Estimating Firms’ Research Quotient (RQ) Anne Marie KnottThe Harvard Business Review article, “The Trillion Dollar R&D Fix” describes a measureof firms’ R&D effectiveness called RQ (Research Quotient) and explains how firms canutilize it to determine the optimal level of R&D investment, R*.At the request of the HBR editors, the article also includes an excel example capturing theintuition behind RQ. This is a very simplified approach to estimating RQ, which leads tosubstantial errors. In fact the raw RQ for the fictitious firm in the Excel example is high bya factor of two. I included a caution in the article comparing the two estimates, but it didn’tsurvive final edit.I have received a number of inquiries from firms indicating they want to estimate RQinternally on a regular basis. Because erroneous estimates of RQ could create worseproblems than the current problem of not knowing RQ, I have created this tutorial.The tutorial describes how firms’ can estimate RQ on a regular basis. In addition togenerating top-level (company-wide) estimates for RQ using “real numbers”, the advantageof internal estimation is the ability (in some cases) to generate division level estimates ofRQ and R*. Such division level estimates 1) inform how R&D should be allocated acrossdivisions, and 2) support internal benchmarking (by identifying which divisions havehigher/lower R&D effectiveness).The tutorial has five sections. Section 1 describes the link to theory; Section 2 defines thefully specified approach to estimation; Section 3 describes the excel exercise and explainswhy it leads to erroneous estimates; Section 4 identifies a commercial means to obtain firmRQs; Section 5 provides links to other RQ resources.1. TheoryThe formal link between firms’ R&D spending and growth come from endogenous growththeory, e.g., Romer 1990. These models rely on a construct of R&D elasticity to define theR&D spending that maximizes firm market value. The principal limitation of growththeory from the firm perspective is that it is concerned with macro-economics and thustypically treats firms as being identical. Accordingly, when the models are tested, R&Delasticity is estimated at the industry level or higher. More recent growth models e.g, Lentzand Mortensen 2008, have accommodated firm heterogeneity, but there hasn’t been a firm-level measure of R&D elasticity to test them directly. RQ is the missing measure.Raw RQ is the “firm-specific output elasticity of R&D”. It is the term, γ, in the firm’s finalgoods production function (equation 1). Gamma is interpreted as the percentage increasein revenues associated with a 1% increase in R&D.Y = K L R S Aα β γ δ ε (1)where:

Y = outputK = capitalL = laborR = R&DS = spillovers (R&D available for free-riding)A = advertisingBecause γ is normally distributed, firm RQ resembles individual IQ. Both capture problemsolving capability. For individuals, IQ is captured as the speed and accuracy of solvingproblems of increasing difficulty--within any given time constraint, individuals with higherIQ solve more problems correctly than those with lower IQ. For firms, IQ is efficiencysolving new problems. For any given level of R&D spending, high IQ firms will generatemore innovations, or for any given innovation, high IQ firms will invest less developing it.Accordingly, the raw values of γ are mapped onto the IQ scale (mean = 100, standarddeviation = 15) to support intuition.2. Fully-specified estimation2.1 ModelWe derive (RQ) by estimating the production function (equation 1) with a randomcoefficients model that allows for heterogeneity in the output elasticity for R&D (as well asall other inputs). A random coefficients model represents a general functional form modelwhich treats coefficients as being non-fixed (across members of a cross- section or overtime) and potentially correlated with the error term. Random coefficient models are thosein which each coefficient has two components: 1) the direct effect of the explanatoryvariable and 2) the random component that proxies for the effects of omitted variables. Theempirical model (Equation 2) is a log transform of equation 1 that models output (value-added, Y) for firm i in year t with random coefficients for all inputs (capital, K, labor, L,R&D, R, spillovers, S, and advertising, A) as well as the intercept:ln Yit = (β0 + β0i ) + (β1 + β1i ) ln Kit + (β2 + β2i ) ln Lit + (β3 + β3i ) ln Rit+ (β4 + β4i ) ln Sit + (β5 + β5i ) ln Ait + εit (2)We estimate Equation 2 using the Stata program, xtmixed. xtmixed fits linear mixedmodels (both fixed effects and random effects) using maximum likelihood estimation. Therandom effects, β_i, are not directly estimated, but we form best linear unbiased predictions(BLUPs) of them (and standard errors) using xtmixed postestimation.2.2 Data and variablesWe estimate firm RQ using moving 7- year panels of all US publicly traded firms engagedin R&D. Data for the study comes from the Compustat industrial annual file. Firm level

data items include (in $MM unless otherwise stated): revenues (Yit), capital as net property,plant and equipment (Kit), labor as full-time equivalent employees (1000) (Lit), advertising(Ait), and R&D (Rit). From these primary data, we derive a secondary measure: firmspecific spillovers (Sit) which is computed as the sum of the differences in knowledgebetween focal firm i and rival firm j for all firms in the respective industry (2-digit SIC)with more knowledge than firm i. Spillovers represent the knowledge firms free-ride on ingenerating their products/processes. Failure to include them in the estimation leads tosubstantial upward bias in gamma.2.3 Getting by without the full compustat datasetIn principle RQ estimation does not require the full set of firms in compustat. Non-focalfirms play two roles: First, they “bootstrap” the production function (the estimation exploitsinformation from all firms to better gauge what the elasticities should be for each input).Second, they help control for year-to-year changes in economic conditions. Third, theyprovide data on the spillover pool.There should be a subset of firms (most likely the set of firms in the two digit industry)sufficient to support reliable estimation. If so, firms could obtain 10K data for these firmsdirectly from EDGAR rather than subscribing to Compustat. Since the reliability of thesesubset estimates varies across firms and industries, we recommend validating resultsagainst the full dataset (either on your own or by providing subset estimates to BerkeleyResearch Group for comparison to the master set of estimates)Note however subset estimation still requires the use of random coefficients to obtain thefirm (or division) specific coefficients.3. Spreadsheet estimation (use for intuition only)As mentioned in the introduction, spreadsheet estimation should only be used fordeveloping intuition. It does not generate reliable estimates. I estimate RQ by combiningdata from all US publicly traded firms reporting R&D. This generates precise estimatesthat control for spillovers between firms and for economy-wide effects like recessions.To generate the spreadsheet estimate, you need several years’ data on revenues, and annualexpenditures on inputs: PP&E (property, plant and equipment), labor, R&D andadvertising. We show these for a fictitious firm in Table 1 (columns 1-5).

Table 1Year 1 234 5 6 7 8 9 10 1990 Revenue PP&E Employee Advertising R&D ln(Rev) ln(PPE) ln(Emp) ln(Adv) ln(R&D) 1991 1992 1484 518 6 219 39 7.30 6.25 1.70 5.39 3.67 1993 1646 522 6 244 47 7.41 6.26 1.81 5.50 3.86 1994 1717 581 6 270 45 7.45 6.37 1.76 5.60 3.80 1995 1642 538 5 243 42 7.40 6.29 1.55 5.49 3.75 1996 1846 7.52 6.28 1.58 5.66 3.80 1997 1991 533 5 287 45 7.60 6.26 1.55 5.60 3.80 1998 2225 7.71 6.31 1.67 5.65 3.82 1999 2541 525 5 272 45 7.84 6.35 1.70 5.85 3.92 2000 2753 7.92 6.39 1.89 5.89 4.03 2001 4010 551 5 285 46 8.30 6.96 2.40 6.16 4.13 2002 4089 8.32 6.98 2.40 6.14 4.14 2003 3903 571 6 349 50 8.27 6.95 2.40 5.86 4.20 2004 4061 8.31 6.83 2.25 5.98 4.20 2005 4144 596 7 362 56 8.33 6.98 2.19 6.12 4.33 2006 4324 8.37 6.96 2.15 6.06 4.43 2007 4388 1054 11 474 62 8.39 6.91 2.03 6.08 4.48 2008 4644 8.44 6.91 2.03 6.11 4.60 2009 4847 1079 11 465 63 8.49 6.88 2.05 6.16 4.68 2010 5273 8.57 6.87 2.12 6.19 4.71 5450 1046 11 352 67 8.60 6.86 2.12 6.21 4.74 5534 8.62 6.89 2.12 6.25 4.78 922 10 397 67 1072 9 456 76 1052 9 429 84 999 8 435 88 1004 8 450 99 976 8 474 108 960 8 486 111 955 8 499 114 979 8 518 119Next, transform each variable into log form (columns 6-10 of the table). To perform theanalysis in Excel, choose “regression” from the Data Analysis tab.1 Designate column 6 asyour dependent variable by highlighting rows 1-21 of that column. Designate columns 7-10 as your independent variables. Then choose “labels” to indicate the first row includesthe variable name. When you have run the analysis, Excel will open a new worksheet withthe regression results. We’ve shown you these results for the fictitious firm in Table 2.Column 2 marked “coefficients” contains the elasticity for each variable in column 1. Theelasticities for PP&E, employees, advertising and R&D for the sample firm are 0.23, 0.15,0.74, 0.43, respectively. These coefficients are not accurate for the reasons discussedpreviously. (When the fictitious firm data is combined with all publicly traded firms in adata set that also includes knowledge spillovers, its coefficients are 0.13, 0.51, 0.18, 0.21,respectively). Thus the coefficients on capital and R&D are high by a factor of 2,advertising is high by a factor of 4, and the labor coefficient is 70% below the fully-specified estimate.1 You need to have loaded the “Analysis Toolpak” to get the Data Analysis tab

Table 2.SUMMARY OUTPUTRegression StatisticsMultiple R 0.990523R Square 0.981137Adjusted R S0q.9u7a6r4e21Standard Err0o.0r70398Observations 21ANOVA df SS MS F Significance F 4 4.12428 1.03107 208.0504 1.42E-‐13Regression 16 0.079294 0.004956Residual 20 4.203573Total CoefficienSttsandard Error t Stat P-‐value Lower 95%Upper 95%Lower 95.0U%pper 95.0%Intercept 0.0705 0.8407 0.0838 0.9342 -‐1.7117 1.8527 -‐1.7117 1.8527ln(PP&E) 0.2322 0.2055 1.1299 0.2752 -‐0.2035 0.6679 -‐0.2035 0.6679ln(Employee)0.1499 0.1633 0.9176 0.3724 -‐0.1963 0.4961 -‐0.1963 0.4961ln(Advertising0.7356 0.1653 4.4498 0.0004 0.3852 1.0861 0.3852 1.0861ln(R&D) 0.4303 0.1089 3.9509 0.0011 0.1994 0.6612 0.1994 0.66124. Alternatives to internal estimationIf you have no need to do divisional estimates, then it probably doesn’t make sense toinvest in compustat and labor to generate RQ annually. Instead, you can obtain RQ and R*on a regular basis via an RQ data subscription.5. Other RQ resources 5.1 Academic articles R&D/Returns Causality: Absorptive Capacity or Organizational IQ RQ and endogenous firm growth 5.2 NSF grants 0965147: Firm IQ: A Universal, Uniform and Reliable Measure of R&D Effectiveness The Impact of R&D Practices on R&D effectiveness (RQ) 5.3 HBR article 5.4 Consulting

# Estimating Firms’ Research Quotient (RQ) - amkANALYTICS

##
**Description: ** Estimating Firms’ Research Quotient (RQ) Anne Marie Knott The Harvard Business Review article, “The Trillion Dollar R&D Fix” describes a measure

### Read the Text Version

No Text Content!

- 1 - 5

Pages: