Innovativeness and new product success: insights from the cumulative evidence

Bijmolt and Pieters (2001) outline a procedure to correct for non-independence violations in meta-analysis when the ratio of coded correlations to moderators is sufficiently large. However, with 14 design variables and a more conservative sample size of 95, sufficient degrees of freedom are not available to model the many possible interactions among the design elements that are required to execute Bijmolt and Pieters’ suggested approach. While any resulting non-independence violations from our using model-level correlations are unlikely to be significant for reasons outlined in the text, readers are nonetheless advised to use due caution when interpreting our findings because non-independence influences could not be explicitly corrected within the meta-analysis.

While it may be otherwise desirable to include the omitted variables along with the measurement and contextual factors in a single regression model for the meta-analysis, the correlation matrix of omitted variables (Table 3) cannot be merged with the correlation matrix of measurement and contextual variables (Table 3). The correlation matrix for omitted variables in Panel A contains correlations for innovativeness with each omitted variable—i.e., some r _I,i* where r* is the correlation between the omitted variable i and innovativeness (I)—as well as performance with each omitted variable—i.e., r _P,i* where r _P,i* is the correlation between the omitted variable i and performance (P). The data in Table 3 are the correlations for the respective design element with the correlation between innovativeness and performance—i.e., some r _c,i** where r _c,i** is the correlation between the design variable i and the actual correlation (c) for innovativeness and performance that was reported in the literature. Hence, the data in the two tables are not comparable and cannot be combined into one matrix as they might be in a meta-analysis of elasticities. Consequently, the appropriate and only available recourse in a meta-analysis of reported correlations is separate analyses of omitted-variable bias and design-element bias.

In a balanced design (i.e., equal ns per category of the moderator) using contrast coding, the intercept represents the grand mean. However, the ns are rarely equal in meta-analyses, and they are unequal here (see Table 1). The resulting intercept when contrast coding is used under these conditions is the unweighted mean, which differs in value from the grand mean (Pedhazur, 1982).