R&D and productivity: using UK firm-level data to inform policy

Abstract

The UK’s business R&D (BERD) to GDP ratio is low compared to other leading economies, and the ratio has declined over the 1990s. This paper uses data on 719 large UK firms to analyse the link between R&D and productivity during 1989–2000. The results indicate that UK returns to R&D are similar to returns in other leading economies and have been relatively stable over the 1990s. The analysis suggests that the low BERD to GDP ratio in the UK is unlikely to be due to direct financial or human capital constraints (as these imply finding relatively high rates of return).

Appendices

Appendix 1: first difference specification

Starting with

$$ \Updelta y_{it} = \Updelta \beta_{t} + \Updelta a + \alpha_{1} \Updelta n_{it} + \alpha_{2} \Updelta k_{it} + \alpha_{3} \Updelta r_{it} + \Updelta u_{it} , $$

where r is the log of R&D stock, hence (omitting the i index)

$$ \Updelta r = r_{t} - r_{t - 1} = \ln \left[ {{\frac{{{\text{RD}}_{t} + (1 - \delta ){\text{RS}}_{t - 1} }}{{{\text{RS}}_{t - 1} }}}} \right] = \ln \left[ {{\frac{{{\text{RD}}_{t} }}{{{\text{RS}}_{t - 1} }}} + (1 - \delta )} \right] \approx {\frac{{{\text{RD}}_{t} }}{{{\text{RS}}_{t - 1} }}} $$

where RD _it is R&D expenditure, RS _it is R&D stock and δ is the rate of depreciation of R&D. Under the assumption that δ and RD _it /RS_it − 1 are close to zero, the Δr _it term is approximately RD _it /RS_it − 1. The parameter α₃ is the elasticity of R&D (i.e. [dY/dRS] × [RS/Y]), hence Eq. 4 can therefore be re-written as

$$ \Updelta y_{it} = \Updelta \beta_{t} + \Updelta a + \alpha_{1} \Updelta n_{it} + \alpha_{2} \Updelta k_{it} + \alpha_{4} {\frac{{{\text{RD}}_{it} }}{{Y_{it} }}} + \Updelta u_{it} $$

(5)

where α₃ is now the gross marginal rate of return to R&D.

Data

2.1 General

The financial data include in the OIPRC database contain sales, profits, wages and salaries, assets, standard industrial classification (SIC), and also R&D expenditure. A major constraint on the size of the sample is the availability of R&D data in financial accounts. The accounting conventions covering disclosure of R&D changed substantially in January 1989 with the introduction of SSAP 13 (which effectively made large and medium sized firms report R&D in their accounts). Given these constraints, and the economic issues of interest, the analysis is confined to the period between 1989 and 2000. A firm is required to have at least 3 years of data to be included in the analysis (this removes 189 observations (4.3%) of the sample). A balanced panel of 86 firms is also constructed for the period 1990–1999.

2.2 Value added, labour and capital

The primary measure of output in the analysis is value added. Value added is defined as total staff costs plus depreciation plus profit before tax. We have calculated and analysed a measure of value added without depreciation and this is highly correlated to our measure (correlation coefficient 0.99). Also, some studies correct for interest paid and impute a value for the user cost of capital. The data available here do not contain interest paid. Some analysis was undertaken using imputed interest payments (based on debt and market rates of interest) and user cost of capital, but the value added measure obtained was similar to the measure used here.

Since the log of value added is always used in the analysis this effectively removes any negative values (this removes 7.8% or 347 observations).

The measure of capital is tangible fixed assets as entered in the balance sheet. The measure of labour input is the number of employees. Although the database contains variables for part-time and full-time employees, the coverage for these is minimal, so we are unable to adjust the labour measure to reflect differences across firms in part- and full-time working. All financial data are deflated with using the best available series.

2.3 R&D

The R&D data is as entered in company accounts. Since there is no published R&D deflators for the UK, a GDP deflator is used where needed (note that the ratio of R&D/value added allows nominal values).

Since the accounting data provides no breakdown for the nature of R&D we base adjustment for ‘double counting’ issue on ONS (2001). This indicates that 10% of R&D costs are capital, 40% are labour costs and 50% are materials. ONS data also indicate that of R&D employees 15% are scientists, 21% are technicians and 64% are administrators. Using ONS Annual Survey of Hours and Earnings (2000) we can find median wages for each of these categories. Assuming that the firm also pays national insurance, pension and related costs equally 37% of wages, we can impute a value for the median R&D worker (£23,868 in 2000 prices). Adjusting this using GDP deflator we can calculate an imputed value for R&D employees in each year, and then subtract this from employment (n). The 10% capital costs were simply deducted from capital stock since we have no information on depreciation rates. In our view this provides a crude adjustment for ‘double counting’. Some sensitivity tests for the median R&D worker salary were carried out and these had little effect on estimates.

2.4 Price deflators

The ONS statistics produces industry-level data only for manufacturing industries and only from 1991 onwards. For the 1980s there are no industry-level deflators available. The value added measured is deflated by the overall manufacturing deflator (gross output measure) for 1988–1990 and then industry-level deflators are used (which are at the 2-digit SIC level). These data are on ONS web-site and are listed under producer prices (MM22). For non-manufacturing the GDP deflator is used (ONS, YBFY). All deflators are in constant 2000 prices.

2.5 Outliers

Plots of the first difference of the dependent variable (log of value added) against the first difference of the log of assets and employment variables, as well as the ratio of R&D to value added (for period t − 1), indicate some outliers. The regression samples exclude high growth firms. Any firms with year-on-year growth in value added of greater than 300% and less than −90%. Also, firms with growth rates of labour, capital or R&D of greater than 200% and less than −50% are excluded. This follows Hall and Mairesse (1995) criteria and leads to 220 firms being omitted (4.4% of the sample). The main effect of omitting these firms is that the coefficients in the instrumental variable estimates become more significant and closer to theoretical predictions. Many papers do not discuss such data issues explicitly, although they often have samples of, say, just large firms, which may well remove the smaller firms with extreme R&D values. As noted, Hall and Mairesse (1995) is an exception, as is Los and Verspagen (2000) who omit any firm with a year-on-year sales growth of greater than 80% (in any year of sample).

2.6 Panels

The analysis below makes use of both an unbalanced and balanced sample, with both being defined for manufacturing and non-manufacturing firms. The Table 6 shows summary statistics for these different samples. Note that in both cases these summary statistics are based on samples that include firms with 3 or more years of data over the sample periods. The manufacturing unbalanced sample has 3,163 observations, which come from 468 unique firms. The median sales revenue is £182 million (the mean is £1.3 billion). There are fewer non-manufacturing firms in the data yielding only 1,288 observations (228 firms) (the median firm has sales of £190 million; mean of £1.1 billion). The non-manufacturing sample contains a wide range of industries. However, the major industries represented are ‘business services’ with 27% of observations (this includes software and data service firms); the wholesale trade industry (19%); electricity, gas and water utilities (14%); and, oil and gas extraction (8%).

Table 6 Summary statistics for R&D

Appendix 2

See Tables 7 and 8.

Table 7 GMM, dynamic panel estimations

Table 8 Olley-Pakes estimations

The above estimates are based on Olley and Pakes (1996). This asserts that η _it equals a function of investment and capital h(I, K), so that Eq. 2 can be written as

$$ y_{it} = \beta_{t} + a + \alpha_{1} n_{it} + \alpha_{3} r_{it} + \alpha_{2} k_{it} + h(I_{it} ,K_{it} ) + \varepsilon_{it} . $$

Using a third order polynomial in investment and capital to proxy for $\alpha_{2} k_{it} + h(I_{it} ,K_{it} )$we can estimate the above and obtain a consistent estimate for α₁ and α₃ (the spillover term is also added). These are shown in the table, along with results from tests on year and industry dummies that are also included. The third order polynomial is denoted by the function ϕ. In the second step we estimate

$$ V_{it} = \alpha_{2} k_{it} + g(\phi_{t - 1} - \alpha_{2} k_{it - 1} ) + \nu_{it} $$

where $V_{it} = y_{it}^{{}} - \hat{\alpha }_{1} l_{it} - \hat{\alpha }_{3} r_{it} - ({\text{spillover and dummies)}}$terms in order to get a consistent estimate of α₂. This is the estimate shown in the table.