Differences in technology innovation R&D performance creation behavior between for-profit institutions and not-for-profit institutions

The present study compares the performance creation behavior between for-profit institutions and not-for-profit institutions within a national technology innovation research and development (R&D) program. Based on the stepwise performance creation chain structure of typical R&D logic models, a series of successive binary logistic regression models is newly proposed. Using the models, a sample of n = 2076 completed government-sponsored R&D projects was analyzed. For each institution type, its distinctive behavior is diagnosed, and relevant implications are suggested for improving the R&D performance.

the qualitative effectiveness viewpoint is underscored in the field of R&D performance evaluation with the consideration of a clear relationship between R&D inputs and crucial performance created by GSPs (Ruegg 2006;KISTEP 2011;MKE·KIAT 2012;STAR METRICS 2014). In national R&D programs' planning and deployment stage, effective government subsidy allocations are demanded by reflecting the performance differences between institution types and between R&D collaboration types (KEIT 2010(KEIT , 2011(KEIT , 2013OMB·OSTP 2012;OSTP 2012).
Regarding some national technology innovation R&D programs, typical R&D logic models were developed such as the Advanced Technology Program (ATP) logic model of the U.S. Department of Commerce (DOC) (Ruegg and Feller 2003) (Ruegg and Feller 2003;Shipp et al. 2005;Ruegg 2006;Hsu and Hsueh 2009;KEIT 2010KEIT , 2011KEIT , 2013. David et al. (2000) investigated the role of government subsidies in the private R&D investments by scrutinizing a total of 33 previous studies from 1966 to 2000. The majority of the papers examined argued for a net complementary effect in which a government subsidy facilitated private R&D investment (17 papers). The remainder described a net expulsive effect in which the government subsidy only replaced the private R&D investment (11 papers). Up to date, a clear agreement has not been reached yet on the relationship between government subsidy and the amount of private R&D investment (Jaffe 1996;Hall 2002).
As one of the critical influencing factors regarding performance enhancement, R&D collaboration between researchers or between research groups may be considered. It is known that the utilization of diverse perspectives through R&D collaboration contributes to the improvement of R&D performance (Van Aken and Weggeman 2000; Reagans and Zuckerman 2001;Caloghirou and Hondroyiannis 2003). Furthermore, R&D collaboration has the advantage of providing an environment where each research entity can combine complementary resources of R&D collaboration participants in terms of allowing access to information and knowledge held within each entity. In general, universities and research laboratories have comparative capabilities for basic and applied research, and for-profit companies tend to focus more on commercialization-oriented R&D projects. In this context, it is recognized that R&D performance can be enhanced readily through R&D collaboration by multiple competent institutions (Kogut 1988;Das and Teng 2000;Hagedoorn et al. 2000;Narula and Duysters 2004;Nieto and Santamaria 2007;Hillman et al. 2009). Specifically, many preceding studies reported that key measures representing the level of technology innovation of the companies were greatly improved by carrying out external R&D collaboration (Fritsch and Lukas 2001;Belderbos et al. 2004;Laursen and Salter 2006;Ortega-Argilés et al. 2009;Chen et al. 2011;Gronum et al. 2012;Berchicci 2013;Esteve-Pérez and Rodríguez 2013;Robin and Schubert 2013).
The Office of Management and Budget (OMB) and the Office of Science and Technology Policy (OSTP) under the U.S. Government emphasized the need for R&D collaboration associated with the 2014 federal R&D budget compilation and execution procedure (OMB·OSTP 2012). In addition, the ATP accepted applications from single companies and joint ventures. For-profit companies could apply as single applicants to receive an award up to $2 million USD over 3 years to cover project costs (Ruegg and Feller 2003). Also, among the initial ATP's 50 completed GSPs, 42 GSPs (84 %) were conducted through R&D collaboration.
Most of the literature used nonnormal statistical models. Because of the inherent scarcity of GSPs achieving R&D performance, and the extreme skew to the right distribution, the literature seems to adopt non-normal techniques such as Tobit regression models to cope with the censored data characteristics, binary logistic regression models to deal with the performance creation success-failure binary characteristic, and so on (Fritsch and Lukas 2001;Laursen and Salter 2006;Berchicci 2013;Robin and Schubert 2013). Related to measuring the relative efficiency and the total productivity changes in R&D programs, some prior studies provided excellent classifications on R&D inputs and the performance factors to be considered in the performance evaluation (Meng et al. 2006;Wu et al. 2006;Sharma and Thomas 2008;Hsu and Hsueh 2009;Guan and Chen 2010;Chen et al. 2011;Park 2014).
Additionally, some research papers were reported associated with the relationship between R&D management and technology innovation. Tan et al. (2015) presented a comparative impact analysis on collaborative research in Malaysia using journal articles published in the 10-year period spanning, the years 2000-2009. Bacchiocchi and Montobbio (2009) estimated the process of diffusion and decay of knowledge from university, public laboratories and corporate patents. Hu (2009) investigated the extent to which East Asia had become a source of international knowledge diffusion and whether such diffusion was localized to the region. Branstetter and Ogura (2005) emphasized on the use of the knowledge generated by university-based scientists. Meantime, related to measuring efficiency of R&D programs, Chen et al. (2004) assessed the R&D efficiency of 31 computer-related companies in Taiwan. They examined the total efficiency, technical efficiency and scale efficiency respectively, and revealed the correlations between inputs and outputs. Osawa and Murakami (2002), Eilat et al. (2008), Hashimoto andHaneda (2008), andCullmann et al. (2012) analyzed the number of patents as one of the output variables for evaluating R&D performance. Also, Kim et al. (2009) argued that the number of patents per R&D expenditure declined with the firm size (i.e., the firm sales) for both pharmaceutical and semiconductor companies. Lamperti et al. (2015) examined the impact of science parks on growth and innovativeness of affiliated firms. They found that both patenting activity and R&D investments were actively sustained by the presence and quantity of research centers within the park. Figure 1 shows the research model of the present study in which important measures of R&D inputs and performance are organized based on the related literature (Tong and Frame 1994;Werner and Souder 1997;McLaughlin and Jordan 1999;Ruegg and Feller 2003;WK Kellogg Foundation WKKF 2004;Meng et al. 2006;Wu et al. 2006;Bitman and Sharif 2008;Sharma and Thomas 2008;Hsu and Hsueh 2009;Guan and Chen 2010;Chen et al. 2011;Park 2014). In Fig. 1, drawn with squares, multiple R&D inputs and performance factors comprises the stepwise chain structure of the research model. Ruegg and Feller (2003) and Hsu and Hsueh (2009) presented representative R&D inputs and performance factors for a GSP-level performance analysis. Some external influence factors were pointed out such as institution types, R&D collaboration, the internal R&D capability and the accumulated knowledge and experience of institutions (Geuna et al. 2003;Stephan 2010). It was reported that intangible R&D inputs (i.e., accumulated research experience, educational and training efforts for human resources by institutions, etc.) demonstrated a positive influence on R&D performance enhancement (Bowman 1992;Parikh 2001;Lee et al. 2005). As described in "Empirical analysis" section, the research model in Fig. 1 is designed as a parsimonious model composed of quantitative characteristics of GSPs analyzed in the present study. In particular, based on the literature such as Ruegg and Feller (2003), Shipp et al. (2005), Ruegg (2006), Wu et al. (2006, Hsu and Hsueh (2009), Guan and Chen (2010), Chen et al. (2011), the key performance factors suggested include published articles, patent applications and registrations, patents used, profited commercialization sales, new employment, and so forth. Also, we can find discussions on the typical R&D performance creation pattern conformed to the stepwise chain structure from the literature.

Research model
The total of eight variables describe the overall characteristics of each observation (i.e., GSP), which included R&D inputs, performance and external influence factors as shown in Fig. 1. For the input variables, three characteristics are considered: R&D Budget (X1), R&D Period (X2) and R&D Workforce (X3). The three performance variables analyzed are Patent Registration (Y1), Sales (Y2) and New Employment (Y3). The present study considers two additional external influence variables, Institution Type (T1) and R&D Collaboration Type (T2). The three performance variables (Y1, Y2 and Y3) are converted into three corresponding binary variables, B1, B2 and B3 respectively, to deal with the sample characteristics as explained in "Empirical analysis" section. For example, for the ith observation, if the condition of Y1 i > 0 is satisfied (i.e., the case of patent registration performance creation success), then the corresponding patent registration performance creation success-failure binary variable B1 i is defined as 1 (i.e., B1 i = 1 if Y1 i > 0 and B1 i = 0 otherwise). Meanwhile, the variables are selected based on the representative factors examined in the aforementioned literature closely related to GSP-level performance evaluation, and the data availability; the reliability of the sample is fully verified beforehand.
Two external influence variables T1 and T2 are defined as follows. First of all, T1 is a 4-level categorical variable. According to the institution type of the ith observation, T1 has four different values: (1) T1 i = L (Large Company), (2) T1 i = U (University), (3) T1 i = R (Research Laboratory), and (4) T1 i = S (Small and Medium-Sized Enterprise, SME). T2 is a 3-level categorical variable. T2 is classified into three separate values based on the R&D collaboration type of the ith observation: (1) T2 i = Sg (Single Institution R&D), (2) T2 i = Cs (R&D Collaboration with the Same Type Institution), and (3) T2 i = Cd (R&D Collaboration with the Different Type Institution). Specifically, X1 is the pure amount of government R&D subsidy, and the institution type, denoted by T1 i = R, refers to government-funded research laboratory only. Furthermore, the four institution types are categorized into two broader institution types: for-profit institution (i.e., large company and SME) and not-for-profit institution (i.e., university and research laboratory). If only the ith observation reports R&D collaboration between the two heterogeneous institution types, then T2 i is equal to Cd (i.e., R&D collaboration between for-profit institutions and not-for-profit institutions). On the other hand, T2 i = Cs means that the ith observation is conducted by R&D collaboration between the homogeneous institution types.
In particular, based on the assumption that R&D performance creation behavior can be different between for-profit institutions and not-for-profit institutions, the present study attempts to divide the whole sample of n = 2076 described in "Empirical analysis" section into two mutually exclusive subordinate datasets as follows: (1) the first dataset of n 1 = 1637 with for-profit institutions, and (2) the second dataset of n 2 = 439 with not-for-profit institutions. In addition to the whole sample, these two subordinate datasets are analyzed separately for more accurate comparisons of the performance behavior between the two types of institutions. Meanwhile, it is not desirable that all the mathematical details are embedded in the narrative at this point. Therefore, not to wade through all the detailed equations, the mathematical details are arranged in the appendix associated with the design of successive binary logistic regression models.

Description of the sample
As mentioned briefly above, the sample analyzed in the present study is a set of completed GSPs within a representative national technology innovation R&D program (i.e., ITIP) administered by the MOTIE with the Korean government over the recent five performance follow-up survey years (2008)(2009)(2010)(2011)(2012). Initially, the sample consisted of 6267 completed GSPs. Even though the completion years of each GSP were slightly different, this sample can be regarded as fully considering the time lag between R&D inputs and the performance.
For the initial sample, the first data collection was carried out by investigating two national R&D databases provided by Korean government agencies such as the Project Management System (PMS) of the Korea Institute for Advancement of Technology (KIAT) and the eR&D of the Korea Evaluation Institute of Industrial Technology (KEIT). For the second data investigation stage, the data obtained from the two databases were verified, and missing data from the first data collection stage were gleaned using the National Science and Technology Information Service (NTIS) database administered by the Korea Institute of S&T Evaluation and Planning (KISTEP) (KISTI 2008;MKE 2008;MST·OSTI 2008). After the first two data collection processes, an offline survey was implemented to verify the data reliability associated with the sales as well as to obtain missing data of the R&D budget in particular. Consequently, the sample of n = 2067 completed GSPs (i.e., 33.13 % of the initial sample) was prepared. Table 1 shows the descriptive statistics of the sample. As seen, the three continuous performance variables, Y1, Y2 and Y3, are severely skewed to the right. For example, Y2 has the largest coefficient of variation (CoefVar), and its CoefVar is equal to 4.71. With respect to the three continuous performance variables, the number of observations with a value greater than zero (i.e., the number of observations achieving performance) are as follows: (1) the short-term, technical output variable Y1, 905 (905/2076 × 100 = 43.59 %), (2) the mid-term, economic outcome variable Y2, 818 (39.40 %), and (3) the socioeconomic impact variable Y3, 560 (26.97 %). An interesting phenomenon was that the number of observations creating performance decreases monotonically when the position of each performance variable moves forward along with the research model's chain path from the starting to the ending points. These proportions exactly coincide with the means of the three binary variables B1, B2 and B3 (Mean = 0.44, 0.39 and 0.27 respectively). The inherent scarcity of observations achieving performance can be confirmed because the medians of these three continuous performance variables are zero. In preparing the sample, the exchange rate of 1000 Won/$1 USD was applied to the raw data to convert monetary units.

R&D input variables' correlation analysis
When a correlation exists among the R&D input variables X1, X2 and X3, the multicollinearity impairs the precision of the estimated regression coefficients as a whole. When a full model including all these input variables together is estimated, the standard errors of the estimated regression coefficients usually tend to be inflated drastically. Therefore, the stability of the estimated regression models cannot be ensured. In a general linear regression analysis, the degree of multicollinearity can be measured by Variance Inflation Factor (VIF). Because of the nonnormal characteristics in the present study, three kinds of correlation coefficients among the input variables are scrutinized: the parametric Pearson's r and the nonparametric Kendall's τ B and Spearman's ρ s ( Table 2). As seen in Table 2, strong correlations exist among these input variables, as expected. In particular, the largest correlation coefficients are found between X1 and the remaining two input variables, and all nine correlation coefficients in Table 2 have their own P values at 0.000***. The asterisk marks, *, **, *** indicate statistical significance at the significance level α = 10 %, 5 %, 1 % respectively. Hereafter, assuming that X1 is a representative input variable, a reduced model is analyzed associated with the input variable X1 only.
Logistic regression analysis: the whole sample (n = 2076) Table 3 shows the results from analyzing three successive binary logistic regression models from Model (1) to Model (3) using the whole sample of n = 2076 observations. For example, in Model (1), the case of the response variable B1 = 1 is defined as the reference case. As for the two external influence variables, the levels of T1 = L and T2 = Sg are defined as the reference levels. According to the research model's performance chain structure, Model (1) is extended to Model (2) by adding a predictor variable B1 that is the response variable of the preceding model, Model (1). Consequently, in Model (3), the two additional predictor variables, B1 and B2, are included compared with Model (1). In Model (3), the levels of B1 = 0 and B2 = 0 are defined as the reference levels.

Model diagnosis
In Table 3, to examine the significance of Model (1) accompanied with a total of six predictor variables, the likelihood ratio test was carried out. In contrast with Model (1), the null hypothesis model including only the intercept term can be estimated. In the likelihood ratio test, the deviance difference between Model (1) and the null hypothesis model is calculated as Chi-Sq (χ 2 ) = D Fitted (1) -D Null = 197.967 and P value = 0.000*** where D Fitted (1) and D Null denote the deviance of Model (1) and the null hypothesis  (3) Kendall's τ A = 0.30. In practice, these measures of association can be referred to in the comparison with the remaining two models, Model (1) and Model (2). In Model (3), all three measures of association are greater than the corresponding values of the two other models, so the prediction capability of Model (3) is better than the two preceding models, Model (1) and Model (2).  Regarding Model (3), among the 848,960 (= 560 × 1516) pairs of (success, failure) observations, the number of concordant pairs (743,362; 87.6 %) is much larger than the number of discordant pairs (96,802; 11.4 %). In this binary logistic regression analysis, a concordant pair indicates that the pair with the predicted probability of the success observation is larger than the predicted probability of the failure observation. Inversely, the discordant pair is defined as the pair whose predicted probability of the failure observation is higher than that of the success observation. In terms of the concordant and discordant pair counts, Model (3) also gains good predictive power on the probability for B3. Among the three models, Model (3) has the largest proportion of concordant pairs (87.6 %), and Model (1) has the second largest proportion (68.0 %). The smallest proportion of concordant pairs (65.6 %) is found in Model (2). Therefore, the two models, Model (3) and Model (1), show better prediction capabilities, which also agrees with the interpretation of the measures of association.
As shown in Tables 4 and 5, all three models have model significance in terms of the likelihood ratio tests. Compared with the corresponding models in Table 3, the three models in Table 4 have reduced degrees of freedom (DF) such as 4, 5 and 6 respectively due to the reduction of the levels of the external influence variable T1. Additionally, in Model (3) of Table 5, the external influence variable T2 is converted into a 2-level categorical variable, so the model has DF = 5 (i.e., the total of five predictor variables). Consistently, based on the measures of association and prediction pair types' proportions, the prediction capability of Model (3) is excellent in Tables 4 and 5.
Here, an interesting point is found that changes of the estimated coefficients of X1 are (+) 0.283*** → (+) 0.028 → (+) 0.021 along with the successive model extension from Model (1) to Model (3). As seen, all three estimated coefficients of X1 have positive signs (+) consistently, but both the absolute values and the Z values of these decrease monotonically according to the successive model extension. In addition, the statistical significance of the estimated coefficient of X1 is confirmed in Model (1) solely, and then X1 becomes statistically insignificant in the two other models, Model (2) and Model (3). Therefore, a phenomenon can be pinpointed that R&D inputs can exert their influence more on the chronologically adjacent short-term, technical output performance factor B1. Afterwards, their influence diminishes against the mid-term, economic outcome B2 and the long-term, socioeconomic impact B3. Table 3 presents three estimated coefficients associated with the external influence variable T1 in Model (1), except for the reference level T1 = L. The level T1 = U has a positive (+) estimated coefficient β T1 U = 0.343 and Z value = 1.65*, which is statistically significant to the response variable B1. Because the level of T1 = U yields a positive (+) estimated coefficient, the odds ratio exp(β T1 U) = 1.41 is larger than one. Thus, patent registration performance creation success probability odds ratio increases 1.41-fold when the level of T1 changes from the reference level T1 = L to T1 = U. Therefore, patent registration performance creation success probability is sensitive to institution types. Specifically, the university institution type shows the best probability, and the three other institution types (i.e., research laboratory, large company and SMS) are not statistically significantly different.

Model estimation (2): external influences
On the other hand, all three estimated coefficients associated with T1 in Model (2) are statistically significant to the response variable B2: (1) the level of T1 = S has a positive (+) estimated coefficient β T1 S = 0.702 and Z value = 5.39***, (2) the level of T1 = R has a negative (-) estimated coefficient β T1 R = −0.927 and Z value = −4.77***, and (3) the level of T1 = U has a negative (-) estimated coefficient β T1 U = −1.770 and Z value = −5.48***. It is noted that there is a monotonically decreasing sequence of exp(β T1 S) = 2.02*** → exp(β T1 R) = 0.40*** → exp(β T1 U) = 0.17***. It implies that the sales performance creation success probability odds ratio decreases significantly when the institution type changes from for-profit institutions to not-for-profit institutions as follows: SME (1st) → Large Company (2nd) → Research Laboratory (3rd) → University (4th). It is verified that sales performance creation success probability is very sensitive to institution types. In particular, among a total of seven predictor variables in Model (2), the level of T1 = S has the largest odds ratio. It means that a certain GSP's sales performance can be greatly enhanced when it is conducted by an SME. In Table 3, Model (3) shows similar results to Model (2). For the response variable B3, the three consecutive odds ratios decrease monotonically, exp(β T1 S) = 1.54** → ex p(β T1 R ) = 0.74 → exp(β T1 U) = 0.39*. Even though there is no statistically significant difference between research laboratory and large company type, new employment performance creation success probability odds ratios decrease along with the changes of institution types as follows: SME (1st) → Large Company (2nd) → Research Laboratory (3rd) → University (4th). As we can see, the ranks are identical to the B2 related odds ratios shown above. Consequently, an SME performs best in the aspect of sales and new employment. Meantime, university is excellent in the short-term, technical output factor (i.e., patent registration).
As for the external influence variable T2, two estimated coefficients are presented in Model (1) of Table 3, except for the reference level T2 = Sg. The level of T2 = Cd has a positive (+) estimated coefficient β T2 Cd = 0.419 and Z value = 2.53**, which is statistically significant to the response variable B1. When we see the odds ratio, the patent registration performance creation success probability odds ratio increases 1.52-fold with the change from the reference level T2 = Sg to T2 = Cd (i.e., exp(β T2 Cd) = 1.52). Thus, patent registration performance creation success probability is affected by R&D collaboration types, and R&D collaboration with the different type institutions can promote this probability.
Likewise, in Model (3), the level of T2 = Cd has a positive (+) estimated coefficient β T2 Cd = 0.516 and Z value = 2.19**, which is statistically significant to the response variable B3. New employment performance creation success probability odds ratio increases 1.68-fold with the change from the reference level T2 = Sg to T2 = Cd (i.e., exp(β T2 Cd ) = 1.68). In summary, R&D collaboration with the different types of institutions promotes the performance of both patent registration and new employment.

Model estimation (3): performance chain
This section presents a comprehensive investigation on how closely the time-ordered previous and subsequent performance factors relate to one another. First, we try to identify the relationship between the response variable B2 and the predictor variable B1 in Model (2) of Table 3. The predictor variable B1 has a positive (+) estimated coefficient β B1 = 0.376 and Z value = 3.79***, which is statistically significant to the response variable B2. Sales performance creation success probability odds ratio increases 1.46-fold when B1 changes from the reference level B1 = 0 to B1 = 1 (i.e., exp(β B1 ) = 1.46). Thus, a higher sales performance creation success probability is detected when an observation creates patent registration performance in advance.
Furthermore, referring to Model (3) in Table 3, it is verified that two predictor variables B1 and B2 are statistically significant to the response variable B3 simultaneously. The predictor variable B2 has a positive (+) estimated coefficient β B2 = 3.603, Z value = 21.65*** and exp(β B2 ) = 36.72. Therefore, new employment performance creation success probability odds ratio increases as much as 36.72-fold when B2 changes from the reference level B2 = 0 to B2 = 1. Most notably, among the eight predictor variables in Model (3), B2 has the largest values of estimated coefficient, Z value and odds ratio. This finding implies that new employment performance creation success probability increases drastically by achieving sales performance beforehand. In the comparison with B2, B1 survives as a significant predictor variable to the response variable B3, even though both the absolute value and the statistical significance of the estimated coefficient are weakened slightly. Specifically, B1 has a positive (+) estimated coefficient β B1 = 0.420, Z value = 3.06***, and the odds ratio exp(β B1 ) = 1.52. Therefore, the preceding performance creation of sales and patent registration can act as a catalyst for creating the subsequent performance of new employment. In summarizing the series of three successive binary logistic regression models analyzed from Model (1) to Model (3) in Table 3, the research model's performance chain structure showing B3 ← B2 ← B1 ← X1 is clearly identified, accompanied with statistical significance. Table 4 shows the primary results from analyzing three successive binary logistic regression models from Model (1) to Model (3) using the partial sample of n 1 = 1637 observations associated with for-profit institutions only. As aforementioned, the two samples analyzed in "Logistic regression analysis: the whole sample (n = 2076)" and "Logistic regression analysis: sample split (1) (for-profit institutions) (n 1 = 1637)" sections are overlapped approximately 80 % (i.e., n and n 1 ), as the results and interpretation in "Logistic regression analysis: sample split (1) (for-profit institutions) (n 1 = 1637)" section are very similar to the explanation in "Logistic regression analysis: the whole sample (n = 2076)" section. Thus, only distinctive features of Table 4 are explained briefly below in the comparison with Table 3.

Model estimation (1): input versus performance
In Model (1) of Table 4, X1 is statistically significant with a positive (+) estimated coefficient β 1 = 0.242 and Z value = 6.44***. Based on the odds ratio exp(β 1 ) = 1.27, we can interpret that patent registration performance creation success probability odds ratio increases 1.27-fold with 1 unit increment in X1. Similar to the results from the analyses using the whole sample of n = 2076, only Model (1) shows the statistical significance of X1 with the expected positive (+) sign. Again, in the case of for-profit institutions, the influence of R&D inputs is confined within the chronologically adjacent short-term, technical output performance factor B1. Table 4 presents only one estimated coefficient associated with the external influence variable T1, except for the reference level T1 = L. In Model (2), the level T1 = S has a positive (+) estimated coefficient β T1 S = 0.644, Z value = 4.76*** and the odds ratio for the response variable B2, exp(β T1 S) = 1.90. Also, Model (3) shows similar results such as β T1 S = 0.372, Z value = 1.87* and the odds ratio for the response variable B3, exp(β T1 S ) = 1.45. So, an SME performs better than a large company for creating sales and new employment performance.

Model estimation (2): external influences
In Table 4, two estimated coefficients are presented related to T2, except for the reference level T2 = Sg. In Model (1), the level of T2 = Cd has a positive (+) estimated coefficient β T2 Cd = 0.359 and Z value = 1.84*, which is statistically significant to the response variable B1. Patent registration performance creation success probability odds ratio increases 1.43-fold with the change from the reference level T2 = Sg to T2 = Cd (i.e., exp(β T2 Cd) = 1.43). Similarly, in Model (3), the level of T2 = Cd has a positive (+) estimated coefficient β T2 Cd = 0.512 and Z value = 1.98**, which is statistically significant to the response variable B3. New employment performance creation success probability odds ratio increases 1.67-fold with the change from the reference level T2 = Sg to T2 = Cd (i.e., exp(β T2 Cd) = 1.67). In the case of for-profit-institutions, R&D collaboration with the different types of institutions enhances the performance of both patent registration and new employment.

Model estimation (3): performance chain
In Model (2) of Table 4, the predictor variable B1 has a positive (+) estimated coefficient β B1 = 0.514 and Z value = 4.87***, which is statistically significant to the response variable B2. Sales performance creation success probability odds ratio increases 1.67-fold when B1 changes from the reference level B1 = 0 to B1 = 1 (i.e., exp(β B1 ) = 1.67). In Model (3), two predictor variables B1 and B2 are statistically significant to the response variable B3 simultaneously. In particular, the predictor variable B2 has a positive (+) estimated coefficient β B2 = 3.733, Z value = 19.79*** and exp(β B2 ) = 41.79. Therefore, new employment performance creation success probability odds ratio increases drastically as much as 41.79-fold, when B2 changes from the reference level B2 = 0 to B2 = 1. Also, among the six predictor variables in Model (3), B2 has the largest values of estimated coefficient, Z value and odds ratio. Even though the influence is diminished slightly compared with B2, B1 remains as a significant predictor variable to the response variable B3. The predictor variable B1 has a positive (+) estimated coefficient β B1 = 0.497 and Z value = 3.40***, and exp(β B1 ) = 1.64. Hence, new employment performance creation success probability is heavily sensitive to the two predecessor performance success-failure within for-profit-institutions' GSPs. As described in detail in "Model estimation (3): performance chain" section, the performance chain structure of B3 ← B2 ← B1 ← X1 is revealed once again.
Logistic regression analysis: sample split (2) (not-for-profit institutions) (n 2 = 439) Table 5 shows the results from analyzing three successive binary logistic regression models from Model (1) to Model (3) using the partial sample of n 2 = 439 observations conducted by not-for-profit institutions only. As explained below, the R&D performance creation behavior of not-for-profit institutions is clearly distinctive from the forprofit institutions' behavior identified in "Logistic regression analysis: the whole sample (n = 2076)" section. Meanwhile, it is noted that 67 observations are eliminated from the partial sample of n 2 = 439, and the remaining 372 observations are used to analyze Model (3) in Table 5. Since the removed observations have T2 = Cs and the response variable B3 = 0 simultaneously, it is not possible to use them for estimating the coefficient of β T2 Cs.

Model structure
As for the two external influence variables T1 and T2, the levels of T1 = U and T2 = Sg are defined as the reference levels in Model (1) of Table 5. Because not-forprofit institutions' observations are extracted solely, the external influence variable T1 becomes a 2-level categorical variable (i.e., T1 = U and T1 = R). The number of observations achieving performance are summarized as follows: (1) proportion (i.e., 53.30 > 40.99 %). On the contrary, the two other proportions are considerably smaller than the for-profit institutions' corresponding proportions (i.e., sales 14.81 < 46.00 % and new employment 11.02 < 31.70 %). In the case of not-for-profit institutions, another feature is that patent registration proportion 53.30 % is relatively high, but the two subsequent proportions drop sharply (i.e., 53.30% → 14.81 % → 11.02 %).

Model estimation (1): input versus performance
In Model (1) of Table 5, X1 is statistically significant with a positive (+) estimated coefficient β 1 = 0.337 and Z value = 7.59***. Based on the odds ratio exp(β 1 ) = 1.40, so we can interpret that patent registration performance creation success probability odds ratio increases 1.40-fold with 1 unit increment in X1. In the case of not-for-profit institutions, there is a distinctive relationship between the R&D input variable X1 and the three R&D performance binary variables B1, B2 and B3. When the models are extended successively as Model (1) → Model (2) → Model (3), the estimated coefficient β 1 is calculated consistently accompanied with statistical significance and the expected positive (+) sign as follows: (+) 0.337*** → (+) 0.110*** → (+) 0.092*. This is clearly different from the for-profit institutions' pattern explained in "Model estimation (1): input versus performance" section. However, both the absolute values and the Z values of these estimates decrease monotonically according to the successive model extension. Consequently, R&D inputs can exert their influence throughout the entire R&D performance chain path from the short-term, technical output factor B1 to the long-term, socioeconomic impact factor B3 via the mid-term, economic outcome factor B2. However, it is noted that R&D inputs' influence reduces gradually when the performance factors move forward along with the chain path.
When the two estimated coefficients are examined related to T2, the statistical significance is found in Model (1) only. In Model (1), the level of T2 = Cs has a positive (+) estimated coefficient β T2 Cs = 0.766 and Z value = 1.95*, and the level of T2 = Cd has a positive (+) estimated coefficient β T2 Cd = 0.607 and Z value = 1.89*. Hence, patent registration performance creation success probability odds ratio increases: (1) 2.15-fold with the change from the reference level T2 = Sg to T2 = Cs (i.e., exp(β T2 Cs) = 2.15), and (2) 1.83-fold with the change from the reference level T2 = Sg to T2 = Cd (i.e., exp(β T2 Cd ) = 1.83). In the case of not-for-profit institutions, patent registration performance can be improved by controlling R&D collaboration types. Specifically, the two R&D collaboration types denoted by Cs and Cd can lead to superior patent registration performance.

Model estimation (3): performance chain
When we examine the relationship between the response variable B3 and the two predictor variable B1 and B2 of Model (3) in Table 5, B2 is statistically significant to the response variable B3 only. The predictor variable B2 has a positive (+) estimated coefficient β B2 = 2.803, Z value = 7.19*** and exp(β B2 ) = 16.50. So, new employment performance creation success probability odds ratio increases 16.50-fold when B2 changes from the reference level B2 = 0 to B2 = 1. Among the five predictor variables in Model (3), B2 has the largest values of estimated coefficient, Z value and odds ratio. Thus, new employment performance can be improved by generating sales performance in advance. However, the odds ratio exp(β B2 ) = 16.50 of not-for-profit institutions is much less than the odds ratio exp(β B2 ) = 41.79 of for-profit institutions. It indicates that the sensitivity of new employment performance creation success probability decreases in the case of not-for-profit institutions.
Unlike for-profit institutions, we cannot derive a statistically significant positive (+) estimated coefficient of the predictor variable B1 in the two models, Model (2) and Model (3) in Table 5. In contrast to our expectation, Model (2) shows the negative (-) estimated coefficient β B1 = −0.769 and Z value = −2.39**. Associated with not-forprofit institutions, two disconnected relationships are detected: B3 ← B2 ← X1 and B1 ← X1 performance chains. Therefore, it is interpreted that not-for-profit institutions do not greatly link patent registration to sales, as compared with for-profit institutions.

Conclusions
Based on typical R&D logic models, various types of R&D performance factors can be evaluated more systematically within national technology innovation R&D programs, and some useful policy implications can be derived for restructuring subsequent R&D programs as well as R&D budget allocations more effectively. However, due to the nonnormal sample characteristics and the difficulty of dealing with the time lag between R&D inputs and performance, related literature is still limited especially associated with empirical analyses presenting both relevant research models and practical implications.
The present study analyzed a sample of n = 2076 completed GSPs within the representative national technology innovation R&D program. In particular, the present study verified differences in R&D performance creation behavior between for-profit institutions and not-for-profit institutions within the program. Methodologically, a series of successive binary logistic regression models was proposed, and the sample was split into two mutually exclusive subordinate datasets to compare R&D performance creation behavior between the two types of institution more accurately. Major results and implications of the present study are summarized as follows. First, the sustainability of the government R&D subsidy was relatively weaker for creating R&D performance within for-profit institutions. On the other hand, the government funds exerted its influence throughout the entire performance chain path of not-for-profit institutions. From the sustainability perspective, it might be desirable to invest more government R&D budget into GSPs managed by not-for-profit institutions. Also, practitioners should carefully select performance-oriented GSPs in building subsequent R&D programs considering this behavioral characteristic. Second, on the whole, because of the strong relationship between the two directly connected performance factors, we need to encourage a sequential performance creation as much as possible to extend a GSP R&D performance creation life cycle. Specifically, the for-profit institutions' performance creation behavior conformed exactly to the stepwise chain structure. In contrast, not-for-profit institutions showed somewhat a discontinuous pattern. In particular, we discovered an undesirable pattern of not-for-profit institutions in which the technical output (i.e., patent registration) was not linked to the economic outcome (i.e., sales). Therefore, intensive efforts should be required to utilize patent registration to promote sales performance more frequently. Third, for-profit institutions achieved higher performance levels of patent registration and new employment simultaneously through R&D collaboration with the different types of institutions (i.e., T2 = Cd). However, not-for-profit institutions showed higher performance of patent registration only through R&D collaboration with the different types of institutions. Thus, for-profit institutions need to take advantage of the collaborative activities with not-for-profit institutions to enhance patent registration performance. This strategy can extend for-profit institutions' performance creation life cycle gradually up to the economic and the socioeconomic performance levels. Meanwhile, both for-profit institutions and not-for-profit institutions should make efforts to generate sales performance through R&D collaboration. Fourth, SMEs performed best in the aspect of sales and new employment, and universities excelled in the performance of patent registration. Thus, each institution type has its own performance factors with competitive advantages. Therefore, it is necessary to select appropriate GSPs for satisfying the program-level objectives effectively.
The present study did not consider other intangible R&D inputs such as inherent R&D capability owned by the institutions and accumulated R&D knowledge and experience through R&D activities in the past. Therefore, these intangible R&D inputs should be reflected in a future modification of the research model. Furthermore, ordinal logistic and Poisson regression models can be incorporated into the successive regression analyses procedure to accommodate more finely categorized values of the performance variables.
on Semiconductor Manufacturing, International Journal of Simulation and Process Modelling, SpringerPlus, and so on. His research interests include R&D performance evaluation, operations and technology management, etc.
In Eq. (5), T1 is replaced by three binary variables T1 U , T1 R and T1 S where T1 U , T1 R and T1 S correspond to the three institution types, university, research laboratory and SME respectively. In the same context, Eq. (4) is converted to Eq. (6). Likewise, in Eqs. (5) and (6), T2 is replaced by two binary variables T2 Cs and T2 Cd where T2 Cs and T2 Cd correspond to the two R&D collaboration types, R&D collaboration with the same type institution and R&D collaboration with the different type institution respectively. The variable conversion related to the two categorical variables T1 and T2 is commonly applied from Model (2) to Model (3) as explained afterward.