(Reprinted from HKCER Letters, Vol.33, July 1995)
On Unemployment Statistics in Hong Kong
In the past few months, the unemployment rate in Hong Kong has been the most frequently quoted figure in the press. The official estimate of the ''seasonally adjusted'' unemployment rate for the second quarter of 1995 is 3.1 percent. This number has been used, misused, criticized, but seldom understood. As concern about rising unemployment grows, the quality of unemployment statistics has become not merely an issue for academic discourse. Perhaps the best demonstration of the latter assertion is Governor Patten's commitment to spend resources towards improving labor market data.
One of the measures proposed by the government following the ''Employment Summit'' in June is to increase the size of the sample drawn in the General Household Survey. The General Household Survey is the official source of labor force and unemployment statistics in Hong Kong. It currently covers 13,500 households (or 32,000 individuals aged 15 or above) in each quarter. While some people have cast doubts on the reliability of the official unemployment figures, most of the alternative unemployment estimates are based on casual empiricism or guesswork rather than on the strict statistical standards followed by the Census and Statistics Department. Some other comments about official unemployment estimates are simply wrong. For instance, it is sometime alleged that official figures understate the unemployment problem since they do not include those who do not seek work because they believe that work is unavailable. In fact, such ''discouraged workers'' are counted as part of the unemployed population in official statistics. Similarly, official unemployment figures also include persons who are without a job and are not available for work because of temporary sickness, though it is sometimes mistakenly believed to be otherwise. The definitions of labor market status in Hong Kong follow the recommendations of the International Labour Organization, and are the same as those found in other parts of the world.
Although the General Household Survey is better than its alternatives, there is still room for improvement. With the existing sample size, the number of unemployed persons enumerated in each quarter is about 400. When these 400 individuals are cross-classified by age, education, or other categories, the sampling variability can be quite large. Increasing the sample size will help mitigate such problems. However, increasing the sample size is not a cost-effective way to increase the accuracy of aggregate unemployment statistics. At the current sample size, the standard error of the estimated unemployment rate is about one-tenth of a percentage point. Doubling the sample size will only marginally reduce the already small sampling error to about 0.07 percent. Efforts to increase the accuracy of unemployment statistics should therefore be targeted at non-sampling errors as well as sampling errors. Reducing the non-response rate, conducting validation studies, and other data quality control measures are options worth exploring.
Reducing sampling and non-sampling errors is costly, but two measures can easily be implemented that will increase the value-added from the data collected in the General Household Survey. The first suggestion is to make the individual records of the survey data publicly available. Currently the data are only available in the form of aggregated statistics published in the Quarterly Report on General Household Survey. One can find out, for example, the number of unemployed persons who are over 60 years old and the number of unemployed persons who have no schooling. But there is no way of finding out how many unemployed persons are over 60 years old and have no schooling. Opening up the individual records of the survey to the public (with appropriate safeguards to protect anonymity) will let researchers dissect the data in ways that are most suitable for the problems at hand. Ultimately this will enrich our understanding of the labor market. Information is a classic example of a public good; once the data is collected, the cost of allowing more people to use it is virtually zero. Allowing researchers access to General Household Survey records will not conflict with the current use of the data by the Census and Statistics Department. In the United States and in Canada, computer readable tapes of government labor market data are routinely released for public use. At little cost to the statistical authorities, research that makes use of those data has brought substantial improvements to our understanding of labor market behavior.
Another proposal that will improve current unemployment statistics is related to the issue of seasonal adjustments. In the Monthly Digest of Statistics and in its regular press releases, the government only reports the ''seasonally adjusted'' unemployment rate. Although unadjusted unemployment figures can be obtained (by direct calculation or from the Quarterly Report on General Household Survey) if needed, public opinion and policy makers have focused almost exclusively on the seasonally adjusted unemployment rate. For example, in the second quarter of 1995, there were 3,079,700 persons in the labor force, of which 90,700 were unemployed. Simple division gives an unemployment rate of 2.9 percent. This is at odds with the widely quoted unemployment figure of 3.1 percent. The divergence arises because the latter figure has been ''seasonally adjusted.'' Unfortunately the seasonal adjustment procedures currently adopted by the Census and Statistics Department are wrong.
In producing the seasonally adjusted unemployment rate, the Census and Statistics Department tries to control for variations in the number of first-time job seekers in the labor market. For example, there may be a disproportionate number of school-leavers who enter the labor market in the summer months. Since job searching takes time, these school-leavers may experience a temporary period of unemployment. If this is really the case, the unemployment rate will be systematically higher during the summer months than in other seasons. Thus a rise in the unemployment rate in summer does not necessarily signal a deterioration in labor market conditions. To make unemployment figures more comparable across the seasons, one might conceivably want to smooth out such seasonal patterns.
A very minimal requirement for any seasonal adjustment method is that it should preserve the mean or central tendency of the original series. In plain language, this means the seasonally adjusted series should not be systematically higher than or lower than the original series over a year. If you want to adjust the unemployment rate in summer downwards to reflect the large number of first-time job seekers, you have to adjust the unemployment rate in other months upwards so that, on average, the unemployment rate is not biased down. The seasonally adjusted unemployment rate used by the Census and Statistics Department fails to meet this criterion. Figure 1 plots the quarterly unemployment rate and the seasonally adjusted unemployment rate from 1983 Q4 to 1995 Q2. From 1987 onwards, the seasonally adjusted unemployment rate has been consistently above the raw unemployment rate. Seasonal adjustments are always upwards---for all seasons! From 1987 Q1 to 1995 Q2, the unemployment rate averages at 1.71 percent, while the seasonally adjusted unemployment rate averages at 1.92 percent. In other words the commonly used seasonally adjusted unemployment figures are biased upwards by 0.21 percentage points. Since the average unemployment rate during the period is 1.71 percent, this means seasonally adjusted unemployment figures are overestimated by as much as 12 percent.
A second minimal requirement for any seasonal adjustment procedure is that it does what it is supposed to do. If the original series contains peaks and troughs that occur regularly in particular seasons, the adjusted series should remove such patterns. On a diagram, the adjusted series should therefore appear ''smoother'' than the original series. A closer look at Figure 1 shows this is not the case. The seasonally adjusted unemployment rate is almost parallel to the raw series after 1986. In effect, the ''seasonal adjustments'' performed by the Census and Statistics Department amount to little more than adding 0.2 percentage points to the unemployment rate every quarter; it is not smoothing out any seasonal regularities.
Although the utility of seasonally adjusted figures remains controversial (e.g., who wants to read the seasonally adjusted temperature?), it is understandable why statistical authorities may sometimes want to smooth their reported figures. Some economic series do exhibit strong seasonal patterns. For instance, the volume of retail sales in Hong Kong typically shows a distinct peak near the Chinese New Year. Paradoxically there is no seasonally adjusted index of retail sales; instead the Census and Statistics Department has chosen to seasonally adjust its unemployment figures. As it turns out, the unemployment rate is not one of those series that displays any systematic seasonal behavior. From 1983 Q4 to 1995 Q2 the average unemployment rate for the four quarters are 2.17, 2.13, 2.21 and 2.20 percent respectively. Such small differences are hardly of any economic significance. A regression of the unemployment rate on lagged unemployment rate and seasonal dummy variables indicates that the seasonal effects are individually and jointly insignificant. Spectral analysis on the frequency domain also fails to reveal regular seasonal peaks. As there is little ground for making seasonal adjustments to unemployment figures, the government is advised to use only the unadjusted unemployment rate and to abandon its attempt to produce a seasonal adjusted unemployment rate.
Besides the General Household Survey, the government also conducts surveys of employers to collect information on labor market conditions. On the employers' side, the counterpart to unemployment is vacancy. Reported vacancies refer to those unfilled, immediately available job openings for which the employer is actively trying to recruit personnel. Vacancy and unemployment are often inversely related (a relationship known as the Beveridge curve). Since vacancy statistics are available at the industry level, they provide richer information about sectoral labor demand conditions than the unemployment rate. Unfortunately vacancy figures are usually released with a time lag of four to five months. The process can be sped up, but the gain will probably be marginal. One way to provide more timely information about vacancies is to compile indices of the number of help-wanted advertisements in newspapers. The cost of compiling the statistics is fairly low, and they can be released with minimal delay. Both Canada and the United Kingdom regularly publish such a ''help-wanted index'' as one barometer of labor market conditions.
A satisfactory understanding of unemployment requires information beyond that contained in current labor market statistics. For example, to determine the cost of unemployment, one has to know not only the number of days a person is without a job, but also how his earnings in the new job differ from what he previously earned. If unemployment is associated with a rise in subsequent earnings (say, as a result of better job matching), the policy implications will be very different from a case where unemployment is accompanied by a drop in subsequent earnings (say, because of human capital obsolescence). As another example, it is important to determine whether unemployment experience is history-dependent. If people who were unemployed at some point in time are more likely to become unemployed again later, the cumulative effects of unemployment will be much greater than a single spell of joblessness. These and other questions cannot be answered without specially designed surveys that follow the same individual's life history over time. Such longitudinal surveys are costly, and useful results will not come out until after several years, but the potential payoffs are also large. If the government has a more than passing interest in the unemployment problem, a commitment to support longitudinal studies will be a cost-effective way of improving our understanding of the labor market.
Dr. Wing Suen is a lecturer in the School of Economics and Finance at the University of Hong Kong.