NPL REPORT IEA 6

ESTIMATING THE PRICE ELASTICITY OF DEMAND FOR NPL’S

SERVICES

MIKE KING, EUGENIO RENEDO

JULY 2020
 NPL Report IEA 6

ESTIMATING THE PRICE ELASTICITY OF DEMAND FOR NPL’S

SERVICES

DR MIKE KING, EUGENIO RENEDO

Analysis and Evaluation, Strategy Directorate

Abstract
This paper applies standard panel data analysis to a country-level panel dataset to estimate
the price elasticity of demand for NPL’s services. The data used consists of NPL internal
administrative information, as well as data coming from reliable and widely used publicly
available databases such as the World Bank’s World Development Indicators database, or the
CEPII database. The analysis finds that the price elasticity of demand for NPL’s services is
greater than one, meaning that these are elastic goods for which the quantity demanded will
vary more than proportionally if the price changes.
 NPL Report IEA 6

© NPL Management Limited, 2020

ISSN 2633-4194

https://doi.org/10.47120/npl.IEA6

National Physical Laboratory

Hampton Road, Teddington, Middlesex, TW11 0LW

Extracts from this report may be reproduced provided the source is acknowledged
and the extract is not taken out of context.

Approved on behalf of NPLML by

Fiona Auty, Head of Government Relations
 NPL Report IEA 6

Contents

List of Acronyms
Preface
Executive Summary
1 Introduction .................................................................................................................... 1
1.1 The National Physical Laboratory ........................................................................... 1
1.2 The public funding of NPL ....................................................................................... 2
1.3 Estimating the price elasticity of demand ................................................................ 2
2 Theory ........................................................................................................................... 4
2.1 Key variables in the market for high accuracy measurement services..................... 4
2.2 National price level and elasticity of demand........................................................... 6
2.3 The effect of distance in international trade ............................................................. 8
2.4 Search costs and the home bias effect ................................................................... 9
3 Data ............................................................................................................................. 12
3.1 Variables ............................................................................................................... 12
3.1.1 Quantity sold .................................................................................................. 13
3.1.2 Metrology expertise of the home NMI ............................................................ 14
3.1.3 Geographical proximity .................................................................................. 15
3.2 Data description .................................................................................................... 15
3.3 Econometric analysis ............................................................................................ 18
3.4 Specification ......................................................................................................... 18
3.4.1 Omitted variables bias ................................................................................... 19
3.4.2 Errors-in-variables ......................................................................................... 20
3.4.3 Simultaneity ................................................................................................... 20
3.4.4 Weighting of observations in the sample ........................................................ 20
3.4.5 Selectivity ...................................................................................................... 20
3.5 Estimation ............................................................................................................. 20
3.5.1 Models tested and interpretation of the coefficients ....................................... 21
3.5.2 Model B1: 2-step Heckman selection model .................................................. 22
3.6 Postestimation ...................................................................................................... 28
4 Economic Impact ......................................................................................................... 33
4.1 Direct benefit to paying customers and welfare for the UK .................................... 34
4.2 The effect of shifts in public funding received by NPL on welfare generated ......... 36
5 Conclusion ................................................................................................................... 39
6 References .................................................................................................................. 40
Annex A: Robustness tests ................................................................................................. 41
Countries that do not purchase NPL’s services regularly ................................................. 41
The effect of UK sales in the estimated elasticity of demand ........................................... 44
 NPL Report IEA 6

Outliers in the dependent variable ................................................................................... 45

Specification of the CMC variable .................................................................................... 46
Negative binomial regression .......................................................................................... 50
Split sample test .............................................................................................................. 52
Income as the dependent variable ................................................................................... 53
Annex B: Testing for endogeneity. ...................................................................................... 55
Annex C: Detailed summary statistics ................................................................................. 57
Annex D: Stata outputs ....................................................................................................... 59
Model A1 ......................................................................................................................... 59
Model A2 ......................................................................................................................... 63
Model A3 ......................................................................................................................... 67
Model B1 ......................................................................................................................... 71
Model C1 ......................................................................................................................... 78
Model C2 ......................................................................................................................... 80
Model C3 ......................................................................................................................... 82
Model C4 ......................................................................................................................... 84
Model C5 ......................................................................................................................... 86
Model D1 ......................................................................................................................... 88
Model E1 ......................................................................................................................... 90
Model F1 ......................................................................................................................... 92
Model G1......................................................................................................................... 95
Model H1 ....................................................................................................................... 100
Model I1 ........................................................................................................................ 101
Model J1........................................................................................................................ 102
More information ............................................................................................................... 103
Contact us ..................................................................................................................... 103
 NPL Report IEA 6

List of Acronyms
A4I Analysis for Innovators

The International Bureau of Weights and Measures (usually referred by its French
initialism, Bureau International des Poids et Mesures) is an intergovernmental
BIMP
organisation through which member states act together on matters related to
measurement science

The Centre d'Etudes Prospectives et d'Informations Internationales is a leading

CEPII
French centre for research and expertise on the world economy.

CMC Calibration and Measurement Capabilities

NMI National Measurement Institute

NMS National Measurement System

NPL National Physical Laboratory

PPP Purchasing Power Parity

 NPL Report IEA 6

Preface
This paper has been prepared by NPL’s Analysis and Evaluation team to support the evidence
this organisation is submitting for the 2020 Comprehensive Spending Review. This document
is addressed to NPL’s government owner: The Department for Business, Energy & Industrial
Strategy. Its objective is to support the case NPL is presenting to justify the benefits of its
public funding.
 NPL Report IEA 6

Executive Summary
This work applies a two-step Heckman selection model to a country-level panel dataset to
estimate the price elasticity of demand for NPL’s services. The data used consists of NPL
internal invoicing information, as well as data on financial and demographic variables coming
from reliable external sources – namely, the World Bank’s World Development Indicators
database, the Centre d'Etudes Prospectives et d'Informations Internationales, the Bureau of
Weights and Measures, and the United Nations. The analysis finds that:
• The central estimate for the price elasticity of demand is 1.24 with a 95% confidence
interval given by [1.09 , 1.39].
• The four regressors in our model (the national price level, GDP per capita, distance
and the CMC of the local NMI) turned out to be highly significant.
The rest of the paper is organised as follows. Section 1 provides context about NPL and
outlines the rationale for its public funding; it also makes the case for estimating the price
elasticity of demand using a country-level panel. Section 2 depicts the conceptual framework
that synthetises the basic economic theory that explains NPL’s international sales of services
and identifies the variables of interest. Section 3 discusses the data and the variables used by
showing some descriptive statistics. Section 4 provides a discussion of the model selection
criteria and presents the estimation results. Section 5 shows how the results found in the
econometric analysis can be used to estimate the welfare generated by NPL. Section 6
concludes.
1 Introduction

1.1 The National Physical Laboratory

NPL is a government owned and funded national laboratory that specialises in metrology (the
science of measurement). It has over 380 state of the art laboratories and more than 800
scientists and engineers, as well as 200 co-supervised postgraduate students working in
partnership with industry and academia. NPL’s goal is to generate welfare for the UK’s society.
To that end, NPL carries out a multitude of activities across a wide range of areas – from
quantum sensing and composite materials, to radiotherapy and emissions monitoring. In a
nutshell, NPL’s value chain can be summarised as follows. Firstly, NPL conducts fundamental
research and performs international measurement comparisons that generate articles in peer-
reviewed scientific journals. This enables the development of cutting-edge measurement
capabilities that support the creation of primary standards and state-of-the-art instrumentation.
This expertise is then used to deliver calibration, testing, and training services to private
businesses, hospitals, and universities. In addition, NPL works closely with Innovate UK to
offer grant-funded collaborative R&D projects which involve many firms and research
organisations.
NPL is part of the National Measurement System (NMS). The NMS is the technical and
organisation infrastructure which ensures a consistent and internationally recognised basis for
measurement in the UK. It has two central objectives:
i. To enable individuals and organisations in the UK to make measurements competently
and accurately and to demonstrate the validity of such measurement
ii. To coordinate the UK’s measurement system with the measurement systems of other
countries.
The following figures give a clearer picture of NPL1:
• 820 scientific and technical staff, as well as 280 administrative and managerial staff.
• A turnover of around £90m; £57m of that revenue in annual NMS2 funding.
• Around 350 articles in peer reviewed journals each year; also, its scientists perform
around £30m of public research work each year3.
• £13m of revenue from sales of measurement services. The R&D performed by NPL
supports the introduction of new and improved calibration services, whose benefits
fan-out down the calibration chain.
• Sells services to around 500 UK-based firms each year. And, the lab’s scientists
collaborate on R&D projects with around 200 UK-based firms each year.

1 These figures are as of 2019. They differ to the ones used in this study which relates to the period
2010-2017.
2 The National Measurement System (NMS) is a network of national laboratories and processes that

provide measurement standards and calibration testing facilities in the UK. It maintains the
measurement infrastructure, represents the position of UK measurement internationally and influences
the development of standards.
3 Based on tax credits claims information.

1
NPL seeks to fulfill its objective of providing excellent metrology science to generate welfare
for the UK. Hence, it is essential for NPL to know its impact. For that reason, NPL
commissioned:
• A survey asking its users about the sales of new products that they feel would not have
been achieved without the support of NPL and the other NMS laboratories. This survey
found that users of the NMS laboratories believe that without NPL's support, their total
annual sales of new products would decrease by at least £470M. Furthermore, they
believe that about £2bn worth of new products might be at risk without this support.
The self-reported nature of these estimates may make the exact size of these benefits
doubtful, but it still provides evidence that such benefits exist.
• An independent econometric study by that found that companies supported by NPL
grow more rapidly than unsupported comparators - on average, supported companies
have around 20 additional employees 2-3 years after working with NPL, when
compared to a matched control group of similar unsupported ones.

1.2 The public funding of NPL

The economic rationale for the existence of a publicly funded organisation like NPL is that
measurement R&D is subject to market failure. Indeed, the private investment needed to
generate innovative measurement capabilities, will always be below the socially optimal level.
This occurs because the benefits that measurement R&D generates will always spill over to
firms who did not contribute, and this creates a strong incentive to free ride. The problem is
particularly acute in the case of the R&D that the NPL undertakes, because advances in
metrology tend to have applications across many sectors. It is this wide applicability that
makes the development of these new tools and techniques particularly susceptible to free
riding. Consequently, it is argued that measurement should really be seen as a public
infratechnology, that is, a technology that provides tools and techniques which can be widely
applied across a number of sectors to enable further innovation. In short, NPL and its partners’
scientific work generates a pool of knowledge that can be accessed and used by any firm.
This fact carries a strong incentive to free ride, and thus, there is a clear need for public funding
to complement measurement R&D funded through private spending.
Another key argument for publicly supporting a specialist laboratory like NPL is that the kind
of metrology research it conducts requires the setting up of large facilities. In such cases, the
fixed costs could be so high that they exceed the private gains to any one company. Therefore,
the facility would never be developed on the basis of individual private funds alone, despite
the total benefits from the capability outweighing the cost.
Lastly, there is an efficiency justification that supports the idea of a publicly funded metrology
laboratory like NPL. Indeed, the high cost and difficulty of maintaining primary standards
makes the calibration chain very efficient. NPL supplies a costly high-level calibration service
to a commercial laboratory, which then calibrates the instruments of a vast number of users
without the need for the calibration laboratory to establish their own primary standard.

1.3 Estimating the price elasticity of demand

Knowing the responsiveness of demand to changes in prices is essential for any business to
assess its potential for revenue growth. Moreover, since NPL is a publicly funded organisation,
a reliable estimate of the price elasticity of demand constitutes the cornerstone for further work
aimed to predict the effect of increasing public support.
The analysis outlined in this paper makes use of a country-level panel dataset involving more
than 100 countries over a period of 17 years (2001-2017) to estimate the price elasticity of

2
demand for NPL’s measurement services. There are several reasons why this country-level
approach was chosen to estimate the price elasticity of demand:
• Over half of NPL’s sales are made overseas. This means that a fairly significant
fraction of NPL’s income is subject to international trade forces.
• Analyses of the market for high accuracy measurement services reveal that relative
prices with respect to other NMIs have experienced hardly no variation over the last
two decades. This means that any trend in prices is common to almost every NMI, and
that price-driven substitution effects must have been almost insignificant. On the other
hand, we know that, in an international trade context, the perceived price by consumers
is different from the selling price. Indeed, the perceived price is also influenced by the
consumer’s local purchasing power parity and the exchange rates between the local
and the selling currency. Thus, variations in inflation and exchange rates can be
exploited to measure the price elasticity of demand for NPL’s services.

3
2 Theory

This section outlines the conceptual framework that synthetises the basic economic theory
that explains NPL’s international sales of services. To that end, the global market for high
accuracy measurement services is described first. Next, a few simple micromodels are
presented. These capture the intuition behind the role played by the various variables of
interest in the subsequent econometric analysis. Hence, the objective of these models is not
to provide a comprehensive analytical framework to be econometrically tested, but to show
the mechanisms whereby several key variables affect international sales of measurement
services.

2.1 Key variables in the market for high accuracy measurement services

NPL and other NMIs offer high quality measurement services. These vary from one NMI to
another. In particular, the accuracy delivered depends on how developed the measurement
infrastructure is in each country. Although all users in the market value accuracy (since it
ultimately constitutes valuable information4), it is not equally appreciated by all of them; for
instance, a manufacturer of aircraft engines is more likely to require greater accuracy than a
mass producer of chocolate bars.
Another important aspect of the market for high accuracy measurement services is the time it
takes for an NMI to provide the service. Indeed, particularly in the case of calibration services,
the assets to be calibrated are extremely valuable. This implies that, while waiting to be being
calibrated, the instruments do not generate value for their owners. Thus, queues have a
substantial effect on consumption patterns.
On the other hand, the global market of high accuracy measurement services is not frictionless
– it is subject to significant transaction costs. Particularly two elements play a major role:
geographical proximity and imperfect information. The former entails that transport costs make
services from distant NMIs less attractive; the latter causes the existence of search costs, that
is, the time and money spent by users who research the market to find a service that satisfies
their needs.
Therefore, based on this theoretical framework, we can think of a worldwide, vertically
differentiated, spatial market for high accuracy measurement services. Total demand in this
market is assumed to depend mainly on:
• The price of the service: Like most goods, high accuracy measurement services are
affected by the law of the demand which states that conditional on all else being equal,
as the price of a good increases, the quantity demanded decreases. However, we are
dealing with a multi-country, multi-currency, international trade framework; thus,
exchange rates and monetary effects play a substantial role. The perceived price is
influenced by the user’s local purchasing power and the exchange rate.
• The time to delivery: Since measurement is at the heart of many production processes,
the longer it takes for the customer to get the service, the higher the opportunity cost.
• The geographical proximity between the NMI and the user: Transport costs
disincentivise sales abroad – the further the distance, the higher the transportation and

4 Accurate measurement is often crucial to R&D activities to develop new products (and therefore the
possibility of increasing market power and commanding a price premium), or new processes that
typically entail a cost-saving improvement to the production process, which leads to a greater efficiency
and an increase in the profit margin.

4
 the insurance cost. Moreover, highly accurate measurement instruments are very
sensitive, and their specification may deviate due to transportation. These two facts
create a bias towards consumption from closer NMIs.
• The variety of services provided by the home NMI: The market of high accuracy
measurement services is a very specialised market. Hence, users face substantial
search costs. Rational consumers will continue to search for a better product until the
marginal cost of searching exceeds the marginal benefit. Thus, users will be more
inclined to buy services from their local NMI (provided that they can find a service that
meets their needs), rather than a distant foreign NMI. This comes as no surprise, since
search costs are lower at the local level because there are fewer barriers to knowing
the NMI’s portfolio (e.g. common language, the possibility to visit the facilities…).
The framework outlined above synthetises the main forces driving the decision of users
worldwide regarding the purchase of high accuracy measurement services from NMIs.
However, one typical element affecting demand is missing: real income. This assumption
implies that the income elasticity of high accuracy measurement services is thought to be
close to zero. The argument for this is that the quantity demanded of these services by any
specific firm, tends to be constant over time. In other words, high accuracy measurement
services are a basic need for most of the users. Indeed, there are effectively two reasons for
demanding these services: either because it is a requirement of the production process or
because it supports R&D activities. The former is consistent with the zero-income elasticity of
demand assumption. The customer is forced to acquire the services of an NMI instead of those
of a commercial laboratory because that level of accuracy is fundamental to run its operations.
The latter is not as straightforward. Although the empirical evidence shows that R&D activities
are highly procyclical , and therefore linked to income growth, it could be argued that research
spending comprises a very small fraction of the company’s resources. Moreover,
measurement is just one of the activities within an R&D project. Hence, measurement R&D
spending must be negligible compare to measurement spending in productive related
activities. Consequently, the zero-income elasticity of demand is considered a valid
approximation.
In addition, the forces that drive demand may not be the same when we aggregate the units
of analysis. In other words, it is not the same to analyse the factors that determine the demand
of a single user, than to study those at the country level. Notably, the number of companies
that need high accuracy measurement services within a country, must be a good predictor of
the quantity sold. In section 4 we will control for this by including GPD per capita as a time-
varying covariate. The intuition behind this is twofold. On the one hand, the richer a country
is, the more sophisticated tends to be its economy, and the higher the demand for high
accuracy services. On the other hand, the number of companies fluctuates with income
growth, i.e. the number of companies is certainly linked to the business cycle. Arguably, the
first aspect (the relationship between economic sophistication and the need for high accuracy
measurement services) is related to GDP, whereas the second one (the link between the
volume of this kind of services sold and the number of companies), pertains to oscillations
around the mean value of GDP per capita. In any case, both situations are addressed by
adding the GPD per capita time-varying covariate in our regression model. Moreover, as a by-
product of adding GDP per capita in the regression model, we are also accounting for any
income effect, should our assumption of zero-income elasticity not be completely valid.
Lastly, note that queue times are impossible to account for in our analysis. Simply this
information is not publicly available, since this consists of internal information of each NMI.
Section 4.1 deals with the econometric issues around the lack of data for this relevant variable
and details the assumptions made to work around it.
The rest of the section provides further theoretical basis to understand the relationship of sales
with the rest of the magnitudes identified as drivers of the demand for high accuracy

5
measurement services. A series of micromodels are presented in order to show the effect of
the national price level, the distance, and the services provided by the local NMI on the
international market of high accuracy measurement services.

2.2 National price level and elasticity of demand

Perceived price is key to understand consumption patterns in international trade; thus,

international trade is driven by purchasing power parities and the exchange rates. A concise
and convenient way to express the effect of these variables is the ratio between the two –
which is commonly known as the national price level. The national price level allows us to
compare the cost of the same bundle of goods across different countries.
We define the national price level as:

𝑆 𝑝ℎ′ ⁄𝑝′
𝐿= = (2.1)
𝑟 𝑟

where 𝑆 is the purchasing power parity relative to the UK, that is, the price ratio of the same
basket of goods in local currency units of the buyer’s home country, 𝑝ℎ′ , and in sterling pounds
in the UK, 𝑝′ ; 𝑟 is the exchange rate expressed in local currency units per pound – the
superscript ′ in both pricing variables in equation 2.1 symbolise that both relate to the
numeraire good, i.e. the representative basket of goods typically bought by the firm; this will
play a role in the model below. Unsurprisingly, if the buyer’s home country is in fact the UK,
then 𝐿 is equal to one.
From the definition of the national price level it follows that increases in the national price level
can be driven either by a higher domestic inflation or a stronger local currency unit. Hence, a
positive relationship between an NMI’s international sales and the national price level is
expected. To show this, the micromodel below is developed to find the analytical expressions
that links the quantity sold of high accuracy measurement services and the national price level.
A rational agent is only interested in changes in the perceived price, no matter whether they
come from fluctuations in selling prices or in the exchange rates. To show this, let’s start by
considering the payoff function5 of the typical user of high accuracy measurement services6.
Naturally, the money spent on measurement services is just a tiny fraction of a company’s
total spending. Hence, the following quasi-linear payoff function serves as a good first
approximation of the value obtained by the user:

𝛱 = 𝑐𝑄 𝜃 + 𝑋 (2.2)

5 This implicitly states the productive nature of high-accuracy measurement services which provide
valuable information that support the development of new products or processes which ultimately yield
higher profits.
6 Without any loss of generality, for the sake of simplicity, in this model we consider the choice faced

by a rational agent when purchasing measurement services either in their local market or in the UK.
Note that the choice of the UK is arbitrary, but irrelevant for the validity of the model. The reason why
we set the choice between the UK and the local country, is because it simplifies the explanation of the
model and it links with the later empirical analysis in this paper, in which this model is used to estimate
the price elasticity of demand for NPL’s measurement services.

6
Where 𝑐 and 𝜃 are positive constants such that 𝑐 > 0 and 0 < 𝜃 < 1, 𝑄 is the quantity
demanded of high accuracy measurement services7, and 𝑋 is the number of units of the
numeraire good. This functional form is considered suitable to model the firm’s payoff function
because spending on measurement services is typically quite small compare to the rest of
spending by the firm. Therefore, the effect on the marginal utility of money is negligible.
Moreover, another key property of the quasilinear function, is that the Marshallian demand
does not depend on income, which is consistent with our previous assumption of zero income
elasticity.
The agent’s spending limit is given by the budget constraint:

𝑀 = 𝑟𝑝𝑄 + 𝑝ℎ′ 𝑋 (2.3)

Where 𝑀 is the income of the agent, 𝑟 is the exchange rate, 𝑝 is the market price of the high
accuracy measurement service in the UK and 𝑝ℎ′ is the price of the numeraire in the home
country8. Note that this specification means that 𝑀 and 𝑝ℎ′ are measured in the customer’s
country currency, and 𝑝 is expressed in sterling pounds.
If we substitute 2.3 in 2.2 and maximize the resulting objective function, we obtain the
Marshallian demand function:

1
𝑐𝜃𝑝ℎ′ 1−𝜃 (2.4)
𝑄∗ = ( )
𝑟𝑝

Now, this equation can be expressed in terms of the national price level and UK prices using
equation 2.1:

1 1
𝑐𝜃𝑝′ 𝐿 1−𝜃 𝑐𝜃 1−𝜃 (2.5)
𝑄∗ = ( ) =( )
𝑝 𝒫

Where 𝒫 = 𝑝⁄𝑝′ 𝐿 = 𝑟𝑝⁄𝑝ℎ′ is the perceived price by the customer – i.e. the number of units of
the numeraire good that the firm can buy in the home country for the amount it would spend
on one unit of the measurement service in the UK.
If we take the logarithms at both sides of equation 2.5, we get:

1
ln 𝑄 = 𝐴 − ln 𝒫 (2.6)
1−𝜃

1
Where 𝐴 = 1−𝜃 ln 𝑐𝜃 is constant.

We can derive both sides of equation 2.6 to get:

7 Implicitly, the heterogeneous nature of measurement services is being ignored here. However, for the
purposes of the model and the empirical analysis later on, we assume that in effect what all NMIs sell
is the time of highly trained scientists and engineers in metrology. More on this on section 3.1.
8 The price notation is such that subscripts denote if the price is referred to the UK (no subscript) or the

buyer’s home country (subscript ℎ), and the superscripts represent if the price is related to the
measurement service (no superscript) or to the numeraire good (superscript ′)

7
 𝛥𝑄 1 𝛥𝒫
=− (2.7)
𝑄 1−𝜃 𝒫

Where the percentage change in the perceived price is given by:

𝛥𝒫 𝛥𝑝 𝛥𝑝′ 𝛥𝐿
= − ′ − (2.8)
𝒫 𝑝 𝑝 𝐿

Now, NPL’s measurement services prices have roughly moved directly in relation to the UK
general price level (i.e. the price of the numeraire good). Consequently, both percentage
changes in prices in equation 2.8 cancel each other out and we are left with the percentage
change in the national price level. In other words, although by using the national price level as
a proxy for the perceived price we might be suffering from measurement error, this is likely to
be quite small because relative changes in the perceived price are very similar to relative
changes in the national price level
Lastly, if we expand the equation 2.6 using the definition of the perceived price 𝒫, we obtain:

1 1 1
ln 𝑄 = 𝐴 − ln 𝑝 + ln 𝑝′ + ln 𝐿 (2.9)
1−𝜃 1−𝜃 1−𝜃

Therefore, this model yields two important conclusions:

• The price elasticity of demand for NPL’s measurement services is equivalent in
magnitude (but opposite sign) to the elasticity with respect to the national price level.
• Relative changes in the perceived price are mostly driven by relative changes in the
national price level because the price of measurement services have approximately
been in sync with the general price level.
These two facts will be exploited in the empirical analysis in section 4.

2.3 The effect of distance in international trade

The gravity model of trade has been very successful when ordering the observed variation in
economic interaction across space in trade flows. The good fit and the tight clustering of
coefficient estimates in the empirical literature, suggest that some underlying economic law
must be at work. Although the driving forces of international trade have been a vivid debate
since the emergence of Ricardo’s comparative advantage theory, the gravity model has often
been used to test hypotheses rooted in many different economic theories.
Normally, in econometric applications the gravity model of trade is specified as follows:

𝛾 𝛾
𝐸𝑖 1 ∙ 𝐸𝑗 2
𝑋𝑖𝑗 = 𝐴 ∙ 𝛾 ∙ 𝜗𝑖𝑗 (2.10)
𝐷𝑖𝑗3

where 𝑋𝑖𝑗 represents the volume of trade from country 𝑖 to country 𝑗, 𝐴 is a constant, 𝐸𝑖 and
𝐸𝑗 represent some proxy variable of the economic size of the countries, typically their gross
domestic products, 𝐷𝑖𝑗 denotes the distance between the two countries, and 𝜗𝑖𝑗 represents an
error term with mean 1.

8
The most common approach to estimate equation 2.10 is to take the logarithm of both sides
to construct a log-log model that can be estimated by OLS:

ln 𝑋𝑖𝑗 = 𝛾0 + 𝛾1 ln 𝐸𝑖 + 𝛾2 ln 𝐸𝑗 + 𝛾3 ln 𝐷𝑖𝑗 + 𝜂𝑖𝑗 (2.11)

where 𝛾0 = log 𝐴 and the negative sign before 𝛾3 has been included in the coefficient.
If we consider the worldwide trade pattern of just one country, the model becomes simpler
because all the observations are referred to that country (effectively, we drop one of the
indexes in equation 2.11, so it becomes):

ln 𝑋𝑖 = 𝛽0 + 𝛽1 ln 𝐸𝑖 + 𝛽2 ln 𝐷𝑖 + 𝜀𝑖𝑗 (2.12)

Where 𝑋𝑖 represents the volume of trade from country 𝑖 to the country of reference, 𝐷𝑖 denotes
the distance between the two countries, and 𝛽0 = 𝛾0 + ln 𝐸 with 𝐸 representing the proxy
variable of the economic size of the country analysed.

2.4 Search costs and the home bias effect

Most markets in the economy are far from being commoditised. This is especially true for the
market of high accuracy measurement services, where the prevalence of product (service)
differentiation is caused by an ample heterogeneity in user preferences and the wide range of
measurement capabilities of the different NMIs. This differentiation provides NMIs with
profitable opportunities and makes them face important strategic decisions in terms of how
and when specialise in certain areas of metrology. The acute differentiation in the market also
leads to search frictions for users, since they have to consider not just price, but a large
number of characteristics of the various services on offer and how these fit their needs. Hence,
users incur a cost of resources (namely money and time) when having to find and compare
the different options at their disposal. Thus, the decision to purchase a specific service or to
keep researching the market is mostly determined by two factors: search costs and the
suitability of the options already considered.
The behaviour of the user of high accuracy measurement services can be rationalised by
considering the following setup. Let’s assume a differentiated market of high accuracy
measurement services where users have different tastes (horizontal differentiation9). In the
market, any potential buyer knows all the NMIs, but does not know the characteristics of the
services they offer (for the sake of simplicity, price is considered as one of the elements
pondered by the user). Users can gather information by sequentially and randomly sampling
the NMIs in the market (the assumption of random search suggests that they know little about
the market before committing to search; thus, it is consistent with the assumption that users
know all the NMIs in the market, but they do not know the characteristics of the services). The
search process has associated a cost (typically it consumes time and resources). However,

9 Horizontal differentiation refers to distinctions in products that cannot be easily evaluated in terms of
quality. This stands in contrast to vertical differentiation, where the distinctions between products are
objectively measurable and are based in the products' respective level of quality. Therefore, in a
horizontally-differentiated market, in general consumers will buy different products even if these have
the same price, since consumption decisions are motivated by individual preferences. In the case of
vertically-differentiated markets all consumers agree on the preference order of goods in the market.
Hence, consumption decisions are determined (fundamentally) by their budget constraints.

9
knowing the characteristics of the services of the first NMI comes at no cost. This assumption
models the fact that it is easier to know what the home NMI offers. Lastly, at any point in the
search process, users can always fall back and buy the variety of any of the NMIs already
sampled at no additional cost.
Let’s start by considering the net present value of future profits10 that a user of high accuracy
measurement services can obtain thanks to those. The objective function of a user 𝑗 is given
by:

𝛱𝑗 (𝑥0 ; 𝑖) = 𝑥0 + 𝑣𝑖 (2.13)

Where 𝑥0 is the quantity of the numeraire good11 the user 𝑗 has, and 𝑣𝑖 is the value the user
gives to a unit of brand 𝑖.
Given that users are assumed to maximise the net present value of profits, they will try to find
the variety for which 𝑣𝑖 is maximal. In other words, they will try to find the service of the NMI
that best suits them. In addition, in order to account for different tastes within users, it is
assumed that the valuations 𝑣𝑖 are realisations of an independent and identically distributed
random variable with distribution function 𝐺. Without loss of generality, we can assume G is
given by a continuous uniform distribution (rectangular distribution) with distribution support
[0,1].
A rational user will stop searching for another service when the marginal cost of searching
exceeds the marginal benefit. Let’s consider a buyer who has already sampled some NMIs
and attaches a valuation 𝑣 ′ 𝜖[0,1] to the best match found. The marginal benefit of keep
searching (i.e. sample one more NMI) is given by the expected benefit of doing so minus the
valuation of the best match up until that point:
1
𝑏 = 𝐄[𝑣|𝑣 > 𝑣′] ∙ 𝑃(𝑣 > 𝑣 ′ ) + 𝑎 ∙ 𝑃(𝑣 < 𝑎) − 𝑎 = ∫ (𝑣 − 𝑣 ′ )𝑑𝑣
𝑣′ (2.14)
1
= (𝑣 ′ − 1)2
2
The marginal cost of one more sample is given by the search cost, 𝑠. Hence, the user will stop
searching when the marginal benefit given by 2.14 is equal to the search cost:

1 ∗
(𝑣 − 1)2 = 𝑠 ⇔ 𝑣 ∗ = 1 − √2𝑠 (2.15)
2

Therefore, the user will stop searching when he finds the service of an NMI that provides a
value equal to or greater than 𝑣 ∗. Note that this means that the higher the valuation the user
assigns to the initial NMI (remember that the information about this first NMI comes at no cost),
the less likely is the participation in the international market. In other words, the better the local

10 As in section 2.1, the fact that the objective function for the representative agent is the net present
value of future profits suggests the productive nature of high-accuracy measurement services.
11 The inclusion of the numeraire good simply seeks to resemble the linear functional form of consumer's

utility that is most commonly used in the economic literature. The numeraire does not play any role in
the model. It only denotes the fact that a separate and additive relationship is assumed between all the
goods that the user can acquire. Effectively, this implies that the user looks for the variety that best
satisfies his needs as a standalone service, without considering any relationship of complementarity or
substitution with the rest of the goods.

10
NMI meets the needs of the user, the less likely it is that the user will decide to survey the
portfolio of foreign NMIs.

11
3 Data

The following section presents the data used to construct the variables used in the
econometric analysis. First the data sources and the variables are presented. Then the
methodological issues around those variables are addressed.

3.1 Variables

This study uses NPL’s internal managerial data as well as data coming from several external
sources. Table 1 summarises the variables and the sources of data used in the econometric
analysis:

Variable Units Symbol Source Description Magnitude

Yearly number of
Number of
invoices issued to all
invoices invoices
users in a country
normalised by per I NPL Quantity sold
divided by the
the population inhabitant
population of that
of the country
country.
Relative national
National price
. (index) PL price with respect to Price
level
the UK.
World Market value of all
Bank the final goods and
GDP per capita £ GDP services produced in Income
a year per person in
millions of pounds.
Distance in km
between London and
Geographical
Distance km D CEPII the most populate
proximity
city in the user's
country.
Number of calibration
Calibration and and measurement
. (index Metrology
measurement CMC BIPM capabilities of the
0-100) expertise
capabilities home NMI of each
country.
Population in millions
United
Population inhabitants POP of people at 1 July Population
Nations
2019.

Table 1: Variables, data sources and magnitudes.

There are several methodological difficulties when it comes to associating suitable proxies
with the magnitudes identified in Table 1. The rest of this subsection analyses those issues
and discusses potential solutions.

12
3.1.1 Quantity sold
Although NPL’s portfolio is quite varied, roughly speaking it can be considered to offer one
unique good: the time and the expertise of highly trained scientists and engineers. In that
regard, NPL can be thought to operate much like a professional services firm which sells the
time and the knowledge of its workforce. However, in other aspects NPL is far from resembling
a typical professional services firm. In particular, there is a fundamental difference when
comparing both business models – say between a law firm and NPL. For the former, its core
business consists of one sole activity: representing its clients in court and providing them with
legal advice. For NPL it is not as straightforward, because NPL’s staff need to maintain the
measurement capabilities required to deliver high accuracy measurement services and
engage in cutting-edge collaborative R&D projects; and this task requires a significant portion
of their time. Hence, unlike a law firm, NPL must preserve a knowledge stock that depreciates
over time in order to meet the requirements of its users. To do so, NPL’s staff carry out a wide
variety of activities. These include conducting international key comparisons, participating in
proficiency testing schemes, maintaining UKAS accreditation for calibration and testing
services, running audits, contributing to standards and protocols, and performing research that
generates articles in peer-reviewed scientific journals. This knowledge is then used to meet
the needs of users. In this sense, NPL could be considered to resemble an orchestra for which
the day of the concert is only the tip of an iceberg of constant work over the months and years.
The show would not be possible without the previous effort put into practice. Similarly, for NPL
both activities (the maintenance of capabilities and the delivery of services) are deeply
entwined. Thus, it is not easy to discriminate the time spent by the staff on one activity or the
other.
On top of that, NPL’s strategic managerial approach focuses on achieving specific revenue
targets for each scientific group. The way this target is achieved varies from one group to
another, and over time. This means that the decisions taken by the group leader on how to
distribute the time of scientists and engineers in the group may be influenced by many factors,
such as long-term perspectives or the idiosyncrasy of the area of metrology. Effectively, this
means that NPL does not have timecard data broken down by customer. Therefore, any ex-
post consideration on the distribution of the time of scientists and engineers would not have
enough comparability between groups and will lack temporal consistency.
Since the ideal variable to proxy the quantity sold (the time distribution of the scientific staff
among jobs) is not available, some other proxy is needed. There are two other alternatives
available to us: the number of invoices issued, and the income generated12. Undoubtedly, both
must hold a positive relationship with the time spent by scientific staff. However, the number
of invoices variable is the preferred option. This is because income is related to quantity sold
through price (𝑅 = 𝑝 ∙ 𝑄); and prices vary substantially between jobs despite the service
delivered being the same. This is because NPL’s prices are set on a cost-plus basis13. This
means that, although the services provided are fairly standardised, these are not always
delivered by technical staff of the same grade – i.e. depending on availability of the workforce
sometimes the service may be delivered by some more senior scientists or engineers, and
this effectively translates into prices. Thus, income does not relate well to time spent on each
job.
Nevertheless, this is a fundamental choice of the analysis that requires further discussion. To
that end, a robustness check has been carried out in Annex A. The test consists of executing
the same model using income as the dependent variable instead of the number of invoices.
We conclude that both models produce very similar estimates; however, the preferred

12 Note that in reality any choice for the dependent variable is normalised by population. This ensures
comparability between countries.
13 For most services provided by NPL there is no competitive market in the UK; which is why NPL deliver

those services. This means that there is no competitive process driving prices and therefore, NPL set
prices on a cost-plus basis.

13
specification is the one that considers the number of invoices as the dependent variable
because we have the aforementioned reasons to believe that these estimates are slightly more
accurate.
3.1.2 Metrology expertise of the home NMI
Another methodological issue revolves around the way to model the CMCs of a country. BIPM
reports a comprehensive table that details the number of CMCs by country broken down into
different metrology areas. These areas encompass many different services depending on the
physical quantities they measure; for example, Acoustics, Ultrasound and Vibration,
Photometry and Radiometry or Chemistry. There are mainly two possibilities to assess the
CMCs of an NMI based on the information reported in the table. Both have been tested and
compared.
The first one adopts a simple probabilistic approach to account for the CMCs of NMIs. The
intuition is straightforward: the more measurement services an NMI offers, the more likely it is
for a company to find a service within the NMI’s portfolio that meets it needs. This assumption
has a major weakness though. Undoubtedly, some metrology areas are more in demand than
others. Moreover, even within a specific area of metrology, some services are more in demand
than others. Hence a probabilistic measure that weights all services equally is not completely
accurate. On the other hand, this approach implicitly assumes that companies do not demand
services from different areas at the same time. Although this may be true for small businesses,
large companies may demand numerous services of different types from the same NMI. They
do so both for getting better prices, and for technical reasons (the services may be entwined
or depend of one another). Again, this means that the probabilistic approach to model CMCs
is not entirely accurate. Nonetheless, keeping those conceptual concerns aside, there is also
another methodological issue to overcome with this approach. Since the probability of
matching – i.e. the probability that an NMI can meet the user’s needs – is defined as the
number services offered by an NMI over the total number of services offered worldwide, it is
crucial to know the total number of distinct services offered by all NMIs together. This cannot
be inferred from the table provided by BIPM, since the services are not uniquely identified;
only the total number of CMCs by area of metrology is reported. In order to approximate this,
a simple method has been used. For each of the nine metrology areas, the country that offers
the greatest number of services has been identified, and this figure has been assumed to be
the total number of services available in that area. The idea behind this reasoning is that if a
country is the world's largest specialist in an area of metrology, its portfolio must be quite
comprehensive. In other words, we are assuming that there are not too many services in that
area that the NMI does not offer. Therefore, both conceptual and methodological difficulties
show that this probabilistic measure has some disadvantages in order for it to be used as
proxy of the CMCs of an NMI. In particular, NPL’s internal sales data suggests that some users
demand services from different scientific groups (or equivalently, they demand services from
different areas of metrology). To what extent this is in fact relevant to the econometric results
must be determined. That is why an alternative method that accounts for the versatility of an
NMI (i.e. how varied its portfolio is) is tested and compared.
For that matter, the geometric mean of the number of CMCs in the nine areas of metrology
reported in the table is also proposed as a convenient proxy for the CMCs of an NMI. This
alternative measure constitutes a comparable measure that represents the metrology
expertise and scope of the NMI’s portfolio. It consists of a continuous dimensionless index that
assesses the development of the NMI in all areas equally and assigns larger values to the
institutes that offer greater variety of services, hence accounting for taste variety among users,
who obtain a greater benefit from those NMIs that provide them with more complete service
bundles.
Both measures of CMCs have been tested, yielding very similar (almost equivalent) estimation
results. The first approach is the preferred one though; basically, because it allows for better

14
interpretation of the estimated coefficients. In any case, the estimation results with respect the
second approach based in the geometric mean is included in Annex A.
3.1.3 Geographical proximity
For each country in the dataset, geographical proximity is measured as the distance in
kilometres between London and the most populated city in the purchasing country. However,
around half of NPL’s income comes from the UK. Hence, we need a measure of the average
distance between NPL and users within the UK; here, we follow the commonly used guidelines
by that define the internal distance as:

𝑑 = 0.67√𝐴⁄𝜋 (3.1)

where 𝐴 is the area of the country.

3.2 Data description

The objective of this analysis is to measure the price elasticity of demand. For that matter, it
is convenient to work with the logarithms of the variables above. Table 2 shows some basic
descriptive statistics of the variables of interest. (Variables in levels are abbreviated in capital
letters; variables in logarithmic form are lowercased).
Relative
Std. within
Variable Mean Std. Skewness Kurtosis Min Max
Dev variation
Dev

I 2.4E-06 8.0E-06 3.3 0.5 6.2 46.3 0.0 8.6E-05

PL 0.5 0.3 0.5 0.3 0.8 2.8 0.1 1.4

GDP 1.5E+04 1.4E+04 0.9 0.4 1.7 7.0 4.1E+02 9.7E+04

CMC 11.6 16.7 1.4 0.0 2.3 8.0 0.0 77.6

D 5167.5 3933.0 0.8 0.0 0.8 3.5 185.8 19147.1

POP 5.9E+07 1.8E+08 3.0 0.0 6.0 40.8 3.3E+04 1.3E+09
i -14.1 2.1 0.2 0.3 -0.4 2.7 -20.3 -9.4

pl -0.7 0.5 0.7 0.4 0.0 2.3 -2.1 0.4

gdp 9.2 1.1 0.1 0.3 -0.6 2.9 6.0 11.5
d 8.2 1.0 0.1 0.0 -0.7 2.9 5.2 9.9
pop 16.3 1.9 0.1 0.0 -0.4 3.9 10.4 21.0

Table 2: Summarised descriptive statistics.

Table 2 provides some insight to understand the dataset14. To begin with, the outcome variable
𝐼𝑖,𝑡 (the number of invoices issued to a country normalised by the population of the country) is
not defined for over 40% of the panel. This is because small and/or less developed countries
lack enough companies that need high accuracy measurement services. This leads to NPL
sales to those countries being rare events. Therefore, the number of invoices is zero for a

14 Annex C contains a more detailed table with summary statistics for all the variables included in Table
2.

15
significant part of the panel, causing a substantial incidental truncation of the outcome variable
once logarithms are taken. This has important consequences on the statistical techniques
needed to analyse the data properly; section 4.1 analyses this matter in depth, and proposes
a two-step Heckman model as the preferred way to account for any selection bias that would
come from this structure of the panel. Moreover, Annex A analyses the effect of removing
countries with too few observations from the sample. The logic behind checking the
robustness of the results when trimming these observations, is that those countries might have
fundamentally different needs than the rest of the sample. These unobserved characteristics
could have a misleading effect when estimating the real price elasticity. As shown in this annex
the effect is not considerable.
Another key feature of the dataset is the distinction between within and between variation.
These measure how far are the variables of interest spread out from their average value,
across countries and over time respectively. Obviously, since distance (𝐷) is a time-unvarying
variable, it will not show any within variation. In addition, although the measurement
capabilities of the home NMI (𝐶𝑀𝐶) and the population of the country (𝑃𝑂𝑃) do vary with time,
they barely do so. For that reason, for both variables only the 2019 figures have been
considered. Therefore, these two will not show any within variation either.
As expected, the time-variant variables show more variation from country to country than over
time for the same unit. Nonetheless, the national price level presents significant variation over
time. This is an important requirement for the analysis proposed to work. Note that a basic
element of the analysis, and so one of the reasons why overseas sales are used to estimate
the price elasticity, is that measurement services relative prices have stayed fairly constant
over the period analysed. Hence, we require enough variation in relative national price levels
to estimate the price elasticity of NPL’s services.
Finally, it is important to assess if there are outliers in the panel and whether they affect the
estimates or not:

Figure 1: Box plot of the dependent variable.

16
The dependent variable shows a neat log-normal distribution. In any case, the identified
outliers were dropped applying the standard approach by . However, the effect of these
outliers in the estimation results is negligible. This is shown in Annex A.

17
3.3 Econometric analysis

The following section constitutes the nucleus of the analysis carried out in this paper. Firstly,
the model of the data-generating process is presented, and the potential sources of bias are
discussed. Secondly, the estimation results for the models proposed are presented; the
models differ on the dynamics considered and whether the Heckman correction is applied or
not – the goal is to build from simple to complex to allow the reader to understand the line of
reasoning followed. Lastly, the postestimation test are run to find the preferred setup.

3.4 Specification

The theoretical population model to be analysed is:

log 𝐼𝑖,𝑡 = 𝛽0 + 𝛽1 log 𝑃𝐿𝑖,𝑡 + 𝛽2 log 𝐺𝐷𝑃𝑖,𝑡 + 𝛽3 𝐶𝑀𝐶𝑖 + 𝛽4 log 𝐷𝑖

(4.1)
+ 𝛽5 log 𝑃𝑂𝑃𝑖 + 𝛾𝑡 𝑌𝑡 + 𝜀𝑖,𝑡

where 𝐼𝑖,𝑡 is the number of invoices issued to country 𝑖 in year 𝑡 normalised by the population
of the country, 𝑃𝐿𝑖,𝑡 is the national price level, 𝐺𝐷𝑃𝑖,𝑡 is the gross domestic product per capita,
𝐶𝑀𝐶𝑖 is a continuous index that rates the measurement capabilities of the country, 𝐷𝑖 is the
distance between the country and the UK, 𝑃𝑂𝑃𝑖 is the population of the country, 𝑌𝑡 is a year
dummy variable, and 𝜀𝑖,𝑡 is the error term which is assumed to be independently and identically
distributed as a normal distribution with mean of 0 and finite variance.
It could be argued that users in different countries might have fundamentally different needs.
However, the proposed specification given by equation 4.1 does not contain an intercept which
is dependent on the unit of analysis – a fixed effect that reflects the unobserved heterogeneity
across countries. The reason for that is that we lack the statistical power as to assign each
country its own intercept. This poses the question of whether we can use the elasticity found
from this international dataset to characterise demand by UK users. We argue that the majority
of sales occur in comparable countries such as OECD countries (94.8%) or EEA countries 15
(89.4%), and thus, there is reason to believe that ignoring the fixed effect is not crucial for the
purpose of the analysis.
On the other hand, the inclusion of the year dummy variable must be discussed. This variable
allows us to control for the effect of year-specific events.
Another important element worth highlighting is the fact that the variable that controls for the
measurement capabilities of the home NMI is introduced in the model in level rather than
logged. This makes the interpretation of the coefficient much more straightforward. Since this
variable is a probabilistic measure of the local NMI offering a service that meets the user needs
(as detailed in section 3.1), the coefficient estimate should be interpreted as the effect on sales
(expected to be negative) of an increase of 1% on such probability. In any case, the
correctness of the specification in level rather than logged has been tested using the approach
by . All the detail can be found in Annex A.
Lastly, note that the ultimate goal of the model given by equation 4.1 is to estimate the price
elasticity 𝛽1 . To ensure that the estimate we get from this specification is close to the real
parameter, any potential source of bias needs to be accounted for. The rest of this subsection
analyses systematically all potential issues that could be biasing the estimation results.

15 European Economic Area.

18
3.4.1 Omitted variables bias
The specification defined by equation 4.1 clearly does not include all the variables that are
responsible for all the variation in the dependent variable. By definition, the effect of these
omitted variables is included in the error term. Consequently, the estimations provided by our
model will not be biased, as long as the error term is not correlated with the regressors.
One relevant omitted variable identified in section 2 which is not explicitly included as a
covariate, is delivery time. Indeed, international sales of high accuracy measurement services
are influenced (presumably to a large extent) by the time that NMIs take to deliver their
services. For instance, we can think of how VSL – the Netherlands’ NMI – and NPL compete
for an Australian customer. From the user’s perspective, both are distant NMIs offering similar
services and prices; hence, it is plausible that the decision between the two is mainly
determined by how much time it would take for the customer to get the service.
So, the question is whether delivery times – or any other omitted variable for that matter – is
correlated to our regressors. In particular, we are interested in knowing whether the national
price level is endogenous, since the ultimate goal of the analysis is to determine the price
elasticity of demand for NPL’s services.
Endogeneity could be originated in either of the two dimensions of our panel. On the one hand,
a certain degree of unobserved time-independent heterogeneity among the countries in our
sample is expected. Unsurprisingly, customers from the US are likely to show different needs
than Nigerian businesses. However, whether this fixed effect is sufficiently correlated with our
regressors as to significantly bias our estimations is another matter. As it will be shown in
section 4.3 this fixed effect is negligible; thus, the proposed approach of pooling all the
observations ignoring any country-specific effect is deemed to be quite accurate.
Alternatively, endogeneity could be the result of the dynamics of the generating process. In
other words, NPL’s sales to other countries may depend not only on the contemporary values
of the variables that determine these sales, but on their recent history too. This persistence
can be caused by the explanatory variables, the unobservables, or both (i.e. we could
introduce in our specification lagged regressors, past realisations of the residuals or lagged
values of the dependent variable which effectively encompasses both). In our case, it is fair to
assume that the inertia is just in the error term; that is, the dynamic persistence comes from
the unobservables. There are two main arguments to motivate this assumption. Firstly, sales
are supposed to react quickly to the time-varying regressors in our specification. That is, in
most markets, a consumer is assumed to react very quickly to variations in relative prices and
changes in their budget constraint. Secondly, NPL’s sales worldwide are subject to
macroeconomic forces that determine international trade; that is, the global business cycle.
These in turn tend to show some persistence. Hence, the dynamics of the system must be
caused by macroeconomic variables in the error term. For that reason, most of the models
tested section 4.2 include lagged values of the residuals of the static regression model to
account for this dynamic behaviour of the population process16.
Endogeneity has been systematically checked in section 4.3 for all the models proposed. To
that end, the approach suggested by is implemented17. The results of this test show that the
assumption of exogeneity is correct (even if country-fixed-effects are ignored) as long as the
dynamics of the population process are taken into account. This is very much expected in the
sense that the idea behind using the Mundlak test is to test whether the time-invariant
unobservable (unobserved heterogeneity) is correlated with our regressors of interest; in

16 (Beck & Katz, 2011) deals with a variety of dynamic issues in the analysis of time-series–cross-
section data that can be addressed by different kinds of dynamic specifications. In particular, the authors
discuss the role of the serially correlated error model in panel data analysis governed by dynamic
processes.
17 A thorough explanation of this method is provided in Annex B.

19
particular with the national price level regressor which gives us in the end the price elasticity
estimate. The fact that we pass the Mundlak test shows that the covariates are uncorrelated
with the error term, which includes the unobserved heterogeneity. As mentioned, this is
expected because although the fixed effect embodies fundamental determinants of trade such
as historical links or language, these are almost certainly not correlated to the national price
level, which is determined by macroeconomic forces at a global scale.
3.4.2 Errors-in-variables
Some of the independent variables might have been measured with errors. This would lead to
inconsistent OLS estimations that underestimate the coefficient (attenuation bias). When
dealing with country-level economic data the reliability of statistical information from less
developed countries is always a concern. However, a significant measurement error seems
unlikely due to the fact that the data utilised comes from a trusted external source such as the
World Bank. In any case, even if some of those less developed countries include
measurement error, they are not among the countries NPL sells the most, thus the relevance
in our regression analysis is assumed minimal.
3.4.3 Simultaneity
Simultaneity bias can be ruled out from the get-go mainly because of the nature of the analysis.
Our goal is to explain NPL sales worldwide through macroeconomic variables that are driven
by forces at a global scale. Even if a hidden macroeconomic variable is affecting both the
dependent variable and one of the regressors, as mentioned before, our specification controls
for it by including a dummy variable for the year which should capture the effect of significant
events.
3.4.4 Weighting of observations in the sample
An analytical weighting correction has been implemented across all estimations conducted,
because the dependent variable consists of the number of invoices issued to a country divided
by its population. Although the change in the results is not noticeable (the same values and
confidence intervals are obtained up to several significant figures), the correction has been
implemented as it is the standard in the empirical economic literature that uses country-level
data.
3.4.5 Selectivity
Finally, selection bias deserves a more detailed explanation, since it is very relevant to our
analysis. Many countries in the dataset, usually small and/or less developed countries, do not
make any purchase in some years, and, thus, NPL does not issue any invoice to those
countries. As a result, the dependent variable is zero for over 40% of the dataset. Moreover,
since equation 4.1 is specified in terms of the logarithm of the dependent variable, a significant
incidental truncation arises. In order to address this issue, a two-step selection correction has
been used to model the individual sampling probability of each observation occurring
(extensive effect), together with the conditional expectation of the dependent variable
(intensive effect). In other words, the approach adopted allows us to disentangle the incidence
of the national price level on the probability of purchase and the level of consumption. The first
stage is represented by the dichotomous choice of whether to purchase NPL measurement
services or not, and the second stage determines the level of consumption once the decision
to purchase is made.

3.5 Estimation

Four different models have been tested based on the population equation 4.1. Models A1 to
A3 consists of uncorrected for selectivity OLS regressions. Model A1 is static. Thus, it consists
of a first naïve approximation to the population process that neglects any persistence.

20
However, as mentioned in section 4.1.1, the generating process is likely to exhibit some
inertia, due to all the macroeconomic unobservables in the error term. For that reason, models
A2 and A3 introduce one lag and three lags of the residuals of A1 respectively. These hold
significant explanatory power over the dependent variable – the latter removing any trace of
serial correlation. Thus, the presumed dynamic structure is confirmed. Lastly, Model B1
consists of the 2-step Heckman correction of A3. This is the preferred model because it
accounts for the dynamics of the population process, as well as the potential bias coming from
selectivity in the sample.
This subsection first introduces the estimations results for the four models, analysing the
economic meaning of each estimated coefficient. Then the preferred model is presented and
analysed in depth to show how the correction proposed by applies to our model and why it is
a reliable method to this context.
3.5.1 Models tested and interpretation of the coefficients
Table 3 summarises the estimations results for models A1 to A3 and model B1:

Model A1 Model A2 Model A3 Model B1

DV = log I P- P- P- P-
Coeff. Coeff. Coeff. Coeff.
value value value value

1.31 0.000 1.25 0.000 1.19 0.000 1.24 0.000

log PL
(0.10) (0.08) (0.07) (0.08)

1.15 0.000 1.23 0.000 1.25 0.000 1.33 0.000

log GDP
(0.08) (0.07) (0.07) (0.08)

-0.02 0.000 -0.02 0.000 -0.02 0.000 -0.02 0.000

CMC
(0.00) (0.00) (0.00) (0.00)

-0.44 0.000 -0.45 0.000 -0.47 0.000 -0.48 0.000

log D
(0.03) (0.02) (0.02) (0.02)

-0.20 0.000 -0.20 0.000 -0.20 0.000 -0.17 0.000

log POP
(0.04) (0.02) (0.02) (0.03)

. . 0.76 0.000 0.36 0.000 0.35 0.000

u (t-1)
. (0.03) (0.05) (0.05)

. . . . 0.33 0.000 0.33 0.000

u (t-2)
. . (0.04) (0.04)

u (t-3) . . . . 0.22 0.000 0.22 0.000

21
 . . (0.04) (0.04)

. . . . . . 0.28 0.047
Λ
. . . (0.14)

P- P- P- P-
F-stat F-stat F-stat F-stat
value value value value

time
7.13 0.000 11.23 0.000 18.59 0.000 17.87 0.000
dummies

R-squared 0.81 0.92 0.94 0.94

Number of
939 801 648 648
obs.

Heckman No No No Yes

Table 3: Estimation results.

The estimated coefficients for our variables of interest are highly significant and make perfect
economic sense. Notably, the price level coefficient is positive. This is consistent with the
definition of the national price level given by equation 2.1. With respect to GDP per capita, the
coefficient is also coherent with the expected income effect, since NPL’s services are
supposed to be normal goods whose demand increases with real income. By contrast, the
negative significant coefficients found for distance and CMCs of the local NMI tally with our
assumption of transactional frictions in the form of transport and search costs. Furthermore,
the statistical significance of the population coefficient suggests a non-linear relationship of
the number of invoices and the population of the country. In addition, the significance of the
coefficients of the lagged residuals confirms the presumption of persistence in the generation
process.
Lastly, there is an additional regressor symbolised by 𝜆 which only applies to model B1. This
is called the Inverse Mills Ratio (IMR), and results from a probit model that tries to reproduce
the decision made by users worldwide about whether to buy NPL’s services or not. Surely,
there are unobserved factors that make the decision of buying NPL’s services more likely, as
well as being associated with higher levels of consumption. In other words, it is expected that
those users who are more inclined to buy NPL’s services, are also the ones that buy more
services. By adding the IMR to the dynamic structure set by A3, we can control for those
unobserved factors that affect both the decision of buying NPL’s services and the level of
consumption; thus, any bias in the rest of estimated parameters can be corrected.
The next subsection analyses in more depth the preferred setup Model B1. It depicts
extensively the application of the Heckman correction to our model and justifies why it is a
suitable method to address the potential selection bias in the analysis. Finally, the estimation
results are interpreted focusing on the existing bias on the uncorrected setup, A1 to A3.
3.5.2 Model B1: 2-step Heckman selection model
The estimates found through the models A1 to A3 might be biased due to selectivity. The
coefficients estimates may not be applicable to all countries (buyers and non-buyers) because
we only observe those that actually make a purchase. In other words, the estimation results

22
may not be representative of the whole population because of the non-random nature of the
observed sample. To correct for this source of bias, we follow the approach proposed by .
Formally, the correction to account for incidental truncation consists of adding an explicit
selection equation to our population equation 4.1. Thus, the model is given by:

log 𝐼𝑖,𝑡 = 𝛽0 + 𝛽1 log 𝑃𝐿𝑖,𝑡 + 𝛽2 log 𝐺𝐷𝑃𝑖,𝑡 + 𝛽3 𝐶𝑀𝐶𝑖 + 𝛽4 log 𝐷𝑖

(4.2a)
+ 𝛽5 log 𝑃𝑂𝑃𝑖 + 𝛾𝑡 𝑌𝑡 + 𝛿1 𝑢1 + 𝛿2 𝑢2 + 𝛿3 𝑢3 + 𝜀𝑖,𝑡

𝑠𝑖,𝑡 = 1[𝜂0 + 𝜂1 log 𝑃𝐿𝑖,𝑡 + 𝜂2 log 𝐺𝐷𝑃𝑖,𝑡 + 𝜂3 𝐶𝑀𝐶𝑖 + 𝜂4 log 𝐷𝑖 + 𝜂5 log 𝑃𝑂𝑃𝑖
(4.2b)
+ 𝜃𝑖 𝑌𝑡 + 𝜅1 𝑣1 + 𝜅2 𝑣2 + 𝜅3 𝑣3 + 𝜑𝑖,𝑡 ]

where 𝑠𝑖,𝑡 = 1 if a purchase is made, and zero otherwise, 𝐼𝑖,𝑡 is the number of invoices issued
to country 𝑖 in year 𝑡 normalised by the population of the country, 𝑃𝐿𝑖,𝑡 is the national price
level, 𝐺𝐷𝑃𝑖,𝑡 is the gross domestic product per capita, 𝐶𝑀𝐶𝑖 is a continuous index that rates
the measurement capabilities of the country, 𝐷𝑖 is the distance between the country and the
UK, 𝑃𝑂𝑃𝑖 is the population of the country, 𝑌𝑡 is a year dummy variable, 𝑢1 to 𝑢3 are the first
three lagged residuals of the static uncorrected model (Model A1), 𝑣1 to 𝑣3 are the first three
lagged residuals of the static specification of the probit model, and 𝜀𝑖,𝑡 and 𝜔𝑖,𝑡 are the error
terms in the population and selection equation respectively, which are assumed to be
independently and identically distributed as a normal distribution with mean of 0 and finite
variance.
An important remark should be made with regard to the differences between the sets of
regressors used in both equations 4.2. Ideally, for the proposed method to work best, the set
of regressors in the population equation should be a strict subset of the covariates in the
selection equation. This means that two conditions must be satisfied: (1) any of the regressors
in equation 4.2a should also be included in equation 4.2b, and (2) some of the regressors in
equation 4.2b should not appear in equation 4.2a.
Technically, the first constraint is not met, whereas the second one is. Indeed, the first three
lagged residuals of the static uncorrected model, 𝑢1 to 𝑢3 , are not included in equation 4.2b.
Conversely, these have been substituted by the first three lagged residuals of the static
specification of the probit model. The reason for this is that, although both sets of residuals
are linked to the unobservables, arguably the latter is more suitable to consider the effect of
the omitted variables on the outcome variable since it does so in probabilistic terms. In any
case, ideally, at least a true additional identifying variable should be included in the selection
equation. This variable would affect the probability of making a purchase, but not (at least to
a large extent) the number of invoices issued. However, all attempts to include identifying
variables linked to the complexity of the country's exports or the relevance of the
manufacturing sector have been unsuccessful. The reason for this is that these variables are
highly correlated with the CMC regressor which assesses the measurement capabilities of the
local NMI – and which holds more explanatory power. Nevertheless, the proposed setup
happens to work quite well in terms of statistical significance and makes the most sense from
a theoretical perspective.
The Heckman two-stage estimation procedure allows us to decompose the total effect of the
covariates of interest (in particular the price level) into an extensive (probability of purchase –
equation 4.2b) and an intensive (level of consumption – equation 4.2a) component.

23
Table 4 summarises the estimation results of the probit model that assesses the probability of
purchasing NPL services (dummy variable 𝑇𝑅𝐸𝐴𝑇𝐸𝐷 = 1) as a function of the original
regressors18.

Probit Model
DV = TREATED
Coeff. P-value

0.91 0.000
log PL
(0.19)

0.91 0.000
log GDP
(0.09)

0.02 0.211
CMC
(0.01)

-0.26 0.000
log D
(0.07)

0.46 0.000
log POP
(0.06)

-0.40 0.229
v (t-1)
(0.34)

-1.22 0.000
v (t-2)
(0.33)

-0.17 0.586
v (t-3)
(0.32)

Pseudo R-
0.44
squared

Number of obs. 1134

Table 4: Probit regression estimation results (extensive effect).

18The estimated coefficients for the year dummy variables have been omitted from this output for the
sake of conciseness. These are included in Annex D.

24
These coefficients have no direct interpretation in terms of marginal effects; they are simply
the values that maximise the likelihood function. Although the signs of the coefficient give the
direction of the effect, their magnitude is in units of the standard-deviation of the errors. There
are two commonly used approaches to approximate the marginal effects for non-linear
models: the partial effects at the average (PEA) and the average partial effect (APE). The PEA
is the marginal effect evaluated at the average value for each regressor; the APE is simply an
estimate of a population-averaged marginal effect:

𝜕𝐄[𝑦|𝒙] 𝜕𝛷[𝒙𝜷]
𝑃𝐸𝐴𝑗 = | = | = 𝛽̂𝑗 ∙ 𝜑(𝒙
̅𝜷̂) (4.3a)
𝜕𝒙 𝒙=𝒙̅ 𝜕𝒙 𝒙=𝒙̅

𝑁
𝛽̂𝑗
̂)
𝐴𝑃𝐸𝑗 = ∑ 𝜑(𝒙𝜷 (4.3b)
𝑁
𝑖=1

where 𝑃𝐸𝐴𝑗 and 𝐴𝑃𝐸𝑗 are the partial effect at the average and the average partial effect for
regressor 𝑗, 𝑦 and 𝒙 are denote our binary dependent variable and the set of covariates in the
probit regression, 𝑁 is the total number of observations, 𝜑(∙) is the standard normal density
̂ is the vector of the parameter estimates (including the intercept) of the probit
function, and 𝜷
regression – with 𝛽̂𝑗 representing a component of that vector. Table 5 below shows both
approximations of the marginal effects19.

DV = Marginal Effect at Means Average Marginal Effect

TREATED Coeff. P-value Coeff. P-value

0.18 0.000 0.17 0.000

log PL
(0.04) (0.04)

0.18 0.000 0.17 0.000

log GDP
(0.03) (0.01)

3.6E-03 0.173 3.4E-03 0.204

CMC
(2.6E-03) (2.7E-03)

-0.05 0.000 -0.05 0.000

log D
(0.01) (0.01)

0.09 0.000 0.09 0.000

log POP
(0.01) (0.01)

19 Again, the estimated coefficients for the year dummy variables have been omitted; these can be
found in Annex D.

25
 -0.08 0.240 -0.08 0.230
v (t-1)
(0.07) (0.06)

-0.24 0.000 -0.23 0.000

v (t-2)
(0.07) (0.06)

-0.03 0.587 -0.03 0.586

v (t-3)
(0.06) (0.06)

Number of
1134 1134
obs.

Table 5: Partial effects for the extensive effect.

Note that both effects are very similar for all of the regressors. These is expected given the
covariates are reasonably normally distributed, and thus, the mean is a good measure of
centrality within the domain of definition of the covariates.
Once the extensive effect is estimated, we can turn our attention to the intensive effect. Our
objective is to know how the variables of interest affect the level of consumption – in particular,
the national price level. If we analyse this through a simple OLS regression, the results will be
surely biased because the factors that determine if users buy NPL’s services, also affect how
much they consume. Hence, the sample we are working with is not representative of the whole
population. In effect, we need to separate out the two decisions made by the user: whether to
make a purchase or not (extensive), and how much to buy (intensive). To that end, we can
use the results of the previous probit estimation by computing the Inverse Mills Ratio. The IMR
is equivalent to the concept of hazard ratio in survival analysis, or force of mortality in
demography and actuarial science. In the context of our analysis, the IMR shows the relative
increase in the likelihood of buying NPL’s services given a unit change in the level of
consumption:

̂)
𝜑(𝒙𝜷
𝜆𝑖,𝑡 = (4.3)
̂)
𝛷(𝒙𝜷

If the estimate of the linear parameter corresponding to the IMR is significant (expected
positive), then we know the unobserved factors that make the decision of buying NPL’s
services more likely, tend to be associated with higher levels of consumption. Therefore,
effectively, including the IMR in the OLS regression allows us to control for those unobserved
factors and correct any bias in the rest of estimated parameters. Table 6 shows the results of
the OLS regression including the IMR obtained from the previous probit setup.

Model B1
DV = log I
Coeff. P-value

1.24 0.000
log PL
(0.08)

26
 1.33 0.000
log GDP
(0.08)

-0.02 0.000
CMC
(0.00)

-0.48 0.000
log D
(0.02)

-0.17 0.000
log POP
(0.03)

0.35 0.000
u (t-1)
(0.05)

0.33 0.000
u (t-2)
(0.04)

0.22 0.000
u (t-3)
(0.04)

0.28 0.047
λ
(0.14)

F-stat P-value

time dummies 17.87 0.000

R-squared 0.94

Number of
648
obs.

Table 6: Heckman's selection model estimation results (intensive effect).

As expected, the IMR is significant and positively signed – which suggests that the error terms
in the selection and primary equations are positively correlated. Therefore, we confirm our
suspicions that factors that make buying NPL’s services more likely are associated with
purchasing more of these goods. This justifies the use of Heckman’s correction in our setup
and provides confidence on the estimation results.

27
Finally, the estimated coefficient on the national price level is now larger than before – that is
if we compare estimates for models A3 and B1. This indicates that the non-random nature of
the sample biases the price elasticity down.

3.6 Postestimation

This subsection analyses the validity of the models tested, particularly that of the preferred 2-
step Heckman selection model. Table 7 shows the results of a link test20 which has been run
to check for misspecification of the dependent variable for the four models tested. The
motivation behind the link test is to assess if a regression model is affected by the so-called
link error, that is, that the dependent variable needs a transformation or link function to properly
relate to the independent variables. To verify this is not the case, the link test regresses the
dependent variable against the original regression's predicted values and the squared values
of this prediction. If the squared prediction regressor is significant, there is evidence that the
model is mis specified; in addition, it is expected that the coefficient for the prediction regressor
is highly significant with a coefficient close to one.

Model A1 Model A2 Model A3 Model B1

DV = log I P- P- P- P-
Coeff. Coeff. Coeff. Coeff.
value value value value

predicted
1.07 0.000 1.31 0.000 1.04 0.000 1.03 0.000
values
(0.20) (0.12) (0.11) (0.11)

squared
0.00 0.706 0.01 0.011 0.00 0.731 0.00 0.768
prediction
(0.01) (0.00) (0.00) (0.00)

R-squared 0.82 0.92 0.94 0.94

Number of
939 801 648 648
obs.

Table 7: Link test.

All the models pass the test. Even in the case of model A2 for which the coefficient of the
squared prediction is statistically significant, the coefficient itself is very close to zero.
However, clearly those models that include three lags of the error term of the static model
(models A3 and B1) behave better. Therefore, the link test provides assurance in the
specification of the preferred model.
In order to provide further evidence of the correct specification of the model, the Ramsey
Regression Equation Specification Error Test (RESET) is conducted. This technique tests
whether non-linear combinations of the fitted values have explanatory power over the
dependent variable. The intuition behind this test is that if the second or the third power of any
of the regressors, or any interaction term between them have any power in explaining the

20The test implemented here is based on an idea of (Tukey, 1949) which was further described by
(Pregibon, 1980).

28
dependent variable, then the model omits relevant variables and may be better approximated
by different functional form. Table 8 reports the results of the Ramsey RESET test for the four
models considered.

Model A1 Model A2 Model A3 Model B1

F-stat P-value F-stat P-value F-stat P-value F-stat P-value

12.02 0.00 6.67 0.00 1.64 0.18 1.52 0.21

Table 8: Ramsey RESET test.

Like the link test, the Ramsey RESET test favours both specifications that account for three
lags of the error term of the static regressions. Moreover, these are the only specifications that
pass the test; hence, the test supports the preferred Heckman-corrected dynamic specification
– model B1.
Two other important aspects to ensure the validity of the specification proposed are whether
there are signs of endogeneity in the models tested, and whether the residuals are serially
correlated. Both issues can be analysed together since they are linked. Indeed, endogeneity
is likely to arise due to the dynamics of the generating process, hence leading to serially
correlated errors if not taken into account. Table 9 shows the results of a test based on which
allows us to test for endogeneity is an alternative to the Hausman test21. Unlike the Hausman
test, the method developed by Mundlak may be used when the errors are heteroskedastic or
have intragroup correlation. Hence, this alternative approach is especially suitable for the
purpose of this analysis.

21 A thorough explanation of the intuition and the algebra behind this method is depicted in Annex B.

29
 Model A1 Model A2 Model A3 Model B1
DV = log I P- Coeff P- Coeff P- Coeff P-
Coeff.
value . value . value . value

-0.04 0.816 1.00 0.000 0.74 0.006 0.77 0.004

log PL
0.19 0.17 0.27 0.26

1.83 0.000 0.28 0.218 0.50 0.094 0.52 0.066

mean (log PL)
0.34 0.22 0.30 0.28

0.85 0.000 1.24 0.000 1.37 0.000 1.47 0.000

log GDP
0.22 0.18 0.22 0.23

-0.10 0.682 -0.02 0.911 -0.14 0.509 -0.16 0.457

mean (log GDP)
0.25 0.19 0.21 0.21

-0.01 0.105 -0.02 0.000 -0.02 0.000 -0.02 0.000

CMC
0.01 0.00 0.00 0.00

-0.40 0.001 -0.45 0.000 -0.46 0.000 -0.47 0.000

log D
0.12 0.03 0.02 0.02

-0.32 0.002 -0.20 0.000 -0.20 0.000 -0.17 0.000

log POP
0.10 0.04 0.02 0.02

P- P- P- P-
F-stat F-stat F-stat F-stat
value value value value

240.3
Time dummies 0.000 93.91 0.000 44.45 0.000 48.04 0.000
3

Panel-level
29.43 0.000 1.57 0.457 3.97 0.137 4.32 0.115
means

within = 0.323 within = 0.14 within = 0.137 within = 0.136

between = between = between = between =
R-squared
0.861 0.982 0.988 0.991
overall = 0.808 overall = 0.917 overall = 0.938 overall = 0.939

30
 Number of obs. 939 801 648 648

Heckman No No No Yes

Table 9: Mundlak test.

After accounting for the dynamics of the generating process the cluster means of the two time-
varying covariates become insignificant. Therefore, this means that our choice to explicitly
model this inertia is appropriate, and particularly our preferred model B2 shows no signs of
endogeneity. These conclusions are reinforced when we test whether the residuals are serially
correlated:

Model A1 Model A2 Model A3 Model B1

DV =
residuals P- P- P- P-
Coeff. Coeff. Coeff. Coeff.
value value value value

0.76 0.000 -0.29 0.000 -0.03 0.453 -0.03 0.444

residuals (t-1)
(0.02) (0.04) (0.04) (0.04)

Table 10: Serial correlation test.

It can be concluded that those models that incorporate three lags of the residuals of the static
model better reflect the dynamics of the generating process.
Finally, it is necessary to make a brief mention about the normality of the residuals of the
models tested. Although the sample is not very wide (just below a thousand observations) in
all cases the residuals have a fairly normal distribution. This can be appreciated in Figure 2
where the histogram of the residuals for all models are plotted, and a kernel density estimation
of the probability density function is compared against the expected normal distribution.

31
Figure 2: Normality of the residuals.

32
4 Economic Impact

This section shows how the price elasticity found in the previous econometric analysis can be
used to estimate the welfare generated by NPL for the UK’s society. It should be emphasized
that the objective here is not to develop a comprehensive and accurate impact analysis, but
to showcase that the price elasticity of demand is a useful tool to asses welfare generated.
Therefore, the reader should not expect the following to be a detailed and complete analysis
of the net present value of the public funding of NPL. In fact, all the figures pertaining to NPL’s
finances and public support received are estimative (although not very far off). Thus, any cost-
benefit ratio presented should not be taken as representative of the current value for money
of NPL public financing.
When it comes to assessing welfare created, there are two main types of benefits generated:
direct and indirect benefits. Direct benefits are those that occur within supported companies.
Effectively, NPL sells the time and the expertise of its scientists and engineers who help users
to develop new products and processes. New products enable supported businesses to
increase their market power and command a price premium. New processes enhance
productivity and competitiveness. In any case, NPL’s services lead to additional earnings,
either through increased sales or costs savings. However, direct benefits are not limited to the
increase in private profits which roughly accounts for about 25% of gross value added (GVA).
The remaining three quarters are earned by workers in the form of higher salaries. Therefore,
GVA is probably about four times the direct return to NPL’s paying customers.
There are wider benefits that arise beyond the supported businesses. These are indirect
benefits that occur because knowledge generated due to the collaboration between NPL and
paying customers spills over to unsupported firms – normally in the same sector. The main
channel through which this knowledge benefits companies that do not engage with NPL is the
movement of workers among firms. The intuition behind this knowledge spillovers is as follows.
First, a company engages with NPL. This allows the supported business to acquire knowledge
which is then drawn on to promote additional sales and/or costs savings. Obviously, most of
this knowledge is effectively obtained by the company’s workers. If these workers decide to
move to other firms (typically in the same sector or industry), they will carry over the acquired
new knowledge. Frontier Economics found that the existing literature estimates that “social
returns, based on spillover benefits from R&D conducted by one agent to the productivity or
output of other agents, are typically 2 to 3 times larger than private returns.” Therefore, in this
analysis we will stick to the lower bound of these estimates and will approximate the indirect
benefits by just doubling the direct benefits.
Lastly, given that 50% of NPL’s income comes from customers who are based in other
countries, the question of whether those sales have a positive impact in the UK arises.
Regarding the direct benefits, it is clear that none of these stay in the UK – most of the GVA
generated happens in the foreign country. However, that is not the case for indirect benefits.
Indeed, NPL plays a fundamental role when it comes to any knowledge generated from
collaborations with non-UK based users spilling over to UK businesses. This is due to the fact
that NPL’s scientific staff gain knowledge as a result of the collaboration. Moreover, these
scientists and engineers tend to be young professionals who are likely to switch jobs given
they are in the early stages of their career. In fact, the stability in NPL's workforce has
decreased in recent times, as has the average age of workers. In this sense, NPL acts as a
platform for all the knowledge generated (including the one developed as a result of
collaborations with overseas companies) to reach all layers of the sectors involved. Hence, in
the following impact analysis we will account for indirect benefits resulting from sales to users
abroad, despite direct benefits being ignored.

33
4.1 Direct benefit to paying customers and welfare for the UK

Companies regard NPL’s services as investment projects that generate profits over time. A
rational user would prioritise projects with a higher payoff and, if no budget constraint is in
place, would buy NPL’s services to the point that the marginal benefit equals the marginal
cost. This suggests a downward sloping aggregate demand curve for NPL’s services: more
profitable projects yield higher future earnings which increases customers’ willingness to pay.
This situation is depicted in Figure 3. The horizontal axis shows the number invoices issued
by NPL, which is a proxy for the volume of services provided. The vertical axis of Figure 3 is
the net present value of the investment projects supported by NPL.

Figure 3: Demand curve for NPL's services.

Therefore, at first approximation the demand curve for NPL’s services is given by:

𝑝(𝑄) = 𝐴 − 𝐵𝑄 (5.1)

Where 𝐴 > 0 and 𝐵 > 0 are constants.

The direct benefit to NPL’s customers is given by the area beneath the curve:

1 1
𝛱 = 𝑝𝑄 + (𝐴 − 𝑝)𝑄 = 𝐴𝑄 − 𝐵𝑄 2 (5.2)
2 2

Hence, in order to estimate the direct benefit to paying customers, we need to know the values
of 𝐴 and 𝐵. To that end, we can make use of the price elasticity of demand:

𝑑𝑄 𝑝0 1 𝑝0
𝜀𝑝0 = =− (5.3)
𝑑𝑝 𝑄 0 𝐵 𝑄0

34
Note that in general the value of the elasticity depends on the level of sales. (In particular, this
is true for the linear approximation we are considering). Thus, the superscripts in 𝜀𝑝 , 𝑝 and 𝑄
denote the fact these are evaluated in the current market equilibrium.
Parameters 𝐴 and 𝐵 can be derived from the price elasticity of demand given the current level
of sales:

1 𝑝0
𝐵=− (5.4a)
𝜀𝑝0 𝑄 0

1
𝐴 = 𝑝0 + 𝐵𝑄 0 = 𝑝0 (1 − ) (5.4b)
𝜀𝑝0

As mentioned in section 2.1, the elasticities of demand with respect to conventional price and
the national price level are equivalent in magnitude but of opposite sign:

𝜀𝑝0 = −𝜀𝐿0 (5.5)

Where 𝜀𝐿0 is the national price level elasticity of demand estimated in section 4.
If we substitute equations 5.4 and 5.5 into 5.2, we get:

1 1 11 1
𝛱0 = 𝐴𝑄 0 − 𝐵(𝑄 0 )2 = 𝑝0 𝑄 0 (1 + 0 ) − 0 𝑝0 𝑄 0 = 𝑅 0 (1 + 0 ) (5.6)
2 𝜀𝐿 2 𝜀𝐿 2𝜀𝐿

Where 𝑅 0 is the revenue made by NPL.

The average commercial revenue made by NPL over the last three years was £35.6m22 – 50%
(£17.8m) of it was generated in the UK. Substituting this figure along with the national price
level elasticity found in the econometric analysis of section 4 into equation 5.6, we get:

1 1
𝛱𝐷 = 17.8 ∙ (1 + ∙ ) = 25.05 [£m]
2 1.24

Hence, it is estimated that NPL’s work with UK businesses generates £25.05m for supported
companies. However, a mentioned at the beginning of this section, this direct return to NPL’s
paying customers corresponds to additional profits rather than gross value added (GVA).
Hence, given that the return to capital (profit) accounts for about 25% of income, the
corresponding increase in GVA is probably about four times the result given by equation 5.6,
that is, £100.20m. Now, if we take into account that the current level of public funding received
by NPL is £83.10m, we end up with a cost-benefit ratio of 1.21.

22 Instead of using the average revenue generated throughout the whole period considered in the
econometric analysis (2001-2017), only the average of the last three years of the dataset has been
taken. This is considered to be more representative of the current situation. All revenue figures have
been conveniently inflated to 2019 prices based on the GDP deflators provided by the ONS.

35
Up to this point, only the direct benefits that take place within supported companies have been
calculated. Given the nature of the work carried out by the scientific staff of NPL, the benefits
spill over to non-supported companies, thus carrying even greater benefits for the UK. As was
also mentioned at the top of this heading, Frontier Economics finds that social returns are
typically 2 to 3 times larger than private returns. Therefore, we can estimate the indirect
benefits by doubling the direct benefits. Moreover, we argued that NPL allows for spillovers to
occur even when the collaboration takes place with overseas users. Therefore, doubling the
direct benefits to all customers is a reasonable approximation of the overall indirect benefits
for the UK. Hence, given that average revenue made by NPL over the last three years was
£35.6m, the direct benefit generated for all customers (UK and overseas) is given by:

1 1
𝛱𝐼 = 35.6 ∙ (1 + ∙ ) = 49.95 [£m]
2 1.24

Now, this figure multiplied by 4 gives us the additional gross value added (GVA), £199.82m.
Lastly, if we double this number, we get the indirect benefits for the UK, £399.64m, which
added to the direct benefits previously computed, yields a total welfare generated of £499.84m
and a cost-benefit ratio of 6.01 when the current level of public funding received by NPL is
considered.

4.2 The effect of shifts in public funding received by NPL on welfare

generated

From an evidence-based policy perspective, it is key to know the effect of a change in the
public funding of NPL and the welfare generated. Public funds allow NPL to hire scientific staff
who support UK companies develop new product or processes. In other words, since NPL is
an organisation that acts as a vehicle to efficiently allocate the public funds needed to
complement private spending in measurement R&D, there is a strong relationship between
public support and NPL’s output. The estimation of such relationship is beyond the scope of
this report though. Nonetheless, under the premise that this relationship exists and is stable,
it is possible to assess the effect on welfare generated of a variation in NPL’s output.
Therefore, we will consider that a reduction in public funding triggers an equivalent reduction
in output. That is, for every 1% less in public funding, a 1% reduction in output is expected.
Figure 4 shows a schematic representation of a shift in output of 20% up (𝐸 ′′ ) or down (𝐸 ′ ).

36
 Figure 4: Changes in welfare generated.

The direct benefits to paying customers is currently given by areas 1 and 2 – this has been
estimated in the previous section. Area 3 corresponds to the additional welfare that would be
created if NPL’s output increases by 20% (triggered by an equivalent increase in funding).
Conversely, area 2 is what would be lost if NPL’s reduces its output by 20% (because of a
20% reduction in funding). Analytically, this can be estimated using equation 5.6, for which
parameters 𝐴 and 𝐵 stay constant. Thus, the first step is to compute both parameters using
the current number of invoices issued in the UK (3,860 invoices), and the average size of an
invoice (£4,624)23. These are given by Table 11:
Parameter Value
𝐴 £8,354
𝐵 £0,97/Invoice

Table 11: Parameters for the demand curve.

Next, the output (number of invoices issued) in the new equilibria must be calculated (𝑄 ′ for
𝐸 ′ , and 𝑄 ′′ for 𝐸 ′′ ), as well as the corresponding level of public funding (𝐹 ′ and 𝐹 ′′
respectively). These consist of a 20% positive and negative variation over the current situation.
Table 12 shows these alternative equilibria.
Parameter Value
𝑄′ 3,088 invoices
𝐹′ £66.5m
𝑄 ′′ 4,632 invoices
𝐹 ′′ £99.7m

Table 12: Alternative equilibria.

23 Again, the last three years of the dataset are considered to be a representative period of the current
situation.

37
Lastly, we can proceed the same way we did in the previous subsection. First, we calculate
the change in private profits generated for the supported UK companies using equation 5.6
and multiply by 4 to get the variation in GVA. This gives us the change in direct benefits.
Secondly, to get the change in indirect benefits we calculate the GVA variation for all users
and multiply by 2. The sum of both variations (direct and indirect benefits) yields the overall
change in welfare and cost-benefit ratios24. For the 20% decrease in funding scenario the
overall welfare is reduced in £77.11m, yielding a decrease in the cost-benefit ratio of −1.16;
for the 20% increase in public funds scenario the welfare increases by £65.60m and the cost-
benefit ratio by 0.66.

24For the sake of simplicity all of these calculations have been omitted since they are equivalent to
those made in section 4.1.

38
5 Conclusion

Our study estimates the price elasticity of demand for NPL’s services, finding that it is well
above 1 at 95% confidence. This supports the idea that NPL’s services are elastic goods for
which the quantity demanded will change more than proportionally if the price changes.
We apply standard panel data analysis to a country-level panel dataset to estimate the price
elasticity of demand. The data utilised consists of NPL internal invoicing information, as well
as data on financial and demographic variables coming from trusted external commonly-used
sources (namely, the World Bank, the Centre d'Etudes Prospectives et d'Informations
Internationales, the Bureau of Weights and Measures, and the United Nations). The proposed
analytical setup allows us to control for several variables that are expected to affect the
number of services sold. All controls in our model, GDP per capita, distance and the
measurement capabilities of local competing NMIs, turned out to be highly significant.

39
6 References

Arrow, K. (1962). Economic welfare and the allocation of resources for invention. In R. R.
Nelson, The Rate and Direction of Inventive Activity: Economic and Social Factors (pp.
609-625). NJ: Press: Princeton.
Barlev, G. (2007). On the cyclicality of research and development. American Economic
Review, 97(4), 1131-1164.
Beck, N., & Katz, J. N. (2011). Modeling dynamics in time-series--cross-section political
economy data. Annual Review of Political Science, 331-352.
Bureau International des Poids et Mesures. (2019, August 27). BIPM Calibration and
Measurement Capabilities - CMCs. Retrieved September 15, 2019, from
https://www.bipm.org/utils/common/pdf/KCDB/KCDB_CMCs.pdf
Davidson, R., & MacKinnon, J. G. (1981). Several tests for model specification in the presence
of alternative hypotheses. Econometrica, 781-793.
Frontier Economics. (2014). Rates of return to investment in science and innovation. London.
Frontier Economics. (2016). The Impact of Public Support for Innovation on Firm Outcomes.
Head, K., & Mayer, T. (2002). Illusory border effects: Distance mismeasurement inflates
estimates of home bias in trade. Citeseer.
Heckman, J. J. (1976). The common structure of statistical models of truncation, sample
selection and limited dependent variables and a simple estimator for such models.
Annals of Economic and Social Measurement, 5(4), 475-492.
HM Treasury. (2018). The Green Book: Central Government Guidance on Appraisal and
Evaluation.
Mundlak, Y. (1978). On the pooling of time series and cross section data. Econometrica:
journal of the Econometric Society, 69-85.
Pregibon, D. (1980). Goodness of link tests for generalized linear models. Journal of the Royal
Statistical Society: Series C (Applied Statistics), 29(1), 15-24.
Ramsey, J. B. (1969). Tests for specification errors in classical linear least-squares regression
analysis. Journal of the Royal Statistical Society: Series B (Methodological), 31(2),
350-371.
Tassey, G. (2004). Underinvestment in public good technologies. The Journal of Technology
Transfer, 3(1-2), 89-113.
Tassey, G. (2008). and measuring the economic roles of technology infrastructure. Econ.
Innov. New Techn., 17(7-8), 615-629.
Tukey, J. W. (1949). One degree of freedom for non-additivity. Biometrics, 5(3), 232-242.
Winning Moves. (2017). National Measurement System: Customer Needs and Impact Survey.
NMS.

40
Annex A: Robustness tests

This annex weighs the effect of some of the decisions adopted throughout the analysis. Particularly with regard what countries to include in the
analysis, and the methodological approaches taken as to how to deal with the distinct nature of the observations of the UK, the treatment given
to outliers, and the definition of the CMC variable.

Countries that do not purchase NPL’s services regularly

One of the fundamental features of the dataset is the notable differences among the countries included. For some of them, the size and
commercial interrelations with the UK ensures a stable flow of purchases over time. For others, usually small less developed countries, buying
services from the NPL is a rare event. This could lead to some sort of bias in the estimation results. Therefore, the effect of excluding countries
that do not have a minimum number of observations has been verified. In particular, the effect of excluding countries that have only one
observation, up to five observations (models C1 to C5), has been systematically tested. The results of this robustness test are included in Table
13, which shows that the effect is not very significant and that is why the main analysis is carried out with all observations.

41
 Model B1 Model C1 Model C2 Model C3 Model C4 Model C5
DV = log I P- P- P- P- P- P-
Coeff. Coeff. Coeff. Coeff. Coeff. Coeff.
value value value value value value

1.24 0.000 1.24 0.000 1.23 0.000 1.17 0.000 1.12 0.000 1.13 0.000
log PL
(0.08) (0.08) (0.08) (0.08) (0.07) (0.07)

1.33 0.000 1.33 0.000 1.35 0.000 1.40 0.000 1.46 0.000 1.46 0.000
log GDP
(0.08) (0.08) (0.07) (0.07) (0.07) (0.07)

-0.02 0.000 -0.02 0.000 -0.02 0.000 -0.02 0.000 -0.02 0.000 -0.02 0.000
CMC
(0.00) (0.00) (0.00) (0.00) (0.00) (0.00)

-0.48 0.000 -0.48 0.000 -0.47 0.000 -0.48 0.000 -0.47 0.000 -0.47 0.000
log D
(0.02) (0.02) (0.02) (0.02) (0.02) (0.02)

-0.17 0.000 -0.17 0.000 -0.17 0.000 -0.16 0.000 -0.14 0.000 -0.14 0.000
log POP
(0.03) (0.03) (0.03) (0.03) (0.03) (0.03)

u (t-1) 0.35 0.000 0.35 0.000 0.35 0.000 0.35 0.000 0.35 0.000 0.35 0.000

42
 (0.05) (0.05) (0.05) (0.05) (0.05) (0.05)

0.33 0.000 0.33 0.000 0.33 0.000 0.33 0.000 0.33 0.000 0.33 0.000
u (t-2)
(0.04) (0.04) (0.04) (0.04) (0.04) (0.04)

0.22 0.000 0.22 0.000 0.22 0.000 0.22 0.000 0.22 0.000 0.21 0.000
u (t-3)
(0.04) (0.04) (0.04) (0.04) (0.04) (0.04)

1.28 0.003 -1.24 0.000 -1.24 0.000 -1.26 0.000 -1.28 0.000 -1.27 0.000
Λ
(0.43) (0.12) (0.12) (0.12) (0.12) (0.12)

P- P- P- P- P- P-
F-stat F-stat F-stat F-stat F-stat F-stat
value value value value value value

time
17.87 0.000 17.87 0.000 18.50 0.000 19.23 0.000 19.16 0.000 18.49 0.000
dummies

R-squared 0.94 0.94 0.94 0.94 0.94 0.94

Number of
648 648 648 648 648 648
obs.

Table 13: Preferred model after trimming countries with too few observations.

43
The effect of UK sales in the estimated elasticity of demand

Around 50% of NPL sales take place within the UK. This could have a significant effect on the
price elasticity estimation results. Hence, the preferred two-step Heckman model has been
run with and without (model D1) the observations of the UK. The results are given by Table
14:

Model B1 Model D1
DV = log I
Coeff. P-value Coeff. P-value

1.24 0.000 1.15 0.000

log PL
(0.08) (0.08)

1.33 0.000 1.23 0.000

log GDP
(0.08) (0.08)

-0.02 0.000 -0.01 0.000

CMC
(0.00) (0.00)

-0.48 0.000 -0.36 0.000

log D
(0.02) (0.02)

-0.17 0.000 -0.30 0.000

log POP
(0.03) (0.03)

0.35 0.000 0.34 0.000

u (t-1)
(0.05) (0.05)

0.33 0.000 0.33 0.000

u (t-2)
(0.04) (0.04)

0.22 0.000 0.21 0.000

u (t-3)
(0.04) (0.04)

1.28 0.003 -1.13 0.000

λ
(0.43) (0.12)

F-stat P-value F-stat P-value

44
 time dummies 17.87 0.000 14.31 0.000

R-squared 0.94 0.93

Number of
648 634
obs.

Table 14: Preferred model with and without UK observations.

As seen, the effect of not considering the UK observations is not significant (around a 10%
variation).

Outliers in the dependent variable

Figure 1 in section 3 shows a box plot of the dependent variable. Two outliers lay beyond
Tukey’s fences (represented by the whiskers); they are not far from the rest of the observations
though. This suggest that they should not have a significant misleading effect on the estimation
of the elasticity of demand. To check that, the preferred model has been tested with and
without these two observations (see Table 15).

Model B1 Model D1
DV = log I
Coeff. P-value Coeff. P-value

1.24 0.000 1.15 0.000

log PL
(0.08) (0.08)

1.33 0.000 1.23 0.000

log GDP
(0.08) (0.08)

-0.02 0.000 -0.01 0.000

CMC
(0.00) (0.00)

-0.48 0.000 -0.36 0.000

log D
(0.02) (0.02)

-0.17 0.000 -0.30 0.000

log POP
(0.03) (0.03)

0.35 0.000 0.34 0.000

u (t-1)
(0.05) (0.05)

45
 0.33 0.000 0.33 0.000
u (t-2)
(0.04) (0.04)

0.22 0.000 0.21 0.000

u (t-3)
(0.04) (0.04)

1.28 0.003 -1.13 0.000

λ
(0.43) (0.12)

F-stat P-value F-stat P-value

time dummies 17.87 0.000 14.31 0.000

R-squared 0.94 0.93

Number of
648 634
obs.

Table 15: Preferred model with and without outliers.

As expected, the estimation results are almost equivalent.

Specification of the CMC variable

This subsection addresses the two methodological issues around the variable that controls for
the capabilities of the local NMI:
• On the one hand, there is the question of which one of the two variables constructed
using the table reported by BIPM to use; either the probabilistic measure of the NMI
being able to fulfil the user’s needs (based on the total number of services provided),
or the geometric mean of the number of capabilities in different areas of metrology,
that tends to favour those NMIs with more variety in their portfolios.
• On the other hand, once the appropriate CMC variable is selected, we have to decide
whether to include this in levels or in logarithmic form in our model.
Therefore, firstly we test which one of the two alternative CMC measures is more suitable for
our model. To that end, both have been used to estimate the national price level elasticity of
demand through the preferred Heckman corrected setup – both models yield very similar
results. This is shown by Table 16 which compares the preferred model B1 that uses
probabilistic CMC variable, against model F1 which includes instead the geometric mean CMC
variable.

Model B1 Model F1
DV = log I
Coeff. P-value Coeff. P-value

46
 1.24 0.000 1.22 0.000
log PL
(0.09) (0.08)

1.33 0.000 1.28 0.000

log GDP
(0.09) (0.08)

-0.02 0.000 -0.01 0.000

CMC
(0.04) (0.00)

-0.48 0.000 -0.49 0.000

log D
(0.02) (0.02)

-0.17 0.000 -0.22 0.000

log POP
(0.03) (0.03)

0.35 0.000 0.35 0.000

u (t-1)
(0.05) (0.05)

0.33 0.000 0.33 0.000

u (t-2)
(0.05) (0.04)

0.22 0.000 0.22 0.000

u (t-3)
(0.05) (0.04)

0.28 0.136 0.25 0.047

λ
(0.18) (0.14)

F-stat P-value F-stat P-value

time dummies 26.69 0.000 17.71 0.000

R-squared 0.94 0.94

Number of
564 648
obs.

Table 16: Preferred model using the two different definitions of the CMC variable.

The fact that both variables produce almost equivalent results reflects that those NMIs that
have a more comprehensive portfolio (greater number of services available to the user) are

47
also those that have a more complete portfolio (well distributed in all the areas of metrology).
This was to be expected, because although some NMIs decide to focus on certain areas of
metrology, it is usual to develop the measurement national infrastructure in all areas equally.
In any case, the probabilistic measure of the CMC has been taken as a reference in the main
analysis, since its interpretation is more immediate.
Once the CMC variable has been chosen, we need to address the question of whether the
CMC variable should be introduced in the model in levels or logged. This affects the estimation
results of the national price level elasticity substantially. Hence, it is essential to determine
whether our specification in levels is appropriate or not. For that matter, once again, the
preferred Heckman-corrected setup (model B1) is compared to an equivalent specification
which only differs in that the CMC regressor is logged instead of in levels (model G1). In
principal, running a link test should be a convenient way to detect any issue on the functional
relation between the dependent variable and the regressor. However, as shown by Table 17,
both specifications pass the test satisfactorily.

48
 Model B1 Model G1
.DV = log I
Coeff. P-value Coeff. P-value

predicted
1.03 0.000 1.03 0.000
values
(0.11) (0.08)

squared
0.00 0.768 0.00 0.784
prediction
(0.00) (0.07)

R-squared 0.94 0.94

Number of
648 648
obs.

Table 17: Link test for the preferred model using the two different definitions of the CMC variable

Hence, another complementary approach to decide whether the CMC variable should appear
in levels or in logarithmic form is needed. Following Davidson and MacKinnon we can test for
this by verifying if the fitted values of the alternative model have any explanatory power in the
original model. Hence, if our choice of including the CMC variable in levels is correct, then the
fitted values of the model that considers this variable in logarithmic form should be insignificant
when added to our model. Table 18 shows the results for this test.

Model B1 Model G1
DV = log I
Coeff. P-value Coeff. P-value

-0.05 0.772 0.80 0.054

log PL
(0.77) (0.41)

0.50 0.470 1.19 0.032

log GDP
(0.47) (0.56)

-0.07 0.548 -0.02 0.033

CMC
(0.55) (0.01)

-0.06 0.646 -0.35 0.040

log D
(0.65) (0.17)

49
 -0.05 0.632 -0.12 0.119
log POP
(0.63) (0.08)

0.09 0.496 0.28 0.063

u (t-1)
(0.50) (0.15)

0.00 0.996 0.21 0.068

u (t-2)
(1.00) (0.11)

0.01 0.908 0.14 0.079

u (t-3)
(0.91) (0.08)

0.14 0.469 0.35 0.110

λ
(0.47) (0.22)

predicted for 0.81 0.023 0.23 0.534

alternative
model (0.36) (0.37)

F-stat P-value F-stat P-value

time dummies 0.28 0.995 0.37 0.979

R-squared 0.94 0.94

Number of
564 564
obs.

Table 18: Davidson-MacKinnon test.

The results are highly clarifying. The fitted values of the alternative model are not significant
when the CMC variable is included in the model in levels. Conversely, when the CMC variable
is entered in logarithmic form, none of the variables of interest are significant and the predicted
values of the alternative are highly significant. This provides assurance that our chosen
specification is adequate.

Negative binomial regression

Although the gravity model of trade has been widely used traditionally to predict and explain
trade flows across countries, there is no consensus about the optimal method to solve the
existence of zero flows. Especially when analysing trade patterns at the product level, the
possibility that two countries do not trade in that specific good is much higher, and therefore
the problem of the existence of zeros in the dataset is more acute.

50
Count models are an alternative approach to the Heckman selection model used in this paper.
In these models the dependent variable is introduced in levels rather than logged. The
negative binomial regression is a generalisation of the Poisson regression model which
loosens the restrictive assumption that the variance is equal to the mean. We can apply this
model to our dataset given that the dependent variable in levels shows significant
overdispersion. Table 19 compares the estimation results for the static specification using OLS
and the negative binomial regression25.

Model A1 Model H1
DV = log I P-
Coeff. P-value Coeff.
value

1.31 0.000 1.25 0.000

log PL
(0.10) (0.13)

1.15 0.000 1.28 0.000

log GDP
(0.08) (0.13)

-0.02 0.000 -0.05 0.000

CMC
(0.00) (0.00)

-0.44 0.000 -0.56 0.000

log D
(0.03) (0.04)

-0.20 0.000 0.33 0.000

log POP
(0.04) (0.03)

P-
F-stat P-value Chi2
value

time dummies 7.13 0.000 112.70 0.000

Pseudo R-
0.81 0.10
squared

Number of obs. 939 1377

Table 19: Negative binomial regression estimation results.

25 Note that neither dynamics of the system (i.e. including past realisations of the error term) nor the
excess of zeroes (by considering for example a zero-inflated negative binomial regression model) are
accounted by this model. The purpose of model H1 is to serve as a robustness check that the estimation
procedure does not have a significant impact on the estimation results, and not as a complete
alternative approach to the Heckman selection model used in the main body of the paper.

51
We can see how the price elasticity found is quite similar to the ones found by the models
tested in the main body of this paper. This provides further assurance that our empirical results
are sound.

Split sample test

The macroeconomic and commercial conditions that govern the exchange of goods at the
international level today are not the same as in 2001. Likewise, the market for high precision
calibration services may have undergo significant changes during the last two decades. Both
factors may influence NPL’s sales in other countries and in the UK. Indeed, it could be that
the estimated price elasticity differs substantially from the actual one at present. In other
words, our analysis comes across an important trade-off. As more years are added to the
sample, we have more variability that allows us to find reliable estimates. However, it could
be that the initial years had little to do with the current situation.
For this reason, a simple robustness test is presented below. It consists of the static
specification (model A1) for a subsample from 2010 onwards. We estimate the static model
instead of the preferred Heckman corrected one because as soon as the dynamics of the
system are taken into account, we inevitably lose observations; this fact plus the reduction in
the sample will leave us with too few observations to play with.
Therefore, the objective of this robustness test is to compare the estimates from both static
models and verify that both estimates are comparable. Table 20 shows the estimation results
for the full sample static model (model A1) and the restricted sample static model (model I1).

Model A1 Model I1
DV = log I P- P-
Coeff. Coeff.
value value

1.31 0.000 1.23 0.000

log PL
(0.10) (0.14)

1.15 0.000 1.27 0.000

log GDP
(0.08) (0.12)

-0.02 0.000 -0.02 0.000

CMC
(0.00) (0.00)

-0.44 0.000 -0.47 0.000

log D
(0.03) (0.04)

-0.20 0.000 -0.24 0.000

log POP
(0.04) (0.05)

52
 P- P-
F-stat F-stat
value value

time
7.13 0.000 1.60 0.121
dummies

R-squared 0.81 0.80

Number of
939 524
obs.

Table 20: Split test for the static model.

It can be seen how both estimates for the coefficient of the national price level regressor are
very similar and within the confidence interval of the other. On the other hand, the year
dummies are not jointly significant (and in fact individually), which could show that the years
in the second half of the sample (after the huge shock of 2008 and 2009) is more comparable
to each other.
In any case, this simple check points to the validity of the estimates obtained.

Income as the dependent variable

A fundamental aspect of the econometric analysis is the choice of the dependent variable. We
have two possible variables to use as our dependent variable: the number of invoices and the
income generated (both normalised by population to ensure comparability across countries).
The objective of this robustness test is to check the impact on the estimation results of using
income instead of the preferred dependent variable, the number of invoices. To that end, the
following specification is tested:

log 𝑅𝑖,𝑡 = 𝛽0 + 𝛽1 log 𝑃𝐿𝑖,𝑡 + 𝛽2 log 𝐺𝐷𝑃𝑖,𝑡 + 𝛽3 𝐶𝑀𝐶𝑖 + 𝛽4 log 𝐷𝑖

+ 𝛽5 log 𝑃𝑂𝑃𝑖 + 𝛽5 log 𝐴𝑣𝑃𝑟𝑖𝑐𝑒𝑖,𝑡 + 𝛾𝑡 𝑌𝑡 + 𝜀𝑖,𝑡

where 𝑅𝑖,𝑡 is income generated from customers of country 𝑖 in year 𝑡 normalised by the
population of the country, 𝑃𝐿𝑖,𝑡 is the national price level, 𝐺𝐷𝑃𝑖,𝑡 is the gross domestic product
per capita, 𝐶𝑀𝐶𝑖 is a continuous index that rates the measurement capabilities of the country,
𝐷𝑖 is the distance between the country and the UK, 𝑃𝑂𝑃𝑖 is the population of the country,
𝐴𝑣𝑃𝑟𝑖𝑐𝑒𝑖,𝑡 is the average income per invoice, 𝑌𝑡 is a year dummy variable, and 𝜀𝑖,𝑡 is the error
term which is assumed to be independently and identically distributed as a normal distribution
with mean of 0 and finite variance.
Note that this specification differs in two ways from the one considered in the main body of the
document: (1) we are using income as the dependent variable instead of the number of
invoices, and (2) we are including the average price per invoice as a regressor. The reason
for including this other regressor is to get the 𝑄 component in income; i.e. given that 𝑅 = 𝑃 ∙ 𝑄,
we know that:

log 𝑅𝑖,𝑡 = log 𝑃𝑡 + log 𝑄𝑖,𝑡

53
Hence, by adding the average price per invoice to the regressors we can approximate log 𝑄𝑖,𝑡 .
We can compare the estimation results for the static version of the model both using the
number of invoices (model A1) and income (model J1)

Model A1 Model J1
DV = log I P- P-
Coeff. Coeff.
value value

1.31 0.000 1.23 0.000

log PL
(0.10) (0.10)

1.15 0.000 1.09 0.000

log GDP
(0.08) (0.08)

-0.02 0.000 -0.01 0.000

CMC
(0.00) (0.00)

-0.44 0.000 -0.36 0.000

log D
(0.03) (0.03)

-0.20 0.000 0.70 0.000

log POP
(0.04) (0.03)

. . 1.01 0.000
log AvPrice
. (0.04)

P- P-
F-stat F-stat
value value

time
7.13 0.000 7.04 0.000
dummies

R-squared 0.81 0.80

Number of
939 909
obs.

Table 21: Static model using both available dependent variables.

Therefore, it is found that both estimates are very similar for all regressors of interest.

54
Annex B: Testing for endogeneity.

This annex discusses the alternative to the Hausman test proposed by . This method is used
in section 4.3 to assess the whether the regressor of interest in our model, the price level, is
correlated with the error term. If so, any the estimation results would be biased.
Firstly, the computation of the test proposed by Mundlak is presented. Then the intuition
behind it is described.
In panel data analysis, the decision of using the fixed effects estimator or the random effects
estimator depends on how time-invariant unobservables are related to the variables in the
model. To assess this, Mundlak’s test follows three steps:
1. The panel-level average of the time-varying covariates are computed.
2. The random effects estimator is used to regress the covariates and the panel-level
means generated in the first step against the dependent variable.
3. Test whether the panel-level means are jointly zero.
If the test rejects the null hypothesis that the coefficients are jointly zero, it suggests that there
is correlation between the time-invariant unobservables and the regressors, namely, the fixed-
effects assumptions are satisfied. On the contrary, If the test does not reject the null hypothesis
that the generated panel-level mean regressors are zero, there is evidence of no correlation
between the time-invariant unobservable and the regressors; that is, the random effects
assumptions are satisfied.
The intuition behind Mundlak’s approach is straightforward. Suppose a linear panel-data
model given by equation B.1:

𝑦𝑖𝑡 = 𝛼𝑖 + 𝑥𝑖𝑡 𝛽 + 𝜀𝑖𝑡 (B.1)

Where the index 𝑖 denotes the panel unit and the index 𝑡 time. 𝑦𝑖𝑡 is the outcome of interest,
𝑥𝑖𝑡 is the set of regressors, 𝜀𝑖𝑡 is the time-varying unobservable (idiosyncratic error), and 𝛼𝑖 is
the time-invariant unobservable (unobserved heterogeneity).
Now, the mean of 𝛼𝑖 conditional on the time-invariant part of the regressors is given by
expression B.2:

𝛼𝑖 = 𝑥̅𝑖 𝛾 + 𝜈𝑖 ⇔ 𝐄(𝛼𝑖 |𝑥𝑖 ) = 𝑥̅𝑖 𝛾 (B.2)

Where 𝑥̅𝑖 is the panel-level mean of 𝑥𝑖𝑡 , and 𝜈𝑖 is a time-invariant unobservable that is
uncorrelated to the regressors.
Hence, if 𝛾 = 0, 𝛼𝑖 and the covariates are uncorrelated. This is precisely what the Mundlak
method tests. This can be seen by substituting B.2 into B.1:

𝑦𝑖𝑡 = 𝛼𝑖 + 𝑥𝑖𝑡 𝛽 + 𝜀𝑖𝑡 = 𝑥̅𝑖 𝛾 + 𝜈𝑖 + 𝑥𝑖𝑡 𝛽 + 𝜀𝑖𝑡 ⇔ 𝐄(𝑦𝑖𝑡 |𝑥𝑖𝑡 ) = 𝑥𝑖𝑡 𝛽 + 𝑥̅𝑖 𝛾 (B.3)

55
Where the last relation relies on the fact that the regressors and the unobservables are mean
independent. Therefore, Mundlak’s approach tests the null hypothesis of 𝛾 = 0.

56
Annex C: Detailed summary statistics

Table 22 provides detailed overall, between and within summary statistics for all the variables
used in the analysis. Variables in level are abbreviated in capital letters; variables in
logarithmic form are lowercased.
Variable Variation Mean Std. Dev Min Max Obs.

overall 8.0E-06 0.0 8.6E-05 N =1717

I between 2.4E-06 6.9E-06 7.5E-10 5.8E-05 n =101

within 4.2E-06 -3.9E-05 6.6E-05 T =17

overall 0.3 0.1 1.4 N =1649

PL between 0.5 0.3 0.2 1.2 n =97

within 0.1 0.2 1.0 T =17

overall 1.4E+04 4.1E+02 9.7E+04 N =1649

GDP between 1.5E+04 1.3E+04 6.2E+02 7.0E+04 n =97

within 4.9E+03 -2.3E+03 4.7E+04 T =17

overall 16.7 0 77.6 N =1394

CMC between 11.6 16.8 0 77.6 n =82

within 0.0 11.6 11.6 T =17

overall 3933.0 185.8 19147.1 N =1683

D between 5167.5 3951.9 185.8 19147.1 n =99

within 0 5167.5 5167.5 T =17

overall 1.8E+08 3.3E+04 1.3E+09 N =1717

POP between 5.9E+07 1.8E+08 3.3E+04 1.3E+09 n =101

within 0 5.9E+07 5.9E+07 T =17

overall 2.1 -20.3 -9.4 N =1025

i between -14.1 2.1 -18.7 -9.8 n =101

within 0.6 -16.7 -11.4 T-bar =10.15

overall 0.5 -2.1 0.4 N =1649

pl between -0.7 0.5 -1.6 0.2 n =97

within 0.2 -1.7 0.0 T =17

overall 1.1 6.0 11.5 N =1649

gdp between 9.2 1.0 6.4 11.1 n =97

within 0.3 8.3 10.2 T =17

overall 1.0 5.2 9.9 N =1683

d 8.2
between 1.0 5.2 9.9 n =99

57
 Within 0 8.2 8.2 T =17

overall 1.9 10.4 21.0 N =1717

pop between 16.3 1.9 10.4 21.0 n =101

within 0 16.3 16.3 T =17

Table 23: Summary statistics.

58
Annex D: Stata outputs

Model A1

Estimation results

59
Wald test year dummies

Link test

Ramsey RESET test

60
Serial correlation test

Normality of the residuals

61
Mundlak test

62
Model A2

Estimation results

63
Wald test year dummies

Link test

Ramsey RESET test

64
Serial correlation test

Normality of the residuals

65
Mundlak test

66
Model A3

Estimation results

67
Wald test year dummies

Link test

Ramsey RESET test

68
Serial correlation test

Normality of the residuals

69
Mundlak test

70
Model B1

Probit model estimations results

71
Partial effects at the average

72
Average partial effects

73
Estimation results

74
Wald test year dummies

Link test

Ramsey RESET test

75
Serial correlation test

Normality of the residuals

76
Mundlak test

77
Model C1

Probit model estimations results

78
Estimation results

Wald test year dummy variables

79
Model C2

Probit model estimations results

80
Estimation results

Wald test year dummy variables

81
Model C3

Probit model estimations results

82
Estimation results

Wald test year dummy variables

83
Model C4

Probit model estimations results

84
Estimation results

Wald test year dummy variables

85
Model C5

Probit model estimations results

86
Estimation results

Wald test year dummy variables

87
Model D1

Probit model estimations results

88
Estimation results

Wald test year dummy variables

89
Model E1

Probit model estimations results

90
Estimation results

Wald test year dummy variables

91
Model F1

Probit model estimations results

92
Estimation results

Wald test year dummy variables

93
Link test

94
Model G1

Probit model estimations results

95
Estimation results

Wald test year dummy variables

96
Link test

Davidson and McKinnon test

97
98
99
Model H1

100
Model I1

101
Model J1

102
More information

Contact us

For more information, contact us:

Senior Economist Mike King , mike.king@npl.co.uk

Economist Eugenio Renedo, eugenio.renedo@npl.co.uk

National Physical Laboratory

Hampton Road, Teddington, Middlesex, TW11 0LW
Web: https://www.npl.co.uk/digital
Tel: 020 8977 3222

103