Skip to content

Micro-data Perspectives on the UK Productivity Conundrum - An Update

Released: 04 October 2013 Download PDF


This article reports some new perspectives on UK productivity up to 2010, using a large dataset assembled from firm-level micro-data. A central finding is that productivity performance over 2008-10 has varied widely, by industry, firm size and a range of other firm-level characteristics. Our evidence suggests that the productivity conundrum is more pronounced in services (excluding the financial and communications sectors) than in manufacturing, and that labour productivity performance in 2010 was weaker among smaller firms than larger firms across all sectors. This is also the case using a broader measure of productivity, taking account of capital inputs at the firm level. Other things equal, firms that export, or are part of a multi-national enterprise, or report higher levels of ICT maturity demonstrate systematically stronger performance over the recession across all sectors than firms which do not have these characteristics. However, we find less evidence that recent productivity performance is related to prior growth rates of firm employment or a measure of firm-level innovation. In common with other micro-data research we find evidence of large and persistent variation in productivity across firms. Our results suggest that the impact of the economic downturn in 2008-09 has been more apparent among high productivity than low productivity service sector firms, while the reverse is the case in the high-tech sector, where high productivity firms have been little affected. There is also some evidence that the level of productivity below which firms exit the industry has fallen over the recession.


We are grateful for financial support from the European Commission (Grant 50721.2013.001-2013.082), to the other 13 national statistical institute consortium members (listed in Appendix A) for data access and to the academic advisers to the consortium: Eric Bartelsman of the Free University of Amsterdam and Patricia Kotnik of the University of Ljubljana and our ONS colleagues John Allen, Joseph Murphy and Bhavik Patel for technical and analytical support.


This article updates an earlier article published in January 2013 which used firm-level micro data to throw some light on to the productivity conundrum. In particular we have made a number of improvements to the underlying dataset and extended the dataset by one year to 2010.

Our point of departure is a unique dataset built from business micro-data as part of the 14-country EU-funded ESSlait project ( This is a very large dataset (the UK component alone runs to 24MB of data) with hundreds of thousands of separate estimates. This article reports a tiny fraction of the outputs, and it is perfectly feasible that we have overlooked some valuable information. The aim of the article is to raise awareness, to encourage feedback and to promote the use of linked micro-data in economic analysis.

The primary focus of the ESSlait project is to examine the impact of information and communication technologies (ICT) on firm performance. However, with some small amendments to the project coding, we can use this dataset to explore a number of lines of enquiry that cannot be addressed using conventional macro-level statistics1. The ESSlait project is an ESSnet project2 on linking of micro-data to analyse ICT impact, and continues work carried out over 2010-12 (Eurostat 2012) and 2006-2008 (Eurostat 2008). The underlying methodological approach is to link multiple micro-data sources to compile ‘micro-aggregated statistics’ (sometimes referred to as meso statistics) designed to inform policy-makers and researchers. Examples of meso statistics are where firm-level productivity estimates are aggregated by two or more categories over time (such as industry and a measure of the firm’s ICT maturity) and statistics on the distribution of firm-level productivity by industry, such as quartile averages. Additionally, the project dataset can be used to analyse firm-level demographics such as the characteristics of firms entering and exiting particular industries, and reports regression results for a standardised set of productivity specifications. An important feature of the ESSlait project and its predecessors is the comparability of outputs across consortium members. This approach is also designed to comply with confidentiality and disclosure control regimes governing the use of micro-data.

The layout of the rest of the article is as follows. Section 1 provides descriptive statistics of the dataset and some notes on interpretation of results.

Section 2 reports time series for labour productivity and total factor productivity up to 2010 by filtering the dataset in different ways, for example by size of firm, and - by merging information from different business surveys – considering interactions between productivity and (a) ICT usage and (b) innovativeness. This section also compares UK micro-aggregated productivity indicators with comparable indicators from other ESSlait consortium countries and reports some regression results.

Section 3 turns the focus to measures of the distribution of productivity across firms. This is one of the central benefits of using micro-data, and allows us to investigate heterogeneity between firms at a point in time and over time. This section also looks at measures of industry dynamics, focusing in particular on reallocation, the dynamic process by which resources flow from less productive to more productive firms. 

There are 3 appendices: a list of consortium members, an overview of the ESSlait dataset and data developments since our previous article, and a list of the full industry breakdown.

Notes for Introduction

  1. As in our previous article, our definition of ‘macro’ includes components within the national accounts framework, such as industry level measures of output, employment and productivity. micro-data refers to data at the level of individual firms and enterprises, typically but not exclusively in the form of their responses to ONS surveys.
  2. ESSnet projects are consortia of national statistical institutes (NSIs) and are aimed at providing results beneficial to the European Statistical System, see

Section 1: Descriptive statistics

Table 1 provides some basic summary statistics in terms of five broad industry aggregates. We focus on these broad sectors partly for reasons of brevity and partly to minimise risks of disclosure. The full dataset contains results at the 2-digit (SIC03) industry breakdown shown in Appendix C.

Table 1: Sector shares

Period averages, 2001-2010

  nv e pay k nobs pay/nv
  % % % % % %
MexElec 18 13 16 13 15 53
EleCom  6 4 5 6 3 55
MServ 50 53 48 45 56 57
NonMar 8 19 17 20 10 129
OtherG 18 11 13 16 16 43

Table source: Office for National Statistics

Table notes:

  1. n/v is nominal value added; e is employment; pay is wage costs; k is real capital stock; nobs is number of observations. Weights are approximate inverse sample probabilities (except nobs which is unweighted)

  2. Sector classification (SIC03, 2-digit) and description:
    MexElec (15-29, 33-37)  Manufacturing (excluding electrical machinery)   
    EleCom  (30-32, 64)  Electrical machinery, Telecommunication services   
    MServ (50-63, 71-73, 90-93)  Market services (excluding telecommunication services)   
    NonMar (70, 75, 80, 85)  Non-market services   
    OtherG (01-14, 40-41, 45)  Other goods
  3. Financial services (65-67) are not included in the dataset     

Download table

As in our January article, the focus of this article will be on MexElec (that is, SIC03 manufacturing other than electrical machinery), EleCom (an amalgam of electrical machinery plus post and telecommunication services and designed to proxy the 'hi-tech' industries) and MServ (market services excluding post and telecommunication services, and excluding the financial sector which is excluded from the survey sample frame). These three sectors respectively account for 18%, 6% and 50% of weighted nominal value-added in the whole project sample. For MexElec and EleCom the shares of value-added are a little larger than their shares of employment, whereas the employment share of MServ is a little larger than its share of value-added. The distribution of the capital stock is broadly consistent with the distribution of value-added and employment. The relatively low share of capital in MexElec probably reflects the disproportionate share of structures in micro-data capital stocks, which is also a feature of macro capital measures. See Appendix B for more information.

Table 1 also shows shares of pay in value-added, which averages 53-57% for the three featured industries. One reason why this is lower than labour’s share of national income recorded in the National Accounts is that the micro-data do not include non-wage labour costs such as employers’ social security and pension contributions. The pattern broadly reflects a priori assumptions about labour and capital shares across the three industries, recalling that – under SIC03/NACE1, EleCom includes labour intensive postal services. However, the pay share is greater than 100% in NonMar (public administration, education and health) reflecting measurement issues for value-added in the surveyed components of these industries. Thus, although the ESSlait dataset contains some interesting information on ICT usage in this sector, we do not report productivity estimates in this article.

Similarly the pay share of the OtherG sector (comprising agriculture, extractive industries, utilities and construction) is rather low. But this aggregate is very heterogeneous, and cutting the micro-aggregated data into its separate constituent parts is beyond the scope of this article. Again we do not report productivity estimates for this aggregate.

Table 2: Sample coverage

Period averages, 2001-2010

    All SZ1 SZ2 SZ3 SZ4
N_BR 000s 133.2 114.7 10.2 6.7 1.7
N_PS % of BR 6.2 2.2 16.5 40.1 85.0
N_PSEC % of BR 0.6 0.0 0.1 3.1 35.2
N_PSIS % of BR 1.2 0.2 1.8 7.2 40.8
N_BR 000s 30.3 27.2 1.6 1.1 0.3
N_PS % of BR 4.8 1.9 15.4 37.3 84.7
N_PSEC % of BR 0.6 0.0 0.1 3.6 39.3
N_PSIS % of BR 1.0 0.1 2.5 7.7 41.7
N_BR 000s 1234.6 1179.2 34.8 16.6 3.9
N_PS % of BR 2.4 1.6 11.9 27.2 79.9
N_PSEC % of BR 0.1 0.0 0.1 2.7 30.8
N_PSIS % of BR 0.2 0.0 0.9 4.5 40.0

Table source: Office for National Statistics

Table notes:

  1. For row headings, see text
  2. SZ1: Size class 1: employment<20    
    SZ2: Size class 2: 20<=employment<50   
    SZ3: Size class 3: 50<employment<250   
    SZ4: Size class 4: 250<=employment   

Download table

Table 2 provides summary information on sample size and overlaps between samples for each of the three sectors. N_BR is the number of firms in the sample universe (BR stands for business register), averaged over the sample period 2001-10. N_PS is the number of firms in the annual production survey (PS) datasets. And N_PSEC and N_PSIS are numbers of firms in the intersections of the PS and E-Commerce (EC) survey, and PS and Innovation Survey (IS) respectively, all expressed as percentages of N_BR.

On average there are around 1.4m firms in the sample frame for these three sectors, but only some 30,000 (2%) in EleCom and 133,000 (10%) in MexElec. More than 1.3m (94%) are in the smallest size class, and less than 6,000 (0.4%) are in the largest size class.

After outlier filtering and other deletions there are about 54,000 observations in the annual PS datasets, of which around 40,000 are in the three sectors which are the focus of this article. Of these, around 4% are in EleCom and 21% in MexElec. Differences in sampling probabilities across the size classes means that the share of the smallest size class falls to 54% while the share of the largest size class rises to 12%.

The EC and IS sample sizes are much smaller than the Annual Business Survey (ABS, which is the main source for our PS datasets, see Appendix B), and in addition, ONS sampling policy means that small firms selected for one survey are automatically not selected for any other survey for several years. This means that the sample size of the PS-EC intersection is much smaller (around 2,700 firms on average), with only tiny numbers of firms in the lowest two size classes. Moreover, since smaller firms are disproportionately found in services, the rate of attrition is larger for the MServ sector.

Sample attrition is not quite as severe for the PS-IS intersection, with slightly better representation for the smaller size classes. On average there are around 2,900 firms in this matched dataset1.

Table 3: Weighted employment shares

Period averages, 2001-2010

  Size Class   Exporter  MNC HGE FO
  1 2 3 4        
  % % % % % % % %
MexElec 18 13 28 41 58 51 31 60
EleCom  11 7 16 65 61 74 23 80
MServ 29 8 12 50 15 42 37 62

Table source: Office for National Statistics

Table notes:

  1. Weights are approximate inverse sample probabilities
  2. SZ1/2/3/4: as Table 2; Exporter: firm is exporter; MNC: Firm is Multi-national; HGE: firm is high growth enterprise; FO: firm is foreign owned

Download table

Table 3 focuses on the PS datasets and reports weighted shares of employment in terms of various firm characteristics. Here and elsewhere in this article, weights are approximate inverse sample probabilities, so these are estimates of the sample universe rather than the selected sample. This is a development from our previous article, when inconsistencies between the ABS sample and the register population meant that some of the sample re-weighted estimates were unreliable.
Although, as shown in Table 2, there are many more small firms than large firms, employment is heavily weighted towards large firms, especially in EleCom. Weighted employment shares of exporters may seem surprisingly high for MexElec and EleCom – a recent ONS article reported that only 11% of registered businesses in the non-financial business sector were engaged in exporting of goods or services. However, the same ONS article reports that larger firms are much more likely to be engaged in exporting than smaller firms. Unsurprisingly the employment share of exporters in MServ is much smaller.

Equally, employment shares of multi-national firms may seem surprisingly high – these firms make up less than 2% of all firms on the business register, and 4-5% of firms in MexElec and EleCom. But as for exporters, larger firms are much more likely to be multi-nationals than smaller firms. The employment share of high growth enterprises (defined as those with more than 10% annual employment growth for three years) is comparatively large in MServ and comparatively small in EleCom.

The employment share of high growth enterprises (defined as those with more than 10% annual employment growth for three years) is comparatively large in MServ and comparatively small in EleCom.

Table 4: Headline micro-aggregated statistics

Average compound growth, 2001-2010

  % % % % %
MexElec 5.3 4.0 -2.4 -1.6 3.5
EleCom  6.2 7.3 -2.8 -0.2 3.2
MServ 2.2 1.3 1.1 -0.5 3.8

Table source: Office for National Statistics

Table notes:

  1. LPV and TFP weighted by product of inverse sample probabilities and employment; E and K weighted by inverse sample probabilities; Pay/E derived from weighted pay and employment

Download table

Table 4 provides some broad information on growth rates of key variables pertaining to productivity. We focus on period averages to highlight differences between the sectors; time series for these series are reported in the following section. Productivity growth has on average been strongest for firms in the EleCom sector, followed by MexElec. Productivity growth has been considerably slower in MServ. This ranking holds both in terms of LPV and TFP, although the dispersion is wider in terms of TFP which has grown faster than labour productivity in EleCom but slower than labour productivity in the other two sectors. Employment growth has been negative over the 2001-10 period in MexElec and EleCom, and the real value of the capital stock has fallen in all three sectors. Average pay has grown at similar rates across the three sectors, although fastest in MServ and least in EleCom.

Since growth in labour productivity can be decomposed as growth of real value added minus growth of employment, these estimates imply growth of real value added of 2.8%, 3.4% and 3.3% for MexElec, EleCom and MServ respectively.

When interpreting the ESSlait productivity statistics it should be borne in mind that micro-data productivity estimates cannot be directly compared with productivity estimates derived from macro data, such as those in ONS’s quarterly Labour Productivity statistical release, for a raft of reasons, including:

  • micro-aggregated estimates of value added and employment will differ from those reported from the ABS and BRES due to differences in weighting and outlier filtering among other reasons. Users should also recall that ABS and BRES have moved to SIC07 with effect from 2008

  • although ABS is an important source for benchmarking value added in the UK National Accounts, it is by no means the only source. And in any event, timing lags mean that ABS is only used in benchmarking annual estimates around 18 months after the year in question

  • ONS labour productivity statistics make use of employee estimates that are benchmarked to BRES but again BRES is not the only source

  • both ABS and BRES survey only the corporate sector and do not capture information on non-incorporated businesses and most of the self-employed

  • macro statistics are subject to balancing adjustments to produce a coherent picture of the economy across the output, expenditure and income approaches

  • the relationship between chained volume and current price measures of value added is much more complex in the macro environment than the approach taken in this article2. Thus even in sectors where there is a close correspondence between ABS measures and macro current price measures of value added, in practice there will be differences in the volume series between the macro and micro-level data.

There are also some important differences between micro-data based estimates of TFP and ONS estimates of multi-factor productivity (MFP)3. The principal differences are that the micro-data measures labour input simply in terms of headcount employment, whereas ONS MFP estimates measure labour input in hours and account for changes in labour quality. And on the capital side, the micro-data measure is a crude aggregation of real stocks of structures, equipment and vehicles, whereas ONS MFP estimates attempt to measure flows of capital services using a more disaggregated asset breakdown.

Notes for Section 1: Descriptive statistics

  1. Questions in the biennial IS refer to innovation activity over the previous 3 years, so for years when IS is not conducted, such as 2009, we match firms in the PS with firms in the 2010 vintage of the innovation survey.
  2. UK macro statistics do not currently have a chained volume dimension in terms of gross output, but only in terms of value-added. Whereas the ESSlait database reports productivity both gross and net of firm-level intermediate consumption.
  3. In the ESSlait dataset, MFP refers to a measure of gross output divided by a weighting of labour, capital and intermediate inputs.

Section 2: Micro-aggregated productivity statistics

This section examines time series of productivity according to various firm level characteristics that are identified within the ESSlait dataset. Unless otherwise stated the data are weighted to provide estimates of productivity for the population of firms with the named characteristics. This is an improvement on the weighting methodology used in our January article, and reflects developments in achieving closer consistency between the continuous business register and the annual PS datasets.

Figures 1-3: Labour productivity by size class

Figures 1-3: Labour productivity by size class
Source: Office for National Statistics

Download chart

Figures 1-3 show time series of labour productivity for MexElec, EleCom and MServ differentiated by size class, as defined in Table 2 above. For MexElec (Figure 1) there is a very clear relationship between size class and labour productivity, and equally clear evidence of a recession effect in 2009, with productivity in all size classes picking up in 2010. However, only in the largest size class  did productivity in 2010 return to something like the pre-recession trend.

A somewhat similar picture is apparent for the two largest size classes in the EleCom sector (Figure 2), in terms of outperforming smaller size classes and recovering from a downturn in 2009. In fact, there is little evidence of a recession-induced downturn in labour productivity in these size classes. By contrast labour productivity fell sharply among firms in the two smaller size classes of the EleCom sector in 2008, and while there has been a decent recovery in size class 2, labour productivity was flat in size class 1 in 2009 and fell further 2010.

In MServ (Figure 3) there is no clear relationship between size class and level of labour productivity, although the ranking is fairly consistent over time, and the dispersion in productivity is less than for the other two sectors. Remarkably, firms in the largest size class are consistently at the bottom of the ranking. This pattern may reflect fewer opportunities to exploit economies of a scale in services, compared with manufacturing. The recession-induced downturn in productivity in 2009 is most pronounced for size classes 2 and 3. But while 2010 witnessed a recovery in productivity for size class 3 (and a muted upturn for size class 4) productivity continued to fall across the two smaller size classes. With the possible exception of size class 4 (where the pre-recession trend was weakest) productivity in 2010 was well below the level implied by the pre-recession trend.

Figures 4-6: Total factor productivity by size class

Figures 4-6: Total factor productivity by size class
Source: Office for National Statistics

Download chart

Allowing for the distribution of capital across firms changes the picture considerably. In MexElec (Figure 4) the ranking is completely reversed compared with Figure 1, suggesting that larger firms use more capital per person employed but extract less value added from their combined labour and capital inputs than do smaller firms. As in Figure 1 there is a clear recession effect in 2009 and a clear rebound in 2010, but unlike Figure 1 none of the size classes have fully recovered ground lost since 2008 in terms of TFP. A recurring pattern is that the steepest downturn in productivity was in the smallest size class.

Larger firms are also bottom of the TFP ranking in EleCom (Figure 5), but the time profile is similar to LPV, with little evidence of a recession-induced downturn. TFP for the middle size classes is remarkably similar, much more so than labour productivity (Figure 2), with modest falls in productivity in 2008 and 2009 but a reasonable recovery in 2010. Once again size class 1 stands out as experiencing the steepest fall in TFP, albeit with a slight recovery in 2010.

As for labour productivity, there is only mixed evidence of a recovery in TFP in MServ in 2010 (Figure 6). In contrast to the MexElec and EleCom sectors there is relatively little difference between productivity in terms of TFP compared with LPV.

Productivity by other firm-level characteristics

One advantage of the linked micro-data approach is that it allows ABS data to be analysed by other firm-level attributes from other sources which can be matched at the firm level using unique firm identifiers. Here we examine productivity trends up to 2010 in terms of a range of firm-level characteristics that can be identified in the ESSlait dataset.

(i) Multi-national corporations

In our previous article we showed that labour productivity was systematically higher for foreign-owned firms than for domestically-owned firms. As discussed in the previous section, for the current round of work we have compiled a panel of multi-national firms, some of which are foreign-owned and some of which are UK-owned.

Figures 7-9: Labour productivity by multi-national status

Figures 7-9: Labour productivity by multi-national status
Source: Office for National Statistics

Download chart

Labour productivity is also systematically higher for multi-national firms (MNCs) than non-MNCs across all three sectors (Figures 7-9). In general, falls in LPV in 2009 and recoveries in 2010 are both more pronounced for MNCs than for non-MNCs.

Corresponding estimates of TFP are available in the chart download component (385.5 Kb Excel sheet) of this release, which can be accessed by clicking on the link below Figures 7-9.

(ii) Exporters

Here we use an export flag to examine the relationship between export status and firm level productivity. As noted above, export status has not until very recently been captured by ABS so we use pooled information from monthly business surveys to populate an export panel which is then merged into our PS annual datasets. For further information see Appendix B. The ESSlait dataset also includes a variable on export intensity (export value as a share of turnover). Further information on export intensity is available from the analytical lead,

Figures 10-12: Labour productivity by exporter status

Figures 10-12: Labour productivity by exporter status
Source: Office for National Statistics

Download chart

In the MexElec sector labour productivity is higher among exporting firms than non-exporters (Figure 10), although the difference is not as pronounced as the productivity gaps in terms of size class and MNC status noted above. Another notable feature is that there was no recovery in productivity in 2010 across the non-exporting cohort.

As in our previous article we find that labour productivity is higher among non-exporters than exporters in EleCom (Figure 11), although the gap is narrower on our current dataset and appears to be narrowing over time. We also continue to see a considerable productivity advantage for exporters in MServ (Figure 12). In all three industries, labour productivity in 2010 is stronger for exporters than non-exporters.

Corresponding estimates of TFP are available in the chart download component (385.5 Kb Excel sheet) of this release (click on the link below Figures 10-12). One feature of interest in that the ranking in EleCom reverses in 2008 and later years, that is, exporting firms outperform non-exporters in terms of TFP.

(iii) High-growth enterprises

High-growth enterprises are defined as firms where employment has grown by at least 10% per year for the previous three years.

Figures 13-15: Labour productivity by high-growth status

Figures 13-15: Labour productivity by high-growth status
Source: Office for National Statistics

Download chart

Figure 13 shows that there is little difference in labour productivity in MexElec between high-growth firms and other firms, either in levels or growth rates. By contrast, the set of high-growth firms in EleCom (Figure 14) has demonstrated consistently higher labour productivity than their counterparts, and LPV grew recovered more strongly in 2010 across these firms. High-growth firms also outperform in terms of LPV in MServ (Figure 15). Corresponding estimates of TFP are available in the chart download component (385.5 Kb Excel sheet) , accessible via the link below Figures 13-15.

(iv) ICT intensity

Under this heading we examine the intersection between the PS and EC datasets. The EC survey collects a large amount of information on business use of ICT. One of the challenges for analysts is to identify summary ICT indicators which are consistently related to productivity. Here we focus on such a summary variable - BROADCAT – which takes different values depending on the proportion of workers that have access to the internet over a high speed connection. Specifically, BROADCAT=1 when this proportion is between 10% and 40%, and BROADCAT=2 when the percentage is between 40% and 90%.

The following results are slightly different from those previously reported. The survey base has been reduced because the BROADCAT variable is derived from the E-commerce survey meaning that that the sample size is smaller than in the previous characteristics which are drawn exclusively from the PS dataset. This accounts for the greater volatility of these estimates.

Figures 16-18: Labour productivity by ICT maturity status

Figures 16-18: Labour productivity by ICT maturity status
Source: Office for National Statistics

Download chart

Nevertheless it is apparent from Figures 16 to 18 that labour productivity is consistently higher among firms where more workers have access to high speed internet. LPV is also considerably more volatile for the BROADCAT=2 cohort of firms, or to put it another way. LPV is less volatile among firms with a lower level of ICT maturity as measured by this summary indicator. A similar pattern holds in terms of TFP, as can be seen in the chart download component (385.5 Kb Excel sheet) , accessible via the link below Figures 16-18.

Note that these measures of ICT maturity are not exhaustive – some firms lie below 10% and some lie above 90%. Moreover, the average level of ICT maturity has increased dramatically over the period 2001-10, such that some firms in the lower maturity cohort in one year will have moved into the higher category in the following year. The section on regression results later in this article reports positive and significant co-efficient on a continuous version of this ICT maturity indicator in OLS estimates of production function specifications, after taking account of other measurable inputs.

(v) Innovation

The ESSlait dataset also provides productivity estimates broken down in terms of firm-level characteristics taken from the biennial Innovation Survey (IS). As with the E-Commerce survey there is a wealth of information in the IS and part of the challenge to analysts is identifying summary indicators which demonstrate consistent relationships with firm performance. This is work in progress.

Figure 19: Labour productivity by innovation status

Figure 19: Labour productivity by innovation status
Source: Office for National Statistics

Download chart

Here we are using the PS-IS intersection, and the small sample size of the overlap produces extreme volatility of some of the weighted productivity estimates in EleCom and MServ. Figure 19 shows labour productivity of the MexElec firms in the overlap, split between those firms that report a new or significantly improved product or service (INPD=1) and those firms which do not (INPD=0). Over the sample period the respective proportions are around 40% and 60%. It is perhaps surprising that there is no evidence of a link between innovative firms on this measure and LPV. Indeed, non-innovators have experienced faster productivity growth over this period. There is also no evidence of different productivity performance during the recession and its aftermath. Neither is a relationship apparent in terms of TFP, as can be seen in the chart download component (385.5 Kb Excel sheet) , accessible via the link below Figure 19.

Comparisons with other countries

Figures 20 to 22 show UK labour productivity compared with weighted average productivity for the rest of the ESSlait consortium over the period 2001 to 2010. (See Appendix A for a list of consortium members. Note that at the time of writing, not all ESSlait member countries have delivered outputs to 2010.) For ease of exposition we have indexed productivity estimates for each country to 2007=100. Weights are based on employment shares across the consortium, by sector.

Figures 20-22: Labour productivity, UK and non-UK average

Figures 20-22: Labour productivity, UK and non-UK average
Source: Office for National Statistics

Download chart

For MexElec (Figure 20), UK labour productivity has grown consistently faster than the average across the rest of the ESSlait consortium and especially in 2008, when UK labour productivity increased while labour productivity fell across the rest of the consortium. Perhaps reflecting this, productivity fell more steeply in the UK in 2009, and the recovery in 2010 was less pronounced than across the rest of the consortium.

For EleCom (Figure 21) there is more evidence of a recession effect on labour productivity outside the UK. While for MServ (Figure 22), strong UK labour productivity growth up to 2007 contrasts sharply with much weaker growth elsewhere. But the corollary is that the fall in labour productivity in 2009 was much more pronounced in the UK.

A broadly similar story holds in terms of TFP, as can be seen in the <<chart download as can be seen in the chart download component (259 Kb Excel sheet) , accessible via the link below Figures 20-22. Note that some members of the consortium are unable to provide estimates of TFP due to the absence of firm level measures of capital inputs.

Regression results

The ESSlait coding generates a large volume of OLS regression results. The main aim is to compare regression coefficients between different project countries (since co-ordinated micro-data regression analysis is very scarce). The regression results for a single country are not intended as causal models – for example they take no account of endogeneity between the dependent variable and regressor variables. Nevertheless the multivariate structure can provides some insights.

Table 5: Sample Regression Output

IND EleCom   MexElec   MServ  
LNK 0.233 *** 0.286 *** 0.246 ***
LNE 0.722 *** 0.757 *** 0.697 ***
HKITPCT 0.518 ** -0.014 0.414 ***
HKNITPCT 0.524 ** 0.215 ** 0.182 ***
AGE -0.004 0.012 ** 0.033 ***
AGE2 0.000 0.000 *** -0.001 ***
BROADPCT 0.543 *** 0.549 *** 0.524 ***
MNC0 -0.068 -0.116 *** -0.182 ***
EXPORT0 0.113 -0.037 -0.181 ***
R-SQD 0.790 0.780 0.776
NOBS 918 4456 9201

Table source: Office for National Statistics

Table notes:

  1. Significance levels- ***=0.1%, **=1%, *=5%
    LNV             Dependent Variable, log value added      
    LNK             Log capital stock      
    LNE             Log employment      
    HKITPCT    Proportion of workers with higher IT qualifications      
    HKNITPCT  Proportion of workers with higher non-IT qualifications      
    AGE             Firm age      
    AGE2           Firm age squared      
    BROADPCT  Proportion of Broadband enabled employees in firms      
    MNC0           Firm is not a multinational      
    EXPORT0    Firm does not export      
    R-SQD         Goodness of fit measure defined as Explained Sum of Squares/Total Sum of Squares       
    NOBS           Number of observations      
    Fixed effects (industry, size class and year) included but not reported       

Download table

Table 5 shows the results of one fixed effects panel regression model which can be interpreted as a generalised production function, with the log of real value added as the dependent variable. The table illustrates the different parameter results obtained for the industry categories EleCom, MexElec and MServ for a number of different variables.

The results of table 5 indicate that across the three industries, capital and labour are the strongest, most significant predictors of output, their coefficients summing roughly equal to one in the case of each industry category, consistent with constant returns to scale. This finding (which generally holds across a number of regression specifications) provides some informal support for the quality of the micro-data estimates for employment and firm-level capital. Moreover the size of the respective coefficients on labour and capital seem plausible.

Having a high proportion of employees with post upper secondary ICT education appears to have a large and significant effect on production in the EleCom sector, which is likely due to reliance of its constituents on IT ability. In the MexElec category, only a very small, statistically insignificant value could be found. In the case of MServ, another industry category where constituents rely on IT ability, the coefficient is highly significant, and only a little smaller than that of EleCom.

The results above contrast with the results for the proportion of employees with post secondary but not IT related qualifications. In that case, while a significant and large parameter is found for EleCom, comparatively small values are found for MexElec and MServ, all three results are significant. This indicates that in the EleCom category, having a high proportion of highly skilled workers with and without advanced IT skills have significant effects on production. For MexElec, while having higher IT skills is deemed insignificant, having non-IT post secondary education does significantly correlate with output. In MServ, while the figure is significant, it is the smallest of the three coefficients, and compared to the figure with IT skills it shows that in MServ, having more employees with higher IT skills contributes to output much more than having higher non-IT skills.

Both the firm age and age squared variables produce relatively small parameters. In the case of those in EleCom, they are both statistically insignificant. However, for MexElec and MServ the parameters are highly significant for both variables, and they do have measurable effect on output. This could be indicative of a “Horndal effect” (Lazonik and Brush, 1985), whereby firms maintain assets and workers for a long period of time, becoming more productive over this period simply by gaining experience. The age coefficient is almost 3 times larger in MServ than in MexElec, suggesting that a typical service firm can produce some 3% more value added simply by being a year older, with no change in factor inputs. The negative sign on the age squared variable suggests that this effect decays over time, but the coefficient is tiny. For MexElec firms, the age squared variable suggests that the impact of aging gets fractionally (but significantly) greater over time.

Having a high proportion of broadband enabled employees seems to positively and highly significantly affect production across all three industry categories. The parameter values are very similar, which is likely to reflect a strong reliance on broadband access across the three industry categories. These coefficients and the units of measurement imply that an increase of 1 percentage point in the share of workers with broadband access is associated with an increase of over 0.5% in real value added. This is an interesting finding, although more work is needed to establish causality.

Being a non-multinational company (note that the regression coding creates a dummy variable for non-MNCs, and non-exporters) appears to have a small negative effect on production. For EleCom this result is statistically insignificant, however for the other categories, it is highly significant. This is consistent with the TFP relationships noted above.

The effect of being a non-exporting company appears to be insignificant for categories other than MServ. For MServ, the value is negative and a reasonable size, meaning that for MServ constituents, being a non exporter will negatively affect output. Again this is consistent with the relationships between TFP and export status noted above.

Section 3: Productivity distributions and market dynamics

This section explores the distribution of firm-level productivity a little further. We begin with a summary measure of industry dynamics known as the Olley-Pakes (OP) coefficient (Olley & Pakes, 1996). We then look at quartiles of productivity, which provides an indication of the overall distribution of productivity across industries, and extend this analysis further to summary statistics of other firm characteristics (such as employment and pay growth) in terms of the firm’s position in the productivity distribution. Lastly we examine productivity in terms of firm-level demographics over the recession.

OP coefficients

The intuition behind OP is that the observed heterogeneity of firm-level productivity reflects a dynamic process (‘reallocation’) whereby increases in productivity at the aggregate (industry) level reflect at least in part a reallocation of output from less productive firms to more productive firms (as opposed to an increase in average productivity of all firms). This implies that, at any point in time, productivity should be correlated with firm size. Olley & Pakes pointed out that this can be quantified very simply, by subtracting an unweighted measure of productivity from a comparable measure weighted by firm size. Effectively, OP extends the analysis of productivity by size class in the previous section.

Figures 23-25: Labour productivity, Olley-Pakes coefficients

Figures 23-25: Labour productivity, Olley-Pakes coefficients
Source: Office for National Statistics

Download chart

Figures 23-25 show the OP coefficient for labour productivity, computed as the difference between weighted and unweighted LPV, both expressed in logs. For MexElec and EleCom the OP coefficients are positive, consistent with the interpretation that reallocation of output from less productive to more productive firms is occurring in these industries. Moreover, the OP coefficient in MexElec increased sharply in 2010, suggestive of a pick up in the reallocation process as the sector recovered from the recession.

The OP coefficient also increased in the EleCom sector but earlier, in 2008 and 2009, and the movement was reflected more in the unweighted productivity measure than the weighted measure.

The OP coefficient is consistently negative in MServ (Figure 25), consistent with earlier evidence (Figure 3) showing lower productivity among firms in the largest size class. This implies that reallocation works to lower aggregate productivity in this sector, possibly reflecting weak competitive pressures in MServ, though it may also reflect other factors such as measurement error and compositional changes. It is also worth noting that the OP coefficient in MServ in 2010 was sharply less negative than in 2009.

OP coefficients can also be computed in terms of TFP, although in this case the weighted productivity aggregate is weighted by both labour and capital inputs. These results (available in the link below Figures 23-25) suggest that once capital is accounted for, the reallocation effect of output going to more productive firms disappears. Across all 3 sectors there appears to be a negative redistribution; output seems to be redistributed to smaller firms. And evidence of a recession-induced change in OP coefficients is less apparent. One possible explanatory factor for the difference between OP coefficients in terms of LPV and TFP is that capital is less mobile than labour across firms in the same sector.

Productivity Quartiles

Another perspective on the distribution of productivity across firms is provided by summary statistics of ranked quartiles. Here we rank firms in an industry by (unweighted) productivity and report summary statistics for each quartile. In the figures in this section, the top and bottom points of each line represent the average of the highest and lowest quartile of firms in that industry and year respectively, with the (unweighted) average for the whole sample shown for comparison purposes. The height of the lines can be seen as a crude summary of dispersion of productivity across firms.

Figures 26-28: Labour productivity, highest and lowest quartiles

Figures 26-28: Labour productivity, highest and lowest quartiles
Source: Office for National Statistics

Download chart

Alongside the evidence in the previous section, Figures 26-28 provide strong evidence of wide heterogeneity in firm-level productivity performance. Remember that the high and low points are averages of productivity in the top and bottom quartiles; a cursory analysis of standard deviations of productivity within these quartiles (a full analysis is beyond the scope of this article) shows that there is a much wider gap between the very highest and lowest performing firms in all sectors.

The three sectors display different reactions to the 2008-09 recession. The least affected industry seems to be MexElec, where the productivity dispersion widened gradually up to 2007 and rather more sharply in 2008, but has then more or less moved in step with the overall average. This suggests that the variance between the most and least productive firms has not changed as a result of the recession.

EleCom firms in the highest productivity quartile do not seem to have been affected by the recession at all. The effect of the recession is more visible in the lowest quartile, with the average productivity of the bottom quartile of firms turning negative in 2008, and some sign of a double dip in productivity in 2010 after a small recovery in 2009.

By contrast the effect of the recession in MServ is most apparent at the top of the distribution, with a clear narrowing apparent since 2008. As in our previous article we find that average productivity among the least productive quartile of firms is consistently negative in this sector. There was some deterioration in 2008 and 2009, but nothing like as much movement as at the top of the productivity distribution.

Similar trends in productivity distributions are also apparent in terms of TFP (available in the link below Figures 26-28).

Further characteristics of productivity distributions

In addition to summary statistics of quartile distributions of productivity (as above) we can examine cross relationships with other variables of interest, along the productivity distribution. Specifically, we can investigate changes in employment and wages in each of the above productivity quartiles. Some caution needs to be exercised in interpreting these estimates. To be able to report changes requires that firms are present in both period t and t-1, leading to considerable sample attrition compared with the simple productivity distribution. For example, for the purpose of reporting LPV quartiles there are approximately 6,000 firms in MexElec in 2010. But of these, changes in employment and changes in wages can be computed for less than 3,000 firms.

Figures 29-31: Employment growth by quartile of labour productivity

Figures 29-31: Employment growth by quartile of labour productivity
Source: Office for National Statistics

Download chart

Figures 29-31 show average percentage changes in persons employed for each of the labour productivity quartiles. There is no obvious relationship between employment changes and productivity quartiles for either MexElec or EleCom. Employment fell across all productive quartiles in MexElec in 2009 and for all but the lowest quartile in 2010. By contrast there was a large fall in employment among less productive firms in the EleCom sector in 2009 before a belated drop in employment across all quartiles in 2010.

But the most interesting sector is MServ, where 2010 witnessed a clear break in the previous pattern of exceptionally strong employment growth over the whole of the sample period, albeit that employment growth remained positive for all but the most productive quartile. The resilience of employment among less productive MServ firms is a consistent feature of the micro-data and a likely contributory factor in the ‘productivity conundrum’.

Figures 32-34: Wage growth by quartile of labour productivity

Figures 32-34: Wage growth by quartile of labour productivity
Source: Office for National Statistics

Download chart

Analysing average percentage changes in wages by labour productivity quartiles (Figures 32-34) suggests a number of interesting trends. Firstly there is a very clear positive relationship between wage growth and productivity quartiles in MexElec and MServ over the whole period. This is much less evident in EleCom, suggesting that labour productivity is less of a driver of wage growth in this sector. (Note however that EleCom looks more like the other sectors if we use TFP as the metric for productivity quartiles).

Secondly, average wage growth is negative among the least productive firms in MexElec and MServ in all years, which may account for some of the resilience of employment growth noted above. Thirdly, there is clear evidence of a recession effect in 2009 in MexElec (where wages fell on average across all but the most productive quartile of firms) and MServ (all quartiles).  In both sectors this was a short lived phenomenon – wage growth resumed in 2010 and was particularly strong in MexElec.

Productivity by demographic status

We conclude this section of the article with a brief look at firm-level demographics. Over time, entry and exit plays a major role in movements in productivity, see Dunne et al (1988) for a quantitative analysis of US manufacturing. Here we report simple averages of labour productivity, distinguishing between continuing firms (CO), entrants (EN) and exiting firms (EX). Note that these characteristics are taken from the continuous business register. Exiting firms cannot be identified for the last year of the sample, and entrants cannot be identified in the first year. For context, in a typical year more than 90% of firms are continuers, with entrants and exiting firms being a few percentage points each.

Figures 35-37: Labour productivity by demographic status

Figures 35-37: Labour productivity by demographic status
Source: Office for National Statistics

Download chart

Figure 35 shows that in MexElec, the average productivity of exiting firms and entrants is consistently lower than that of continuing firms. For entrants this partly reflects the fact that new firms are typically smaller than incumbents. However, drilling down in the ESSlait dataset reveals that the average size of exiting firms is broadly comparable with the average size of incumbents. The drop in productivity of exiting firms in 2008 and 2009 is somewhat counter-intuitive. Other things equal one would expect a recessionary shock to drive better performing firms out of the market compared with normal turnover in the industry, whereas these estimates suggest that exiting firms during the recession were on average markedly less productive than exiting firms in earlier years. This finding is even more apparent for EleCom and MServ, perhaps suggesting that competitive pressure to exit (sometimes referred to as ‘creative destruction’) has weakened in recent years, despite the recession. This is turn may reflect increased forbearance by lenders. However, the sample size for exiting firms is relatively small in all sectors so caution needs to exercised in interpreting these results.

Background notes

  1. Details of the policy governing the release of new data are available by visiting or from the Media Relations Office email:


  1. Appleton J and Franklin M, 2012, ‘Multi-factor Productivity – Indicative Estimates, 2010’, Available at:
  2. Bartelsman E, Haltiwanger, J and Scarpetta, S, 2009, ‘Cross-country differences in productivity: the role of allocation and selection’, NBER Working Paper 15490, November 2009. 
  3. Crawford C, Jin W and Simpson H, 2013, ‘Productivity, Investment and Profits during the Great Recession: Evidence form UK Firms and Workers’, Fiscal Studies, vol. 34, no. 2, March 2013
  4. Criscuolo C, Haskel J and Martin R, 2003, ‘Building the evidence base for productivity policy using business data linking’, Economic Trends, volume 600, November 2003
  5. Dunne T, Roberts M and Samuelson L, 1988, ‘Patterns of firm entry and exit in US manufacturing industries’, Rand Journal of Economics, Vol 19, No 4, Winter 1988
  6. Eurostat (2008), ‘ICT impact assessment by linking data from different sources – Final Report’, Available at:
  7. Eurostat (2012), ‘Final Report: ESSnet on Linking of Micro-data on ICT Usage’, Available at:
  8. Field S and Franklin M, 2013, ‘Micro-data perspectives on the UK productivity conundrum’, Available at:
  9. Lazonick, W and Brush, T (1985), ‘The “Horndal Effect” in Early US Manufacturing’, Explorations in Economic History, Vol 22, pp 53-96
  10. OECD, 2013, ‘Indicators of productivity and competitiveness: Improving coherence and applying micro-data sources’, STD/CSTAT(2013)2.
  11. Olley G and Pakes A, 1996, ‘The Dynamics of Productivity in the Telecommunications Equipment Industry, Econometrica, Vol 64, No 6
  12. Robjohns J, (2006). ‘ARD2: the new Annual Respondents Database’, Economic Trends, No 630, May.

Appendix A: List of participants in ESSLAIT project

UK  Office for National Statistics
Sweden  Statistics Sweden
Netherlands Statistics Netherlands
Norway  Statistics Norway
Italy  National Statistical Institute of Italy
Germany Statistisches Bundesamt
France  Institut National de la Statistique et des Études Économiques
Denmark Statistics Denmark
Ireland  Central Statistics Office
Slovenia Statistical Office of the Republic of Slovenia
Austria  Statistics Austria
Finland  Statistics Finland
Luxembourg Service Central de la Statistique et des Études Économiques
Poland  Central Statistical Office of Poland

Appendix B: The ESSlait dataset

The primary source of the data used in this article is the annual structural business survey, which since 2008 has been known as the Annual Business Survey (ABS) in the UK. ABS collects information from approximately 50 thousand firms on a number of economic variables (e.g. turnover, value-added, employment costs, capital investment etc). Additionally, the ESSlait project draws on firm level data from other business surveys, principally an annual survey of business use of ICT, known as the E-commerce (EC) survey, and the Community Innovation Survey (IS) which is conducted every two years and collects a range of information on innovation, such as whether firms have introduced new products or processes.

In addition to these primary sources, the ESSlait dataset merges in other firm-level information which is not collected in any of ABS, EC or IS. Chief among these is employment, which plays a key role in the ESSlait project framework (as a core measure of firm size) and, from 2008, comes from the annual Business Register and Employment Survey (BRES)1. Additionally, two export variables (an exporter/non-exporter flag and a measure of export sales as a share of turnover) are merged in from a panel built from successive editions of the monthly business survey (MBS) and we also merge in a dataset on firm-level capital stocks. The derivation of these datasets is described further below. We use questions on the IS to populate variables on firm-level skills. The combined dataset of annual firm-level economic variables is referred to in this article as the PS dataset.

The ESSlait project architecture also utilises a continuous register dataset containing basic information (on employment, firm age, and 1/0 flags for whether the firm is foreign owned and/or part of a multi-national enterprise) on the universe of firms of which the ABS and its predecessors are samples. This register is derived from similar sources, but is not identical to either the Inter-Departmental Business Register (IDBR) or the Business Structure Dataset (BSD), both of which have been used by other micro-data researchers. The principal differences are that the IDBR covers all parts of the economy, whereas our register is confined to those sectors that are sampled in ABS (and its predecessors), and the BSD is available in terms of enterprise units and local units, whereas our register is, like ABS, structured in terms of reporting units2.

The register dataset serves two main functions. First, to provide information on firm demographics (entry and exit). And second, to allow weighted estimates to be computed for any sub-category of micro-aggregated data, including data from two or more merged datasets. It is important to realise that the ESSlait dataset does not use reported sample weights for ABS or any other source dataset, but rather endogenous weights computed within the project coding. 

Information on whether the firm is part of a multi-national enterprise is merged into the register dataset from a panel of multi-national firms constructed from the Annual Survey of Foreign Direct Investment (AFDI). AFDI is also used to supplement administrative data on foreign ownership, such that firms in receipt of inward direct investment are deemed to be foreign owned. The addition of a multi-national flag is a development from the dataset used in our previous article.

In the final register dataset there is some overlap between multi-national and foreign ownership flags but not an exact correspondence – some UK-owned firms are multi-nationals and many firms recorded as foreign owned are not recorded as multi-nationals. In the results sections below we report estimates in terms of the multi-national flag rather than (as in our January article) the foreign ownership flag because we believe the new variable to be a more robust indicator.

Transition from ABI1/2 to ABS/BRES

In the ONS micro-data research environment (known as the virtual micro-data laboratory, VML) ABI1 and ABI2 survey results up to 2007 were used to compile a set of annual datasets known as the Annual Respondents Database (ARD) (Robjohns 2006). This process used coding developed over a number of years by VML researchers to map survey questions to a standardised set of ARD variables and carry out other tasks such as data-cleaning, assignment of industry codes and so forth. Extracts from the ARD datasets were used to compile PS datasets up to 2007 in earlier phases of ESSnet work. But the ARD coding is not robust to the transition from ABI1/2 to ABS and BRES so we (and other micro-data researchers) have had to do our own processing to create ARD-like datasets from 2008 onwards from the ABS and BRES source data.

Parallel to this, the VML team compiled a register specific to the ARD, known as the ARD Register Panel. Again these datasets were used to compile the register datasets for the previous phases of micro-data linking work, and again compilation of these datasets broke down after 2007. For 2008 and later years we have carried out work in the VML to replicate the ARD Register Panel from the ABS and BRES sample frames.

One specific issue is that whereas ABI1 and ABI2 were drawn from the same sample frame, ABS and BRES have slightly different sample frames and are drawn from the IDBR at different times. BRES datasets for 2008 and later were not available in the VML at the time of compiling the register dataset for our January paper. For this update we have converted BRES to a reporting unit basis and used this as source for firms that are in both the ABS universe and the BRES survey sample. We have also gone back to fix missing records on foreign ownership, fixed a discontinuity between the universe and the (ABI) sample in 2006, and made a continuous age variable for all years.

A further issue concerns outlier filtering. In ABS there is an explicit outlier marker (used to suppress observations in weighting ABS responses to population totals), but investigation showed that many of the records marked as outliers had perfectly plausible productivity characteristics, while other firms not marked as outliers had implausible characteristics. So a pragmatic approach of selective outlier filtering was used, in which records with nominal productivity (value-added and turnover based) more than 6 standard deviations from the sample mean were deleted, plus further deletions where quality assurance of the micro-aggregated results revealed implausible time series properties of productivity or other metrics.

Capital stocks

Coding to build a dataset of firm-level capital stocks has been completely re-run since our January article. In particular, we have benchmarked starting values to 2-digit capital stock estimates derived from an exercise in 2012 to derive volume indices of capital services for the purpose of estimating multi-factor productivity (MFP), see (Appleton & Franklin 2012)3. This has led to a completely revised set of firm-level capital stock data, but one where the distribution of capital across industries is a priori more plausible than previously (see next section on descriptive statistics). Building firm-level capital stock dataset is not a trivial task and requires a number of assumptions to be made. Accordingly, we judge these estimates to be less robust than our measure of firm-level employment. And more work remains to be done. In particular, we would like to expand the asset breakdown to reflect more closely the breakdown used in ONS’s MFP framework (and note that this differs from the asset breakdown used to compile capital stock estimates in the current National Accounts). Secondly, we are planning to introduce more “triangulation” between asset stocks at the micro-data level and the equivalent series at the macro level. Thirdly, we plan to combine micro-data capital stock estimates with information on firm demographics to derive estimates of capital scrapping.

Industry classification

Like other international datasets such as EUKLEMS and the OECD STAN database, the ESSlait industry taxonomy is classified in NACE1, equivalent to SIC03 in UK terms. ABS and other ONS business surveys moved to the SIC07 industry taxonomy in 2008. The ABS survey data for that year and for 2009 are dual coded, but for 2010 (and later years) ABS contains only SIC07 industry codes. The project uses dual industry coding to analyse mappings between industry classifications and derive weights (based on employment) where SIC07 codes are mapped one-to-many from SIC03. These weights are used to proportion new firms back to SIC03. Continuing firms in such industries are coded to their original SIC03 coding.

Export variables

Perhaps surprisingly, ABS and its predecessor surveys have only included questions on exports (and imports) of services since 2007, and only introduced questions on exports (and imports) of goods since 2011. This means that we have had to look for other sources to populate the export variables used in the project architecture. Following user feedback after the January article, we have reviewed the compilation method for creating the export flag, as described below. This has meant that the number of firms classified as exporters differs from our previous article.

There are two main difficulties in deriving a panel of exporting firms from monthly business surveys. Firstly the MPI (Monthly Production Inquiry) was replaced in 2009 with the MBS (Monthly Business Survey), this has led to some inconsistencies in survey questions. Secondly, only large firms are continually surveyed, small and medium sized firms are surveyed sporadically. This can cause problems when deciding whether or not a firm is an exporter in years when the firm has not been surveyed.

To create a consistent export panel, questions 40 (“what is this firm’s turnover?”) and 49 (“value of exports included in turnover”) are taken from the MPI and MBS. This creates a panel of firms showing turnover and exports for each month of the year. We create a yearly turnover to export ratio by taking the sum of responses to both questions throughout the year and then taking the ratio of exports to turnover. This is done for all years, the individual years panels are merged to create a complete panel. This shows two things; the yearly ratio of firms’ exports to turnover, and whether a firm has been surveyed in any given year.

The complete panel also demonstrates the sporadic nature of small and medium sized firms being surveyed. To create an export flag for small and mid size firms, we calculate how many observations show that the firm is exporting (by having an export ratio which is not equal to zero).  We then divide the number of observations for which a firm is exporting by the total number of observations on the firm. This creates an exporter probability.

If this probability equals one then it suggests that the firm is an exporter, therefore for any years for which data are missing we replace with the average export ratio. Any firms for which the exporter probability is zero we interpret as indicating that the firm is not an exporter, and so we replace missing observations with zero. If the probability is between zero and one we cannot conclude whether the firm exports. Therefore, we do not alter the observations for which the firm has reported being an exporter and do not replace any of the missing observations. 

Firm-level productivity measures

As in our previous article, in this article we focus on two measures of productivity at the firm level: labour productivity and total factor productivity. Labour productivity is denoted as LPV and is computed from firm-level measures of real value-added and employment. Note that real value-added is derived using industry value-added deflators, since no firm-level deflators are available. Value added is recorded in the UK micro-data in £k, employment in actuals and deflators are indexed to 2005=100, so LPV in the figures and tables in this article are £k per person employed in prices of 2005.

Firm-level total factor productivity (TFP) is computed by dividing real value-added by a weighted index of employment and capital, where the weights are derived from industry shares of labour and capital remuneration as computed from the micro-data. Units of TFP in the figures and tables have no direct intuition because the units of labour and capital are not comparable. Increases in TFP imply that real value-added is rising more (or falling less) than weighted factor inputs and vice versa4

Notes for Appendix B: The ESSlait dataset

  1. Prior to 2008 the information collected in ABS and BRES was collected through two versions of the Annual Business Inquiry, known as ABI1 and ABI2. See below for more on data issues resulting from the move from ABI1/2 to ABS and BRES.
  2. For more information on the differences between enterprise units, local units and reporting units, see Criscuolo et al, 2003. For more information on coverage of the ABS, go to the ABS index page on the ONS website.
  3. Capital services and MFP are on SIC07, so benchmarks have been converted back to SIC03.
  4. The project database also contains results for two further productivity metrics, where real value added is replaced with real turnover. For more information on results using these alternative productivity metrics, please contact the authors.

Appendix C: Full industry breakdown

EUK Industry Definitions:




Total Economy








__Food, Beverages and Tobacco






__Pulp, paper, publishing


___Pulp, paper and paper


___Publishing and Printing


__Refining, chemicals, and rubber


___Refining and chemicals


___Rubber and plastics




__Metals and Machinery


___Basic metals


___Fabricated metal


__Machinery and Equipment






____Office, accounting, computing and scientific machinery


____Electrical Equipment


____Electronic Equipment


__Motor Vehicles and Transport Equipment


___Motor vehicles, trailers and semi-trailers


___Transport Equipment


__Misc Manufacturing






_Market Services


__Trade, Hotels, Restaurants


___Trade, Hotels, Restaurants


____Sale, and repair of motor vehicles and motorcycles; retail sale of fuel


____Wholesale trade , except of motor vehicles and motorcycles


____Retail trade, except of motor vehicles; repair of household goods




__Transport and Communications








__Real Estate and Bus Services


___Real estate activities


___Renting of machinery and other Bus services


___Computer and related activities


___Research and development


_Social Services








__Personal Services


___Personal Services excl media


___Media activities


Content from the Office for National Statistics.
© Crown Copyright applies unless otherwise stated.