As a response to your suggestion and reviewer’s comment (response to reviewer Point # 1), there have been several changes made to the sample. First, throughout the paper we now use 194 unique observations. Each firm accounts for a single entry. The number of observations increased because we changed the variables used in the analysis, and the new set of variables was available in 20 more observations.
Table 2 was added to explain how we arrived at the 194 observations.
Per your suggestion to compare our sample to the total population, we have added Table 3, Panel B, of demographic data about our sample. This table shows, and we discuss this in the text, that the firms in our sample are larger than the average firms in COMPUSTAT. We state that one of the limitations of this study is that we have underrepresented smaller firms, so extrapolating our results to such firms may not be possible.
Unlike most surveys, response rates are not a problem in this situation because firms pay to participate in the CAPS studies. Therefore, once they become part of the organization, they participate. Thus, the "non-response bias" that occurs with other surveys is not a problem here. However, there is a self-selection bias because firms choose to participate. By comparing the sample and industry means, we have identified how the self-selection bias affects our sample.
2: Measures
a: Confirmatory Factor Analysis
Thank you for pointing out the need for using confirmatory factor analysis to validate our measurements for product strategy and organization design. We agree that this would be an ideal way. However, we did not do so for the following reasons.
1. Confirmatory factor analysis is commonly used in studies that have relied on survey responses to confirm construct validity and insure that the measurement does not suffer from perceptual bias. In contrast, measurement of the constructs in this study is based on archival data (i.e., CAPS and COMPUSTAT). While using archival data mitigates perceptual bias, it presents a sample size problem in our context. As you have correctly pointed out, the main concern we are facing by using a holdout sample for testing the validity of the constructs is the size of our sample.
34
2. However, we did follow very closely your suggestion to identify the empirical proxies for measuring product strategy based on the prior literature (see pp. 12-17). In doing so, we believe the measurement now is grounded in the prior literature and is explicit on the selection criteria (see page 13) and thus allows for replications and making comparisons to the past and future studies. As for the proxies for the design, we are restricted to the archival data set collected in the CAPS survey. Again, we now tie the measurement to the prior literature.
To make sure that the readers are aware of this orientation, we make it clear that our use of the factorial analysis is for data reduction rather than hypothesis testing (Harman 1967, Kim and Mueller 1978). Specifically, the factor analysis is for reducing the dimensionality of the proxies (for measuring product strategy and design structure) and with the resulting factor scores to be used in the later analyses when testing the main hypotheses.
3. The main focus of the study is to investigate the performance contingency effect between product strategy and design rather than validating their construct validity. In fact, given the vast amount of the literature on both constructs and the disparity in the measurement approaches (e.g., survey vs.
archival) in the literature, future studies may find this investigation to be worthwhile. Granted that our measurement may suffer in precision for lacking a confirmatory analysis treatment, we believe the noise in the measurement could bias against finding the results (i.e., the interaction effects and the functional form) we hypothesized.
b. Measure for Product Strategy
In essence, your comments point to the need to tie the measurement of product strategy to the prior literature. We agree and have revised the measurements extensively to make them more credible and more focused.
To better measure Product Strategy, we reviewed the prior literature extensively to insure that the empirical proxies we now use in the revised manuscript are rooted in the prior literature. As indicated in the revised manuscript, our conceptualization of product strategy is based on Porter’s framework (see page 12-13) and is now being proxy by six measures: R&D Propensity, Advertising & Administrative to Net Sales, Relative Gross Margin, Market to Book Ratio, New Capital Investment to Sales, and Asset Utilization (see pages 13-14).
In addition, the new measure of Relative Gross Margin is a ten year moving average of the differences of a firm’s gross margin against industry mean gross margin. It thus removes the industry effects on margin.
Concerning your observation that some of these measures are correlated with ROA, we agree. In fact, by resorting to the use of “objective” rather than “subjective”
measures of product strategy, it is unavoidable. However, we are comfortable with this tradeoff in that the main focus of our study is on the interaction rather than
35
the main effect (see footnote 7, p. 14).
We have also addressed your concerns on the coefficient of variation and cash flow predictability. We no longer used these two measures to proxy product strategy.
Justify the use of ROA per Accounting literature
We have reviewed the accounting literature to identify papers that are studying similar questions to ours using ROA as the dependent variable. Balakrishnan, Linsmeier, and Venkatachalam, (1996) make the argument that ROA is an appropriate performance measure when the phenomena being studied can improve a firm's gross margins and profits but may require additional assets to produce the benefits. In such cases, looking solely at income or asset utilization measures of manuscript, we have discussed measures exclusively in the Measures for Product Strategy and Measures for Organization Design at Purchasing Management Level sections. We believe that this will eliminate your concern.
d. Editorial consistency – done.
e. Control Variables
Thank you for your comments on the control variables. The revised manuscript now contains a much broader set of control variables as indicated by the prior literature, including those at the environmental and industry levels (see p. 19 and Appendix 1).
Analysis
a. Cross-sectional and Time series data
Your observation concurs with the reviewer’s comment (point # 1) that our earlier analysis suffered from a problem of pooling both cross-sectional and longitudinal data in the analysis. In theory, a panel data approach would be the ideal model for the analysis (Green, Ch. 16, 1993). Unfortunately, our sample is not a balanced panel (i.e., same number of years of observations from each firm); and the technique requires a minimum of two years of observation for each sample firm. Thus, under the panel analysis many of our sample firms will have to be dropped. This will further reduce the sample size. Given that we have a small sample to start with, an unbalanced panel approach is deemed impractical. For the same reason, it is impossible to have a change model.
Alternatives to a formal panel analysis are as follows. 1) Taking the mean value of the independent and dependent variables of different years of a sample firm and make it into one entry for each firm [Lang & Lundholm, 1993]; or 2) using just a single year
36
for each firm (e.g., using the most recent observation available in the sample for each firm) for the analysis.
In this revision, we have taken the first approach. Although in the current text we only report findings from the “mean value” approach; qualitatively, results from the “most eliminated/added some variables based upon the reviewer's comments. As a result, the number of complete data sets increased from 174 to 194 firms.
b. Lag Effects:
There are two difficulties with presenting a lag model for our analysis. First, there are no clear event dates from the data set. Hence, we do not know when the firm made its strategic and design decisions. Thus, we are unable to hypothesize how long (or even if) we should lag the performance measures from the independent measures. For example, if the firm has adapted its strategy and design in the past, then the effect of those changes would already be seen in the current performance measures. Second, if firms only recently made those choices, we should not find the hypothesized performance congruency effect. Thus, using an un-lagged model biases against our results. Because of these reasons, we believe that using data from the same year is appropriate.
However, to evaluate the robustness of our findings, we also ran the model using a one-year lag model, and the results did not change (see sensitivity analyses on p. 25).