R&D alliances and the effect of experience on innovation: a focus on the semiconductor industry.

contract structures, evaluate performance and

manage differences in corporate cultures. Coordination

across firm boundaries is always challenging and,

therefore, skills gained in improving this coordination

likely are gained from any type of alliance. (p. 1012)

This concern was, in any case, tested by considering only previous R&D alliances. Results do not exhibit any significant differences.

Given the nonlinear relationship of EXPERIENCE and the independent variable, the natural log is used as the independent variable.

Control Variables

There might be other factors influencing the patenting activity of a firm other than alliance managerial competence. For instance, it is possible that firms with greater experience are actually more capable firms in R&D. To control for several factors that might be influencing the results, several variables are included here. These variables will allow us to isolate other effects on the independent variable, thus allowing us to capture the hypothesized relationship. Control variables selected here are basically the same as those used by Sampson (2005), but they include some differences that are noted when appropriate.

Prealliance firm and partner patents (FIRM PATENT and PARTNER PATENT). These variables allow us to rule out other factors that may influence firm innovation. Particularly, the number of (firm's and partner's) prealliance patents has been used to capture independent technological capabilities (Patel & Pavitt, 1997), technological acquisitions, R&D spending, and a firm's propensity to patent (Trajtenberg, 1990). Thus, the inclusion of prealliance patents will allow us to control for firm individual R&D and technological efforts. This of course means that both firms' prealliance patent must be accounted for. Hence, two variables, FIRM PATENT and PARTNER PATENT, are included. Prealliance patent count was obtained from USPTO Web site for a 4-year period (1993 to 1996). These variables are not weighted because Trajtenberg (1990) found that "unweighted patent counts are more highly correlated with R&D spending and other R&D inputs than citation-weighted patent counts and, as such, are better measures of R&D inputs." For these variables once again, the total patenting activity of the focal and the parent firm were considered.

Alliance scope (SCOPE). This variable is intended to capture the breadth of R&D collaborative projects. R&D alliances have a wide range of purposes; they can have a very specific and narrow objective (e.g., a very specific software or application) or be broader and more general (e.g., to develop next-generation semiconductors). Thus, an alliance will exhibit different characteristics based on the breadth of the agreement. With this in mind, three dummy variables are selected using information in the SDC database. We look at the alliance description field in this database and assigned each alliance to one of three categories as described in Table 2. Because alliances with intermediate breadth are the most common, we omit SCOPE (Intermediate) from the empirical analyses.

Multilateral alliances (MULTILATERAL). This is also a dummy variable that discriminates alliances between two (bilateral) and alliances between three or more partners. Multilateral alliances might be an indication of larger objectives, and we expect them to be associated with greater returns. Thus, this possible effect is controlled, too. The number of total multilateral alliances for the sample period is 7, as represented in Table 1.

International alliances (INTERNATIONAL). When firms from different nations engage in an R&D alliance, there are some difficulties that they have to overcome (e.g., coordination efforts, cultural differences, etc.). These difficulties have an important impact on the alliance effectiveness, and they are here accounted for by the dummy variable INTERNATIONAL that equals 1 if the firms are from different countries and 0 otherwise.

Other concurrent alliance(s) (OTHER ALLIANCE). If a firm is involved in one or more alliances other than the focal alliance, that firm has to distribute its efforts to successfully coordinate activities with more than one partner. We expect this to have a negative impact on R&D collaboration. Once again, a dummy variable is constructed to control for this possible effect. OTHER ALLIANCE is a dichotomous variable that has a value equal to 1 if the firm is involved in one or more other alliances for the sample period (1997 to 1998), 0 otherwise.

Diversity of technological capabilities (TECH DIVERSITY). One of the main characteristics of collaborative projects is the fact that partners learn new capabilities from the ally, and thus, collaborative benefits increase as partners' capabilities differ because partners with dissimilar capabilities have much more to learn from each other. This must be controlled for to isolate the hypothesized relationship. Following Jaffe (1986) and Sampson (2005), we construct a measure of technological diversity based on the distribution of each firm's patents across different USPTO classifications. For each patent of a certain firm, we use the codifications contained in the class field in the USPTO Web site. This allows for the creation of a multidimensional vector of patents assigned to firms, [F.sup.s.sub.i] = ([F.sup.s.sub.i] ... [F.sup.s.sub.i]), where [F.sup.s.sub.i] represents the number of patents assigned to firm i in class s. With this information a composite index can be calculated as:

TECHDIVERSITY = 1 - ([F.sub.i][F.sub.j]')/[square root of ([F.sub.i][F.sub.i]')([F.sub.j][F.sub.j]')], i [not equal to] j

A value of 1 indicates the highest diversity possible between partners; 0 indicates high similarity.

It has been suggested that by using absorptive capacity, firms must have some capabilities in common to benefit from newly acquired knowledge. Thus, high technological diversity might actually hinder collaborative efforts. Following this line of arguments, the square of TECH DIVERSITY is used in the model.

Statistical Method

Because the dependent variable is a count of citation-weighted patents, this is undoubtedly a nonnegative, integer value. These types of variables also have high frequency of 0 and small integer values. Therefore, to take these characteristics of the sample into consideration, a negative binomial model is selected because 0 and small values are "naturally incorporated into the model" (Hausman et al., 1984).

The negative binomial is an extension of the Poisson model, but it relaxes the assumption of the equality of mean and variance. Therefore, it is a preferred method when Poisson estimation is inappropriate because of overdispersion. Overdispersion is said to be present when the variance is greater than the mean. This is the case in our study.

As such, the following model is used:

Pr[PATENT=p] = [e.sup.[lambda]][[lambda].sup.p]/p! (1)

where [lambda] is [e.sup.[beta]]X + z], [beta] is a vector of independent variables, and X is a vector of parameters. (6)

This particular model has independent errors, and parameter estimates do not require correction because maximum likelihood estimates are unbiased and consistent even if the assumption of independence is violated (Greene, 1993). Nevertheless, observation exhibits a lack of independence because for each firm there is the possibility of multiple observations if the firm participated in multiple alliances. This calls for a correction of the standard errors. A technique by Huber (1967) has been used in the past to correct for this lack of independence. This technique involves the creation of "super observation" calculated as the sum of the likelihood scores for each firm involved in more than one alliance. This value is then used to calculate the standard error.

Results

As reported in Table 1, 137 firms are included in the sample, 86 are involved in only one alliance, 7 are involved in two alliances, and 7 are involved in three or more alliances. Following Sampson (2005), we created a separate observation for each alliance a firm is involved in, leading to 182 observations. Table 3 shows descriptive statistics of the variables.

The estimation of the general model is set out in Model 1 of Table 4. This is the original model, and it considers the logarithm of alliance experience, EXPERIENCE (LOG), as the focal independent variable. From this table we can see that the count of alliance experience has a positive and significant relationship with firm innovation (p < .01). Using the median values of the independent and control variables, we evaluated the model E[PATENT] = [lambda] = [e.sup.[beta]'X], and then we increased the number of prior alliances in one, keeping the rest of the variables unaltered. This exercise produced a change in the expected number of patents equal to 2.56; however, this value was not significant at the .05 level.

We also considered the possibility of a linear relation between EXPERIENCE (COUNT) and PATENT, namely, a regression model was obtained (Model 2) using the original alliance experience variable. Table 4 shows again that the hypothesized relation is statistically significant at the .01 level.