Subsidies and Time Discounting in New Technology

Adoption:
Evidence from Solar Photovoltaic Systems
Olivier De Groote and Frank Verboven

November 2018

Abstract

We study a generous program to promote the adoption of solar photovoltaic (PV)

systems through subsidies on future electricity production, rather than through upfront
investment subsidies. We develop a tractable dynamic model of new technology adop-
tion, also accounting for local market heterogeneity. We identify the discount factor
from demand responses to variation that shifts expected future but not current util-
ities. Despite the massive adoption, we …nd that households signi…cantly discounted
the future bene…ts from the new technology. This implies that an upfront investment
subsidy program would have promoted the technology at a much lower budgetary cost.
(JEL C51, Q48, Q58)

De Groote: Toulouse School of Economics, University of Toulouse Capitole, Toulouse, France,

olivier.de-groote@tse-fr.eu. Verboven: KU Leuven, Leuven, Belgium and CEPR, London, UK,
frank.verboven@kuleuven.be. Acknowledgements: we would like to thank the Editor and three anonymous
referees for helpful comments, and are also grateful to Peter Arcidiacono, Estelle Cantillon, Allan Collard-
Wexler, Tobias Klein, Ashley Langer, Fabio Miessi, Guido Pepermans, Stef Proost, Mathias Reynaert, Paul
Scott, Yutec Sun, Christopher Timmins, Jo Van Biesebroeck, Frederic Vermeulen and participants at LCM
(Riberão Preto), EEA (Mannheim), ECORES (Brussels), IIOC (Philadelphia), EARIE (Maastricht) and
seminars at Duke University, NYU Stern/Columbia, Tilburg University, University of Düsseldorf, University
of Rome Tor Vergata, Université Saint-Louis, University of Antwerp and McGill/HEC Montréal. We also
thank Guido Pepermans for his help in obtaining the PV data and Iris Grant for excellent research assistance.
Finally, we thank KU Leuven Program Financing and FWO for …nancial support.
1 Introduction
Many countries have relied on subsidies to promote the adoption of renewable energy tech-
nologies for electricity production, such as wind power and solar photovoltaic (PV) systems.
The generous support has often been motivated on the grounds that there is not only an
environmental externality (CO2 emissions from fossil sources), but also a technology market
failure (insu¢ cient incentives to innovate and adopt a new technology). The subsidies for
the green technologies often consist of a combination of investment subsidies, which are paid
upfront at the moment of installation, and production subsidies, which are paid in the future
when the systems are producing the electricity (or equivalently, a combination of investment
and production tax credits, as reviewed for the U.S. in Murray et al. (2014)).
In this paper we investigate the incentives to adopt a new green technology, and the
role played by upfront investment and future production subsidies. The adoption decision
involves a fundamental trade-o¤ between the immediate investment costs and the future
bene…ts from electricity production. The successful adoption of the new technology thus
depends on how much households discount future bene…ts, and on the extent to which
subsidies apply to the upfront investment costs or the future electricity production. We
study a generous program for residential solar PV systems, running in Belgium during 2006–
2012, and responsible for a particularly high adoption rate compared with other countries.1
The program relied heavily on future production subsidies in the form of Green Current
Certi…cates (GCCs), which were committed for up to 20 years. The program was similar
to the German feed-in tari¤ system and several other European programs. U.S. programs
also involve production subsidies, but tend to rely more heavily relied on upfront investment
subsidies.2 Interestingly, the GCC subsidy program revised its conditions many times at pre-
announced dates. These revisions typically consisted of reductions in the future production
subsidies, and applied only to new adopting households. The considerable variation in the
future subsidies enables us to identify the households’implicit discount factor in a reliable
way. Furthermore, because the program mainly consisted of future production subsidies
instead of upfront investment subsidies, it potentially enabled the government to shift the
…nancial burden to future electricity consumers. Based on the estimated discount factor, we
can assess how costly this was.
1
Belgium ranked third in the European Union with a total capacity of 240 Watt peak/capita at the end
of 2012 (Eurobserv’er 2013).
2
In the U.S. there were federal tax credits of 30%, and several states took additional measures. For
example, the famous California Solar Initiative (CSI) had a budget of $2.2 billion and aimed to install
1.9GW of solar PV capacity. Combined with the federal tax credits, the investment subsidies could amount
to 50% of the cost of a solar PV system. Source: https://en.wikipedia.org/wiki/California_Solar_Initiative.

1
 To estimate how households discount the future bene…ts of a new technology, we develop a
dynamic discrete choice model, where in each period households face the decision to adopt the
new technology or to postpone their investment. We …rst develop a model to be estimated
with aggregate, country-level data. Next, to evaluate the robustness of our …ndings, we
extend the model to account for rich forms of persistent observed and unobserved local
market heterogeneity in a tractable way. As discussed further below, our approach does
not require specifying an explicit stochastic process for the expected state transitions, which
would be particularly di¢ cult for a new technology.
We obtain the following main …ndings. First, although the program led to a massive
adoption of solar PV systems, households signi…cantly discounted the future bene…ts from
the new technology. They use an implicit real interest rate of 15% in evaluating these future
bene…ts, which is much above the real market interest rate of about 3%. Put di¤erently,
this implies a considerable undervaluation of the future bene…ts from electricity production:
consumers are willing to pay only approximately 0:5 euro upfront for one euro of discounted
future bene…ts from electricity production. Our …nding of considerable time discounting is
robust with respect to various assumptions about households’expectations on the value of
current and future PV systems. It can either be interpreted as intrinsic consumer myopia
or as mistrust in the government’s commitment to pay out the future subsidies. This raises
speci…c policy concerns, at least from a budgetary and distributional perspective. Upfront
subsidies instead of future production subsidies would have reduced public expenditures by
e1:9 billion (or 51% of the amount spent). This is a saving of more than e700 per household,
a very large number given that only 8:3% of the households had adopted a PV at the end of
the program. We conclude that there was a high public cost in shifting the subsidy burden
to future households, as they pay for the subsidy through higher electricity prices.
Our paper makes several contributions. First, we contribute to the empirical intertem-
poral choice literature, which studies how consumers value future payo¤s. Much of this work
focuses on the important question whether consumers undervalue future energy cost savings,
as this could be responsible for the so-called energy e¢ ciency gap (for overviews, see All-
cott and Greenstone (2012) and Gerarden et al. (2017)). After Hausman’s (1979) seminal
contribution, the recent evidence ranges from moderate undervaluation to correct valuation,
see for example Allcott and Wozny (2014) and Busse, Knittel and Zettelmeyer (2013). All
this evidence is based on energy-saving investments of existing, mature technologies (such as
cars). Furthermore, this previous work has focused only on the decision how much to invest
in energy cost savings, and it ignored the timing dimension of adoption. This approach may
be reasonable for mature technologies where households simply replace their current prod-
ucts. However, it is unrealistic in new markets when new energy-saving technologies are just

2
introduced, when prices are quickly decreasing and quality is increasing. In these circum-
stances, consumers do not only face a traditional investment problem. They must also decide
on the timing of their investment, as it can be bene…cial to postpone adoption even if it is
already pro…table to invest now. Our paper …lls this gap and considers consumers’valuation
of future payo¤s (energy cost savings) when adopting an entirely new technology, which en-
tails both an investment problem and a timing problem (to take advantage of future changes
in production subsidies and investment costs). Our evidence suggests that time discounting
may be much stronger in this case, with important implications for policy programs.
To incorporate the timing decision, we develop a dynamic discrete choice model that
captures the optimal stopping problem in the spirit of Rust (1987). The discount factor
now plays a double role: it in‡uences both how much households value the future bene…ts
of their investments, and how much they are prepared to wait for better investment oppor-
tunities. The …rst is inherent in every investment decision, but does not necessitate the use
of a dynamic model as it can be treated as a static model with discounted bene…ts. The
second is particularly important for new technologies because they are often characterized
by increasing quality and decreasing prices. This aspect does require a dynamic model: post-
poning a bene…cial investment can be optimal, and a static model may underestimate the
sensitivity to monetary incentives (Gowrisankaran & Rysman 2012). The dynamic discrete
choice literature has stressed that the discount factor is nonparametrically unidenti…ed; see
Manski (1993), Rust (1994) and Magnac and Thesmar (2002). Magnac and Thesmar (2002)
and Abbring and Daljord (2017) show how identi…cation may be obtained through appro-
priate exclusion restrictions. Intuitively, the discount factor can be identi…ed from demand
responses to variation that shifts expected discounted future utilities but not the current
utilities.3 In our setting we obtain identi…cation from variation in the future bene…ts across
products and over time, relative to the investment costs which enter only the current utility
of adopting. In particular, we exploit large variation in the future GCC subsidies, which
were revised many times on pre-announced dates and implied a guaranteed stream of payo¤s
for a …xed number of years.
Second, we contribute by proposing a novel method to estimate a dynamic technology
adoption model with aggregate data, and we also show how to extend this model to account
for local market heterogeneity in a tractable way. The main advantage of our approach is that
it is not necessary to specify an explicit stochastic process for the expected state transitions
of the future investment costs and bene…ts. Specifying such a process would be particularly
di¢ cult for a new technology, in our setting especially because of the highly idiosyncratic,
3
Magnac and Thesmar (2002) impose an exclusion restriction on the current value functions. Abbring
and Daljord (2017) suggest to directly impose an exclusion restrictions on primitive utilities.

3
nonstationary nature of the PV subsidy scheme. Our aggregate adoption model amounts to
estimating an Euler equation with GMM. We start from Hotz and Miller’s (1993) inversion
approach, which writes the ex ante value function as the utility of choosing one alternative,
plus a correction term. We exploit the fact that technology adoption is a terminating action
in our setting (see Arcidiacono and Ellickson (2011) for a particularly clear exposition).
Similar to Scott (2013) we write the expected next period ex ante value function as the
realized value function plus a prediction error, which is uncorrelated with any variables
known by the household at the time of the adoption decision. We then show how to invert
the demand model to solve for the unobserved error term, using a similar approach as in
Berry (1994) for static choice models with aggregate data. Conditional on the discount factor,
this gives rise to a linear regression equation, where the current adoption rate depends on
current and next period prices, as well as the next period adoption rate. One can use a
standard nonlinear GMM estimator to also estimate the discount factor and account for the
endogeneity of several variables.
We subsequently suggest a modi…ed approach to account for rich forms of observed and
unobserved household heterogeneity at the local market level.4 In our setting these markets
are highly disaggregate (almost 10,000 local markets with on average only 295 households
per market). This implies an excessive number of zero (or very low) adoptions, which inhibits
us from inverting the demand model and obtaining a regression equation at the local level.
We suggest an approach to deal with this within a GMM framework: we combine aggregate,
country–level moments (where zero or low adoptions do not occur) with micro-moments at
the local market level to account for household heterogeneity.5 We include demographic
variables, interacted with price and capacity size, and a large set of local market …xed e¤ects
to control for unobserved heterogeneity that can be identi…ed from variation across markets.
Although household heterogeneity is important in explaining adoption behavior, it does not
a¤ect our conclusions for the discount factor, and our policy implications.
Third, our work relates to a recent literature using dynamic models to study the adop-
tion of PV systems and the role of government policies. Burr (2016) estimates a dynamic
4
Other dynamic adoption models with aggregate data have ignored persistent heterogeneity (Melnikov
2013), or allowed for it through random coe¢ cients (Gowrisankaran and Rysman (2012)) or unobservable
types in the population (Scott 2013).
5
Broadly speaking, the static discrete choice literature follows two approaches to deal with zero market
shares within a GMM framework. The …rst approach consists of aggregating and adding micro-moments, as
done in Quan and Williams’s (2018) nested logit with cross-market random e¤ects, Nurski and Verboven
(2016), and our application. An alternative approach, developed by Gandhi et al. (2017), consists of
constructing moment inequalities (starting from the Laplace rule of succession to obtain an initial choice
probability estimator that does not have zeros).

4
adoption model of PV systems for California, and also …nds that an upfront subsidy program
encourages adoption more than future production subsidies. Langer and Lemoine (2018) use
the California experience to study optimal dynamic subsidy paths, which may a priori be
increasing or decreasing (depending on the distribution of household valuations, time dis-
counting and the rate of technological progress). Finally, Feger et al. (2017) use data on
both PV adoption and electricity usage in the Canton of Bern to study optimal tari¤ design,
showing how upfront installation subsidies may be combined with variable and …xed elec-
tricity fees to obtain a given solar energy production target. These studies rely on speci…c
assumptions regarding time discounting, and they require a full model of the state space
with the stochastic process of expected future state transitions. In contrast, we focus on
identifying the discount factor based on rich variation in upfront investment costs and fu-
ture payo¤s. Furthermore, we rely on weak assumptions on the households’expected future
investment opportunities and account for endogeneity of investment costs. Finally, we show
how to incorporate rich forms of observed and unobserved household heterogeneity.6
The rest of the paper is structured as follows. Section 2 describes the datasets and insti-
tutional background. Section 3 speci…es the model that can be estimated with only aggregate
data, and also its extension to account for local market heterogeneity. Section 4 discusses
the empirical results, performs a detailed sensitivity analysis and derives policy implications.
Finally, we conclude in section 5.

2 Industry background
In this section we describe the market of residential photovoltaic (PV) systems. We begin
with a brief description of the available datasets. We then discuss the technology and the
various sources of costs and bene…ts of installing PV systems. Finally, we provide descriptive
statistics on the magnitude of the costs and bene…ts during the considered period, and on
the evolution of the number of adopters of the new technology.

2.1 Datasets
Our main dataset contains information of all installed PVs during 2006-2012 across the
region of Flanders, Northern part of Belgium covering about 60% of the total population.
6
The above papers and our own consider the impact of subsidy policies on the demand side dynamics.
Two papers have focused on the supply side dynamics. Bollinger and Gillingham (2014) consider the role of
learning by doing in solar panel installation, whereas Gerarden (2017) looks at the incentives to invest in
improved technical e¢ ciency of solar panels.

5
For each installed PV, we observe the time of adoption, the location and the capacity size
(but not the brand or other PV characteristics). We will analyze this dataset at the level
of …ve capacity size categories (0–2kW; 2–4kW; ... 8–10kW) at a monthly frequency. We
…rst consider the aggregate level of Flanders (covering about 2:7 million households) and in
an extension consider the disaggregate local market level (which divides the entire region in
9,182 statistical sectors, with an average of 295 households per statistical sector).
We combine the information from this main dataset with several additional datasets.
First, we collected information on the price quotes of 2,659 PV systems adopted during May
2009 until December 2012. Since we observe only the capacity and time at which PVs were
adopted and since the price quotes in any case mainly depend on the PVs’capacity and less
on the brand, we aggregate the price information to the median price for each month and
each of the …ve di¤erent capacity size categories.7
Second, we have information on the bene…ts from adopting PVs, including the public
support measures in the form of Green Current Certi…cates (GCCs), electricity cost savings
from net metering, and tax bene…ts. Finally, for our extension to the disaggregate local
market level, we collected detailed socio-demographic information, such as income, household
and house characteristics. In the Appendix we provide further details on the data sources
and the data construction.

2.2 Technology and public support measures

A PV system consists of solar panels, which absorb sunlight and convert this into electricity.
One can distinguish between residential and commercial PV systems. Residential PV systems
are usually installed on top of a roof and typically have a capacity size no larger than 10
kilowatt (kW). Commercial PV systems may also be on the top of a roof or they may be
grount-mounted, and they generally reach much larger capacity sizes than residential PV
systems.
Our focus is on residential PV systems, with capacity limited to 10 kW. In Flanders, a
PV system produces 0.85 MWh per year for each kW of capacity (CREG 2010). All resi-
dential PV systems are connected to the grid, so that households do not need to synchronize
their electricity consumption and production, or use batteries to store excess production.
Households pay an upfront investment price for a PV system, and they receive two main
sources of future bene…ts from installing a PV system: Green Current Certi…cates (GCCs)
and electricity bill savings from net-metering. We discuss these elements in turn.
7
An OLS regression of the price/kW on monthly time dummies and dummy variables for the di¤erent
capacity size categories results in an R2 of 72.3%.

6
Investment price The investment price is the price households have to pay for a PV
system, including all additional costs. As discussed above, we construct a price measure per
month for each of the …ve capacity size categories. In 2006 and 2007 households could apply
for a 10% investment subsidy for PV installations.8 Furthermore, there was a general tax
credit of 40% for renewable energy investments, including PV installations. The maximum
allowed tax credit varied over the period, ranging from e 1,200 in 2006 to e 3,600 in 2011
(and since 2009 households could transfer the remaining amount to the following three years
if their house was built at least …ve years ago). In 2012 the tax credits for PV installations
were abolished. Finally, PV installations that were built in houses of at least …ve years old
also bene…ted from a reduced VAT rate of 6% instead of 21%.

Subsidies from Green Current Certi…cates (GCCs) The Flemish government has
actively promoted the adoption of PV systems through the program of tradable GCCs.
Households obtained a GCC for each MWh of electricity production through their PV system,
and they could sell these to the distribution system operators (DSOs) at a guaranteed price
for a …xed number of years. This guaranteed price was substantially above the market price
of GCCs. At the start in 2006, the program was very generous, paying e450 per MWh for a
legally guaranteed period of 20 years. The program became less favorable to new adopters
in 2010, and it was subsequently gradually phased out. By the end of 2012, new PV adopters
received a guaranteed price of only e90 per MWh for a period of 10 years. In January 2013,
the government introduced a so-called banding factor. This restricted the number of GCCs
per MWh, and e¤ectively led to an abolishment of the entire GCC system in February 2014.9
From the point of view of PV adopters, the GCCs are a subsidy for future electricity
production. The DSOs were responsible to buy these GCCs at the contracted price. They
subsequently resell them at the prevailing market price to the electricity suppliers, who are
required to purchase a su¢ cient amount every year to meet their renewable energy sources
requirements. The GCCs are thus a cost to both the DSOs and the electricity suppliers, and
these costs are eventually passed on to retail electricity prices. As such, the GCC subsidy
scheme is not …nanced through taxes, but rather through increased electricity prices to all
consumers.
8
The subsidizable investment cost was capped at 7000e per kWp and a maximum subsidizable capacity
of 3kW.
9
The idea of the banding factor was to limit the number of GCCs for every produced MWh, in such a
way that the net present value of installing a PV would essentially be zero at the prevailing market prices of
PV systems. Since the prices of PV systems continued to drop, the net present value soon became positive
even without GCCs, so that GCCs were e¤ectively abolished in February 2014.

7
Electricity cost savings from net metering Households with a PV system with a
capacity limited to 10 kW bene…t from a net-metering principle. This means that they
have to pay only for their net annual electricity consumption, i.e. their consumption after
subtracting the annual electricity production generated by their PV system and transmitted
on the grid.10 Hence, in addition to the subsidies from GCCs, a second main source of
bene…ts from installing a PV system is given by the annual electricity bill savings, i.e. the
PV’s annual electricity production multiplied by the retail price of electricity.
Access to the grid was initially o¤ered without any charge. In July 2015, the DSOs were
able to introduce an annual grid fee of 92e/kW. This came after a long public debate and
several legislative procedures. The grid fee enabled the DSOs to partly …nance their cost of
the GCC subsidies, aiming to avoid further electricity price increases to all consumers.

2.3 Evolution of costs, bene…ts and adoption

Figure 1 summarizes of the costs and bene…ts of a PV system of 4kW.We calculate future
bene…ts in present value terms using a real interest rate of 3% and an expected life time
of 20 years and we convert all prices to 2013 prices. The gross purchase price (net of any
investment tax cuts) dropped from e 21,700 in May 2009 to e 8,800 at the end of 2012.11
The present value of future bene…ts was highest in 2009 and rapidly decreased afterwards.
The most important bene…ts came from the GCCs. They provided a present value of e
20,000 until January 2010, and subsequently declined until they almost disappeared at the
end of 2012. Bene…ts from tax cuts were also high, especially from 2009 on, but they
were removed in 2012. Finally, the bene…ts from net-metering (i.e. electricity cost savings)
formed a fairly stable source of bene…ts. These bene…ts became the most important reason to
adopt PVs since the end of 2012, but only because other bene…t components decreased over
time. A comparison of the total bene…ts (shaded area) with total costs (black line) shows
that adoption was pro…table during the entire period in net present value terms, especially
between 2009 and the middle of 2012.
Figure 2 shows the evolution of the monthly number of new adopters between January
2006 and December 2012. Vertical lines indicate drops in the GCC prices, as typically
announced a few months in advance. Despite the positive gap between bene…ts and costs
throughout the sample, the number of new adopters remained very low until 2009. This
10
Note that there is no reimbursement in case a household would produce more electricity than it consumes
on an annual basis.
11
The price data we collected starts in May 2009. We therefore also estimate the model from May 2009
on. For descriptive purposes, we also show a predicted price variable in Figure 1 (based on the German price
index).

8
Figure 1: Costs and bene…ts of 4kW PV in EUR 2013, discounted at market interest rate

10000 20000 30000 40000 50000

EUR
0

2006 2007 2008 2009 2010 2011 2012 2013

Net metering Tax cut

Subsidy GCC
Upfront investment cost*
Upfront investment cost*: predicted values using German price index EUPD before May 2009
Real interest rate used to calculate present values = 3%

may be because households did not fully value the bene…ts or because they postponed their
adoption in anticipation of better future investment opportunities. From 2009 onwards
the number of new adopters started to increase to reach a sharp peak just before the …rst
announced drop in the GCC price in January 2010. There was again a gradual increase in
the number of adopters in 2010 with a new peak just before the second drop in the GCC
price in January 2011. The same pattern of gradual increases and peaks just before a next
announced drop in the GCC price has been repeated several times until the beginning of
2013 when the GCC policy changed drastically and became less generous. This adoption
pattern illustrates the dynamic nature of the households’decision problem to adopt a PV
installation. Households postpone the adoption of a PV to wait for prices to drop, but they
also anticipate the announced drop in the GCC price and thus in the expected bene…ts of
their investment.
Figure 3 shows the cumulative number of adopters over the considered period, broken
down into our …ve groups of capacity size: 2kW, 4kW, 6kW, 8kW and 10kW. This shows
a gradual long-term increase in the number of adopters, with several kink points around

9
 Figure 2: 2006-2012: Time series of new PV adoptions and drops in nominal GCC price

20000
New PV adoptions (monthly)
5000 100000 15000

2006 2007 2008 2009 2010 2011 2012 2013

Drops in nominal GCC price*

* GCC price corrected for banding factor when installed after 2012

the time of new GCC schemes. The 4kW and 6kW systems were the most popular choices
for a PV. This is because households bene…t from net-metering only for the production
that is below their household consumption. In practice, an average household consumes
3.5MWh per year, while a 4kW system produces about 3.4 MWh per year, so that net
metering bene…ts of larger PV systems are of value only to households that are su¢ ciently
larger than average. Nevertheless, there is a shift towards PV systems of larger capacity:
whereas in January 2010 the market share of PV systems of 8kW and 10kW was only 12%,
it reached 18% by 2013. This shift could be a result of time discounting (when both the
upfront investment costs and the GCC subsidies drop). Our model will therefore take into
account the choice of capacity size.
By the end of 2012, the cumulative number of adopters had reached 220; 464, amounting
to an adoption rate of 8:3% of the households (or 8:4% of the number of buildings). The
total capacity of residential PV systems had at that time reached 1; 057MW, or 5% of total
electricity capacity in Belgium.12
Adoption rates vary widely within the region, as illustrated in Figure 4. Adoption rates
are very high (over 20%) in rural areas often in the west and east parts of the region.
12
According to the US Energy Information Administration, Belgium had a total installed electrical capacity
of 21,000 MW in 2012.

10
 Figure 3: 2006-2012: Time series of total adoption of PVs of di¤erent capacity
100000
40000 60000 80000
Total PV adoptions
20000
0

2006 2007 2008 2009 2010 2011 2012 2013

2kW 4kW 6kW 8kW 10kW

Conversely, adoption rates are extremely small in cities such as Ghent (west of center) and
Antwerp (north of center), or the areas around Brussels (south of center). Various socio-
demographic factors may explain this variation, such as average household size, house size
and income. In an extension of our aggregate demand model, we will take into account the
role of these socio-demographic characteristics.

Figure 4: PV adoption rates in Flanders

Adoption data: VREG, household data: ADSEI census 2011

In sum, this overview shows there is considerable variation over time in the adoption of
PV systems, and this variation appears to be related to the variation in investment costs
and the future bene…ts, in particular the pre-announced changes in GCC policies.

11
3 The model of technology adoption
We …rst specify a dynamic adoption model that can be estimated with aggregate market data
and no household heterogeneity (apart from an i.i.d. taste shock): we describe the adoption
decision (subsection 3.1), derive the estimating equation (subsection 3.2) and discuss estima-
tion and identi…cation (subsection 3.3). We subsequently show how to extend the approach
to estimate the model at a highly disaggregate local market level. This makes it possible to
account for both observed and unobserved heterogeneity across households (subsection 3.4).
Note that we will not need to specify whether the adoption decision is a …nite or in…nite
horizon problem. Nor will we need to explicitly specify how consumers expect the states to
evolve in the future. This is because we can estimate the parameters of the model without
having to solve it, using Hotz and Miller’s (1993) CCP approach with …nite dependence, and
because we need to assume only rational expectations on state transitions by modeling the
expected ex ante value function as the realized ex ante value function plus a prediction error
as in Scott (2013).

3.1 The adoption decision

In a given period t a household i may either choose not to adopt a PV, j = 0, or it may
choose to adopt one of the available PV alternatives, j = 1; : : : ; J. In our application, the
PV alternatives refer to systems with di¤erent capacity sizes. A key feature of the model is
that the adoption decision (j 6= 0) is a terminating action.13 Not adopting (j = 0) gives the
option of adopting at a later period, when the price for a given size may have decreased, or
when the …nancial bene…ts may have increased or decreased.
In each period a household obtains a random taste shock "i;j;t , which we assume to follow
a type I extreme value distribution. Let vi;j;t be the conditional value of household i for
alternative j at period t, i.e. the expected discounted utility from choosing j at t before the
realization of the random taste shock "i;j;t . In general, one can decompose vi;j;t = j;t + i;j;t ,
where j;t is the mean utility and i;j;t is the individual-speci…c utility. In this and the
next two subsections, we set i;j;t = 0, so that vi;j;t = j;t . This implies that there is no
household heterogeneity except for the extreme value distributed taste shocks "i;j;t . This
leads to a particularly easy to interpret and tractable estimating equation. The downside of
this approach is that heterogeneity is then assumed to be uncorrelated over time and over
alternatives. In practice, households with a high valuation for adopting now are also likely
to have a high valuation for adopting in future periods. Furthermore, households may have
13
Replacement is not a viable option because households would lose the generous GCC subsidies. In
practice, replacement is indeed extremely rare.

12
correlated preferences for systems with similar capacity sizes. In the …nal subsection 3.4,
we overcome this by allowing for both observed and unobserved heterogeneity at the local
market level in i;j;t .
Assume that in each period t households choose the alternative j that maximizes random
utility vi;j;t + "i;j;t . This will give rise to a choice probability, or approximately an aggregate
market share, for each alternative j in each period t. Before deriving this, we …rst describe the
conditional value of adoption (vi;j;t , j = 1; : : : ; J) and the conditional value of not adopting
(vi;0;t ) in period t.

Conditional value of adoption (vi;j;t , j = 1; : : : ; J)

The conditional value of adoption is particularly simple. Since this is a terminating action,
we can write it as the expected discounted utility of adoption. We specify vi;j;t = j;t as
follows:
vi;j;t = j;t = xj;t pj;t + j;t , j = 1; : : : ; J (1)
where xj;t is a vector of characteristics of alternative j at period t, pj;t = pj;t ( ) is the price
variable as a function of the monthly discount factor , and j;t is the unobserved quality of
alternative j at period t. In our speci…cation, xj;t will contain only a set of …xed e¤ects for
the alternatives.14 The price variable is the sum of the upfront investment price (pIN V
j;t ) and
the discounted future ‡ow bene…ts from GCC subsidies (pGCC j;t ) and electricity cost savings
EL
from net metering (pj;t ):
G E
G Rt E R
1 1
pj;t = pj;t ( ) pIN
j;t
V
( ) G
pGCC
j;t E
pEL
j;t (2)
1 1
| {z } | {z }
G E
t

G E
where and are monthly adjusted discount factors, speci…ed as:

G
= (1 )(1 ) (3)
E
= (1 )(1 + #) ;

i.e. the monthly discount factor adjusted for a depreciation parameter , the in‡ation rate
and the trend in real electricity prices #. We now discuss the three terms in (2) in more
detail.
The …rst term in (2), pIN V
j;t , is the real upfront net investment price of the PV system j
at period t, i.e. the real gross investment price minus tax cuts (taxcutj;t ) spread over up to
14
In the Appendix, we also include time e¤ects to assess whether our included variables are su¢ cient to
explain the time variation in adoption.

13
4 years ( = 1; : : : ; 4):
X
4
12
pIN V
j;t ( )= pGROSS
j;t taxcutj;t : (4)
=1

Before 2009, there was a tax cut only in the …rst year, capped at an indexed maximum
amount. Since 2009 any remaining tax cuts could be shifted to the following three years, so
that the last three terms in the summation in (4) become non-zero.15
The second and third terms in (2) capture the discounted future bene…ts from electricity
production: pGCC j;t and pEL
j;t are ‡ow variables measuring the monthly bene…ts from the …xed
subsidies from the GCCs and the electricity savings associated with the PV system. Both
pGCC
j;t and pELj;t are essentially prices per kW at period t (pt
GCC
and pEL
t ), multiplied by
the capacity size kj of the alternative j (in kW) and a factor that translates PV capacity
in monthly electricity production ( 0:85
12
M W h=kW ), following CREG, VEA and 3E (2010).16
The parameters G t and
E
are capitalization factors that convert the monthly bene…ts for
Rt months of GCCs and RE months of electricity savings into present value terms using
G

the adjusted monthly discount factors G and E . According to (3), these are the monthly
discount factors net of any depreciation. The parameter captures physical deterioration
of electricity production, whereas is the monthly in‡ation rate (because GCCs are …xed in
nominal prices, while our model is in real prices) and # captures a trend in real electricity
prices. As we make several assumptions in constructing the price variable, we provide a
detailed sensitivity analysis in section 4.3.17

Conditional value of not adopting (vi;0;t )

The conditional value of not adopting is the ‡ow utility in period t, u0;t , plus the option
15
This possibility only applied to houses older than 5 years. Furthermore, a reduced VAT from 21% to 6%
applied to houses older than 5 years. We account for this by taking a weighted average of the VAT rate and
tax cuts over new and old houses (where 91% is the fraction of old houses).
16
As pointed out earlier in footnote 10, households are not reimbursed for electricity production that is
higher than their annual consumption. The imputed electricity savings pEL j;t may therefore be too large for
high capacity options when households are small. As a robustness check, we also estimated a model under
the (extreme) assumption that electricity savings are equal to the smallest capacity, and this gave similar
results.
17
In our main speci…cation we assume a yearly physical deterioration rate of 1%, = 1:011=12 1 (following
Audenaert et al., 2010), a yearly in‡ation of 2%, = 1:021=12 1, and estimate a yearly growth in electricity
prices of 3.4%, # = 0:0028148. We assume RE = 240 months (the expected lifetime of a PV, following
CREG, 2010), and based on the GCC schemes announced by the government we set RtG = 240 months for
January 2006 - July 2012, RtG = 120 for August 2012 - December 2012, and RtG = 180 months for January
2013.

14
value of waiting. More precisely,

vi;0;t = 0;t = u0;t + Et V t+1 ; (5)

where V t+1 is the ex ante value function, i.e. the continuation value from behaving optimally
from period t+1 onwards, before the random taste shocks are revealed. With a type I extreme
value distribution for the random taste shocks "i;j;t ; the ex ante value function V t+1 has the
well-known closed-form logsum expression:

X
J
V t+1 = 0:577 + ln exp ( j;t+1 ) ; (6)
j=0

where 0:577 is Euler’s constant (the mean of the extreme value distribution).

Random utility maximization

With random utility maximization, we obtain the following choice probabilities or predicted
market shares for each alternative j = 0; : : : ; J at period t:

exp ( j;t )
Sj;t = sj;t ( t ) PJ : (7)
j 0 =0 exp ( j 0 ;t )

As in Berry (1994), we can equate the predicted market shares sj;t ( t ) to the observed market
shares Sj;t because of the inclusion of unobserved qualities j;t for every product and period.
The aggregate market share of alternative j 6= 0 is measured as Sj;t = qj;t =Nt , i.e. the
actual number of adopters of j at t, qj;t , divided by the potential number of adopters at
period t, Nt . Since adoption is a terminal action, the potential number of adopters is the
total number of households N minus the number of households that adopted in the past, i.e.
Pt 1 PJ PJ
Nt = N =1 j=1 qj; . The aggregate market share of not adopting is S0;t = 1 j=1 Sj;t

3.2 Estimating equation

The aggregate market share equation (7) involves two complications. First, the conditional
value for not adopting 0;t , as given by (5), involves the expected future value term Et V t+1 ,
which is recursively de…ned by (6). Second, the conditional value for adopting, j;t , contains
the unobservable product quality term j;t , which enters nonlinearly. We now show how
to solve both complications, by obtaining an analytic expression for the future value term
Et V t+1 and by inverting the market share equation.

Expected ex ante value function

15
The expectation operator before V t+1 in (5) integrates over uncertainty about the next period
state variables, i.e. the vector ! t = (u0;t+1 ; 1;t+1 ; :::; J;t+1 ). The usual approach speci…es
an explicit stochastic process of the state transitions.18 Following Scott (2013), we instead
decompose Et V t+1 into the realized ex ante value function V t+1 and a short run prediction
error t V t+1 Et V t+1 . We assume that households’expectations are on average correct,
such that t is mean zero. We can then write (5) as

0;t = u0;t + (V t+1 t ): (8)

Specifying a ‡exible prediction error, avoids having to make arbitrary assumptions on how
households expect the future states to evolve. This is particularly important in this applica-
tion because the GCC subsidies were revised many times and it is unclear how this in‡uenced
households’expectations.19
The ex ante value function V t+1 , as given by the logsum formula (6), recursively depends
on future value functions (through the term 0;t+1 ). Hotz and Miller (1993) show how to
write V t+1 in terms of the conditional choice probabilities (CCPs). This is particularly
convenient when the decision problem has a terminal action, as is the case in our set-up for
any adoption decision j = 1; : : : ; J.20 We can then take the next period CCP for any arbitrary
terminating choice, so we take the CCP of alternative j = 1, as given by s1;t+1 ( t+1 )
P
exp ( 1;t+1 ) = Jj=0 exp ( j;t+1 ). After rewriting and taking logs, this gives:
XJ
ln exp ( j;t+1 ) = 1;t+1 ln s1;t+1 ( t+1 );
j=0

which can be substituted in (6) to obtain the following expression for the ex ante value
function at t + 1:
V t+1 = 0:577 + 1;t+1 ln s1;t+1 ( t+1 ): (9)
As discussed in Arcidiacono and Ellickson (2011), expression (9) has an intuitive interpre-
tation. The ex ante value function (at t + 1) is essentially equal to the utility of choosing
option j = 1 plus the mean of the Type I extreme value distribution (0.577) plus the CCP
correction term ln s1;t+1 ( t+1 ) 0. The CCP correction term adjusts for the fact that
18
For example, if the states follow a Markov process with density f (! t+1 j! t ), we have Et V t+1 =
R
V t+1 (! t+1 ) f (! t+1 j! t ) d! t+1 .
19
Burr (2016), Feger et al. (2017) and Langer and Lemoine (2018) also model PV adoption with a
dynamic model. As in Rust (1987), they solve the model during estimation, which requires assumptions on
how households expect the state variables to transition over time.
20
This is a particular example of a simpli…cation that occurs because of …nite dependence (Arcidiacono
& Miller 2011). An alternative action that quali…es for …nite dependence is the renewal action, as in Scott
(2013).

16
j = 1 may not be optimal, so that the expected utility is on average higher than that of
adopting j = 1 (unless s1;t+1 ( t+1 ) = 1).
We can now substitute (9) in the mean utility from not adopting (8) to obtain:

0;t = u0;t + (0:577 + 1;t+1 ln s1;t+1 ( t+1 ) t)

= ( 1;t+1 ln S1;t+1 t) (10)

where the second equality follows from normalizing u0;t + 0:577 = 0 and from the fact
that the CCP at the realized mean utilities is equal to the observed market share (S1;t+1 =
s1;t+1 ( t+1 )). Note that this di¤ers from other applications that use the Hotz and Miller
(1993) inversion as this usually requires predicting the CCP in a …rst stage. Here we apply
the Hotz and Miller (1993) inversion on aggregate data with the realized ex ante value
function. This implies that the CCP is simply the observed market share of j = 1 in the
next period.21

Market share inversion

To invert the market share equation, we follow the approach of Berry (1994) for estimating
static discrete choice models with aggregate market share data. Using the market share
expressions (7), divide Sj;t for each j = 1; : : : ; J by S0;t and take logs to obtain:

ln Sj;t =S0;t = j;t 0;t , j = 1; : : : ; J: (11)

Substitute the expressions for the mean utilities (1) and (10) in (11), and rewrite to obtain
the following main estimating equation:

ln Sj;t =S0;t = (xj;t x1;t+1 ) (pj;t p1;t+1 ) + ln S1;t+1 + ej;t ; (12)

where
ej;t j;t ( 1;t+1 t) (13)
is the econometric error term. In the static case where = 0, this is Berry’s standard
aggregate logit regression for the number of new adopters on current prices and other control
variables. To gain further intuition when > 0, assume there is only one adoption alternative
j = 1. The estimating equation can then be written as:

S1;t =S1;t+1
ln = (x1;t x1;t+1 ) (p1;t p1;t+1 ) + e1;t :
S0;t
21
As discussed further in subsection 3.4 below, in the model that accounts for local market heterogeneity
we will need to predict the CCPs in a separate …rst stage, because the number of adopters can be very small
at the local level.

17
 With close to 1, this is essentially a regression for the change in the number of new adopters
on the change in price and possibly other characteristics. Intuitively, with forward-looking
consumers one may expect that the number of current period adopters is small relative to
the next period adopters when the next period price drop is large. As pointed out by Scott
(2013), one may think of the estimating equation (12) as analogous to an equilibrium Euler
equation in continuous decision problems (re‡ecting indi¤erence between adopting now and
tomorrow in probability terms).22

3.3 Estimation and identi…cation

The estimating equation (12) contains the price variable pj;t , which is given by (2). This
depends on the upfront investment price pIN V
j;t , the future …nancial bene…ts from GCCs
pGCC
j;t and electricity savings pEL
j;t , and it is a non-linear function of the discount factor .
To …x ideas, …rst consider the case in which is known and all variables are exogenous,
i.e. uncorrelated with the error term ej;t . In this case, it is possible to estimate (12) using
a simple linear OLS regression for the di¤erenced adoption variable ln Sj;t =S0;t ln S1;t+1
on the di¤erenced product characteristics xj;t x1;t+1 and the di¤erenced price variable
pj;t p1;t+1 .
Now consider the more general case where has to be estimated and some of the variables
may be correlated with the error term ej;t . Notice …rst that the estimating equation (12)
is non-linear in because of the way it enters the price term (2), so a non-linear estimator
is necessary. More importantly, several variables in equation (12) give rise to endogeneity
concerns. Recall that, according to (13), the error term ej;t consists of the households’
prediction error t and the demand shocks j;t and 1;t+1 . As discussed in Scott (2013),
the prediction error t is by construction uncorrelated with any variables known by the
households at time t, so it does not give rise to endogeneity concerns. In contrast, the
demand shocks give rise to endogeneity issues that are similar to those in static discrete
choice demand models. First, pj;t contains the investment price variable pIN V
j;t , which may be
correlated with the error term if …rms charge higher prices when demand is high. Second,
pj;t also contains the electricity price variable pEL j;t . This may also be correlated with the error
term to the extent that the GCC subsidies were …nanced through higher electricity prices.
Third, the next period adoption rate ln S1;t+1 may be correlated with the error term, since
it contains the next period demand shock 1;t+1 .23
22
Gandal et al. (2000) obtain a related equation in a model with di¤erent distributional assumptions
about heterogenous consumers, who decide when to adopt new hardware-software.
23
Note however that ln S1;t+1 does not imply an extra parameter, because also enters (non-linearly)
through the price term pj;t ( ) of equation (2). This avoids the need to …nd a extra instrument for the next

18
 To account for these problems we construct an instrument vector zj;t that is uncorre-
lated with the error term, and estimate the model using GMM with the following moment
conditions:
E (zj;t ej;t ) = 0 (14)
We include the following variables in our instrument vector zj;t . First, we include a price
index of Chinese PV modules on the European market, pM j;t
OD
. Since these modules are the
most important cost component of PV installations, the price index pM j;t
OD
is expected to
IN V
be correlated with the endogenous upfront investment price variable pj;t , and as a cost
shifter it is reasonable to assume it does not directly in‡uence demand. The price index
of Chinese PV modules thus provides a strong and valid instrument to identify the price
coe¢ cient . Second, we include the contractually …xed future bene…ts from the GCC
subsidies pGCC
j;t as an instrument. As discussed in section 2, this variable refers to the main
source of future bene…ts from adopting a PV. There is considerable variation in pGCC j;t across
alternatives and over time, even in the short run as the bene…ts showed discontinuous drops in
several months. The variable pGCC
j;t thus provides a strong instrument to identify the discount
factor , i.e. how households trade o¤ upfront investment costs with future bene…ts. After
also adding the exogenous xj;t to the set of instruments, the model is identi…ed. However,
to improve e¢ ciency, in a second stage we use an approximation to optimal instruments
(Chamberlain 1987), as applied in static aggregate discrete choice models by Berry, Levinsohn
and Pakes (1999) and Reynaert & Verboven (2014). We explain this in Appendix A.2.
The dynamic discrete choice literature has stressed that the discount factor is nonpara-
metrically unidenti…ed (Manski (1993) and Rust (1994)), but identi…cation can be obtained
with appropriate exclusion restrictions (Magnac & Thesmar (2002) and Abbring and Daljord
(2017)). More precisely, the discount factor can be identi…ed if there are two or more states
that a¤ect expected future utilities but not current utility. In our setting, the upfront invest-
ment costs a¤ect only current utility, whereas the GCC subsidies, electricity bill savings and
tax credits a¤ect the future bene…ts. A …rst concrete source of identi…cation comes from the
variation in the generosity of the GCC subsidies to new adopters over our sample period (as
documented in Figure 1). Note that “static”models of intertemporal choice, which abstract
from the timing decision and focus only on the investment decision, implicitly use a similar
identi…cation strategy. For example, after Hausman’s (1979) contribution, a detailed liter-
ature on the car market focuses on how households trade o¤ future fuel cost savings against
higher upfront purchase prices, without explicitly modeling the timing of the purchase de-
cision; see Verboven (2002), Allcott and Wozny (2013) and Busse, Knittel and Zettelmeyer
(2013). This work also relies on variation in expected future energy costs, relative to upfront
period market share.

19
car prices.24 In dynamic choice models, Lee (2013) uses a related identi…cation approach in
an application on the timing of hardware purchases (video game consoles) when there are
future bene…ts from new software (games). He makes use of variation in the time until new
games arrive, and assumes the discount factor for the timing of adoption is the same as that
for the valuation of upfront costs versus future bene…ts.25
A second possible source of identi…cation comes from the dynamics of the model. In
Figure 2 we see large increases in adoptions just before a change in GCC subsidies, even
though this change is irrelevant for the utility of adopting a PV in that month. This is
because subsidies for new adoptions change at well-known pre-announced dates. These
future changes do not directly impact the utility of adopting today but they do change the
option value if households choose not to adopt. In the sensitivity analysis we attempt to
estimate a separate discount factor relating to this option value to evaluate which source of
variation is mainly responsible in identifying the discount factor.

3.4 Accounting for local market heterogeneity

The previous subsections provided a framework to study the adoption of PV systems at the
aggregate country level. To evaluate the robustness of our …ndings, in this subsection we show
how to extend the empirical analysis to account for observed and unobserved heterogeneity
at the very disaggregate level of M = 9182 local markets, where each market m consists
of on average 295 households.26 We match information on the number of adopters in each
market m for each alternative j in each period t to several demographic characteristics. This
enables us to include a rich set of demographics to interact with the price and capacity
size in the utility speci…cation. We also include local market …xed e¤ects to control for
unobserved heterogeneity. As such, we relax the assumption that valuations of solar systems
are i.i.d. across alternatives and over time. We allow households in di¤erent markets to
have persistent di¤erences in their valuation for adopting versus not adopting, and to have
correlated preferences for systems with similar capacity sizes.
One approach to incorporate this heterogeneity would be to specify (12) at the local
market level. The dependent variable would become ln Sm;j;t =Sm;0;t , the error term would
become em;j;t and the regression would include local market demographics and interactions
with product characteristics. However, at a highly disaggregate level the number of new
24
This work mainly relies on cross-sectional variation between cars of di¤erent fuel e¢ ciency, whereas we
rely mainly on variation over time.
25
Related approaches to identify the discount factor in dynamic choice models are (Dube et al. 2012),
(Bollinger 2015) and (Yao et al. 2012).
26
Within these disaggregate local markets, we continue to assume consumers are homogeneous.

20
adopters is often zero, so that the log market share terms are not de…ned (both the dependent
variable ln Sm;j;t =Sm;0;t and the next period CCP term ln Sm;1;t+1 ). Scott (2013) addresses
this issue using a smoothing procedure for all market share terms (in a dynamic model
with replacement instead of terminating actions). We instead take an alternative approach,
which combines the moment conditions from the aggregate model (involving aggregate shares
ln Sj;t =S0;t ) with a set of micro-moments at the local level, that arise from the likelihood of
observing new adoption levels in local markets. This is similar to the static discrete choice
literature, for example Berry et al. (2004) and Quan and Williams’s (2018) (although we
continue to rely on …rst-stage predictions for the next period CCP term as is common in the
dynamic discrete choice literature (Arcidiacono and Ellickson (2011)).
Our speci…cation incorporates unobserved heterogeneity in the decision whether to adopt
through a rich set of local market …xed e¤ects (in addition to observed heterogeneity for the
decision which alternative to adopt through demographic interactions with product char-
acteristics). Alternative approaches to account for unobserved heterogeneity would be to
estimate random coe¢ cients, similar to Gowrisankaran and Rysman (2012), or a …nite mix-
ture of unobserved types in the population as in Scott (2013), based on the EM algorithm
of Arcidiacono and Miller (2011). While a random coe¢ cients model gives more ‡exibility
in modeling heterogeneity for the valuation of product characteristics, it requires an explicit
speci…cation of the state transitions, and it does not make e¢ cient use of the rich local mar-
ket heterogeneity we observe. A mixture of unobserved types would be di¢ cult to identify
in our context, since households do not make repeat purchases so that we cannot infer their
types from correlations in their decisions over time.
The basic set-up is as before, except that we now observe adoption decisions at the local
market level m and we can match this with an H 1 vector of household demographics
Dm . In each period t a household i living in market m chooses its preferred alternative
j = 0; 1; : : : ; J, where j = 0 is the option not to adopt (yet).
As in the aggregate model, the conditional value of adoption vi;j;t (j = 1; : : : ; J) is still the
expected discounted utility, because adoption is a terminal action. The di¤erence with the
aggregate model is that vi;j;t no longer consists of only a mean utility term j;t (given by (1)
and including an unobserved quality term j;t as before). It now also includes an individual-
speci…c component i;j;t = m;j;t , which depends on demographics of the local market m (but
households are identical within these disaggregate local markets). More precisely, we have:

vi;j;t = j;t + m;j;t

= xj;t pj;t + j;t + wj;t m, j = 1; : : : ; J (15)

where m;j;t wj;t m and wj;t is a 1 K vector of characteristics of the PV alternatives

21
(which is allowed to di¤er from xj;t entering j;t ). We specify the K 1 vector m = Dm ,
where is a K H parameter matrix with interaction e¤ects to be estimated. The vector
of characteristics wj;t will include a constant, the additional capacity relative to a reference
capacity (we take j = 1, which is the 4kW alternative), and the price variable. The vector
of household demographics Dm includes dummy variables for each local market m, but also
income, household size, house size, etc. We will not estimate all the interaction e¤ects in
, so we constrain some of these coe¢ cients to be zero. We interact the constant with
local market dummy variables, and price and capacity with a selection of the household
demographics. In a sensitivity analysis we also consider heterogeneity in the discount factor.
We explain this approach in Appendix A.4.2.
The conditional value of not adopting vi;0;t is

vi;0;t = um;0;t + Et V m;t+1 ;

where the ex ante value function is now speci…c to market m and given by
X
J
V m;t+1 = 0:577 + ln exp (vi;j;t+1 ) :
j=0

Assume that t V m;t+1 Et V m;t+1 , i.e. the expectational error is common across local
markets, so there is only aggregate uncertainty. As shown in Appendix A.3, the logit choice
probabilities in market m are then given by:
exp(vi;j;t )
sm;j;t = PJ
j 0 =0 exp(vi;j 0 ;t )
exp(vi;j;t vi;0;t )
= PJ
1 + j 0 =1 exp(vi;j 0 ;t vi;0;t )
exp(ej;t + w ej;t m + ln sbm;1;t+1 )
= PJ (16)
1 + j 0 =1 exp(ej 0 ;t + w
ej 0 ;t m + ln sbm;1;t+1 )

where we de…ne ej;t j;t ( 1;t+1 ej;t wj;t

t ) and w w1;t+1 , and where sbm;1;t+1 is a
predicted value of the next period choice probability of j = 1 (derived from the data before
estimating the model).
We explain estimation of the model in Appendix A.3. In short, we combine moments at
the aggregate level with micro-moments at the local market level in one GMM estimator.
These micro-moments are the scores from the likelihood function of the model, using (16).
The scores relating to the demographic variables can be interpreted as moment conditions
that match the observed covariances between the demographic variables and product charac-
teristics to the model’s predictions. The scores relating to the local market (m) …xed e¤ects

22
can be interpreted as matching the total number of adopters in each market at the end of
the sample to the predicted number. Finally, instead of using Berry’s market share inversion
to match the monthly adoption rates at the aggregate level, we estimate product-time (j; t)
…xed e¤ects.27 The aggregate moment is similar to the one in the aggregate model (14,
13, 12), with the di¤erence that (j; t) …xed e¤ects replace the inversion of aggregate market
shares, which will not coincide unless local market heterogeneity is fully absent). Note that
the model therefore still allows for unobserved quality and prediction errors at the aggregate
level.

4 Empirical results
We …rst discuss our main …ndings with a focus on the estimated discount factor (subsec-
tion 4.1). We then perform a sensitivity analysis with respect to several speci…cation choices
(subsection 4.2). Next, we assess the sources of time discounting by estimating the model un-
der alternative assumptions about how future payo¤s enter utility (subsection 4.3 ). Finally,
we use the parameter estimates to consider the budgetary impact of an alternative policy
to promote PV adoption with upfront investment instead of future production subsidies
(subsection 4.4).

4.1 Main …ndings

Table 1 provides summary statistics of the included variables and instruments for the sample
on which we estimate the model (May 2009 – December 2012). The …rst panel shows
summary statistics for the number of adopters. At the aggregate country level, we observe the
number of adopters for 5 levels of capacity during 44 months, resulting in 220 observations.
At the disaggregate level, we observe the number of adopters for 9182 local markets, resulting
in more than 2 million observations. The average number of adopters per capacity level is
901 at the country level, and it has always been positive for every capacity and month. At
the local market level, the average number of monthly adopters is evidently much smaller
at 0:10. Because of the highly disaggregate level, the number of adoptions is zero for many
local markets. The median number of adopters for a capacity level/month/local market is
actually zero.
The second panel presents information on the components of the price variable. This
27
This di¤ers from Quan and Williams (2018). They have a very large set of local interactions (markets m
times alternatives j), which induces them to develop a random e¤ects approach. We instead have additive
…xed e¤ects for each of the local markets m and each alternative-time combination (j; t).

23
shows for example that the investment price of a PV has on average been 20; 700e, with
a large standard deviation both because of falling prices over time and large di¤erences
depending on the capacity size. The third panel shows the excluded instruments, i.e. the
variables that do not enter the model directly but are correlated with the endogenous invest-
ment cost and electricity price. Finally, the fourth panel of Table 1 shows information on the
household characteristics for the cross-section of 9; 182 local markets. This shows for example
that the household size is on average 2:47, but varies between 1 and 6. Similarly, median
yearly income is on average 24; 000 EUR, and varies between 4; 800 and 51; 800 across the
statistical sectors.

24
 Table 1: Summary statistics
Variable Notation Mean Std. Dev. Min Median Max Obs.
Adoptions
Country level qj;t 901.1 1309.58 4 311.5 7226 220
Local market level qm;j;t 0.10 0.41 0 0 26 2,020,040

Price variable (in 103 EUR)

Investment cost pGROSS
j;t 20.70 10.85 4.82 19.61 50.82 220
Monthly GCC subsidies pGCC
j;t 0.14 0.08 0.01 0.13 0.35 220
Monthly electricity bill savings pELj;t 0.09 0.04 0.03 0.09 0.17 220
Tax cut year 1 taxcut1j;t 2.63 1.62 0 3.69 3.69 220
Tax cut year 2 taxcut2j;t 1.83 1.57 0 2.44 3.36 220
Tax cut year 3 taxcut3j;t 1.20 1.50 0 0 3.36 220
Tax cut year 4 taxcut4j;t 0.55 1.11 0 0 3.36 220

Excluded instruments
Module price (103 EUR) pM
j;t
OD
7.81 5.01 1.06 6.56 23.27 220
Oil price (EUR / barrel) pOIL
t 68.37 12.10 40.69 71.20 88.37 44

Local market variables (Nm and Dm )

Households Nm 295.26 320.88 1 191 3608 9,182
Pop. density (104 inhab / m2 ) 0.16 0.24 0.00 0.09 2.89 9,182
Average house size 5.93 0.64 1.85 5.96 9 9,182
Average household size 2.47 0.34 1 2.49 6 9,182
Average house age (decades) 5.19 1.49 0.37 5.07 11.3 9,182
Median income (104 EUR) 2.40 0.36 0.48 2.40 5.18 9,182
% home owners 0.77 0.17 0 0.82 1 9,182
% higher education 0.26 0.11 0 0.25 1 9,182
% foreign 0.06 0.09 0 0.03 1 9,182
Notes: The total number of observations is 2,020,040 = 44 time periods x 5 capacity choices x 9,182
local markets. All prices are corrected for in‡ation using the HICP and set to prices of January 2013.
Half-yearly electricity prices extrapolated using cubic spline interpolation, missing values on local market
level replaced by averages within the 308 municipalities (642 markets for median income and between
0 and 146 markets for other variables).

Table 2 shows the empirical results. We begin with a discussion of speci…cation (1) and
(2), which are estimated with country-level data and do not account for household hetero-
geneity, following the regression equation (12). Both speci…cations include …xed e¤ects for
each capacity size using the most popular 4kW system as the base. As a point of comparison,
speci…cation (1) is the static version of the model (often estimated in other contexts), i.e.
we set = 0 in equation (12) so that the next period terms drop out, and at the same time

25
keep in the price variable, as given by (2) and (3). Speci…cation (2) is the full dynamic
version of (12), where we set the terminating action j = 1 to the base capacity level of 4kW.
The investment price coe¢ cient is negative and statistically signi…cant, meaning that
consumers responded positively to the decline in investment prices of PV systems. The mag-
nitude of the investment price coe¢ cient is smaller in absolute value in the static speci…cation
than in the dynamic speci…cation ( 0:318 versus 0:470). This appears to be consistent with
Gowrisankaran and Rysman’s (2012, p. 1176) interpretation: “a static estimation applied to
a durable good purchase decision with falling prices will then result in mismeasurement that
may tend to bias the price coe¢ cient toward zero.” The di¤erence in the price coe¢ cient
between the static and dynamic speci…cation is however less pronounced in our application,
because the falling investment prices are occasionally interrupted by sharp drops in subsidy
bene…ts.
The estimated (real) discount factor measures the valuation of the future bene…ts relative
to the investment price. The monthly discount factor is very similar for both speci…cations,
and di¤ers signi…cantly from 1. It is more informative to convert the monthly discount factor
into an annual implicit interest rate. The results show that the real implicit interest rate
is 14:82% in the …rst speci…cation (standard error of 2:28%), and a similar 15:09% in the
second speci…cation (standard error of 3:43%). These estimates are much higher than market
interest rates on risk-free or moderate risk investments (even though our estimates are in
real terms, while the market rates are in nominal terms). For example, the interest rate on
mortgages ranged between 3.6% and 5.3% in the period 2006-2012.28 Moreover, between
2009 and 2011, the government subsidized loans for environmentally friendly investments,
so that the e¤ective interest rates at which households could borrow would be even lower.29
This then suggests that consumers discount the future bene…ts of new technologies such
as PV installations much more than has been observed in recent work on mature technologies
such as the car industry. The high implicit interest rate implies that consumers are willing
to pay only 0:5 euro upfront for one euro of total discounted future bene…ts from electricity
production.30 Put di¤erently, if consumers would have been more forward looking, the
generous GCC subsidy policy would have led to an even faster adoption of PV systems. In
subsection 4.3, we will investigate the sources of these high implicit interest rates.
28
Source: National Bank of Belgium (http://stat.nbb.be). Monthly averages of …xed rates on new contracts
for durations over 10 years.
29
https://…nancien.belgium.be/nl/particulieren/belastingvoordelen/groene_…scaliteit/groene_leningen
RE
30
One (real) euro of production bene…ts is valued at A( ) = 1 1((1(1 ) )) . We obtain the cited number as
the ratio of the bene…ts at the estimated household discount factor over the bene…ts at the market discount
factor, i.e. A(0:9884)=A(0:9975) = 0:5, where 0:9975 = 1:03 1=12 at the market interest rate of 3%.

26
 Before turning to this, we discuss the results of speci…cation (3), which is estimated with
local market data and accounts for rich patterns of household heterogeneity. The investment
price coe¢ cient changes somewhat (from 0:470 to 0:604), which can be explained by
the inclusion of an interaction variable for median income with price. This interaction e¤ect
shows that high income households tend to be less price sensitive, so that for the average
income the price coe¢ cient is close to the estimate from the aggregate model.
Most importantly, the estimated discount factor remains almost identical when we ac-
count for household heterogeneity. The implied annual implicit interest rate is 15:00% (com-
pared with 15:09% in the model without heterogeneity). So also in the richer model there is
evidence of considerable time discounting in adopting the new PV technology.
Finally, the coe¢ cients for the household characteristics interacted with the capacity of a
PV usually have an intuitive interpretation. As expected, large households, households living
in large houses or in areas with a low population density especially value a large capacity.
High income households, highly educated people and home owners tend to adopt smaller
PVs. Foreigners and households living in older houses tend to invest in larger PVs.31

31
In De Groote et al. (2016), we estimate descriptive models with a more elaborate set of demographic
variables.

27
 Table 2: Empirical results
(1) (2) (3)
Static Dynamic + micro-moments
Price sensitivity in 103 EUR ( ) -0.318*** (0.074) -0.470*** (0.098) -0.604*** (0.100)
Monthly discount factor ( ) 0.9886*** (0.0016) 0.9884*** (0.0025) 0.9884*** (0.0024)
12
Annual interest rate (r 1) 14.82%*** (2.28%) 15.09%*** (3.43%) 15.00%*** (3.42%)

Control variables ( )
Alternative-speci…c constant
Common constant -8.169*** (0.483) -1.422 (16.374) 3.633 (16.880)
2kW -1.909*** (0.231) -1.828*** (0.562) -1.214* (0.724)
6kW -0.388 (0.241) -0.512 (0.595) -1.225 (0.753)
8kW -2.248*** (0.459) -2.452** (1.158) -3.926*** (1.473)
10kW -2.356*** (0.670) -2.602 (1.683) -4.878** (2.159)

Local market variables ( )

Interactions with constant Local market …xed
e¤ects included
Interactions with capacity di¤ erence
Pop. density (104 inhab / m2 ) -0.689*** (0.029)
Average house size 0.057*** (0.009)
Average household size 0.124*** (0.016)
Average house age (decades) 0.011*** (0.002)
Median income (104 EUR) -0.066** (0.030)
% home owners -0.075** (0.038)
% higher education -0.128*** (0.041)
% foreign 0.383*** (0.040)

Interaction with price

Median income (104 EUR) 0.049*** (0.007)

Obs. macro moments (JxT) 220 220 220

Obs. micro moments (MxJxT) 0 0 935,440
Notes: For all models, standard errors are clustered across alternatives within 44 time periods. In the third
model, for the micro moments at the local market level we additionally cluster across time periods within each
of the 4252 local markets. Instruments are approximations of optimal instruments (Chamberlain, 1987).
Standard errors of r and common constant obtained via delta method. *** p<0.01, ** p<0.05, * p<0.1

28
4.2 Sensitivity analysis
We now perform a sensitivity analysis with respect to several modeling choices. For simplicity,
we mainly focus on the aggregate adoption model, because the estimates of the implicit
interest rate were very close to the disaggregate model with household heterogeneity.

Speci…cation choices We …rst consider the sensitivity of our results with respect to the
terminating action, j = 1, in the implementation of the CCP approach. In the above
analysis we set j = 1 to the capacity level of 4kW, which is the most popular capacity level.
In principle, one can do a sensitivity analysis by taking each of the …ve possible capacity
choices as the terminating action. To explore this more formally, we create a GMM estimator
with moments for each of the …ve possible terminal actions. As shown in the …rst column
of Table 3, this results in a comparable estimate of the implicit interest rate (16:62%) and
a reduced standard error of 1:03%. Since this approach yields …ve times as many moments
than parameters, it is possible to perform a test of overidenti…cation restrictions. Hansen’s
J is 31:7 with a P-value of 0:2858, so we cannot reject the hypothesis that the model is
correctly speci…ed. Table A5 in Appendix shows the empirical results when we take each
of the …ve possible capacity choices as the terminating action. The empirical results are
very similar across the …ve di¤erent models, with estimated implicit interest rates varying
between 11:94% and 16:99%.32
We also considered a speci…cation with a di¤erent assumption of the potential market
size. Our base speci…cation assumed that the potential market is equal to the total number
of households, but many households might not be able to install a solar panel because of a
bad roof orientation, too much shadow or because they live in an apartment building. We
therefore consider an alternative speci…cation where the potential market size is only 10%
of the total number of households (only slightly above the adoption rate of 8.3% observed
at the end of the subsidy period). Table 3 shows that the estimated discount factor remains
very similar.
Next, we considered a speci…cation with a time trend and seasonal dummy variables.
According to Table 3, the coe¢ cients of these variables are insigni…cant and only slightly
a¤ect the estimated implicit interest rate (change from 15.09% to 13.52%). This con…rms
that the main variables in our structural dynamic model (investment cost and future bene…ts)
32
We also considered a speci…cation where we take a varying capacity choice j as the terminating action
(rather than keeping it …xed as in standard approaches). This becomes close to a “…rst-di¤erences” model,
and it results in imprecisely estimated parameters. This di¤erences speci…cation thus appears to eliminate
relevant long-term variation relating to the choice of the capacity sizes. However, when we add …rst-di¤erences
moments to our other moments, our parameters are again precisely estimated and robust.

29
explain the variation over time rather well.

Table 3: Robustness: speci…cation choices

All terminal choices 10 % potential market Time controls
3
Price sensitivity in 10 EUR ( ) -0.422*** (0.046) -0.471*** (0.098) -0.439*** (0.117)
Monthly discount factor ( ) 0.9873*** (0.0007) 0.9883*** (0.0025) 0.9895*** (0.0016)
12
Annual interest rate (r 1) 16.63%*** (1.03%) 15.13%*** (3.44%) 13.52%*** (2.17%)

Control variables ( )
Alternative-speci…c constant
Common constant -10.152 (11.278) 1.557 (16.345) -610.0 (1,017.3)
2kW -2.045*** (0.129) -1.834*** (0.562) -1.614*** (0.421)
6kW -0.282** (0.136) -0.507 (0.595) -0.721 (0.460)
8kW -2.021*** (0.262) -2.442** (1.158) -2.879*** (0.881)
10kW -1.989*** (0.399) -2.587 (1.683) -3.250*** (1.260)

Time controls
Linear trend 1.172 (1.985)
Spring -0.177 (0.470)
Summer -0.047 (0.493)
Fall -0.021 (0.358)

Hansen’s J (p-value) 31.736 (p= 0.2854) Exactly identi…ed Exactly identi…ed

Obs. macro moments 220 220 220
Notes: Standard errors clustered within 44 time periods. Instruments are approximations of optimal instruments
(Chamberlain, 1987). Standard errors of r obtained via delta method. *** p<0.01, ** p<0.05, * p<0.1

Finally, we re-estimated the model under alternative assumptions regarding the evolution
of expected future electricity prices. Our base speci…cation assumed a constant annual growth
in electricity prices of 3.4% (# = 0:0028148). Alternative assumptions such as zero growth
or a growth of 6.8% had a negligible impact on the results.

Separate discount factors for di¤erent utility components To shed light on the
sources of variation behind the identi…cation of the discount factor , we relax some of
the parametric restrictions in our structural dynamic model. On the one hand, directly
enters the regression equation (12), capturing the valuation for the option value of postpon-
ing adoption. On the other hand, also enters (12) indirectly through the present value
term pj;t ( ), as given by (2), capturing the valuation of the future GCC bene…ts and other
bene…ts (electricity cost savings and future tax credits). Table 4 shows two speci…cations

30
where we attempt to empirically distinguish between di¤erent discount factors for separate
components.33
The …rst column of Table 4 estimates a separate discount factor for the valuation of all
future …nancial bene…ts (entering pj;t ( )) and for the option value of postponing adoption
(entering (12) directly). The …rst discount factor is close to the estimate from our base model.
The second discount factor is considerably lower, but is also estimated very imprecisely, and
a Wald test does not reject the hypothesis that both discount factors are equal. This indicates
that, in our application, the discount factor is mainly identi…ed from variation in the future
bene…ts relative to the upfront investment costs, instead of from variation in the option value.
The second column of Table 4 estimates a separate discount factor for bene…ts that do not
come from GCC subsidies. The discount factor relating to the GCC bene…ts is larger than the
one relating to other bene…ts. But the former is estimated more precisely than the latter, and
a Wald test does not reject that both discount factors are equal. In sum, both speci…cations
show that identi…cation of the discount factor mainly comes from variation in the GCC
bene…ts, and that it is di¢ cult to separately distinguish between various components.34

Heterogeneity in the discount factor The model with local market heterogeneity al-
lowed for heterogeneity in the valuation of price and capacity, but not in the discount factor
. A homogenous discount factor allows for a transparent interpretation and counterfactual
simulation. Accounting for heterogeneity in the discount factor is more complex than ac-
counting for heterogeneity in the valuation of price or capacity, because it involves interacting
the unobserved expectational errors ( t ) with local market demographics. In Appendix A.4.2
we explain a procedure to solve this problem, and show the empirical results for a ‡exible
speci…cation in which a rich set of demographics in‡uences the valuation of price, capacity
and the discount factor. Figure 5 plots the distribution of the implicit interest rate, implied
by our estimates. This shows that there is some heterogeneity, but the interest rate of 90%
of households falls within a narrow range of 13:34% to 16:87%.

33
To estimate the additional parameters, we update the approximation of optimal instruments and add
additional instruments in the …rst stage. The …rst column adds the next month value of the cost instrument
and the GCC bene…ts, and the second column adds the cost instrument multiplied with the tax cut rate,
and the oil price multiplied with the capacity.
34
We also considered two speci…cations to control for possible measurement error in the electricity savings.
In one speci…cation we estimate a separate parameter for electricity savings. This gives imprecise estimates,
while the estimated discount factor for the GCC bene…ts remains robust. Another speci…cation assumes
every household receives at most the electricity savings of the smallest PV system (2kW). Also in this case
the estimated discount factor is robust. This con…rms that identi…cation of the discount factor mainly comes
from variation in the GCC bene…ts.

31
 Table 4: Robustness: separate discount factors for di¤erent utility components
Option value and bene…ts GCC and other
3
Price sensitivity in 10 EUR ( ) -0.279*** (0.108) -0.771*** (0.145)
Monthly discount factor on GCC bene…ts 0.9870*** (0.0033) 0.9933*** (0.0011)
Monthly discount factor on non-GCC bene…ts 0.9870*** (0.0033) 0.9031*** (0.0476)
Monthly discount factor on option value 0.4328 (0.3451) 0.9933*** (0.0011)

Control variables ( )
Alternative-speci…c constant
Common constant -4.365 (2.862) 0.056 (0.191)
2kW -2.120*** (0.375) -3.602*** (0.575)
6kW -0.164 (0.387) 1.088** (0.498)
8kW -1.809** (0.751) 0.723 (0.957)
10kW -1.717 (1.096) 1.915 (1.375)

Wald test discount factors di¤erent (p-value) 2.56 (p= 0.1095) 3.58 (p= 0.0584)
Obs. macro moments 220 220
Notes: Standard errors clustered within 44 time periods. Instruments are approximations of optimal instruments
(Chamberlain, 1987). The …rst regression assumes that the discount factor in the valuation of investment bene…ts
is the same for all components but the valuation of option value can be di¤erent. The second regression allows for
a di¤erent valuation of non-GCC bene…ts. *** p<0.01, ** p<0.05, * p<0.1

32
 Figure 5: Heterogeneous interest rate

Implicit real interest rate, resulting from the estimated distribution of the discount
factor, as explained in the Appendix section A.4.2. Discount factor estimated
at the local market level and weighted by number of potential adopters at start of the
sample period.

33
4.3 Sources of time discounting
Before turning to the implications for the government’s GCC policy, we consider various
possible sources of the considerable time discounting we have estimated. We look into this
by assessing the impact of the various assumptions we made in section 3.1 when constructing
the up-front investment price and the future bene…ts. We use the aggregate adoption model,
because the estimates of the implicit interest rate were very close to the disaggregate model
with household heterogeneity, and because it is computationally much faster so that a very
detailed sensitivity analysis is possible.
We distinguish between three alternative explanations for the high estimate of the im-
plicit interest rate: the durability of the PV technology, consumer expectations about the
government’s commitment, and intrinsic consumer undervaluation or myopia.

Durability of the PV technology A …rst explanation for the high implicit interest rate
is that the durability of the PV technology is lower than assumed in our main speci…cation,
so that the future bene…ts are in practice lower. Figure 6 shows how the estimated implicit
interest rate varies as we change the assumptions on the durability of the PV technology:
the life expectancy R and the yearly deterioration rate . The vertical lines denote the
assumptions made in the base model.
The left part of Figure 6 shows that the estimated implicit interest rate remains robust if
we increase the PV’s life expectancy R above the assumed value of 20 years or if we reduce
it by several years. We estimate a low, market-oriented implicit interest rate only under
unrealistically low values for the life expectancy, say 5 years or shorter. Such low levels
may be relevant if the value of a PV is not su¢ ciently capitalized in house prices. However,
Dastrup et al. (2012) show that this is not the case based on evidence for California.
According to the right part of Figure 6, the estimated implicit interest rate decreases as we
assume a higher value for the deterioration rate in the production of electricity. However,
even an unrealistically high deterioration rate of 5% annually does not bring market interest
rate within the con…dence interval of our estimates.
We conclude that the estimated implicit interest rate would become close to market inter-
est rates only under unrealistic assumptions regarding the durability of the PV technology.

Consumer expectations about government’s commitment A second explanation for

the high implicit interest rate is that consumers may fear that the government will not ful…ll
its commitments to the subsidy policy. The government had guaranteed the net metering
principle for the life time of a PV (assumed to be 20 years), and had similarly guaranteed
the payment of the GCC subsidies for a …xed number of years (10 to 20 years, depending on

34
 Figure 6: Estimated implicit interest rate under di¤erent investment assumptions
.2

.2
.15

.15
.1

.1
r

r
.03 .05

.03 .05
0

0 5 10 15 20 25 30 0 5 10 12 15
Life expectancy of PV in years (R/12) % Yearly deterioration ( (1+λ) -1 )

Estimate 90% ci

Note: vertical line indicates assumption used in the baseline model

the date of installation). Figure 7 shows how the estimated implicit interest rate varies as
consumers expect a di¤erent duration for net metering bene…ts or GCC subsidies, i.e. when
we either change the value of RE or RtG in (2).35
Changes in expectations about the duration of net metering do not a¤ect the estimated
implicit interest rate. In contrast, a change in expectations about the duration of the GCC
subsidies does have an impact on the results. If consumers fear that the government will
remove the 20 year subsidy program already after 5 years, the estimated interest rate comes
close to market rates. Hence, one could in principle rationalize consumer behavior if they
expect that the government will breach the contract by removing the subsidies after a short
period. We note however that such a breach in contract would have legal consequences and
has in fact not actually occurred.
35
A breach in both contracts is equivalent to the change in the lifetime of a PV, which we considered
earlier in Figure 6.

35
Figure 7: Estimated implicit interest rate under di¤erent beliefs for the duration of gover-
ment’s commitments

.2

.2
.15

.15
.1

.1
r

r
.05

.05
.03

.03
0

0 5 10 15 20 25 30
E
0 5 10 15 20 25 30G
Expected years govt commits to net metering (R /12) Expected years govt commits to GCC subsidies (R t /12)

Estimate 90% ci

Note: vertical line indicates assumption used in the baseline model

A more realistic scenario to account for consumers’concerns about the lack of government
commitment is the introduction of a grid fee, i.e. an access fee to transfer the generated
electricity to the network. In July 2015, the government in fact introduced such a fee, after
an earlier failed attempt in 2013 (declared illegal by a Brussels Court of Appeal). The
annual fee amounted to about e92 per kW of capacity (hence an annual fee of e368 for
a household with the most common capacity of 4kW). In principle, the government could
have introduced an even higher grid fee, but the incentive to do so is limited as it would
discourage new adopters who would also have to pay the grid fee (and do not bene…t from
any subsidy program in contrast to old adopters).
Since the grid fee was introduced well after the last month of our sample (December
2012), our base speci…cation assumed households did not anticipate such a fee. However,
rumors (and failed government attempts) may have in‡uenced consumer expectations. Note
that the expectation of a grid fee does not a¤ect the estimated discount factor if the expected

36
level and time frame of introduction remain constant over time (as alternative assumptions
on the level or time frame would then be absorbed in the choice-speci…c …xed e¤ects). We
therefore assess how the estimated discount factor changes when the expected time frame
or level of the grid fee changes. The left part of Figure 8 shows the estimated interest rate,
assuming that consumers expect the grid fee of e92 per kW to be introduced in di¤erent
months since January 2013 (the …rst month after the end of our sample). This shows that
the estimated interest rate slightly increases as the expected introduction of the grid fee
moves closer to January 2013. The right part of Figure 8 shows the estimated interest rate,
assuming consumers expect a grid fee in July 2015 (when it actually happened) for various
possible levels of this grid fee. We see that for larger values of this expected grid fee, the
estimated interest rate increases. Intuitively, accounting for consumer expectations of an
earlier and/or higher grid fee results in higher estimates for the interest rate, because with
such expectations forward-looking consumers would have adopted earlier to take advantage
of the months without a grid fee.

Figure 8: Estimated implicit interest rate under di¤erent beliefs about grid fee
.2

.2
.15

.15
r

r
.1

.1
.03 .05

.03 .05
0

2013m1 2018m1 2023m1 2028m1 2033m1 0 50 100 150 200

Start date grid fee of 92 EUR Amount grid fee in EUR 2013, introduced July 2015

Estimate 90% ci

Note: vertical line indicates assumption used in the baseline model

37
Uncertainty or intrinsic undervaluation (“myopia”) In the above we assessed whether
uncertainty about future payo¤s (e.g. because of uncertainty about government’s commit-
ment) could be responsible for the high implicit interest rates.36 A remaining explanation
for the high implicit interest rate would be that this is evidence for intrinsic undervaluation
or consumer myopia. It is then still interesting to ask where such myopia might come from.
A …rst possibility is that consumers take into account only the future GCC subsidies but
fail to take into account the tax cuts. Another possibility is that consumers correctly value
the bene…ts only up to the pay-back period, and undervalue the bene…ts after that. The
pay-back period is that time when all collected bene…ts are equal to the investment costs.
This number is often quoted in advertising or media coverage, so it may be an important
source of information for households who cannot do a net present value calculation. Figure
9 shows how the estimated implicit interest rate varies if consumers do not correctly account
for the tax cuts or for the bene…ts after the pay-back period.
To assess the role of an incorrect valuation of the tax cuts, we multiply the tax cut
bene…ts by a parameter between 0 and 100%. The estimated implicit interest rate remains
high even for quite severe undervaluation of the tax cuts. Hence, a failure to take into account
the tax cuts may partly explain household myopia, but the high interest rate is also due to
the undervaluation of the GCC bene…ts.
To assess the role of the payback period, we multiply the bene…ts after the payback period
by another parameter between 0 and 100%. The estimated implicit interest rate becomes
close to the market interest rate only for strong undervaluation after the payback period
(less than 20% of the actual bene…ts).

In sum, our …nding of a high implicit interest rate remains robust after using more con-
servative assumptions regarding the durability of the PV technology. Potential explanations
for the substantial time discounting are consumer distrust in the government’s commitment
to provide the GCC subsidies for up to 20 years, or intrinsic consumer myopia, for example
stemming from a failure to take into account bene…ts after the payback period.
36
Another source of uncertainty may be the expected amount of electricity production, which may vary
from year to year. However, annual statistics show that the 95% con…dence interval bounds for the total
number of hours of sunshine are only 4.6% higher or lower than the average, implying uncertainty is limited
over a 15- or 20-year period.

38
 Figure 9: Estimated implicit interest rate under consumer myopia

.2

.2
.15

.15
r

r
.1

.1
.05

.05
.03

.03
0

0
0 10 20 30 40 50 60 70 80 90 100 0 20 40 60 80 100
% Valuation of tax cuts % Valuation of benefits after payback period

Estimate 90% ci

Note: vertical line indicates assumption used in the baseline model

4.4 Upfront investment subsidies instead of future production sub-

sidies
Our …nding that consumers use a real implicit interest rate of 15% when deciding to adopt
a PV system has an important policy implication. One may ask the question whether the
government could not have achieved the same level of adoption at a lower budgetary cost
by removing the future GCC subsidy program and instead paying an equivalent upfront
subsidy. It could then borrow the required amount to …nance the upfront subsidy on the
capital market at the long run government bond real interest rate of 3%.
More precisely, according to the utility speci…cation (2) and (3), a household who adopts
a PV system j at time t perceives a net present value from the GCC subsidy during RtG
months of G
P ERC 1 ((1 )(1 ) )Rt GCC
N P Vj;t = pj;t ;
1 (1 )(1 )
where the estimated monthly discount factor = 0:9884 corresponds to an implicit annual

39
interest rate of r = 12 1 = 15:00% (according to our baseline results of Table 2). The
P ERC
government could thus have paid out the households’ perceived amount N P Vj;t as an
upfront subsidy program and obtained the same adoption rate. Because the government in-
stead spread the subsidies over the next RtG months, the net present value at the government
bond interest rate rgov = gov12 1 = 3% amounted to
RtG
ACT U AL
1 (1 )(1 ) gov
N P Vj;t = pGCC
j;t :
1 (1 )(1 ) gov

Hence, the government could have reached an identical number of adopters with an upfront
P ERC ACT U AL P ERC
subsidy N P Vj;t and saved the amount N P Vj;t N P Vj;t for a household that
adopts PV system j at time t. Summing this over all adopters and all PV systems, we …nd
that the cost of the actual subsidy program was e 3:79 billion in net present value terms,
while the cost of an upfront subsidy program would have been only e 1:87 billion (actualized
to 2013). Hence, the government could have achieved the same adoption rates at only 49%
of the current subsidy costs, amounting to a saving of e 1:92 billion (with a 90% con…dence
interval of [e 1:48– e 2:22] billion).37 This is a saving of more than e 700 per Flemish
household, which is a very large number given that only 8.3% of the households had adopted
a PV by December 2012. Note that savings might have been even larger if the government
would also have abandoned the net metering principle (future bene…ts through electricity
cost savings pEL
j;t ) in favour of an even larger upfront subsidy. However, such a policy may
create incentive problems, since households may be induced to invest in PVs even if they do
not have good investment conditions (such as a good roof orientation).38
How large should the upfront subsidy be to obtain these budgetary savings? The answer
to this question depends on the speci…c point in time, because the generosity of the GCC
subsidy program ‡uctuated over time. The blue line on Figure 10 plots the evolution of the
required upfront investment subsidy to avoid the expensive GCC system, as a percentage
of the investment price of an average sized PV of 4kW in each month.39 This shows that
the required investment subsidy varies between 37% and 51% over the period 2006-2011,
37
To calculate the con…dence interval, we take 1000 draws of which, as a GMM estimate, is normally
distributed with mean of 0:9884 and standard error of 0:0024. We calculate the government loss for each
draw of to obtain a distribution of this loss. Note that the estimated savings would also vary if we apply
the point estimates in our robustness analysis. In particular, they would be lower under the estimates of
the second speci…cation of Table 4 (which estimates a separate discount factor for GCC versus other future
bene…ts).
38
Savings may also have been larger if the government would also have followed an upfront subsidy policy
for the equally important commercial users (capacity size higher than 10kW). This would however require
further investigation, since it is possible that commercial users have a lower implicit interest rate.
39
The required percentage subsidy is slightly larger for larger PVs and slightly smaller for smaller ones.

40
 Figure 10: Counterfactual investment subsidy

1
.2 .3 .4 .5 .6 .7 .8 .9
Percentage of investment cost
.1
0

2006 2007 2008 2009 2010 2011 2012 2013

Counterfactual subsidy
Tax cut+counterfactual subsidy

Note: investment cost extrapolated before May 2009 using predicted values from price index EUPD

but drops to 15% at the end of the program. The red line shows the total required upfront
subsidy, i.e. including the tax credit which the government already applied.40 The total
upfront subsidy required to avoid the expensive GCC system varied around 55% in the …rst
half of the period. It then increased to around 80% until the end of 2011. Afterwards, it
coincides with the other line as the tax cuts were abolished. In sum, large upfront investment
subsidies (of up to 82%) are required to obtain the large budgetary savings from removing
the GCC subsidy program. While this might seem paradoxical, it simply illustrates how
generous the GCC system was.
This is because GCC subsidies are proportional to the capacity of a PV, while investment costs exhibit small
returns to scale.
40
In 2006 and 2007, the Flemish government also applied a small investment subsidy. We included this in
the tax cut component of this graph.

41
5 Conclusion
This paper studied the incentives to adopt a new renewable energy technology for electricity
production, and the role played by upfront investment and future production subsidies. We
considered a generous subsidy program for solar PV adoption, and exploited rich variation at
pre-announced dates in the future subsidy conditions. Although the program led to a massive
adoption of solar PV systems, we …nd that households signi…cantly undervalued the future
bene…ts from the new technology, which has important budgetary and distributional impli-
cations. The government could have saved 51% or e 1:9 billion by giving upfront investment
subsidies, and it essentially shifted the subsidy burden to future electricity consumers.
We contribute to the literature on how consumers discount future energy costs. Recent
evidence points to moderate undervaluation to correct valuation for energy saving invest-
ments of existing, mature technogies (such as cars). Our …ndings indicate that consumers
may discount the future bene…ts more when adopting an entirely new green technology.
We adopted a tractable dynamic model of technology adoption, and several directions
of future work are possible. First, in our sensitivity analysis we found little heterogeneity in
discounting across consumers. If such heterogeneity is more important, subsidization policies
would have additional distributional e¤ects, and may need targeting to consumers with a
low discount factor. Another path of research is to extend the model to account for peer
e¤ects, which may provide a rationale for a subsidy path that is declining over time.
Third, it would be interesting to use our framework to study the adoption of new tech-
nologies in other applications. Regarding renewables, we focused on residential PV adoption,
and further work could investigate whether commercial PV adopters discount future bene…ts
in the same way. It would also be interesting to apply our framework to other countries or
regions, or to other renewable technologies, such as wind power, to analyze how di¤erent
subsidy schemes may in‡uence the outcomes.

References

Abbring J H & Daljord Ø (2017). Identifying the Discount Factor in Dynamic Discrete
Choice Models, Working paper .
Allcott H & Greenstone M (2012). Is There an Energy E¢ ciency Gap?, Journal of Economic
Perspectives 26(1), 3–28.

42
Allcott H & Wozny N (2014). Gasoline Prices, Fuel Economy, and the Energy Paradox,
Review of Economics and Statistics 96(5), 779–795.
Arcidiacono P & Ellickson P B (2011). Practical Methods for Estimation of Dynamic Discrete
Choice Models, Annual Review of Economics 3, 363–394.
Arcidiacono P & Miller R A (2011). Conditional Choice Probability Estimation of Dynamic
Discrete Choice Models With Unobserved Heterogeneity, Econometrica 79(6), 1823–
1867.
Audenaert A, De Boeck L, De Cleyn S, Lizin S & Adam J F (2010). An economic evaluation of
photovoltaic grid connected systems (PVGCS) in Flanders for companies: A generic
model, Renewable Energy 35(12), 2674–2682.
Berry S (1994). Estimating Discrete-Choice Models of Product Di¤erentiation, The RAND
Journal of Economics 25(2), 242.
Berry S, Levinsohn J & Pakes A (1999). Voluntary Export Restraints on Automobiles: Eval-
uating a Trade Policy, American Economic Review 89(3), 400–430.
Berry S, Levinsohn J & Pakes A (2004). Di¤erentiated Products Demand Systems from a
Combination of Micro and Macro Data: The New Car Market, Journal of Political
Economy 112(1), 68–105.
Bollinger B (2015). Green technology adoption: An empirical study of the Southern Califor-
nia garment cleaning industry, Quantitative Marketing and Economics 13(4), 319–
358.
Bollinger B K & Gillingham K (2014). Learning-by-Doing in Solar Photovoltaic Installations,
Working paper .
Burr C (2016). Subsidies and Investments in the Solar Power Market, Working paper .
Busse M R, Knittel C R & Zettelmeyer F (2013). Are Consumers Myopic? Evidence from
New and Used Car Purchases, American Economic Review 103(1), 220–56.
Chamberlain G (1987). Asymptotic E¢ ciency in Estimation with Conditional Moment Re-
strictions, Journal of Econometrics 34, 305–334.
CREG (2010). De verschillende ondersteuningsmechanismen voor groene stroom in Belgie,
Technical report.
Dastrup S R, Gra¤ Zivin J, Costa D L & Kahn M E (2012). Understanding the Solar Home
price premium: Electricity generation and Green social status, European Economic
Review 56(5), 961–973.

43
De Groote O, Pepermans G & Verboven F (2016). Heterogeneity in the adoption of photo-
voltaic systems in Flanders, Energy Economics 59, 45–57.
Dube J P H, Hitsch G J & Jindal P (2012). The Joint Identi…cation of Utility and Discount
Functions From Stated Choice Data: An Application to Durable Goods Adoption,
Working paper .
Eurobserv’er (2013). Photovoltaic Barometer, Technical report.
Feger F, Pavanini N & Radulescu D (2017). Welfare and Redistribution in Residential Elec-
tricity Markets with Solar Power, Working paper .
Gandal N, Kende M & Rob R (2000). The Dynamics of Technological Adoption in Hard-
ware/Software Systems: The Case of Compact Disc Players, The RAND Journal of
Economics 31(1), 43.
Gandhi A, Lu Z & Shi X (2017). Estimating Demand for Di¤erentiated Products with Zeroes
in Market Share Data*, Working paper .
Gerarden T, Newell R G & Stavins R N (2017). Assessing the Energy-E¢ ciency Gap, Journal
of Economic Literature 55(4), 1486–1525.
Gowrisankaran G & Rysman M (2012). Dynamics of Consumer Demand for New Durable
Goods, Journal of Political Economy 120(6), 1173–1219.
Hausman J A (1979). Individual Discount Rates and the Purchase and Utilization of Energy-
Using Durables, The Bell Journal of Economics 10(1), 33.
Hotz V J & Miller R A (1993). Conditional choice probabilities and the estimation of dynamic
models, The Review of Economic Studies 60(3), 497–529.
Kwan C L (2012). In‡uence of local environmental, social, economic and political variables
on the spatial distribution of residential solar PV arrays across the United States,
Energy Policy 47, 332–344.
Langer A & Lemoine D (2018). Designing Dynamic Subsidies to Spur Adoption of New
Technologies, Working paper .
Lee R S (2013). Vertical Integration and Exclusivity in Platform and Two-Sided Markets,
American Economic Review 103(7), 2960–3000.
Magnac T & Thesmar D (2002). Identifying dynamic discrete decision processes, Economet-
rica 70(2), 801–816.
Manski C F (1993). Dynamic choice in social settings: Learning from the experiences of
others, Journal of Econometrics 58(1), 121–136.

44
Melnikov O (2013). Demand For Di¤erentiated Durable Products: The Case Of The U.S.
Computer Printer Market, Economic Inquiry 51(2), 1277–1298.
Murray B C, Cropper M L, de la Chesnaye F C & Reilly J M (2014). How E¤ective are US
Renewable Energy Subsidies in Cutting Greenhouse Gases?, American Economic
Review 104(5), 569–574.
Newey W K (1990). E¢ cient Instrumental Variables Estimation of Nonlinear Models, Econo-
metrica 58(4), 809.
Nurski L & Verboven F (2016). Exclusive Dealing as a Barrier to Entry? Evidence from
Automobiles, The Review of Economic Studies 83(3), 1156–1188.
Petrin A (2002). Quantifying the Bene…ts of New Products: The Case of the Minivan,
Journal of Political Economy 110(4).
Quan T W & Williams K R (2018). Product Variety, Across-Market Demand Heterogeneity,
and the Value of Online Retail, RAND (forthcoming).
Reynaert M & Verboven F (2014). Improving the performance of random coe¢ cients demand
models: The role of optimal instruments, Journal of Econometrics 179(1), 83–98.
Rust J (1987). Optimal Replacement of GMC Bus Engines: An Empirical Model of Harold
Zurcher, Econometrica 55(5), 999–1033.
Rust J (1994). Structural estimation of Markov decision processes, Handbook of econometrics
4(4).
Scott P T (2013). Dynamic Discrete Choice Estimation of Agricultural Land Use, Working
paper .
Silva J M C S & Tenreyro S (2006). The log of gravity, The Review of Economics and
Statistics 88(4), 641–658.
Verboven F (2002). Quality-based price discrimination and tax incidence: evidence from
gasoline and diesel cars, RAND Journal of Economics pp. 275–297.
Yao S, Mela C, Chiang J & Chen Y (2012). Determining consumers’discount rates with …eld
studies, Journal of Marketing Research 30(3), 447–468.

45
A Appendix
A.1 Data construction
As discussed in the text, the main dataset contains information of all installed PVs across
Flanders during 2006–2012. We combine this dataset with various additional datasets on
prices, investment tax bene…ts, electricity prices, GCCs and socio-demographic data at the
local market level.

A.1.1 PV installations

The main dataset comes from VREG, the Flemish regulator of the electricity and gas market.
The data records the following three key variables for every new PV installation: the adoption
date, the size of the installation and the address of the installation. We aggregate the data to
the monthly level, distinguishing between …ve categories of capacity sizes: 2kW, 4kW, 6kW,
8kW and 10kW. Each category includes all capacity sizes up to the indicated maximum.
For example, a capacity size of 6kW refers to all capacity sizes between 4kW and 6kW.
To focus on residential solar panels, we exclude all installations with a capacity size larger
than 10kW. This is a commonly used cut-o¤ point for distinguishing between residential and
non-residential PVs (see e.g. Kwan (2012)). Furthermore, systems of more than 10kW do
not qualify from the same public support measures in Flanders.
Our main model aggregates the number of installations to the level of the entire region
of Flanders. The extended model considers the highly disaggregate level of the statistical
sector, as de…ned by ADSEI, the Belgian statistical o¢ ce. The region has 9,182 statistical
sectors, with on average 295 households. To organize the data at the level of the statistical
sector, we use of a geographic dataset from ADSEI that assigns street addresses of each
installation to statistical sectors.

A.1.2 Gross investment price

We obtained price information of PV systems from two independent sources: an internet fo-
rum, zonstraal.be, where consumers posted their quotes; and a website, comparemysolar.be,
which contains historical data. This resulted in a dataset of 2,659 o¤ers from May 2009
until December 2012. To construct a monthly price index for each of the …ve capacity size
categories (between 2kW and 10kW), we proceeded as follows. For each month and each
size category we take the median price per watt, multiplied by the size of the category. If
there are less than ten price observations in a given month and category (usually the less
popular 8kW and 10kW PVs), we consider the median to be insu¢ ciently accurate. As a

A1
price measure for these cases, we use the prediction from a quantile regression model for the
median price per watt on monthly …xed e¤ects, capacity …xed e¤ects and capacity interacted
with a linear time trend.
To combine the price information with the data on PV installations per month and per
size category, we assume there was a time lag of two months between the posted prices and
the actual installment. In some months, especially when subsidies would drop in the near
future, consumers reported the expected waiting time together with the posted price o¤er.
If such information on the announced waiting time was available, we use this instead of the
assumption of a two month time lag.

A.1.3 Public support measures

We obtained information of public support measures from various sources.

Investment tax credits Tax credits fall under the competence of the Belgian Federal
government. Information on a doubling of the tax credit ceilings comes from the o¢ cial
document “Programmawet”of 28 December 2006, and announcements on the website of the
government agency VEA before and after this publication.41 Information on spreading tax
cuts or splitting bills over multiple years comes from newspaper articles42 and the Economic
Recovery Plan of the Federal Government (March 2009). Details about the abolishment
of the tax cut were found on the o¢ cial website of the …nance department of the federal
government.43 Information on the VAT rules also can be found on this website.44
We combine this information with the price data to compute the net investment price,
as described more formally in section 3.1.

Net metering and Green Current Certi…cates (GCCs) Information on retail

electricity prices comes from Eurostat. These data are half-yearly, and we transform it to
monthly data using cubic spline interpolation. We multiply the electricity prices with the
expected electricity production to compute the expected electricity cost savings from net
metering, as described more formally in section 3.1.
41
Announcements on the doubling of the tax credit ceiling on 6 and 16 December 2006 and information on
the increase from 2000 to 2600e between 1 and 21 March 2007 on VEA’s website energiesparen.be. Historic
copies from this website are on Internet Archive (https://web.archive.org).
42
Gazet Van Antwerpen: “Zonnepanelen zijn tot drie keer …scaal aftrekbaar”, 19 Mei 2008; Het Nieuws-
blad: “Belastingvoordeel klanten nekt installateurs zonnepanelen”, 13 December 2008
43
http://www.min…n.fgov.be/portail2/nl/current/spokesperson-11-11-30.htm, consulted 14 May 2014.
44
http://min…n.fgov.be/portail2/nl/themes/dwelling/renovation/vat.htm, consulted 14 May 2014.

A2
 Information on the background and start of the GCC policy relating to PVs in 2006 comes
from the website of the Flemish energy regulator VREG (www.vreg.be) and from o¢ cial
documents and government information brochures.45 The price of a GCC was guaranteed
for a …xed period, but it was initially expected that GCCs could continue to be sold at
the (much lower) market price for the entire life time of the PV system. The renewal of
the energy decree in 2012 (Flemish Energy Decree, 30 July 2012) no longer allowed for the
possibility to obtain GCCs after the expiration of the …xed period with the guaranteed price.
In practice, this does not change much because the life expectancy of PV systems (about 20
years) is close to the …xed period with the guaranteed price.
Information on the …nancial details of the GCC policy comes from the Belgian energy
regulator CREG (2010). Announcements of new subsidy policies were gathered from newspa-
pers. The …rst change in policy was announced in February 2009 (De Standaard, 7 February
2009, p2) for PVs installed from 2010 on. The second change was announced in June 2011
(De Standaard, 6 June 2011, Economie p12) for PVs from July 2011 on. The third change
was announced in May 2012 (De Standaard, 26 May 2012) for PVs installed from August
2012 on and the …nal change was in July 2012 (Degree proposal amending the Energy Decree
of 8 May 2009 (6 July 2012) and Energy decree 8 May 2009, changed 30 July 2012) for PVs
installed from 2013 on.
Based on the information from these sources, Table A1 provides an overview of the policy
support measures during the period 2006–2012 (and the …rst months of 2013). Figure 1 in
the text makes use of this information to express the various subsidies in present value terms.

45
See the Flemish Energy Decree, changed on 6 July 2012, KB 10 February 1983, changed by the Flemish
government on 15 July 2005, 16 June 1998: “Besluit van de Vlaamse Regering tot wijziging van het koninklijk
besluit van 10 februari 1983 houdende aanmoedigingsmaatregelen voor het rationeel energieverbruik.” The
latter also included information about the investment subsidies of which more information was found in a
government brochure “Subsidieregeling voor elektriciteit uit zonlicht” (2005).

A3
 Table A1: PV support policy Flanders: 2006-2013/06
Date of investment GCC Subsidy Tax cut on investment
Price Duration Percentage Ceiling
(EUR) (years) (EUR 1988)
2006 450 20 10% 40% 1000
2007 450 20 10% 40% 2600*
2008 450 20 0% 40% 2600
2009 450 20 0% 40% 2600 x 4**
2010 350 20 0% 40% 2600 x 4**
2011/01-2011/06 330 20 0% 40% 2600 x 4**
2011/07-2011/09 300 20 0% 40% 2600 x 4**
2011/10 - 2011/12 270 20 0% 40%*** 2600 x 4***
2012/01 - 2012/03 250 20 0% 0% 0
2012/04 - 2012/06 230 20 0% 0% 0
2012/07 210 20 0% 0% 0
2012/08 - 2012/12 90 10 0% 0% 0
2013/01-2013/06 21.39**** 15 0% 0% 0

*Announced as 2000 but changed to 2600. New announcement made: 18 March 2007.
** If house > 5years old, the tax cut could be spread over 4 years. Announced March 2009.
*** Contract had to be signed before 28 November 2011. Announced on the same date.
**** Corrected for banding factor

A.1.4 Socio-demographic characteristics

For the disaggregate model at the local market level we collected socio-demographic infor-
mation per statistical sector. This data is freely downloadable from the website of ADSEI,
the Belgian Statistics O¢ ce. We used population data for each statistical sector in 2011
to create the following variables: population density, average house size (number of rooms),
average household size, average house age, median income, % of home owners, % with a
higher education degree and % foreign (people who do not have the Belgian nationality).
For con…dentiality reasons, some variables are not reported when the number of households
in the statistical sector is very small. This applies to a small subset of statistical sectors. In
these cases, we use the average of the municipality to which the statistical sector belongs.

A4
A.1.5 Exogenous instruments

Two variables we use do not directly in‡uence the adoption decision of households, but
we use them as instruments for endogenous variable that do a¤ect the decision. The …rst
exogenous instrument is the price index for Chinese Crystalline PV modules of "pvxchange"
that is available on their website. The prices are per kW so we multiply them by the kW of
each category to create pM j;t
OD
. In the discussion on optimal instruments, we also added the
oil price as an additional exogenous instrument. The price of crude oil was obtained from
Thomson Reuters Datastream. As with other price variables in the model, we correct for
in‡ation by using the HICP.

A.2 Optimal instruments

We estimate the model using an approximation of Chamberlain’s (1987) optimal instruments.
While any set of exogenous instruments leads to consistent estimates, more e¢ cient and
stable estimates can be found using approximations to optimal instruments . In this section
we discuss the optimal instruments in the model that uses only macro data, i.e. ignoring
local market heterogeneity. In the next section, which provides details on how we estimate
the model when local market data are added, we discuss how we adapt optimal instruments
in this case.

De…ning the parameter vector = ( ; ; ), the conditional moment conditions are

E (ej;t ( )jzj;t ) = 0
where
ej;t ( ) = ln Sj;t =S0;t (xj;t x1;t+1 ) + (pj;t ( ) p1;t+1 ( )) ln S1;t+1 (17)
The optimal instrument matrix of Chamberlain (1987) for a single-equation GMM estimator
is:
1
gjt (zjt ) = Djt (zjt )0 jt

with jt = E[(ej;t )2 jzjt ]

@ej;t ( )
Djt (zjt ) = E zjt
@ 0
@ej;t ( ) @ej;t ( ) @ej;t ( )
= E zjt E zjt E zjt
@ @ @ 0
In our approximation, we follow Newey (1990) and set jt = ; i.e. we ignore potential
heteroscedasticity. Moreover, since is a scalar in the single-equation GMM estimator, we
can also replace it by the identity matrix.

A5
 We now derive the optimal instruments for these various parameters. First, for the linear
parameter vector we simply have:

@ej;t ( )
E zjt = E [xj;t x1;t+1 jzjt ] = (xj;t x1;t+1 ) : (18)
@ 0

The optimal instrument for is therefore just a di¤erence term for the exogenous variable
xj;t , where is substituted by an estimate b in a …rst stage using non-optimal instruments.
For the other linear parameter we have

@ej;t ( )
E zjt = E [pj;t ( ) p1;t+1 ( )jzjt ] = E [pj;t ( )jzjt ] E [p1;t+1 ( )jzjt ] : (19)
@

In this expression the conditional expectation of price is

E [pj;t ( )jzjt ] = E pIN V

j;t ( )jzjt
G
t ( ) E pGCC
j;t jzjt
E
( ) E pEL
j;t jzjt
X
4
12
= E pGROSS
j;t jzjt E taxcutj;t jzjt
=1
G
t ( ) pGCC
j;t
E
( ) E pEL 0
t jzjt kj (20)

where the capitalization factors G t ( ) and

E
( ) are de…ned in (2) and depend on the dis-
EL 0
count factor . pj;t is the electricity price per MWh, multiplied by kj ; the monthly electricity
production of a PV with capacity kj : The optimal instrument for thus also depends on
for which we use an estimate b in a …rst stage using non-optimal instruments. In contrast
with the optimal instrument for , it is now also necessary to compute several conditional
expectations, namely for the upfront investment cost of a solar panel, the future tax cuts
and the electricity price. The predicted gross investment cost E pGROSS
j;t ( ) zjt is obtained
from a constant elasticity model, using a Poisson regression and logarithmic regressors (see
Silva and Tenreyro (2006)). Based on this predicted value we can also calculate the predicted
future eligible tax cuts E taxcutj;t jzjt . The predicted electricity price E pEL
t jzjt is simi-
larly obtained using the oil price as an exogenous regressor. We show the regression results
in Tables A2 and A3. Note that any misspeci…cation in‡uences only the optimality of our
instrument set and not the consistency of the structural estimates of our model.
Finally, the optimal instrument for the nonlinear parameter is

@ej;t ( )
E zjt = x1;t+1 E [ ln S1;t+1 j zjt ]
@
@pj;t ( ) @p1;t+1 ( )
+ E zjt E [p1;t+1 ( )j zjt ] E zjt (21)
:
@ @

A6
In the above expression the expected value of the derivative of price with respect to is

@pj;t ( ) X
4
12 1
E zjt = 12 E taxcutj;t jzjt
@ =1
@ G
t ( ) @ E
( )
pGCC
j;t E pEL 0
t jzjt kj
@ @

where the derivatives with respect to the capitalization factors Gt ( ) and

E
( ) are easily
computed from (2) and (3). The optimal instrument for therefore depends on all para-
meters = ( ; ; ), for which we obtain a consistent …rst stage estimate using non-optimal
instruments. There is also an additional expectation term for the CCP term, i.e. the log
of the predicted next period market share of alternative 1, E [ln S1;t+1 jzjt ]. We obtain this
from a linear regression on several variables, similar to the prediction of the …rst stage of
an IV regression, as shown in Table A4. Note that by using future values of exogenous
instruments, we assume that these variables are not correlated with the demand shock or
prediction error at time t. Therefore, they must be known at time t. Since we are using only
one and two month leads, we believe this is a reasonable assumption as new policies were
announced several months ahead (see section A.1).
To summarize, our …nal estimation procedure takes the following steps:

Estimate a GMM model with instruments pM

j;t
OD GCC
; pj;t and xj;t to obtain an initial
consistent estimate of ; and

Compute the conditional expectations for the investment price, the electricity price
and the CCP term using the regression models

Estimate the GMM model again, but now using the approximation of optimal instru-
ments, as given by (18), (19) and (21), after substituting (20) and the initial consistent
estimates of ; and .

A7
 Table A2: Estimation results for electricity price
Variables E pEL
t jzjt

Log of oil price 0.183***

(0.018)
Constant 4.599***
(0.073)

Observations 44
Poisson regression model of exponential conditional mean
Standard errors in parentheses, clustered within time period
*** p<0.01, ** p<0.05, * p<0.1

Table A3: Estimation results for PV investment price

Variables E pGROSS
j;t jzjt

Log of PV module price x kW 0.499***

(0.063)
4kW 0.202***
(0.021)
6kW 0.310***
(0.031)
8kW 0.400***
(0.039)
10kW 0.468***
(0.045)
Log of GCC bene…ts 0.112*
(0.058)
Constant 4.631***
(0.316)

Observations 220
Poisson regression model of exponential conditional mean
Standard errors in parentheses, clustered within time period
*** p<0.01, ** p<0.05, * p<0.1

A8
 Table A4: Estimation results for CCP correction term
Variables E [ln s1;t+1 jzjt ]

PV module price x 4kW in t+1 -0.001**

(0.001)
PV module price x 4kW in t+2 0.001
(0.001)
GCC bene…ts of 4kW in t+1 0.116***
(0.019)
GCC bene…ts of 4kW in t+2 -0.054***
(0.019)
Oil price x 4 kW in t+1 0.006
(0.009)
Oil price x 4 kW in t+2 0.003
(0.008)
Constant -12.995***
(2.485)

Observations 44
OLS regression model of linear conditional mean
Standard errors in parentheses, clustered within time period
*** p<0.01, ** p<0.05, * p<0.1

A9
A.3 Estimation of model with local market heterogeneity
Section 3.4 speci…ed the model with local market heterogeneity. We estimate this model
using a GMM estimator that combines macro and micro-moments at the local market level.
This is in the spirit of the static discrete choice literature, as in Petrin (2002) and Berry et
al. (2004), and applied to local market data in Nurski and Verboven (2016).
First, we explain how one could proceed when the discount factor is known, i.e. does
not need to be estimated. In this case it is possible to estimate the impact of local market
heterogeneity and of the mean utility determinants in two separate steps. Second, we explain
how to proceed if the discount factor is not known, i.e. needs to be estimated. This also
includes a discussion of how we implement optimal instruments and some …nal estimation
details.

A.3.1 Estimation when the discount factor is known

Step 1. Maximum likelihood estimation including …xed e¤ects ej;t

In this step we construct the likelihood function of observing the local market adoption
data, and we maximize this likelihood function with respect to the parameters, including
a large set of alternative/time …xed e¤ects ej;t , de…ned below. We …rst make use of the
Hotz-Miller inversion to obtain an expression for vi;0;t that is parallel to that of (10) above:

vi;0;t = (vi;1;t+1 ln sm;1;t+1 t) : (22)

Note that this assumes that a household’s prediction error is common across local markets,
i.e. t V m;t+1 Et V m;t+1 . We then use the expressions for the conditional values vi;j;t and
vi;0;t , as given by (15) and (22), to write the choice probabilities as:

exp(vi;j;t )
sm;j;t e; = PJ
j 0 =0 exp(vi;j 0 ;t vi;0;t )
exp(vi;j;t vi;0;t )
= PJ
1 + j 0 =1 exp(vi;j 0 ;t vi;0;t )
exp(ej;t + w ej;t m + ln sbm;1;t+1 )
= PJ (23)
1 + j 0 =1 exp(ej 0 ;t + w
ej 0 ;t m + ln sbm;1;t+1 )

where we de…ne ej;t j;t ( 1;t+1 ej;t wj;t

t ) and w w1;t+1 , and sbm;1;t+1 is a predicted
value of the next period choice probability of j = 1, discussed in the next paragraph. The
current choice probabilities sm;j;t e; are thus a function of the alternative/time …xed

A10
e¤ects ej;t (collected in the vector e) and of the local market interaction e¤ects m (collected
in the parameter matrix ).
Note that the right hand side of (23) also depends on a predicted value of the next period
probabilities sbm;1;t+1 , which are treated as data from a …rst-stage estimation. In contrast to
the model with only aggregate data, we no longer accurately observe the CCP correction
term ln sm;1;t+1 directly due to the small number of households in each statistical sector m.
In many local markets adoption rates are zero, so that the CCP correction term would be
unde…ned. We therefore use a …rst-stage prediction of the CCP correction term, sbm;1;t+1 ,
based on a ‡exible logit. We include local market …xed e¤ects, capacity …xed e¤ects for each
time period, capacity-speci…c e¤ects for each demographic, and capacity-time-speci…c e¤ects
for the demographics that enter the price parameter. We then use the parameters of this
model to calculate the predicted market shares for j = 1 in every time period and use the
predictions in t + 1 in the conditional value functions at time t.
The maximization problem of the log likelihood function is then
X
T X
M X
J
max ln L(e; ) = qm;j;t ln sm;j;t (e; )
e;
t=1 m=1 j=0

where qm;j;t is the observed number of households in local market m that adopt (j = 1; :::J) or
choose not to adopt (j = 0) at period t. This is similar to a maximum likelihood estimator
that sums over individual data but since ln sm;j;t (e; ) is identical for each household in
market m, we can multiply it by the number of households that make each choice. Note that
this contains a potentially large number of parameters, because of the set of alternative/time
…xed e¤ects ej;t (J T ), but also a large number of parameters in due to the inclusion of
local market …xed e¤ects.

Step 2. Instrumental variables regression of ej;t

The second step is an instrumental variable regression of the estimated …xed e¤ects ej;t
j;t ( 1;t+1 t ) after substituting the expressions of j;t and 1;t+1 based on (1). This
gives the regression
ej;t = (xj;t x1;t+1 ) (pj;t p1;t+1 ) + ej;t for j = 1; :::J (24)

where ej;t was already de…ned before for the aggregate model as ej;t j;t ( 1;t+1 t ).
The IV regression then imposes the following moment conditions

E (zj;t ej;t ) = 0

Hence, this regression is very similar to the aggregate model. In the disaggregate model the
dependent variable consists of the estimated …xed e¤ects ej;t from the …rst step, while in

A11
the aggregate model the dependent variable, including the correction term, was ln Sj;t =S0;t
ln S1;t+1 . Price is given by (2), based on the imposed value of , and the instruments are
the same as the ones used before in the aggregate model (though one can reduce the number
of instruments, since the discount factor is treated as known).

Simultaneous GMM
Given the known discount factor , this two-step approach yields consistent estimates of all
parameters, but in the second step standard errors need to be corrected because the ej;t are
estimated values. Alternatively, this model can be estimated at once using a GMM estimator
that combines the scores of the likelihood function of the …rst step (micro-moments), with the
moment condition that is imposed by the IV regression of the second step (macro-moment).
The stacked vector of sample moment conditions is then
!
@ ln L(e; )=@(e; )
g(e; ; ; ) = PT PJ e; ;
t=1 j=1 zj;t ej;t

The score ln L(e; )=@(e; ) has an intuitive expression for the demographic parameters and
the …xed e¤ects:
@ ln L(e; ) X
M
qm;j;t
= Nm;t sm;j;t (e; )
@ej;t m=1
Nm;t

@ ln L(e; ) XT XM X
J
qm;j;t
= Nm;t sm;j;t (e; ) w h
ej;t Dm
@ h t=1 m=1 j=1
Nm;t

where Dm h
is demographic characteristic h in the vector Dm and h is a K 1 vector for
demographic characteristic h (one of the columns in ). The scores @ ln L(e; )=@ej;t (for
each j and t) are essentially conditions that the observed country-level market shares should
be equal to the predicted country-level market shares. The scores @ ln L(e; )=@ h (for
each demographic h) are moment conditions that the observed sales-weighted demographic
interactions should be equal the model’s predictions. Since we include dummy variables for
each local market in the ‡ow utility of a PV, it essentially also introduces a moment condition
that matches the total number of adoptions at the end of the sample predicted by the model
with that observed in the data. The GMM estimator minimizes g 0 W g with respect to the
parameters, where W is the weighting matrix.

A.3.2 Estimating the discount factor

When is known, a two-step procedure is possible because no parameter estimated in the

second step, enters the estimation in the …rst step. If also has to be estimated, this

A12
is no longer the case. The discount factor enters the local market shares directly as the
coe¢ cient in front of the CCP term (see (23)), but also implicitly in the interaction e¤ects of
demographic variables with the price variable. We therefore proceed with joint estimation.
The stacked vector of sample moment conditions then also depends on the discount factor
!
@ ln L(e; ; )=@(e; )
g(e; ; ; ; ) = P e; ; ;
j;t zj;t ej;t

Similar to the aggregate model, we now also need an extra instrument in zj;t to identify the
discount factor.

Optimal instruments
We again make use of the approximation to optimal instruments we discussed in section
A.2. However, due to the variation of the CCP correction term across local markets, the
error term, and therefore also the optimal set of instruments, is di¤erent. From (24) it follows
that the error term is now
ej;t (e; ; ; ) = ej;t (xj;t x1;t+1 ) + (pj;t ( ) p1;t+1 ( )) (25)
Notice the di¤erence with (17): ej;t has replaced ln Sj;t =S0;t ln S1;t+1 .Therefore the
derivative of the discount factor no longer depends on the CCP so that (26) replaces (21) in
the construction of the optimal instrument vector:
" #
@ej;t (e; ; ; )
E zjt = x1;t+1 (26)
@
@pj;t ( ) @p1;t+1 ( )
+ E zjt E [p1;t+1 ( )j zjt ] E zjt :
@ @

Estimation details
Our main speci…cation includes a full set of local market …xed e¤ects in . We then
exclude the local markets where adoption never occurred, because with the local market
…xed e¤ects these markets do not add any information to the likelihood function which we
use to construct the micro-moments of the model. To reduce the number of …xed e¤ects and
speed up the estimation procedure, we use a random sample of 50%. We also estimated an
alternative speci…cation with all local markets, but with a reduced number of 308 …xed e¤ects
at the municipality level and with household characteristics interacted with the constant.
This gave similar results to the speci…cation with a full set of local market …xed e¤ects.
To correct for the fact that within a local market observations are not independent over
time, we cluster the moments in the calculation of the covariance matrix. We also cluster
the macro moments within time periods.

A13
A.4 Additional results for robustness checks
A.4.1 Alternative terminal actions for CCP approach

Table A5: Robustness: terminal action

Terminal action: Terminal action: Terminal action:
2kW 4kW (used in paper) 6kW
Price sensitivity in 103 EUR ( ) -0.351*** (0.113) -0.470*** (0.098) -0.513*** (0.102)
Monthly discount factor ( ) 0.9870*** (0.0032) 0.9884*** (0.0025) 0.9906*** (0.0016)
12
Annual interest rate (r 1) 16.99%*** (4.68%) 15.09%*** (3.43%) 11.94%*** (2.10%)

Control variables ( )
Alternative-speci…c constant
Common constant -0.983 (15.425) -1.423 (16.38) -4.575 (19.325)
2kW -2.111*** (0.457) -1.828*** (0.562) -1.199** (0.531)
6kW -0.193 (0.484) -0.513 (0.595) -1.162** (0.565)
8kW -1.847** (0.942) -2.453** (1.158) -3.742*** (1.097)
10kW -1.747 (1.372) -2.605 (1.684) -4.507*** (1.592)

Hansen’s J (p-value for endogeneity) Exactly identi…ed Exactly identi…ed Exactly identi…ed
Obs. macro moments (JxTx terminal choices ) 220 x 1 220 x 1 220 x 1
Obs. micro moments (MxJxT) 0 0 0
Terminal action: Terminal action: Terminal action:
8kW 10kW All (joint estimation)
Price sensitivity in 103 EUR ( ) -0.542*** (0.112) -0.505*** (0.111) -0.422*** (0.046)
Monthly discount factor ( ) 0.9885*** (0.0018) 0.9882*** (0.0020) 0.9873*** (0.0007)
12
Annual interest rate (r 1) 14.85%*** (2.46%) 15.27%*** (2.81%) 16.62%*** (1.03%)

Control variables ( )
Alternative-speci…c constant
Common constant -2.599 (16.429) -1.270 (18.673) -10.158 (11.278)
2kW -1.734*** (0.416) -1.832*** (0.429) -2.044*** (0.129)
6kW -0.628 (0.432) -0.518 (0.448) -0.282** (0.136)
8kW -2.663*** (0.849) -2.453*** (0.879) -2.022*** (0.262)
10kW -2.890** (1.246) -2.591** (1.288) -1.990*** (0.399)

Hansen’s J (p-value for incorrect speci…cation) Exactly identi…ed Exactly identi…ed 31.726 (p= 0.2858)
Obs. macro moments (JxTx terminal choices ) 220 x 1 220 x 1 220 x 5
Obs. micro moments (MxJxT) 0 0 0
Notes: Standard errors clustered within 44 time periods. Instruments are approximations of optimal instruments
(Chamberlain, 1987). Standard errors of r obtained via delta method. *** p<0.01, ** p<0.05, * p<0.1

A14
A.4.2 Heterogeneous discount factor

This section …rst explains how we extend our model of local market heterogeneity to incor-
porate heterogeneity in the discount factor. Next, we present the empirical results.

Approach With a local market-speci…c discount factor m, the predicted local market
shares are given by the following generalization of (23):
exp(em;j;t + w ej;t m + m ln sbm;1;t+1 )
sm;j;t = PJ (27)
1 + j 0 =1 exp(em;j 0 ;t + w
ej 0 ;t m + bm;1;t+1 )
m ln s

where
em;j;t = (xj;t (28)
m x1;t+1 ) m (pj;t ( m ) m p1;t+1 ( m )) + j;t m 1;t+1 t :
| {z }
em;t

ej;t no
Note that we explicitly write a local market speci…c price coe¢ cient m , therefore w
longer contains interactions with the price variable. Suppose the discount factor is the
following function of H 1 vector of household characteristics Dm :

m = g ( 0 + Dm )
exp( 0 + Dm )
= ;
1 + exp( 0 + Dm )
where are parameters measuring how the discount factor varies with household charac-
teristics. This allows for a very ‡exible speci…cation of m and ensures that m 2 (0; 1), even
with continuous variables in Dm .
Apart from the non-linearity through which m enters (also through the term pj;t ( m )),
the key issue relates to the term em;t entering (28). This term contains interactions between
the market-speci…c discount factor m and the expectational error t . One approach would
be to discretize the vector of household characteristics Dm to D possible realizations or
“demographic groups”, d = 1; : : : ; D. One can then absorb the em;t with …xed e¤ects by
period t and group d, allowing us to also control for expectational errors t (d) by period t
and group d.
To make better use of the rich and continuous variables in Dm we also follow an alternative
approach. Let the term em;t be given by the following function of household characteristics
et (Dm ) g( 0 + Dm ) 1;t+1 t (Dm )
where t (Dm ) is a di¤erentiable function of Dm , re‡ecting an expectational error that may
vary across markets by demographics. We approximate et (Dm ) using the following …rst-
order Taylor expansion for et (Dm ) around the mean of Dm , which we normalize to 0:
et (Dm ) g( 0) 1;t+1 t (0) + ret (0)Dm ;

A15
where ret (0) is the 1 H gradient for each t at Dm = 0. A typical element of ret (0) is reht ,
yielding t-speci…c parameters to be estimated as interactions with each of the demographics
h
Dm . The main bene…t of this Taylor expansion is that et (Dm ) now depends linearly on Dm
in each time period.
We add the following scores as micro-moments to identify the discount factor parameters
h
(elements of ) and the parameters reht :

@ ln L XT X M XJ
qm;j;t @
sm;j;t (e; )
m;j;t 0 h
= Nm;t g ( 0 + Dm ) Dm
@ h t=1 m=1 j=1
Nm;t @ m

@ ln L XM XJ
qm;j;t
= Nm;t sm;j;t (e; ) Dm
h
;
@reht m=1 j=1
Nm;t

where m;j;t ( m) is the di¤erenced value function that enters the choice probabilities (27).

Findings Table A6 shows the empirical results. We allow for a very ‡exible speci…cation in
which the valuation of price, capacity and the discount factor depends on all demographics.46
This ‡exible speci…cation mainly aims to document the role of heterogeneity in the discount
factor, as summarized in Figure 5 and the corresponding discussion in the main text. The
coe¢ cients themselves are di¢ cult to interpret on a stand-alone basis, because we include
a large set of demographics in all valuation terms, which show multicollinearity and may
also capture other location characteristics. For example, home owners tend to have a higher
discount factor. Households with a higher income tend to have a lower discount factor,
perhaps because they have better investment opportunities or because the home ownership
variable also captures the impact of wealth.

46
We also considered a speci…cation where we do not rely on the Taylor approximation but instead discretize
the vector of household characteristics into eight groups according to below/above average income, percentage
foreigners and population density. The resulting distribution of the implicit interest rate is discrete but
otherwise comparable to our more ‡exible approach, with most mass at 14.7% and 90% of households has a
rate between 12.8% and 15.2%.

A16
 Table A6: Empirical results with heterogeneous discount factor
Interactions with Price sensitivity Index of monthly discount
capacity di¤erence in 103 EUR ( ) factor ( )
E¤ect at mean of demographics -0.487*** (0.105) 4.468*** (0.223)
Pop. density (104 inhab / m2 ) -0.738*** (0.076) -0.077*** (0.030) 0.010 (0.052)
Average house size 0.108*** (0.033) -0.034* (0.018) -0.058* (0.033)
Average household size -0.157* (0.094) -0.118*** (0.028) 0.157** (0.066)
Average house age (decades) -0.014 (0.013) -0.004 (0.006) 0.016 (0.011)
Median income (104 EUR) 0.187** (0.085) 0.097*** (0.024) -0.173*** (0.064)
% home owners -0.973*** (0.224) -0.178*** (0.062) 0.632*** (0.185)
% higher education -0.027 (0.175) 0.020 (0.085) -0.038 (0.146)
% foreign -0.126 (0.153) 0.172 (0.107) 0.456** (0.193)

Alternative-speci…c constants YES

Local market …xed e¤ects YES
Local market expectational errors YES

Obs. macro moments (JxT) 220

Obs. micro moments (MxJxT) 935,440
Notes: Demographic variables demeaned. Standard errors are clustered accross alternatives within 44 time periods.
For the micro moments at the local market level we additionally cluster across time periods within each of the 4252
local markets. Instruments are approximations of optimal instruments (Chamberlain, 1987).
*** p<0.01, ** p<0.05, * p<0.1

A17