30
the advertising/sales ratio. The second and third variables may be seen as picking up
di¤erent dimensions of the same source of market power (product di¤erentiation). The …rst
and third ones enter the equation lagged one periodto avoid simultaneity biases.
Three variables are included to perform as indicators of a high sensitivity of demand
with respect to product quality and/or product quality with respect to R&D expenditure.
These variables are: the ratio of highly quali…ed workers to total employment, the number
of product innovations reportedby the …rm dividedby the number ofworkers (lagged), and
adummy variable that takes the value one if the market is considered to be in expansion.
Five variables are included to give an account of di¤erent aspects of set-up costs. We
…rstly include the average industry patents (excluding the patents obtained by the …rm,
a classic formal technological opportunities measure), the …rm capital/sales ratio, and a
dummy variable indicating whether the …rm is an occasional R&D performer. We expect
the two …rst variables to act as direct indicators of high …xed costs of R&D linked to spe-
ci…ctechnological product requirements. The occasional character oftheR&D performance
may be seen, instead, as an (indirect) indicator of an easy set-up of tecnological activities.
On the other hand, a mergers dummy variable gives account of signi…cant changes of the
…rm scale through “external” growth. Finally, we include a dummy variable representing
concentrated markets (the variable takes the value one for markets with less than 10 com-
petitors) interacted with the size of the …rm. This variable tries to account for the fact
that relevant set-up costs must be measured in terms relative to the …rm scale. Big …rms
in concentrated markets are likely to experiment smaller set-up costs ratios.
Inaddition, wehavefoundthresholdsinpracticetobesensitivetoasmall listof the…rms,
…rms’ market and …rms’ technology characteristics, all represented by dummy variables.
The list includes the presence of foreign capital, a big dimension of the product market
(national or international as opposed to local or regional), to be located in an autonomous
community with strong spillovers, to be an exporting …rm, to have a product sensitive to
quality controls and to have a technologically sophisticated production process. All these
variables are likely to reduce relative set-up expenses, and some of them will also enhance
the demand for quality. Moreover, it will turn out to be very important to include a set of
16
37
dummy variables of size, measured according to the number of employees, to control for a
strong remaining threshold size e¤ect.
Moreover in both equations we include a set of 18 sector dummies, to control for per-
manent di¤erences arising from activities, and a set of time dummies, which we include
constrained to have the same e¤ects in both equations. Details on all the employed vari-
ables can be found in Appendix 2.
Table 8 reports the results of our preferred estimate of the model. As explained in
Appendix 1, estimationis carriedoutby specifying the likelihoodofadecision andane¤ort
equation, fromwhich the coe¢cients and standard errors of the threshold equation may be
deducted. Blank spaces in the table denote the coe¢cients which have been constrained
to zero in some of the equations. Notice that, in this case, the values of the coe¢cients
consigned in the other two columns coincide in absolute value.
The estimate is robust to changes, its predictive power sensible, and the coe¢cient and
statistics look reasonable. We brie‡y comment these characteristics in turn. The estimate
of Table 8 is obtained by constraining the coe¢cient ¯ and the market power variables to
havethesamevalueinthedecisionande¤ortequationsof amoregeneral speci…cation. This
amounts toexcluding themfromthethresholdequation andit canbe statistically accepted
under a likelihood ratio test of 1.82
19
. In fact, the ¯ values obtained in the two equations
are very close before constraining it (1.65 and 1.48), and we take this as proof of validity
of the speci…cation. In particular, alternative speci…cations basically lead to maintain the
rest of e¤ects unaltered with an increase of the di¤erence between the unscontrained ¯’s.
On the other hand, theinclusionofa dummy variableindicating likelycompetitionchanges
(inferedfrom…rmreportedprice movements attributedtomarketchanges) does not change
the basic results without becoming statistically signi…cant.
We can evaluate the goodness of the …t of the model according to its predictions. Recall
thatthemodel predicts thatthe…rmwill engageinR&Dactivitieswhenthe di¤erence be
¤
¡
b
e
is positive. Table 8 (bottom) reports the results of comparing the model predictions with
19
Instead, the constraintthat the ¾’s ofthetwo equationsare the sameis rejected-when imposed jointly
with theconstraint on thecoe¢cient¯- bya likelihoodratio test value of 5.98.
17
32
the actual observations in the sample for three subgroups of …rms: stable R&D performers,
occasional performers and …rms never observed performing R&D. The model accurately
predicts the zero-one variable that denotes the presence of expenditures for the …rms that
neveroralwaysengageinR&Dactivities(97%and85%ofobservationscorrectlypredicted).
Themodel is, however, muchless accurateinpredictingthe yearly activity of the occasional
performers. Predictioncontinues tobequitegoodwhenR&Dexpendituresarezero(71%of
observations correctly predicted), but assigns erroneouslynegative predictions inhalf of the
cases in which the …rms show occasional expenditures. This is hardly surprising if we take
into account the highdegree of arbitrariness of some …rmaccounting practices inallocating
costs over time and the lack of dynamic structure of the model.
Thekeyvariable, expectedsubsidy, isincludedintheform¡ln(1¡
b
½e), andmusttherefore
attract a ¯ coe¢cient around unity. The value e¤ectively obtained is 1.58, which indicates
ahighe¢ciency of public funds. This estimate gives sensible results on the e¤ect of subsi-
dies, which in particular are very close in magnitude to the comparable subsidy e¤ects on
company …nanced expenditure reported in recent papers (see the detailed analysis of the
next section).
Onthe other hand, theinterpretationoftheresultsobtainedcanbedoneas follows. Mar-
ket power is con…rmed as a determinant of e¤ort, while it seems to have a non-signi…cant
e¤ect on thresholds. The variables aimedat indicating ahighquality-sensitivity of demand
or expenditure-sensitivity of quality show more mixed results. They present signi…cant
positive e¤ects on e¤ort, but we are not able to pick up with some precision the expected
negative e¤ects on thresholds. However, this is compensated by the role that similar vari-
ables play in explaining thresholds (quality controls and technological sophistication). As
expected, high set-upcosts clearly appear toincrease optimal e¤ort and thresholds, but the
scale e¤ect associated with a concentrated market and a big size also lessens this impact.
Finally other …rm characteristics such as having foreign capital, a big market (domestic
or by inclusion of markets abroad), or bene…ts steeming from geographical spillovers, help
to reduce thresholds. In addition, after controlling for all these variables, it remains an
important size e¤ect by which big …rms experience smaller thresholds. This points out the
18
30
permanence of a problem of indivisibility of resources to set up R&D activities, indepen-
dently of the industry or …rm type, explaining asigni…cant part ofsmall …rms’ problems to
undertake these activities.
6. Pro…tability gaps and subsidy e¤ects
Model predictions and parameter estimates can be used to evaluate pro…tability gaps
and the e¤ects of subsidies in a number of ways, which we have explained in detail in sec-
tion 3. In this section we …rstly report the results of computing pro…tability gaps, or the
di¤erences between optimal e¤orts in the absence of subsidies and the …rms’ idiosyncratic
thresholde¤orts. This is done using all the correctly predicted observations.Thenweassess
the potential and actual roles of subsidies in R&D decisions. We …rst report the results of
computingthe trigger subsidies, or thevalueofsubsidies that wouldinducenon-performing
…rms to undertake R&D activities, for all the (correctly predicted) non-expenditure obser-
vations. But we also evaluate the impact of actual subsidies on R&D decisions, by looking
at the …rms that would cease to perform R&D if subsidies were eliminated. The number of
…rms abandoning R&Dinthe absence ofsubsidies turns outto be very small, perhaps abit
surprisingly, and we check the robustness of this result. Next we focus on the e¤ort e¤ects
of subsidies. We employ the ¯ estimate, jointly with the ½
e
estimates and the observed
½for the (correctly predicted) …rms performing R&D that e¤ectively receive subsidies, to
assess the impact of subsidies in private expenditure. Finally, we compare our estimates
with other recent results.
Table 9 reports the distribution of the estimated pro…tability gaps, and Figure 1 depicts
95% of their values (the graphic leaves 2.5% of observations unrepresented in each tail).
Pro…tability gaps show a skew distribution, with a long tail of negative values, and some
concentration of observations around the zero value that presents a greater frequency of
positive gaps. Positive gaps represent 30% of total observations and their mean is about
1%, while negative values average an absolute value of 3.7%. More than 85% of positive
values lie in the interval (0,1.5), while less than 75% of negative values lie in the broader
19
Documents you may be interested
Documents you may be interested
