Don't "waste" your time! The effects of time series errors in management research: The case of ownership concentration and research and development spending

Journal of Management, Winter, 1993 by Donald D. Bergh

Studies using longitudinal data are becoming popular in management research. Theincreasing frequency of such approaches is not surprising, as they offer advantages over contemporaneous, cross-sectional studies (Kimberly, 1976). At the same time, however, researchers have frequently omitted the special circumstances which accompany longitudinal data (Podsakoff & Dalton, 1987). Indeed, many management researchers either ignore the unique assumptions of longitudinal analysis altogether, or they assume that longitudinal effects are stable, and all that is needed is an accounting of statistical assumptions (Bergh, 1993). As a consequence, important research questions are left unanswered, and some studies cannot be replicated or extended.

In an effort to redirect longitudinal analysis in management research, this study tests the effects of committing "time series errors." The explicit goals of this research are to: (1) demonstrate what can go wrong in longitudinal analysis; (2) show how results may depend upon how time effects are operationalized; and (3) provide guidance as to how management scholars can refrain from "wasting" time in their longitudinal research.

To best meet these goals, it is necessary to first define and describe the time series errors in management research. Next, data from a panel of 183 Fortune 500firms (1985-1988) are used for illustrating how these errors can develop. From these results, suggestions for future research endeavors involving longitudinal analysis are then offered.

Time Series Errors in Management Research

Time series errors arise when researchers do not recognize longitudinal data assumptions and/or when analytical techniques do not include components for timechanges (i.e. conducted with "time in mind").(1) Four specific types of such errors are possible. To illustrate what these errors are, and to show how they can arise, I will review some studies on ownership concentration and research and development (R&D) spending. Focusing on this particular research stream is useful for several reasons, including: (1) the

longitudinal linkage between     ownership concentration and R&D spending has
been analyzed differently across   the studies; (2) the longitudinal methods
used for investigating this subject   are representative of the majority of
those appearing elsewhere in management   research; (3) the authors are clear

in how they handled longitudinal issues; (4)the data can be accessed and measured similarly, thereby facilitating comparisonwith earlier study; and (5) there is lack of consensus in the linkage between ownership and R&D. Thus, considering studies on ownership concentration and R&D provides an opportunity for illustrating a wide spectrum of methodological issues in longitudinal analysis. Before the studies on ownership concentration and R&D are discussed, it is emphasized that they were not chosen to illustrate effective or ineffective treatment of longitudinal effects. Rather, these studies provided a forum for demonstrating how time series errors can potentially occur. No attribution of methodological irregularity is made to any single study, as it is not my intention to assign problems or errors to any of the studies reviewed here. Moreover, examples of time series errors could be found elsewhere, and accordingly, there were no unique problems with the studies on ownership structure and R&D spending.

Time Series Error Type 1 (TS 1)

The most frequent of all time series errors, TS 1's arise when researchers fail to adjust for violations in the assumptions underlying their longitudinal data, and when they use analytical techniques which are not "time-wise". The main

assumptions in longitudinal analysis are autocorrelation or "serial correlation"(when covariance between errors is not equal to zero), and heteroscedasticity (variance of error terms is not equal)(2) (See Vogt,

1993). Time-wise analyticalTABULAR DATA OMITTED methods acknowledge both the dynamic characteristics of longitudinal data (such as shifts and changes in variables over time), as well as its statistical assumptions.

Graves' (1988) examination of ownership holdings and R&D spending (per employee)helps in showing how a TS 1 error might arise. In this study, 22 firms in a single industry (computers) were followed over a ten-year period. These "panel" data (when the same subject is observed at two or more

points) were then        "pooled" or lumped together into one group. In
essence, a sample of 22 firms    with 10 distinct years of observations each

was made into a larger sample with one observation apiece. Ordinary least squares (OLS) regression served for the analysis. Findings indicated a negative relationship existed between ownership and R&D. From these results, Graves concluded that increases in ownership power were associated with decreases in R&D. This "pooling" of longitudinal data, followed by OLS assessment is very common in management research. In fact, Graves' particular methodological decisions arenot peculiar in any way from several other studies. Regardless, however, there are three dilemmas with such treatments of longitudinal data.