Analysis of Preference of Incentives to Innovation of Dominican Manufacturing and Service Firms

In this paper is analyzed the structure of preference of incentives to innovation of Dominican manufacturing and services firms. The analysis of preference was carried out using a Conjoint Analysis. In total 326 firms were surveyed across the country. According to the main findings, Dominican firms prefer combinations of incentives to minimize tax liabilities but also to reduce uncertainty related to innovation activities. In terms of preference, no statistically significant differences between manufacturing and service firms were found.


Introduction
This paper aims to analyze the structure of preferences of incentives to innovation of manufacturing and service firms of the Dominican Republic.The main research questions were: 1) What is the structure of preferences of incentives to innovation of Dominican firms?2) Are there statistical differences in the structure of preference based on firms´ activities and other firms´ characteristics?This is a context-specific research, which means that some firms' characteristics such as size, location, tax regime and others are specific to the Dominican Republic, in the context of the Latin American and Caribbean economies.The fieldwork took place from October 2012 to May 2013, and the sample consisted of 326 firms.This research was funded through a public interagency partnership between the National Office of Industrial Property known as ONAPI, the Dominican Fund for Social and Economic Research known as FIES (grant number 0452-20), and the National Competitiveness Council (C.N.C.) under the contract number CO 231/08/2012.Also, the Dominican Republic Association of Industries and the ATABEY Innovation Centre were part of the partnership and provided logistical and technical support.In this research there were no ethical or economic conflicts.

The Dominican Economy
The Dominican Republic is a country located in the Caribbean Region and occupies two thirds of the Hispaniola Island, which is shared with the Republic of Haiti.It has a population of around 10 million of inhabitants, and a surface of close to 49,000 square kilometers (ONE, 2011).According to the Word Bank, it is a middle income country and the largest economy of Central America and the Caribbean, with a GDP of US$61.16 billion in 2013(World-Bank, 2015).Until 2010, the country experienced one of the highest growth rates in Latin America and the Caribbean (CEPAL, 2008).In terms of human development, the Dominican Republic is considered a high human development country (PNUD, 2014), in spite of its widely recognized structural distortions in income distribution and deep social inequalities (Attali, 2010).
Around two thirds of the economic activity consists of services, in which telecommunications and tourism play a key role.The industrial sector represents one third of the economic activity, with an important contribution of free zone firms (Banco-Central, 2013), and with a limitation in competitiveness and value creation through innovation that affects the long term possibilities of growth (Hausmann et al., 2011).

Why a Conjoint Analysis?
In ex-ante and ex-post evaluation of public policies, several techniques have been used through time.These techniques include the Delphi Method, which in general terms is based in the consensus of expert groups (Hsu & Sandford, 2007), the Multi-criteria Analysis with a more complex quantitative approach in determining the objective function which equilibrates benefits and costs (Dooley, Smeaton, Sheath, & Ledgard, 2009), and techniques based on opinions of experts, such as the SWOT analysis (strengths, weaknesses, opportunities, threats), which can be categorized as context-based approach (ORR, 2011).The CA provides the opportunity to combine in one technique the capabilities of groups of experts, the survey and sample design, the design of experiments and the choice modeling (Jordan J. Louviere, 1988).

The Conjoint Analysis
The Conjoint Analysis (CA) has been used successfully in marketing studies, in the field of health services, in transportation and infrastructure projects, in prospective studies for energy projects (Paul.E. Green & Srinivasan, 1990), and over the last 25 years it has been intensively used in the fields of environmental economic and in the valuation and management of natural protected areas (Alpízar, Carlsson, & Martinsson, 2001).The CA relies on the approach developed by Lancaster called the "new consumer theory", in which consumers derive utility from the characteristics or objective attributes of the goods or services such as price, size and design, instead of the goods themselves (Lancaster, 1966).The other theoretical component is called "random utility theory", which basically states that the decision-making process is conditioned by the random or unobservable characteristics that lie behind the choices or the preference, such as personal tastes, education or incomes, in the case of consumers (Boxall, Adamowicz, Swait, Williams, & Louviere, 1996).
The CA is classified as a measure of dominance, which consists in numerical assignments to analyze the degree of preference between certain objects (J.J. Louviere, Hensher, & Swait, 2010).There are different types of CA to allow alternative options of analysis of preferences such as: 1) the discrete selection of an option against other competitive options; 2) more dichotomous choices such as "Yes" or "No"; 3) sorting options (order of most preferred to least preferred), among other (Boyle, Holmes, Teisl, & Roe, 2001).
All attributes and their levels were defined and explained in the clearest manner possible, in order to reduce ambiguity and communicate the purpose of the CA to firms.In terms of design, the matrix in table 1 would produce a complex factorial defined as 25x32x4=1,152 of possible combinations or profiles, which is an impractical amount of combinations to be elicited.In order to reduce the number of profiles in an efficient way, it was necessary to carry out a fractional orthogonal design, which consists on finding the optimal number of combinations (DeShazo & Fermo, 2002).
A design is considered orthogonal when it has three characteristics: 1) there is no correlation of attributes and levels; 2) is balanced, meaning that the levels of each attribute appear with equal frequency in the design, and 3) it has a minimum overlap, which means that a level of an attribute is not repeated in a series of alternatives.This must result in a balance of utility of the choice sets or profiles, implying that the expected utility of each alternative within a set of selection is the same (Huber & Zwerina, 1996).To generate an orthogonal fractional design, the ORTHOPLAN procedure of SPSS® 20 was used, allowing a full profile design with orthogonal fractioning and principal effects.The exercise resulted in 18 optimal profiles or choice sets: sixteen of these were used to be elicited by firms and two were used for simulations.

Fiscal incentives in STI policy
For the purpose of this research, incentives can be defined as policy instruments to support changes in the behavior of firms, encouraging them to move towards certain objectives of public policies (Scotchmer, 2004).One of the roles of policy institutions in an NSI is precisely to promote innovative behavior in firms through incentives (Edquist & Johnson, 1997), by offering basically two types of incentives: monetary and non-monetary incentives.The former includes systems of intellectual property protection (Scotchmer, 2004), and the latter includes fiscal incentives and other mechanisms of direct public funding such grants and subsidies (Rivas Sánchez, 2007).The fiscal incentives provide facilities to ease tax liabilities and can operate in two areas: on the tax base (property subject to taxation) and on the tax debt, which is the amount payable as results of tax liabilities incurred in a given period (Rivas Sánchez, 2008).
According to the evidence available in the context of the OCED economies, tax incentives to innovation have a positive effect on the innovative behavior of firms (Guellec & van Pottelsberghe de la Potterie, 2002).In the context of Latin America and the Caribbean, two contributions have special relevance: the contribution of Park on the diffuse impact of tax incentives to innovation in LAC economies (Park, 2002), and the evidence of the crowdingout effect of government support to activities such as R&D (Dominguez, 2008).
In the LAC region, research on STI policies and particularly on tax incentives and their effect on the innovative behavior of firms, has been disperse, and has been undertaken mainly from the perspective of the supply side of policies and in the context of regional cooperation and funding agencies (Vonortas, 2002).The specific contribution of this research is to explore the perspective of firms´ preferences of incentives to innovation in the specific context of the Dominican Republic.

Defining attributes and levels
Two experts workshops were conducted in order to define attributes and levels, which is a recommended approach when there isn´t a clear background of the combination of attributes and levels to be presented as choice sets (P.E. Green, Krieger, & Wind, 2001) In expression 1, yi represents the ranking of preferences of different profi les to which the fi rm is exposed; α represents the constant generated by the method of OLS; βij represents the utilities or part-worths associated with different levels of attributes; and xij represents the presence or absence of a given level of attribute in the generated profi les.In simple terms, fi rms would select profi les or combinations of levels, taking into account that certain levels are more appealing than others and these levels will neutralize those with low attractiveness that otherwise would not have been selected (Huber & Zwerina, 1996).
The part-worths can reduce their own contribution to zero, which would indicate the lowest preference of some levels, and therefore, the minimum contribution to the expected total utility.They also could achieve the maximum score indicating the highest level of preferences, which implies that the selection of a particular profi le would be conditioned by the attractiveness of a particular level of an attribute in relation to other levels present in the profi le (Huber & Zwerina, 1996).

Segmentation of preferences
In the CA, the disaggregation of preferences is called "segmentation".It helps to analyze differences in preferences inside the collected sample, and can be done in two ways: a priori and post hoc (Picón Prado & Varela Mallou, 2000).In the former, the number of segments, their sizes and characteristics are defi ned in the design stage of the research, based on the literature review of similar cases and on the experience of the researcher.In the latter, the number of segments and their characteristics are defi ned after the sample is obtained, according to one or more classifi cation variables or by cluster analysis (Rivera Deán, González Tabares, Martín Santana, Oñate García, & Sánchez Fernández, 2004).
In the post hoc segmentation, a separately CA is performed to the resulting groups to compare the differences in preferences between them (Ramírez Hurtado, Rondán Cataluña, & Guerrero Casas, 2007).In the case of this research and given its exploratory nature, the two types of segmentations were carried out: an a priori segmentation based on fi rms activities, and a post hoc segmentation based on cluster analysis.

Forecasting preferences
The procedure CONJOINT of SPSS uses three models to estimate the probability of choice: 1) the total utility model or TU; 2) the Bradley-Terry-Luce model or BTL and, 3) the logit model which has been used in several types of Conjoint Analysis (Paul.E. Green & Srinivasan, 1990).In order to

Control variables
The control variables were defi ned to support the analysis of preferences, and refer to basic characteristics of fi rms.With the collaboration of experts of Santo Domingo and Madrid, 10 categorical control variables were defi ned: 1) Region (Santo Domingo Metro, North region, East region and South); 2) Localization (urban or rural); 3) Tax regime (ordinary and free zones); 4) Age of fi rms (young, adult and mature); 5) Capital composition (full national; 10%-50% national; 50%-75% foreigner; full foreigner); 6) Firm size based on the number of workers (small, medium and large); 7) Activity (manufacturing, service); 8) Main market (local, national, regional, international); 9) Technological level of product/services (high level, medium high level, medium low level, low level); and 10) Product/services life cycle (short: <1 year, medium: 1-3 year, long: > 3 years).Two of them ("ages of fi rms" and "technological levels of product/ services") were taken from literature.The former was taken from Berger & Udell (1998), which classifi ed age of fi rms into three groups: 0-4 years, 5-24 years and more than 25 years.The latter was taken from the classifi cation of manufacturing industries according to their technological intensity, in which products could be classifi ed as: high-tech, medium-high technology, medium-low technology and low technology (Hatzichronoglou, 1997).
In the case of the variable "fi rm size" the Dominican defi nition, which is based on number of workers, does not fi t with international standards such as the OECD criteria (OECD, 2005).According to the Dominican Law, micro business are composed of 1-15 workers, small business of 16-60, medium businesses of 61-200 and large fi rms of more than 201 workers (R. Guzmán & Ortíz, 2007).The fi nal survey was piloted on 40 fi rms, which evaluated the survey and suggested changes that were taken into account for the fi nal fi eldwork.

Analysis of preferences
For the analysis of preferences, the CA procedure selected was the decompositional approach.This approach is based on the premise that different levels of attributes make a partial contribution (part-worths) to the total utility, and the obtained scores are equivalent to the regression coeffi cients and indicate the relative importance of each attribute (P.E. Green et al., 2001).Then, the structure of preferences is estimated by an ordinary least square (Sen), which has a linear and additive form that can be represented as follows:

Sample size
As it has been stated, 326 fi rms were sampled from a universe of 6,877 fi rms.This universe of 6,977 fi rms was taken from the database provided by the Dominican Association of Industries (AIRD for its acronyms in Spanish) and by the National Tax Bureau known in Spanish as "Dirección Nacional de Impuestos Internos".The sample design chosen was a simple random design with proportional allocation by regions and activities (Scheaffer, Mendenhall, & Ott, 2007).The confi dence level was of 95% and the margin of error of 5%.
According to literature, the sample size for CA studies tends to vary in a range between 100 to 1000 subjects, and the most typical samples sizes tend to be in a range of 300 to 550 cases, usually for commercial studies (Alpízar et al., 2001).Therefore, the reached sample has the typical range for CA studies.

Results
In regards to their general characteristics, the surveyed fi rms are distributed as follows: 66.3% are located in the metropolitan area of Santo Domingo, and the remaining 33.7% distributed in the other regions; 93.6% of the fi rms are located in urban areas and the remaining 6.3% are located in rural areas.Based on the number of workers, in the Dominican context 82.5% of the fi rms are considered small, 12% are considered medium-size and 5.5% are considered large, refl ecting the composition of the Dominican industrial sectors (R. M. Guzmán, 2011).In regards to the tax regime, 90% of fi rms operate in the regular tax system while 10% operate in the free zone tax regime.In regards to their age, 62.6% of fi rms are considered "adults" (between 5 and 24 years), 31.6% are mature (over 25 years) and only 5% are young fi rms (0 to 4 years).According to their activity, 57% are manufacturing fi rms and 43% belong to the service sector.
Concerning the capital structure, 86.2% of the fi rms are 100% Dominican, 7.4% have a variable mix of Dominican and foreign capital and the reining 6.4% are of foreign capital.Related to the target market, 89.5% of the companies are focused on the domestic market and the remaining 10.5% target the international market.In relation to the degree of sophistication of products/services, 15% of the surveyed fi rms affi rmed to offer products/services with high tech content, 34.7% stated to offer products with a medium-high level of tech content, 20.9% indicated to offer products with a medium-low level of tech content, and 29.4% affi rmed to offer products/services with low level of tech content.perform a simulation of probabilities of choice, the last two profi les generated during the orthogonal design were utilized.The TU model estimates the probability of choice assuming that the profi le with the highest total utility will be the most likely to be chosen (MacFadden, 1980).The selection of the profi le with the highest utility follows a relatively simple binary approximation which can be represented as indicated in expression 2 (Ramírez Hurtado et al., 2007).
The BTL model estimates the probability of choosing by comparing the profi les with the highest utility in relation to the others, and then averaging the balance of utility of respondents and also the probability of choice of all profi les (Huber, Wittink, Fiedler, & Miller, 1993).The functional expression of the BTL model can be represented as: In expression 3, J represents the total number of profi les.In other words, the probability of choosing a specifi c combination of levels is given by the utility provided for the sum of all combinations.The third model used by the CONJOINT procedure, the logit model, is the most popular approach in choice experiments (Hanley, Wright, & Adamowicz, 1998).The logit model assumes that preferences are linear, and unlike the BTL it uses a natural logarithm of the utilities to estimate the probabilities of choice.The functional expression can be represented as follows: In expression 4, the probability that profi le Lni was selected by the fi rm n depends on the observable part of the expected utility function of the selected profi le, related to the set of alternatives that makeup the set of options S. In this function, µ is a scale parameter associated with the distribution of the utility function, theoretically equal to 1 (Hanley, Mourato, & Wright, 2001).
In regards to the life cycle of products, 25.5% of the surveyed fi rms affi rmed to offer product/services of a short life cycle (<1 year), 21.5% affi rmed to offer product/services of medium life cycle (1-3 years), and 39.6% stated that they offer products/ services of a long life cycle (> 3 years).

Aggregate analysis of preferences
The aggregate analysis of preferences is intended to provide an overall view of the structure of preferences of manufacturing and service fi rms, which provides an answer to the fi rst research question.The CONJOINT procedure offers the following outputs: 1) a description of the factors (attributes), 2) correlations of observed and estimated preferences, 3) the part-worths or partial contributions of level of attributes, and 4) the relative importance of attributes as shown in table 2 & 3.

Segmentation of preferences
Two segmentations were made: a priori and post-hoc, in order to provide an answer to the second research question.
The fi rst one shows the preferences based on activities and the second shows the preferences based on cluster analysis.The segmentation based on activities is shown in tables 5 and 6.
Table 4 shows the most important results of the research, which are the structure of preferences of incentives to innovation.The part-worths are equivalent to the coeffi cients of a regression model, in which the positive scores indicate a higher expected utility and the negative ones indicate an aversive expectation or rejection (Boyle et al., 2001).In the table 4 the three most valued levels by fi rms are highlighted in bold, and the three most rejected in italic.
An interesting fi nding to be highlighted is the positive valuation of the collaboration between fi rms with universities and research centers, in contrast with the unwillingness to collaborate with other fi rms.In regards to the relative importance of attributes by themselves, chart 1 shows the preferences of the surveyed fi rms.
Chart 1. Relative importance (%) of attributes at the aggregated level of fi rms
Related to the post-hoc segmentation via cluster analysis, two combined procedures were used to enforce the segmentation process, following several authors: k-means and discriminant analysis (Sánchez & Gil, 1998).As a result, two clusters were defi ned: 186 fi rms were included in the fi rst cluster and 137 in the second cluster, and 3 fi rms were disregarded.One of the most relevant differences between cluster 1 and cluster 2 is that in cluster 1 there are more service fi rms and medium sized-fi rms than in cluster 2; and in cluster 2 there are more manufacturing fi rms and smallsized fi rms than in cluster 1. Table 7 shows the part-worths corresponding to clusters 1 and 2. (Insert table 7) Table 5 shows the correlations among observed and estimated preferences of both manufacturing and service fi rms.In both types of fi rms the correlation indicates that the variables of preference have a good fi t, but slightly better in manufacturing fi rms than in the case of service fi rms.
Regarding to the most important output shown in table 6, it can be appreciated that, although the levels of attributes report some differences between manufacturing and service fi rms, the total utility is very similar, as it has been reported in other CA studies (Ramírez Hurtado et al., 2007).In table 6, the attributes with the highest scores are emphasized in bold, and the worst ranked in are emphasized in italic.Some scores show interesting fi ndings, such as the β1 level, where the utility is visibly lower for service fi rms than for manufacturing fi rms, and the β3 level, which generates more utility for manufacturing than for service fi rms.
Chart 2 shows the relative importance of attributes for both types of fi rms.
Examining the attributes, it is obvious that both types of fi rms have a converging structure of preferences.However, given the observed differences in the structure of preferences in table 6 and despite the similarities shown in chart 2, the central question is whether the differences in preference of attributes between both types of fi rms will become statistically signifi cant.
Chart 2. Relative importance (%) of attributes for manufacturing and services fi rms   2000).The key issue here is that despite the similarity of the structure of preferences of the two segmentation groups, it may not make sense to compare them because they are the result of a different decision making process which cannot be transferred between groups (Geanakoplos, 1996).Related to fi rm´s characteristics and the preferences of incentives to innovation, and based on the MANOVA test (Steven, 1980), in the case of manufacturing fi rms the characteristic "size" seems to affect the election.In the case of service fi rms, two characteristics seem to affect the election: tax regime and capital composition (Annex 3).

Forecasting preferences
As it has been stated before, 18 profi les of level of attributes were generated using the CONJOINT procedure.Two of the 18 profi les were not evaluated by fi rms and were used to codify a simulation on the probability of choice by using the three models explained before: the Total Utility model, the BTL model and the logit model.Table 8 shows the probability of choice based on the estimations of the three models, and table 9 shows the composition of the analyzed profi les based on the aggregate level of analysis of preferences.
The differences in the structure of preference in cluster 1 and 2 are quite obvious in the coeffi cients for both positive and negative part-worths.Table 7 shows the attributes with the highest part-worts in bold and the worse ranked in italic.Two levels of attributes stand out: in the score of β5 the part-worths for cluster 1 is positive and for cluster 2 is negative, and it is the opposite for β16, in which the score is negative for cluster 1 and positive for cluster 2. This highlights the differences in structure of preferences despite the fact that the total utility is similar.Chart 3 shows the relative importance of attributes for the clusters.
In regards to the statistical differences in the structure of preferences of attributes for innovation, the ANOVA test (Annex 2) shows statistically signifi cant differences on preferences in the two clusters contrasting with the results of the a priori segmentation.
The a priori segmentation did not fi nd statistical differences between groups taking into account their activities.This does not imply that such differences don't exist, but it does indicate that these differences are not statically signifi cant.However with the post-hoc segmentation such differences were found, given the more stylized way in grouping fi rms by Chart 3. Relative importance (%) of attributes for clusters 1 and 2 According to the three probabilistic models in table 8, profile 1 was the most the selected given its specific composition of levels of attributes.The two profiles share the same part-worths for the first four levels, but after the fourth level strong differences arise (Table 9).At the aggregate level of preferences and in the context of the Dominican Republic, the simulation indicates that firms will prefer those combinations that maximize direct public funding.
In regards to the probabilities of choice based on firm´s segmentation, the results are quite similar to the aggregate level of preferences, and are shown in Table 10.

Regarding the use of the Conjoint Analysis
Given the exploratory nature of this research, the selection of a CA based on ranking was correct, as well as the decompositional and partial contributions approaches.The results were consistent with the literature on CA as well as with the somewhat sparse literature on incentives to innovation in LAC countries.
The CA is not a methodological panacea for ex-ante evaluation of STI public policies on innovation, but can supplement other qualitative approaches such as focus groups, the Delphi method, and opinion surveys.As it has ben stated before, the advantages of the Conjoint Analysis methods is their capability as choice modeling tools.

Policy implications
The analysis of preferences of incentives to innovation provided the opportunity to explore in depth and in different levels, the structure of preferences of Dominican firms.
At the aggregate level of preferences, the most preferred attributes (tax exemption, depreciation and public funds) could anticipate a "crowding-out" effect of the public funding of the business expenditure in R&D and other innovative activities.This means that the private efforts on innovation could be offset by the government dominance, which was highlighted by Park in the context of LAC countries (Park, 2002).Although in LAC countries the private support to R&D and other innovative activities has increased in recent years (RICYT, 2010), the fact is that the public sector still is the main funder of R&D (Arocena & Sutz, 2001).
The structure of preferences that were found, could express a deep cultural background and dependency of public funding and also a learning path to take into account for STI policy making in the Dominican Republic, which is a late comer country in regards to STI policies in the context of Latin-American countries.
Identifying and analyzing the underlying patterns that could explain the preferences of Dominican firms is beyond the scope of this research, as these are related to the complex dynamics of learning in developing countries, including the technological paradigms and trajectories, which, in turn, depend on the STI institutional context (Breschi, Malerba, & Orseingo, 2000).
It is important to remember that the surveyed firms develop their activities embedded in a particular STI institutional context with defined rules and placed restrictions on what firms can and cannot do, conditioning the possibilities of learning and the incorporation of new knowledge and innovations (Nelson & Nelson, 2002).
In such institutional context, informal learning processes probably have more impact on the innovativeness of firms than the formal dynamics of R&D.In the case the Dominican Republic, the innovativeness of firms could be conditioned by factors ranging from availability of human capital, technological infrastructure, linkages with universities and research centers to financing (Metcalfe & Ramlogan, 2008).
The ideological and cultural context related to STI policies of a developing economy such as the Dominican Republic may significantly condition the preferences of incentives to innovation of firms, and take the shape of mental models or conceptual maps that lead the decision making process (Denzau & North, 1994).

Table 2 .
Factors in the estimated aggregated model

Table 3 .
Correlation of the expected preferences

Table 2
Hurtado et al., 2007)or attributes used to evaluate preferences, indicating that all attributes are discrete and orthogonal, which means that the full profi le design was effi cient.Table3indicates that observed and estimated preferences are correlated, which means that the variables of preference have a good fi t (RamírezHurtado et al., 2007).

Table 4 .
Part-worths of levels of incentives to innovation at aggregated level of firms)

Table 6 .
Part-worths of level of incentives of manufacturing and service fi rms J. Technol.Manag.Innov.2015, Volume 10, Issue 2

Table 8 .
Probabilities of choices of the simulation profiles at the aggregated level