly impact regression results. The dissonance between household survey and payroll survey is greater in Canada than in the US: monthly first differences in the American series are correlated at 0.48 over 1991–2012, while differences in the Canadian series have a correlation of only 0.33. However, the two series for federal employment are harmonious: monthly changes in CANSIM2 federal employment are correlated at 0.89 with those of the SEPH federal employment series even though the CAN-SIM2 covers a universe 40% larger.

In the subsequent empirics, I used LFS data for total employment. For federal employment, I relied on CANSIM2 data from 1981 to 2000 and SEPH data from 2001 to 2016.

The Republic of Korea has held a census of population and housing census every 5 years since 1970; prior to that date, there was some irregularity in the periodicity of the census. Statistics Korea has provided data on the number of enumerators and full-time government employees who were recruited to work on each census from 1990 to 2010 (personal communication, 2013, 2016). However, not all of those recruited in fact worked; some of them remained on a waiting list. The Comprehensive Plan of the 2015 Census reported that expected hiring was much lower than the earlier numbers (see Table 5).

The Korean census takes place in November. Statistics Korea indicates that the temporary workers are chosen and screened in September and some of them may work in October or December (personal communication). The 2015 Census Comprehensive Plan offers a detailed schedule of the enumerators’ training schedule. For the purposes of the Economically Active Population Survey (EAPS), they are only counted as employed if they are working as of the reference day, usually the 15th of a given month.

Employment data for Korea come from EAPS, which provides monthly data since 1963. However, data disaggregated to the sector of interest – “Public Administration and Defense; Compulsory Social Security” – are only available from 1992. As the Korean data are not seasonally adjusted, I used the X-12-ARIMA utility to adjust the data. Korea has used government hiring very aggressively in the major crises of 1998 and 2009, and those large shifts in public employment may bias nearby seasonal adjustment factors.

The Public Administration data show clear increases in November for the censuses of 2005, 2010, and 2015. However, for 1995 and 2000, the series is flat through the census months. This could be due to the timing of the census or due to a change in whether temporary census workers are counted as Public Administration employees. With insufficient knowledge of the earlier censuses, I elected to use Korean data only subsequent to the 2000 Census.

The 2015 census may prove to be the last traditional census in the Republic of Korea: in 2016, Statistics Korea introduced a register-based census, augmenting a 20% long-form field sample with “register” data from other public records for the rest of the population (Statistics Korea, 2018).

Data on census employment are available earlier in Britain than in any other country and provide an unparalleled view of the historical evolution of the labor intensity of enumeration. Thus, I included this section even though a lack of data granularity prevents me from using British data in the empirical part of this paper.

The 1841 census of England and Wales was the first to require “that the business of numbering the people should be completed ‘in one day’, in order to obviate the chance of inaccuracy from omissions or double entries to which the extension of the inquiry over a greater period might have given rise”. In order to execute the census in a single day, it was necessary to appoint “no less than 35,000” enumerators (1841 Census of Great Britain, 1843).

British records, even in recent decades, do not give details of the hiring, training, and employment period for temporarily employed enumerators. The employment totals for every census of England and Wales from 1841 to the present, however, are recorded in Table 6. The number of enumerators remained roughly constant from 1841 to 1931, grew in the post-war period, and dropped to about 31,000 in 2011. Plans for the 2021 census anticipate that only 17,000 field staff will be hired, since most Britons will complete the survey online (HM Government, 2018, p. 76).

Without data on the timing of census employment, and with only quarterly data published on public employment in Britain, I did not use the census of England and Wales in regression analysis. Abstracts and general reports do give rich qualitative information about British censuses, although. The 1991 General Report, for example, discusses the interaction between census and other employment, noting both that “most field staff jobs could be carried out as a part-time activity by people already in regular full-time employment” and “strenuous, but unsuccessful, attempts were made” to allow recruits on unemployment insurance to keep their unemployment benefits despite working temporarily for the census (Office for National Statistics, 1995, p. 49).

Japan has undertaken a population census every 5 years since 1920, including a post-war census in late 1945. Matsuda (1981) put the early population censuses in the context of a broader effort to create national and imperial statistical accounts. Japan’s government employment data are not published at high frequencies, so Japan’s censuses cannot be used in this paper’s empirics.

However, the Statistics Bureau of Japan did note the number of enumerators employed in three recent censuses. In fact, Japan has the largest census labor force recorded: 830,000 enumerators conducted its 2000 and 2005 censuses. The labor force dipped to “about 700,000” for the 2010 census (Statistics Bureau of Japan, “Population Census,” https://www.stat.go.jp/english/data/kokusei/index.html)

The regression results are imprecisely estimated. In my most accurately estimated regression, the point estimate is 0.80, indicating mild crowding out. The 95% confidence intervals of various specifications can never reject 0.80. The confidence intervals are consistent with the recent multiplier literature summarized in Ramey (2019). Regressions GMM1 and GMM2, which use all the data available, point to mild crowding-out effects and cannot rule out mild Keynesian effects or severe crowding-out effects.

The regressions are reported in Table 9 and summarized in the following, and the first-stage IV results are reported in Table 10.

The regressions suggest that Canadian and Korean data lead to a lower multiplier estimate, interpolated (pre-1973) US data lead to a higher estimate, and post-1989 US data give an intermediate result close to the overall estimate. Thus, whether one uses all the information available (Regressions GMM1 and GMM2) or only the most accurate information available (Regression GMM4), the point estimate is similar. It is of note that the biased OLS estimates exhibit relatively little bias relative to the instrumental GMM estimates, and the direction of bias is not uniform. In no case can the GMM results reject a multiplier of 1; in only Regression GMM4 can it reject a multiplier of 0. In all but one case, however, multipliers higher than 2.01 can be rejected.

In the following subsections, I have reported a variety of robustness checks.

The effects of large-scale temporary employment may develop gradually. However, due to the rapidly mean-reverting nature of census employment, I could not estimate the 1-month-forward effect of government hiring without controlling for the contemporaneous effect.

The lagged results should be interpreted with caution: due to the temporary nature of census employment, measurement errors are not independent over time. When my data under-report the number of census workers added in a peak month, it will also underestimate the number of workers let go the following month.

I added lagged federal employment growth as an endogenous variable and lagged census employment growth as an instrumental variable to test the hypothesis that employment effects might emerge a month later. For Canada and Korea, the small and brief sample created interpretive and statistical problems; I thus reported only the following specifications in Table 11:

In all three cases, the cumulative 2-month multiplier has a point estimate between 0.9 and 1.5, though very imprecisely estimated. The US data suggest that lagged effects are near zero. Unreported regressions with two lags showed that a negative multiplier was typically associated with the second lag, washing out some gains from the first lag. None of the lagged estimates is precise.

Government spending may have larger net effects during periods of economic slack. This makes sense in the context of large-scale temporary hiring. To test whether high unemployment is associated with smaller multipliers, I identified “high unemployment” periods in both absolute and relative terms.

Four censuses took place with pre-census unemployment above 8% (US 2010; Canada 1986, 1991, 1996). Using only months in which unemployment exceeds 8%, the multiplier is 0.24 or 0.40, depending on the inclusion of time controls and non-census months. Values this low are not statistically rejected by broader specifications, but the point estimates are lower than baseline estimates, not higher as expected.

Alternately, five censuses took place with pre-census unemployment above the national average (US 1950, 2010; Canada 1986, 1991, 1996). In Korea, unemployment has been very steady and low throughout the short period I studied, and all three censuses have taken place at times of roughly average unemployment. I limited the sample to include months in which unemployment is above the national average in the US and Canada. In this specification, multipliers are 0.34 or 0.35, depending on whether time controls and non-census months are included. Again, the estimates are lower but not statistically different from the baseline specifications.

Another form of asymmetry in multipliers that could arise in the data is between upward and downward changes in federal hiring. The baseline regressions assume that the changes are symmetric, but that is not necessarily the case, especially for contemporaneous effects. Unfortunately, I found that allowing multipliers to differ between upward and downward movements in federal hiring leads to widely varying results under different specifications. In regressions that use all the available data, the multiplier on increasing federal hiring is much larger than the multiplier on decreasing federal hiring. However in the regression that uses only high-quality US data, the result is reversed: the multiplier is much larger for negative movements in federal hiring. I found that differences in either methodology or data inclusion can vastly change the results. In all cases, SEs are large.

When I used calendar months as pay and reporting periods in my interpolation of the 1950–1970 US censuses, regressions have lower multipliers than when using the baseline series.

In first-stage regressions, I found that the alternative series closely matches changes in federal employment and GMM gives a coefficient of 1.00. The better first-stage fit leads to smaller SEs. However, due to shifts in census employment gains from 1 month to another, these regressions may be misspecified in the absence of lagged terms for federal and census employment growth.

Replicating regression LGMM3 with alternative data, I arrived at a contemporaneous multiplier estimate of 0.97 and a forward multiplier estimate of 0.45, the sum of which is almost equal to the estimates from regressions GMM3 and LGMM3. Thus, although the contemporaneous multipliers are lower than in regressions using the baseline data, the difference between the baseline and alternative estimates is accounted for by larger multipliers on lagged federal employment growth.