This article describes new estimates of labour inputs to economic production, compares the new estimates with existing estimates of labour inputs and reviews whether the new estimates revise the statistical record on labour productivity. For the first time, consistent estimates of labour input are available in terms of hours, workers and jobs for up to 64 industries, and split between employees and the self-employed. ONS believes that the new estimates have advantages over existing estimates of labour inputs for the purpose of measuring labour productivity and – subject to user feedback - intends to implement a work programme to replace the existing series with series compiled using the methodology described in this article.
The existing productivity system compiles labour inputs in the form of jobs and hours, referred to respectively as productivity jobs (PJ) and productivity hours (PH), for the whole economy and a range of industries to produce productivity estimates of output per job (OPJ) and output per hour (OPH), as published in the quarterly Labour Productivity release. This release also includes estimates for output per worker (OPW) at the whole economy level only. ONS regards OPH as a more inclusive measure of labour productivity than OPJ, because it implicitly takes account of change in working patterns (reflected in average hours per job). The concepts of jobs and workers differ (a) because some workers have more than one job and (b) because 'job' is a more ambiguous statistical concept than 'worker'.
PJ and PH are hybrid estimates using business survey data on the industry distribution of employee jobs, combined with Labour Force Survey (LFS) data on overall employment, hours of work and the industry distribution of self-employment. The business surveys (known as STES - short-term employment surveys) are fully integrated into ONS data capture of economic statistics, for example from monthly production surveys which feed into GDP(O). By contrast, LFS is a household survey, reliant on respondents knowledge of what industry they work in (and their employment status), and not integrated with ONS mainstream economic statistics.
There are a number of motivations for the work described in this article. First, there are advantages in switching the emphasis from jobs (and OPJ) to workers (and OPW). As noted above, 'job' is an ambiguous statistical concept, and ONS currently publishes two separate series of jobs by industry – PJ as described above and workforce jobs (WFJ)1. Methodological differences between PJ and WFJ are well documented and are not repeated here2. However, despite the fact that the PJ system was built explicitly for the purpose of constructing labour productivity estimates, there is evidence that some users construct their own productivity estimates using WFJ estimates for labour input, and evidence also of confusion among users when – as can happen – PJ and WFJ estimates provide conflicting information.
Second, the existing ONS productivity system has scope for a number of methodological improvements which have been addressed in the work described in this article, including:
Full articulation of the split between employees and the self-employed
Consistency of industry coverage between sources and outputs
An experimental means of “nowcasting” estimates for the latest quarter, one month earlier than estimates are available in the existing system
Improvements to seasonal adjustment, properly reflecting the seasonality of the source data.
These improvements are discussed further in the following section.
Third, the existing productivity system has been developed piecemeal over the last decade or so and has become excessively complex, time-consuming to maintain and prone to human error. The new system is far more streamlined.
Fourth, the new system meets requirements under the European System of Accounts (ESA) in terms of the granularity and timeliness of data delivery, which the existing system does not. Now that UK estimates provided to Eurostat are in line with the ESA requirements, comparability with other Member States labour input data is possible, as well as the generation of EU aggregates.
The layout of the rest of the article is as follows. The next section describes key features and methodological changes in the new system. We then review results, first in terms of labour inputs (hours, jobs and workers) and then in terms of the implications for labour productivity. Finally we discuss next steps and how this work relates to other ONS productivity work. User feedback is always welcome and can be addressed to email@example.com.
Industry level WFJ estimates are published by ONS in a number of high-profile publications including Labour Market Statistics and the Blue Book. Industry level PJ estimates have not been routinely published but will be included as a reference table in future Labour Productivity quarterly releases.
Further information on the differences between PJ and WFJ is available in the quarterly Labour Productivity release and in the corresponding Quality and Methodology Information paper.
The principal innovation of the new system is that it allocates labour input across industries on a headcount (or worker) basis, in addition to the allocation on a jobs basis as in the present system. This headcount allocation uses LFS data on hours worked in first jobs and second jobs which are already collected within the current productivity hours system. A stylised example is shown in Table 1:
|1: Breakdown by jobs|
|Industry||1st job||2nd job||Total||1st job||2nd job||Total||1st job||2nd job||Total|
|2: Breakdown by workers|
|Industry||"1st job"||"2nd job"||Total||"1st job"||"2nd job"||Total||"1st job"||"2nd job"||Total|
The upper panel of the table shows stylised LFS data for 2 industries in the form that the data are collected currently, that is, in terms of jobs and hours. Data points collected through the LFS are shown as bold text in the table, with other cells derived by transformation. Note that Industry 1 has a relatively high share of second jobs (even though average hours in second jobs are lower than in Industry 2). This pattern is based on real LFS data, albeit scaled to assist interpretation.
The lower panel of Table 1 shows how the data in the upper panel can be transformed in terms of workers. In this part of the table the columns headed “1st job” and “2nd job” should be interpreted as weights. There are several points to note:
The total number of workers is equal to the number of workers with first jobs (100 in this stylised example). The total number of jobs is larger (103), but while we know something about hours worked in second jobs by industry, we do not know anything about the hours worked by workers with second jobs1.
Total hours worked by industry is invariant between the two breakdowns2.
Workers’ labour input in the “1st job” column are allocated between industries using the overall share of hours worked in first jobs relative to total hours worked. That is, the industry allocation of workers’ “1st jobs” is directly proportional to the allocation of 1st jobs in the upper panel of the table
Workers’ labour input in the “2nd job” column are allocated between industries using the industry share of hours in second jobs, conditioned on the overall share of hours worked in second jobs. Equivalently, hours worked in second jobs are converted to worker weights using the average hours worked across all industries (30.0/36.1=0.83 and 27.5/36.1=0.76)
As can be seen, the resulting distribution of average hours per worker has the property that it is higher for both industries than the equivalent hours per job. But the increase is larger in Industry 1, reflecting the greater importance of second jobs in that industry
However, looking at the industry distribution of workers versus jobs, it can be seen that the worker share of Industry 1 is lower than the job share (10.67% versus 12/103 = 11.65%), reflecting the lower weight of “2nd job” when weighted by hours worked.
In the present productivity system, estimates of employees and the self-employed by industry are aggregated together and jointly benchmarked to LFS control totals for jobs and hours. This process of jointly constraining employees and the self-employed was introduced in December 2010, reflecting concerns that the employee/self-employed split in LFS was not consistent with the employee jobs estimates derived from business surveys. However, the use of this methodology means that the employee/self-employed split cannot easily be recovered from the constrained and seasonally-adjusted jobs and hours estimates. This is a particular issue in the case of hours, because average hours differ between employees and the self-employed but these differences are lost in the current methodology.
In the new system there is a fully articulated split between employees (abbreviated to EE below) and the self-employed (SE), in terms of workers, jobs and hours worked for up to 64 industry components3. This is effectively similar to current practice since December 2010 in that we jointly constrain EE and SE to LFS, with the unconstrained EE industry allocation of jobs coming from STES. However, it is important to stress that we are neither (a) relying on STES EE jobs and treating SE as the residual nor (b) relying on LFS SE and scaling STES EE jobs. Since STES estimates of employee jobs typically differ from LFS estimates in aggregate terms as well as industry distribution, this means that the EE and SE estimates produced by the new system are not consistent with EE jobs from business surveys, and also not consistent with the source LFS data on the SE. Moreover, since we impose aggregate LFS constraints in terms of jobs, workers and hours, additivity requires that we also iterate away, at the industry component level, from LFS source data on average hours worked.
|1: Current productivity system|
|2: Breakdown by workers|
This is illustrated in Table 2, where as in Table 1 the upper panel represents the present productivity system and the lower panel represents the revised system. Table 2 takes the stylised data in Table 1 to represent employees and adds some stylised data to represent the self-employed. As in Table 1 bold entries represent input data, while other cells are populated endogenously.
In the present system, estimates of employee and self-employed jobs and hours are (separately) aggregated together and constrained to their respective LFS benchmarks. As shown in the table, it is possible to derive implied average hours from the constrained totals. However, constrained estimates for EE and SE separately are not computed in the present system4. The lower panel of Table 2 shows how constraints are applied jointly to employees and the self-employed in the new system, yielding a full set of constrained estimates. Although not shown in Table 2, the new system also generates a breakdown of jobs between employees and the self-employed.
One further improvement is that members of HMF are now treated as employees, whereas they are effectively proportioned over EE and SE in the current productivity system5.
The new system explicitly includes the whole industry distribution A-U in SIC07 terms. No STES employee jobs data are available for industries T or U, so we use LFS for these industries. Although jobs and hours in industries T and U are small, it is nevertheless an inconsistency of the current system to effectively allocate jobs and hours in T and U proportionately over the rest of the industry spectrum, A-S. This is one reason for the large differences in jobs and hours estimates for industry R-U shown in the reference table (461 Kb Excel sheet) published with this article.
Under the terms of the European Statistical System, ONS is required to deliver estimates of employment (in terms of workers) and hours worked for the latest quarter, both at the A10 industry breakdown, to Eurostat before STES employee jobs estimates become available at t+75 days. To meet this requirement we have developed a system for “now-casting” the latest quarter using LFS information on employee jobs to generate provisional estimates of workers, jobs and hours approximately 45 days after the end of each quarter.
This is work in progress. As noted earlier in this article, STES and LFS employee job estimates differ significantly at the industry component level, and provisional analysis suggest that simple estimation techniques such as using LFS growth rates or changes in shares may generate acceptable estimates for some industries but are unlikely to provide consistently accurate estimates for the latest quarter. In addition to different time series properties of the STES and LFS employee jobs series themselves, there is a further complication insofar as STES and LFS aggregate employee series evolve differently over time. Thus even though LFS is de facto the only source for self-employment, forecast errors will arise due to the constraining process outlined above where the post-STES EE/SE weights differ from the LFS EE/SE weights.
|Hours (weekly, m, NSA)|
|Workers (000s, NSA)|
|Jobs (000s, NSA)|
Nowcasts based on source data as of June 2013 except for STES estimate for 2013Q1
Table 3 illustrates the forecast errors in terms of hours worked, workers and jobs at the 11 industry breakdown for 2013Q16. For each labour input, “nowcasts” use LFS annual growth rates as the source for unconstrained EE jobs, while “STES” use actual STES estimates. In general it can be seen that the shares of labour input across industry categories are very close, and unsurprisingly, differences show a common pattern across each labour input. For example, the nowcasts underestimate the share of labour in industry K in 2013Q1, and overestimate the share of labour in industry OPQ.
But the picture looks less satisfactory in terms of percentage differences between the nowcasts and the STES-based estimates. In terms of hours worked, 7 out of the 11 industry categories are different by at least 1 percentage point, and there are very large differences for industries BDE and K. Other things equal these differences would be reflected in seasonally adjusted estimates of labour input and labour productivity estimates that used the nowcasts.
The current system applies seasonal adjustment to the aggregated and constrained jobs and hours estimates, and compiles productivity estimates (OPJ, OPH and OPW) by dividing these into seasonally adjusted GVA estimates. By contrast, the new system carries out seasonal adjustment separately in terms of employment/workers and average hours per worker (and separately for EE and SE in each case). This avoids the methodological inconsistency of applying independent seasonal adjustment to (i) jobs series and (ii) hours series which, as illustrated in Table 1, are compiled from, and hence not independent of, the jobs estimates. In the new system, seasonally adjusted estimates of hours worked are compiled as the product of seasonally adjusted workers and seasonally adjusted average hours per worker, which correctly separates the sources of seasonality. Again this is done separately for the EE and SE components.
Some of the new labour input estimates produced under the above methodology have been transmitted to Eurostat and are available on their website, with comparable international series: epp.eurostat.ec.europa.eu/tgm/table.do?tab=table&tableSelection=1&labeling=labels&footnotes=yes&layout=time,geo,cat&language=en&pcode=teina305&plugin=1,
All of the new estimates are also available in the reference table (461 Kb Excel sheet) published alongside this article.
This section compares results in terms of hours and jobs with the corresponding estimates compiled under the current productivity system1. We also look at movements between employees and the self-employed within the new estimates, and we compare the new worker series with jobs by industry.
Figures 1 and 2 show total hours from the new series and the corresponding PH series in levels and quarterly changes. The principal source of the differences is seasonal adjustment as outlined above – both series are constrained in NSA terms to the same LFS constraint2. Taking the period as a whole, the new series is distinctly less volatile than the existing PH series, although there are some quarters when the change in the new series is larger than in the PH series.
Figure 3 provides an overview of the industry distribution of hours in the two systems, in terms of differences in (annual) shares of hours by broad industry group. Generally the scale of differences is very small. The new system allocates a little more labour input to services and a little less to construction than the current hours system. And there is a trend for the new system to allocate fewer hours to manufacturing over time.
Differences in hours between the new and existing systems are larger at a more disaggregated industry breakdown as shown in the reference table (461 Kb Excel sheet) published alongside this article. And the share difference in construction shown in Figure 3 translates into a maximum difference of almost 10% in the quarterly series.
As noted above, the new system provides a fully articulated breakdown of hours, jobs and workers between employees and the self-employed. Shares of self-employment in total hours are shown for the whole economy and selected industries in Figure 4. It can be seen that, while the share of self-employed hours has increased in recent years, the increase has been fairly muted – less than 1 percentage point between 2008 and 2012 – and the self-employed share of hours is significantly lower than in the 1990s.
In terms of broad industry categories, the share of self employment hours worked in construction fell sharply during the recession but has recovered over 2010-12. By contrast, the share of self employment hours worked in the extractive and utility industries (ABDE) rose strongly between 2007 and 2010 but has fallen back in 2011 and 2012. The share of self-employed hours in the service sector has increased by about 1 percentage point since 2008.
At the whole economy level, there are negligible differences between new and existing estimates of seasonally adjusted jobs, both in levels and quarterly changes, and there is almost no difference in the volatility of the 2 series. Figure 5 shows differences in shares of jobs by broad industry category. Compared with the equivalent hours data in Figure 3 above, there are smaller differences in job shares, but this representation of the data disguises some more significant differences in the industry components, as shown in the reference table (461 Kb Excel sheet) published alongside this article.
Figure 6 shows movements in shares of self-employed job. Compared with Figure 4 the increase in the share of self-employed jobs over the last 5 years is a little more pronounced than the equivalent movement in hours. This reflects evidence that average hours worked by the self-employed have fallen as the number of self-employed jobs has risen.
The existing productivity system does not produce estimate of workers below the whole economy level, so it is not possible to make direct comparisons with the new estimates. Here we examine whether there are any significant differences between the new worker estimates and the corresponding jobs estimates. In general, workers and jobs move very closely together in the new system (although of course the level of workers is always lower than the corresponding level of jobs). In terms of quarterly growth rates, differences between the new worker and job estimates are typically smaller than the corresponding differences between the new job estimates and existing job series, as shown for the construction industry in Figure 7.
Figure 8 summarises differences between shares of workers and jobs from the new estimates. Differences in shares are not large and have narrowed over time. Overall there is a slight increase in the share of workers in manufacturing compared with the share of jobs, and a corresponding reduction in the share of workers in services. Within services, the worker share is lower than the jobs share in sections I, OPQ, R and S but a little higher in sections G, H and K.
Movements in self-employment on a worker basis are very similar to movements on a jobs basis as in Figure 6.
Currently, labour productivity estimates (OPW, OPJ and OPH) are compiled by dividing seasonally adjusted measures of economic output (chained volume measures of gross value added –GVA) by seasonally adjusted measures of labour input (abbreviated to SA/SA below). It follows that, under this approach, the differences in labour input measures reported in the previous section would flow through to differences in labour productivity estimates. In general, this means that whole economy labour productivity measures would be little affected, with somewhat more noticeable changes to OPH than to OPJ due to the revised method of producing the SA hours estimates, viz by seasonally adjusting the intermediate average hours series. Differences in labour productivity would be a little more pronounced below the whole economy level. Again these would be larger in terms of OPH than OPJ. The industry breakdown of OPW would be very similar to that of OPJ.
An example is shown in Figure 9 for manufacturing output per hour, in which new SA hours estimates are divided into the same SA output series as used in OPH as published in March 2013. Clearly there are differences in quarterly growth rates, but the overall time series is not materially affected and the volatility of the new estimates is fractionally lower than that of the existing OPH series.
In this section we take the analysis one stage further by reporting some provisional findings from compiling labour productivity metrics entirely from NSA data (for GVA and labour inputs), then subjecting the resulting estimates to seasonal adjustment. We refer to these estimates as “end of pipe” (EOP) seasonally adjusted estimates. This is more work in progress. For example, seasonal adjustment of the whole economy OPH series currently breaks down from 2010Q4, when there is a discontinuity in the LFS constraint.
An example of the end of pipe methodology for OPH in manufacturing is shown in Figure 10. To the naked eye there is little difference in the volatility of the two series (in fact, the standard deviation of quarterly changes in the end of pipe series is a little larger than that of the current series). But there are sizable differences in certain quarters. For example, in the latest quarter (2013Q1) OPH in manufacturing fell by 2.5% according to the EOP estimate, compared with a fall of 0.4% as reported in the labour productivity release.
One interesting feature of this exercise is that OPH estimates for industries K and M+N are found to be non-seasonal (although the numerator and denominator series in both cases are seasonal). This might suggest that the seasonality of output and labour input is sufficiently correlated to cancel out in terms of labour productivity. However, as shown in Figure 11 OPH for industry K is considerably more volatile than the current SA series, and a similar picture holds for industries M+N.
End-of-pipe seasonal adjustment makes little difference at the whole economy level in terms of OPJ (Figure 12). For manufacturing (13) the end-of-pipe estimates show reduced volatility in some periods such as 2002 compared with SA/SA, although across the whole time series the end-of-pipe series is slightly more volatile in terms of quarter on quarter changes.
More work is needed to test for differences between end-of-pipe productivity estimates for OPJ and OPW. However, given the strong correlation of the movements in jobs and workers we would not expect significant differences.
Subject to user feedback, ONS intends to replace the existing Productivity Jobs and Productivity Hours estimates with jobs and hours estimates produced using the revised methodology described in this paper. Again subject to user feedback, and further quality assurance work, we further intend (i) to switch the focus in the labour productivity quarterly release from OPJ to OPW and (ii) to move to end-of-pipe seasonal adjustment, as described in the previous section.
However, we do not plan to make any changes to the next labour productivity release scheduled for 27 September 2013.
ONS operates a process of continuous improvement, and the changes described in this article do not mark the end of the road. ONS will consult with users over any further methodological changes or other developments to labour productivity statistics.
ONS publishes estimates of quality adjusted labour inputs (QALI), most recently in July 2013. QALI is benchmarked to Productivity Hours by industry and also to industry-level income constraints consistent with Sectional Unit Labour Costs (see below). One option for future development of QALI is to benchmark separately to the employee/self employed hours splits from the new system. However, this would ideally be accompanied by separate income constraints, which would require further work. But as a minimum, assuming that users agree with the next steps as outlined above, further QALI estimates would be benchmarked to new hours estimates employing the methodological changes described in this article.
ONS has recently developed estimates of sectional unit labour costs (SULCs) which anticipate some of the methodological improvements described in this article. Specifically, SULCs utilise industry-level estimates of employee jobs to derive income constraints for periods later than are available through the ONS Supply and Use tables, and seasonal adjustment is carried out “end of pipe”. At present, SULCs are experimental and are published outside the quarterly labour productivity release. Subject to progress as outlined in the next steps section above, we would expect to seek accreditation of SULC estimates as national statistics and to incorporate them in the labour productivity quarterly release.
Details of the policy governing the release of new data are available by visiting www.statisticsauthority.gov.uk/assessment/code-of-practice/index.html or from the Media Relations Office email: firstname.lastname@example.org