An Analysis of Product Lifetimes in a Technologically Dynamic Industry

Barry L. Bayus Kenan-Flagler School, University of North Carolina, Chapel Hill, North Carolina 27599

he conventional wisdom that product lifetimes are shrinking has important implications for Ttechnology and product planning. However, very limited empirical information on this topic is available. In this paper, product lifetimes are directly measured as the time between product introduction and withdrawal. Statistical analyses of desktop personal computer models introduced between 1974 and 1992 are conducted at various product market levels. Results indicate that (1) product and product model lifetimes have not accelerated, and (2) manufac- turers have not systematically reduced the life-cycles of products within their lines. Instead, the products of firms that have entered this industry in the more recent years tend to be based on previously existing technology, and, not surprisingly, these products have lifetimes that are shorter than those of established firms. Implications of these findings are discussed. (Product Life-Cycles; Personal Computer Industry; Failure Time Analysis)

1. Introduction (1986), the technology S-curve captures the relationship Conventional wisdom holds that product life-cycles are between the amount of resources (e.g., people, dollars) getting shorter over time. Based on statements in the required to generate improvements in technological popular press, such as ‘‘marketing consultants say performance. In the early stages of technology devel- product life-cycles are shortening every year’’ (Alsop opment, a relatively large infusion of resources is 1986), ‘‘I can’t document it, but every industry we look needed to obtain incremental advancements in perfor- at seems to be undergoing shorter life-cycles’’ (Fraker mance. As the technology becomes more familiar and 1984), and the usual ‘‘product life-cycles are getting widely used in products, performance improvements shorter’’ claim in a large number of business articles, can be achieved with many fewer resources. During this one would conclude that this is indeed a widespread period, technology development can proceed swiftly. In phenomenon. This perception is especially evident in many instances, the resulting performance improve- technologically dynamic industries like personal com- ments are embodied in new product models, which are puters (e.g., Hof 1992, Verity 1992, Cimento et al. 1993, made available in rapid succession (Meldrum 1995). Steffens 1994). Believing that product lifetimes are Thus, the dynamics associated with the general pattern shrinking in their industries, many firms have imple- of technology development can give rise to the idea that mented new programs to accelerate the development of product life-cycles are getting shorter. new products (e.g., Fraker 1984, Dumaine 1989, Patter- If true, shrinking product lifetimes have important son 1993). In fact, reducing the development time for implications for technology management and product new products and getting products to market faster has planning. For example, shortening product lifetimes become an extensively discussed topic in the academic generally requires an increase in the percentage of rev- and business press (e.g., Rosenau 1988, Stalk 1988, Du- enues that should be spent on R&D and new product maine 1989, Patterson 1993, Stalk and Webber 1993). development (e.g., Pierz 1995). The length of the prod- A possible source of this popular notion relates to the uct life-cycle also affects the amount of time available technology S-curve concept. As discussed by Foster to recover an original R&D investment (e.g., Goldman

0025-1909/98/4406/0763$05.00 Copyright ᭧ 1998, Institute for Operations Research and the Management Sciences MANAGEMENT SCIENCE/Vol. 44, No. 6, June 1998 763

3b2a 0005 Mp 763 Wednesday Jun 03 01:11 PM Man Sci (June) 0005 BAYUS Analysis of Product Lifetimes

1982), and possibly to catch up to the technology leader time of a product. The classical product life cycle has four (e.g., von Braun 1990, 1991). Should a firm be a market stages (e.g., Levitt 1965, Cox 1967, Polli and Cook 1969): pioneer of new that have shrinking prod- an initial introduction period of slow sales growth, a period uct lifetimes, or a follower (e.g., Levitt 1966, Kerin et al. of rapid growth in sales, a maturity period in which sales 1992, Schnaars 1994)? How is corporate decision mak- level off and are relatively stable, and a decline period in ing affected when products have market lifetimes that which sales drop off. Conceptually, the length of the prod- are becoming shorter (e.g., Eisenhardt 1989)? uct life-cycle (i.e., product lifetime) is the time between Empirically, it is very difficult to rigorously examine introduction and withdrawal from the marketplace. De- product lifetimes, since detailed data for the entire pending on the level of product market analysis (see Fig- product life-cycle and at all the various product market ure 1), expected product lifetimes will vary. For example, levels are generally difficult to acquire. Consequently, product technology lifetimes generally are longer than very limited empirical information on product lifetimes product model lifetimes; product model lifetimes usually for any particular industry is available in the literature. are longer than brand model lifetimes; and product cate- The few published studies on this topic generally have gory lifetimes rarely are observed. Note that the under- taken indirect approaches using only aggregate data at lying product life-cycle for a product model is constructed the industry or product category level. For example, by aggregating sales across all brand models of that prod- Qualls et al. (1981) examine sales growth rates for sev- uct (across all manufacturers). eral durable product categories, Bayus (1992) studies To reach the conclusion that product life-cycles are first-time buyer diffusion rates for various consumer shrinking, the business press generally has used an in- durables, and Grubler (1991) investigates household direct measure of product lifetimes. A common ap- penetration rates of several products. See Bayus (1994) proach is to compare the household penetration rates for a review of empirical efforts in this area. (to some arbitrary level) for several product categories Unlike these prior studies, this paper provides an in- with the year of product introduction. Given that prod- depth analysis of product lifetimes for different levels uct lifetimes at the category level are relatively long, this of product market aggregation in one technologically indirect approach seems reasonable. However, any con- dynamic industry. Using data on all the product models clusions obtained across industries, product categories, sold by each manufacturer in the industry, product life- or product forms may not necessarily be appropriate times of personal computers over the period 1974–1992 within a particular product form (i.e., product technol- are investigated. Product lifetime is directly measured ogies, product models, or brand models). as the time between product introduction and with- Generally speaking, a firm’s set of key competitors drawal, and censored data are handled using the statis- can be identified within a specific product category or tical framework of failure time analysis. product form. Shrinking lifetimes at the product tech- In the next section, a general discussion of the various nology or product model level imply a certain type of product market levels that can be defined and analyzed is competitive environment faced by all firms in that in- given. Background on the personal computer industry dustry—an environment in which the opportunity win- and the data empirically studied are discussed in the third dow in which to obtain a sufficient return on invested section. This is followed by a description of the statistical resources (including product development and market- model used for the empirical analysis of product lifetimes ing expenses) is getting smaller over time. In this case, in the fourth section. Estimation results are then presented, getting to market quickly is paramount, since there is and the implications of these empirical findings for firms only a short time in which a firm can earn revenues. and consumers are discussed in the last section. On the other hand, shortening product lifetimes at the brand model level largely is influenced by an individual firm’s decisions. Although a firm may perceive that 2. Product Market Levels and speed-to-market is imperative, a simple but powerful Analysis Issues analysis by von Braun (1990, 1991) shows that there are Generally speaking, the product life-cycle concept is a de- many conditions in which it is not in the firm’s best scription of the evolution of unit sales over the entire life- interest to actively shrink its product lifetimes. The

764 MANAGEMENT SCIENCE/Vol. 44, No. 6, June 1998

3b2a 0005 Mp 764 Wednesday Jun 03 01:11 PM Man Sci (June) 0005 BAYUS Analysis of Product Lifetimes

Figure 1 Levels of Product Markets

MANAGEMENT SCIENCE/Vol. 44, No. 6, June 1998 765

3b2a 0005 Mp 765 Wednesday Jun 03 01:11 PM Man Sci (June) 0005 BAYUS Analysis of Product Lifetimes tradeoffs between time-to-market, product perfor- (1992) and Steffens (1994) provide an excellent histori- mance, and development costs must therefore be care- cal review of the personal computer industry. fully balanced (e.g., rushing to market may lead to First-generation microprocessors (e.g., Intel’s 4004) higher than expected development costs and/or prod- were used in calculator applications and were generally uct performance which is lower than anticipated). Stalk incapable of performing the tasks associated with per- and Webber (1993) also note that blindly speeding new sonal computing. The first personal computers were de- products to market can lead to a ‘‘strategic treadmill’’ veloped around second-generation 8-bit processors.2 where companies are ‘‘condemned to run faster and These CPUs included Intel’s 8080 (introduced in 1974) faster but always staying in the same place competi- and 8085 (introduced in 1976), Zilog’s Z80 (introduced tively’’ (e.g., Japanese consumer electronics companies in 1975), and Mostek’s 6502 (introduced in 1975). are eliminating 25 models of VCRs and 19 models of Second-generation personal computers were based on televisions, and automobile manufacturers are extend- third-generation 16-bit microprocessors. These CPUs in- ing their product cycles from 4 to 5 years). cluded Intel’s 8086 (introduced in 1978), 8088 (intro- In the remainder of this paper, the following analyses duced in 1979), 80286 (introduced in 1982), and are conducted for one specific product form. Motorola’s 68000 (introduced in 1978) and 68010 (intro- (1) Brand model information across all manufactur- duced in 1979). Third-generation personal computers ers is aggregated and then analyzed to determine if are based on CPUs with a 32-bit architecture (i.e., product technology or product model lifetimes are fourth-generation processors). Full 32-bit processors in- shrinking over time. clude Intel’s 80386DX (introduced in 1985), 80386SX (2) The brand model lifetimes within a firm’s prod- (introduced in 1988), 80486DX (introduced in 1989), uct line are analyzed to determine if brand model life- 80486SX (introduced in 1991), and Motorola’s 68020 (in- times within manufacturers are shrinking over time. troduced in 1986) and 68030 (introduced in 1989). (3) Brand model lifetimes across manufacturers are an- As this brief discussion indicates, the personal com- alyzed to determine whether competing firms are pursu- puter industry provides a technologically dynamic set- ing strategies involving different product cycles over time. ting in which to study product lifetimes. In addition, it is generally believed that the product lifetimes of per- 3. The Personal Computer Industry sonal computers are getting shorter over time (e.g., Hof A personal computer can be defined as a general- 1992, Verity 1992). Although the calculation details are purpose, single-user machine that is microprocessor lacking, Cimento et al. (1993) also report that the lifetimes based and can be programmed in a high-level language. of personal computers introduced by U.S. firms declined The central processing unit (CPU) is the ‘‘brain’’ of the from around 2.5 years in 1988 to 1.5 years in 1991. computer. It contains the arithmetic and logic compo- nent, as well as the core memory and unit for 3.1. Data the computer. As such, it controls all system operations, Using the available brand model information from In- including the external peripheral equipment. Since CPU ternational Data ’s Processor Installation chip design determines the overall computer power and Census, a database composed of annual unit sales (do- 3 performance, it is the single most important factor used mestic and international), the year of product intro- by manufacturers to define a brand model.1 Langlois duction, and the installed CPU for all the desktop per-

1 As discussed in Steffens (1994), manufacturers generally use unique 2 brand model names for personal computers with different CPUs and This information was assembled with the assistance of Nancy Pressel incur significant expenses with the production and launch of each from Intel Corporation. model. Also note that separate brand models for each of the different 3 As discussed in the previous section, product lifetime is directly re- speeds within a particular microprocessor family are often used. Mul- lated to the potential revenues that can be earned. Thus, to obtain a tiple memory, display, sound, and communication configurations typ- more complete measure of the time in which a brand model is in the ically are then possible within any brand model and can be changed market, sales across domestic and international markets are studied in at the time or purchase, or even later. The few cases in which a brand this paper (see also Langlois 1992). Although not reported here, anal- model name could not be associated with a unique CPU chip were yses of product lifetimes defined using only domestic sales resulted in deleted from the database analyzed in this paper. similar conclusions to those to be presented in §5.

766 MANAGEMENT SCIENCE/Vol. 44, No. 6, June 1998

3b2a 0005 Mp 766 Wednesday Jun 03 01:11 PM Man Sci (June) 0005 BAYUS Analysis of Product Lifetimes sonal computer brand models introduced between gradual increase to a peak in the annual number of 1974 and 1992 was developed. Any desktop model new entrants and new product introductions fol- with a CPU chip that could not be uniquely identified lowed by a decline over time. This general pattern is (e.g., due to proprietary manufacturer technology) consistent with that of a maturing industry that has was eliminated from the database. The resulting data established a dominant product design (e.g., see Ro- set includes information on over 600 manufacturers segger and Baird 1987, Utterback 1994). and almost 2800 brand models. The annual number Figure 3 shows the total annual desktop personal of new entrants and new brand model introductions computer industry sales between 1974 and 1992, along between 1974 and 1992 is given in Figure 2. Each is with sales for each product technology generation. Total characterized by the same pattern, in which there is a personal computer sales have continued to grow over

Figure 2 Annual Entrants and New Product Introductions in the Personal Computer Industry

Figure 3 Annual Personal Computer Sales by CPU Technology Generation

MANAGEMENT SCIENCE/Vol. 44, No. 6, June 1998 767

3b2a 0005 Mp 767 Wednesday Jun 03 01:11 PM Man Sci (June) 0005 BAYUS Analysis of Product Lifetimes

Figure 4 Annual Personal Computer Sales by CPU Chip

this period, reaching an sales level of almost 15 million In this case, the ‘‘event’’ of interest is the withdrawal units in 1992. Sales within each technology generation of a product from the marketplace. Since the assembled resemble the usual product life-cycle pattern. For ex- personal computer database includes all desktop per- ample, sales of first-generation personal computers sonal computer product introductions and withdrawals (based on 8-bit CPUs) grew slowly from fewer than 100 between 1974 (i.e., the beginning of this industry) and units in 1974, reached a peak of almost 6 million units 1992, the only situation in which an event will not be in 1984, and have generally declined thereafter. As observed is when a product is withdrawn after the ob- shown in Figure 4, total annual sales for personal com- servation period covered by the database (i.e., after puters based on various microprocessor chips also fol- 1992). In other words, some observations may be right- low the general product life-cycle shape. Note that these censored. diagrams contain truncated data, so one cannot neces- The duration time studied is the length of the product sarily conclude that the product technology and prod- life-cycle. For the available data, this duration time is uct model lifetimes are shrinking over time (e.g., 8-bit calculated as the year of product withdrawal (deter- and 16-bit personal computers continued to be sold af- mined as the year after introduction in which total do- ter 1992). mestic and international annual product sales is zero) less the year of product introduction. For example, the product lifetime for IBM’s PC XT desktop model (intro- 4. Estimation Method duced in 1983 and withdrawn in 1987) is four years. To analyze product lifetimes in the personal computer The usual approach associated with duration time industry, an event-history approach based on the statis- analysis is used to statistically model product lifetimes tical framework of failure time analysis is used (e.g., (e.g., see Helsen and Schmittlein 1993). Letting Tk be the Cox and Oakes 1984, Helsen and Schmittlein 1993). The duration time for product k, the hazard rate h(t)is advantage of this approach is that it takes into account h(t) Å f(t)/[1 0 F(t)], (1) both the occurrence and the timing of an event while estimating the effects of explanatory variables. In ad- where f(t) is the probability density function and F(t)is dition, truncated data can be accommodated within the the cumulative distribution associated with the random estimation procedure. Analyses were performed using variable Tk. Here, h(t) is the conditional likelihood that the SAS LIFEREG procedure (SAS Institute 1988). product withdrawal occurs at duration time t, given

768 MANAGEMENT SCIENCE/Vol. 44, No. 6, June 1998

3b2a 0005 Mp 768 Wednesday Jun 03 01:11 PM Man Sci (June) 0005 BAYUS Analysis of Product Lifetimes that it has not occurred in the interval (0, t). Using a ables. This test statistic is distributed as a chi-square standard assumption (e.g., see Helsen and Schmittlein (with degrees of freedom equal to the number of ex- 1993), the duration time p.d.f. is specified as a Weibull planatory variables) and is computed as twice the dif- distribution ference in model log-likelihood values.

b f(t) Å abteb01 0at , a, b ú 0, (2) 5. Analysis and Results with hazard rate h(t) Å abtb01. As discussed in the introduction, the key research ques- To estimate the effects of explanatory variables the tion is whether product lifetimes are systematically get- accelerated failure time model is used. As noted by Cox ting shorter over time. Following prior empirical studies and Oakes (1984), this model is equivalent to the pro- (e.g., Qualls et al. 1981, Bayus 1992), product life-cycle portional hazards regression model for the assumed lengths at various product market levels are statistically Weibull duration time distribution function. The accel- related to the year of product introduction to test this erated failure time model assumes that the effects of hypothesis. If product lifetimes are indeed shrinking, a explanatory variables on a duration time distribution is negative coefficient for product introduction year is ex- multiplicative (i.e., linear in logarithms) on the duration pected. The next three sections present analyses for time. In line with most applications, this scale function product lifetimes defined at various product market is assumed to be of an exponential form, eX=u, where X levels. is a vector of explanatory variables and u a vector of unknown parameters. As noted by Helsen and Schmitt- 5.1. Product Technology and Product Model lein (1993), uj represents the constant proportional effect Lifetimes of xj on the hazard rate. These coefficients can be inter- As suggested by Figures 3 and 4, directly analyzing preted as a quasi-elasticity (i.e., the percentage change product technology or product model lifetimes will not in the hazard rate for a unit change in the corresponding really be feasible because these life-cycles are relatively explanatory variable xj) by calculating 100[exp(uj) 0 1]. long (i.e., products with ‘‘old’’ technology continue to

Thus, if t0 is a duration time sampled from the base- be sold after 1992). Instead, the time to peak sales is line distribution, then the model to be estimated takes examined in this section (therefore, the duration time in the form Equation (3) is the year of peak sales less the year of product introduction). (log(T) Å l / X؅u / s log(t ), (3 0 From Figure 3 it is clear that product technology is where the parameters to be estimated are the vector u, not accelerating: the time to peak sales for 8-bit personal the intercept l, and the scale constant s. This model computers is 11 years, whereas the time to peak sales assumes that the log duration times are normally dis- for 16-bit machines is 14 years (the 32-bit personal com- tributed, thus allowing for increasing and decreasing puter sales peak is sometime after 1992). This finding is hazard rates across duration times. If the scale param- also consistent with Intel’s actual microprocessor de- eter estimate is less than one, then the hazard rate is velopment schedule (see Hof 1992, 1995). Contrary to nonmonotonic. For the assumed Weibull baseline dis- reports by the popular press (e.g., Hof 1995), Intel does tribution, the mean product life-cycle length for a model not seem to have accelerated the availability of new with no covariates can be estimated as elG(1 / s). technology because new microprocessors continue to be The model parameters are estimated by maximum introduced in approximately three-year increments, likelihood using a Newton-Raphson algorithm. Esti- with volume shipments occurring a year after introduc- mates for the standard errors of the parameters are com- tion. In many cases, estimates for future microproces- puted from the inverse of the observed information ma- sors are subject to revision (e.g., earlier estimates of the trix. The ratio of a coefficient to its standard error ap- introduction dates for the 586/Pentium underestimated proximates a t-statistic for a test of significance. Overall the actual dates; see also Brandt 1993). model fit is assessed using the likelihood ratio test, An analysis at the product model level was con- which compares the estimated model with explanatory ducted by estimating the relationship between the variables against a model without explanatory vari- time to peak sales associated with the various CPUs

MANAGEMENT SCIENCE/Vol. 44, No. 6, June 1998 769

3b2a 0005 Mp 769 Wednesday Jun 03 01:11 PM Man Sci (June) 0005 BAYUS Analysis of Product Lifetimes

Figure 5 Time to Peak Sales by CPU Chip

introduced between 1974 and 1992, and the product competing firms are not purposively shortening their model year of introduction (e.g., see Figure 4). As sug- own brand model lifetimes but are systematically pur- gested by Figure 5, the relationship between product in- suing strategies involving different product cycles, troduction year and time to peak sales is insignificant. shrinking product lifetimes would be observed across Since some observations are censored (i.e., the sales for manufacturers if the more recent entrants have rela- some product models did not peak before 1992), this tively short product lifetimes as compared to estab- can be formally confirmed by estimating the duration lished firms. To better understand observed brand model in Equation (3) (details are not reported here but model lifetimes across firms, two additional explana- are available upon request). This finding indicates that tory variables are considered.4 the time to peak sales for product models is not shrink- First, observed product lifetimes across manufactur- ing over time. ers might be related to the entry timing of firms, since 5.2. Brand Model Lifetimes Within Manufacturers later entrants into this industry tend to introduce per- The statistical model in Equation (3) was estimated for sonal computer models that have shorter life-cycles the brand model lifetimes within 20 of the top market than the products of firms already in the industry. For share leaders but for brevity is not reported here (details example, Table 1 (where the manufacturers are sorted are available upon request). The mean brand model life- in ascending order by year of firm entry) suggests that times for these manufacturers is reported in Table 1 (the earlier entrants have larger mean product lifetimes than mean brand model lifetime across all manufacturers is later entrants (e.g., compare the average of the first five 3.67 years). With few exceptions (as noted in Table 1), and last five). As noted by Steffens (1994), these later product introduction year is not statistically related to product lifetime. This finding indicates that the leading 4 As noted by one reviewer, relating other variables such as cumulative personal computer manufacturers are not systemati- industry sales or number of competing firms to observed product life- cally shortening their own brand model lifetimes. times would also be of interest. However, these variables are mono- tonically increasing between 1974 and 1992 and thus are highly cor- 5.3. Brand Model Lifetimes Across Manufacturers related with product introduction year, a key variable in this study. A The duration model in Equation (3) was also estimated complete analysis of the firm’s product line decision is left for future for brand model lifetimes across all manufacturers. If research.

770 MANAGEMENT SCIENCE/Vol. 44, No. 6, June 1998

3b2a 0005 Mp 770 Wednesday Jun 03 01:11 PM Man Sci (June) 0005 BAYUS Analysis of Product Lifetimes

Table 1 Mean Brand Model Lifetimes for Selected Personal Computer introduce personal computers that have shorter lifetimes Manufacturers (in ascending order by year of firm entry) than the products of firms already in the marketplace.

Mean Product Second, product lifetimes may be related to the prod- Life Cycle Length ucts that firms are introducing. With the exception of Manufacturer Country of Origin (Years) the initial years surrounding the birth of this industry, new and established firms can choose to introduce Groupe Bull France 3.39 products based on various available technologies. In Hewlett Packard U.S.A. 3.63 particular, firms must decide between introducing IBM U.S.A. 5.47 Tandy U.S.A. 3.74 products based on the most advanced technology avail- Apple U.S.A. 4.38 able, or based on an already existing technology. If a Northgate U.S.A. 2.861 next-generation product is available, it is expected that Digital Equipment U.S.A. 3.13 a new product based on older technology will have a NCR U.S.A. 3.42 shorter market lifetime than it might have had if the NEC Japan 4.63 Epson Japan 3.36 next-generation technology were not available. For the Sanyo Japan 3.19 total sample of personal computers, the first product of Compaq U.S.A. 3.742 new firms tends to be based on older technologies Gateway 2000 U.S.A. 4.87 (given that a next-generation technology was available, Adv. Logic Research U.S.A. 3.34 46 percent of new firms introduced a personal computer Dell U.S.A. 2.95 model based on a preexisting technology generation, Samsung Korea 2.452 Packard Bell U.S.A. 3.89 whereas only 24 percent of established firms introduced Acer Taiwan 2.99 a personal computer model with old technology; z Hyundai Korea 4.04 Å 22.5, p õ 0.01). This, combined with the entry pattern Goldstar Korea 3.97 of new firms shown in Figure 2, suggests that the like- lihood of introducing products with old technology 1 Although not reported here, a positive and significant coefficient for in- troduction year is exhibited by this manufacturer’s products. (and thus shorter product lifetimes) has increased over 2 Although not reported here, a negative and significant coefficient for time. To capture the possible effect of firms introducing introduction year is exhibited by this manufacturer’s products. products based on preexisting technology, a variable defined as the year in which the next-generation CPU technology is available less the year of product intro- entrants are faced with a maturing industry in which duction is used. Note that this variable can take on pos- distribution is shifting from direct sales force activity itive or negative values. For example, IBM’s PC XT was (e.g., dealers, value-added resellers) to mass merchant introduced in 1983 with a 16-bit 8088 CPU; thus, the time outlets (e.g., mass merchants, superstores, consumer before next generation technology is available for this prod- electronics, mail order). In addition, there is usually a uct model is 2 years (i.e., 1985, the year in which 32-bit limited supply of new components (such as CPUs) in- CPU technology was available, less 1983). A positive corporating the latest technology (e.g., Brandt 1993, coefficient for this explanatory variable is expected if Steffens 1994). Thus, many later entrants choose to com- products based on ‘‘old’’ technology have shorter life- pete by changing their product model offerings more times than products with the most advanced technology often than established firms to appeal to mass distrib- available. utors and/or to update their products as key compo- Table 2 reports the estimation results for two subsam- nents become available (see Ziegler 1995, Armstrong ples: first products of new firms and products of estab- 1996, Choi 1996). To test this hypothesis, brand model lished firms. First products of new firms represent a spe- lifetime is statistically related to the firm’s time of entry cial case, since multiple generations of technology are into this industry. Firm entry year for each product is not available at every time point. In particular, ‘‘old’’ defined as the year in which the manufacturer sold its technology is not available during the early years in the first desktop personal computer. A negative coefficient evolution of this industry (for new firms, 70 percent of for firm entry year is expected if later entrants tend to all the products introduced between 1974 and 1985 were

MANAGEMENT SCIENCE/Vol. 44, No. 6, June 1998 771

3b2a 0005 Mp 771 Wednesday Jun 03 01:11 PM Man Sci (June) 0005 BAYUS Analysis of Product Lifetimes

Table 2 Estimation Results for Personal Computer Brand Model that the time before next-generation technology is available Lifetimes (standard errors in parentheses) is the dominant factor in explaining brand model life-

Products of times across firms selling desktop personal computers. First Product of Established New Firms Firms 6. Implications and Conclusions 1 2 Intercept 72.020 30.041 The statistical results across manufacturers are consis- (15.458) (12.410) tent with the popular press reports that personal com- Product Introduction Year 00.0361 0.002 (0.009) (0.007) puter lifetimes have declined over time. However, this Firm Entry Year na3 00.0161 empirical observation is not due to an underlying ac- (0.004) celeration in product technology or product model life- Time Before Next Generation times, nor is it due to individual firms systematically 1 1 Technology is Available 0.035 0.027 reducing the life-cycles of products within their lines. (0.008) (0.004) Scale Parameter 0.5301 0.4941 Instead, the first products of firms that have entered (0.019) (0.012) this industry in the more recent years tend to be based Log-Likelihood Ratio Statistic 39.751 50.971 on previously existing technology, and, not surpris- Mean Product Lifetime ingly, these products have life-cycles that are shorter (Years) 3.89 3.52 than those of established firms. This is also consistent with recent reports in the business press (e.g., Hardy 1 Significant at 0.01 level. 2 Significant at 0.05 level. 1995). 3 For this subsample, firm entry year is equivalent to product introduction As with any empirical study, the findings in this year. paper are limited to the specific industry, product form, and time period analyzed. Still, there are some interesting implications that can be noted. First, the based on the most advanced technology available, sales opportunity window in which to obtain a financial whereas only 44 percent of all the products introduced return on invested resources in this industry is not getting between 1986 and 1992 had the newest technology; z shorter over time. For firms in the personal computer Å 10.5, p õ 0.01). Also, note that firm entry year is the industry, this finding suggests that there may be no same as product introduction year for new firms. need to increase continually the level of R&D expen- As shown in Table 2, the duration model provides a ditures. Instead, a constant level of new product de- very good fit to the underlying data as evidenced by the velopment activity is appropriate (see Meyer and Ut- significant log-likelihood ratio values. For new firms, terback 1993). Second, firms in this industry are not sys- the coefficients of product introduction year and time before tematically shrinking the lifetimes of products within their next generation technology is available are statistically sig- own lines. This finding suggests that firms in the per- nificant at better than the 0.01 level. However, the co- sonal computer industry are balancing the costs and efficient for product introduction year is not statistically returns associated with their own individual product related to the length of product life-cycles for estab- strategies (see also Lambert 1996). Third, the compet- lished firms. Instead, product lifetimes for established itive environment in this industry is complex, with mul- firms are related to the entry timing of firms and the tiple generations of technology being available in the mar- products that are being introduced. Together, these re- ket at any time. This finding implies that purchase be- sults indicate that the more recent entrants into this in- havior in the personal computer market is complex dustry have products with shorter lifetimes than firms (e.g., for a variety of reasons, some market segments that have been in this industry for some time, and prod- may prefer products with old technology, even ucts based on ‘‘old’’ technology have shorter lifetimes though newer technology is available). than products with the newest technology. Using the Given that firms in the personal computer industry coefficient estimates in Table 2 to calculate the quasi- are not systematically reducing the lifetimes of prod- elasticity transformation 100[exp(uj) 0 1], it is also clear ucts within their lines, it should not be surprising that

772 MANAGEMENT SCIENCE/Vol. 44, No. 6, June 1998

3b2a 0005 Mp 772 Wednesday Jun 03 01:11 PM Man Sci (June) 0005 BAYUS Analysis of Product Lifetimes

Figure 6 Average Manufacturer Product Line Length

the average product line length for personal computer greater confusion over the alternatives that are available manufacturers has steadily increased over time (see (e.g., Carlton 1994, Hays 1994, Wildstrom 1996). This is Figure 6). Thus, firms are not consistently withdraw- especially the case for personal computers because the ing products from the market at the same rate as they incremental ‘‘value’’ of new products over existing are introducing new products. Moreover, the busi- products may not be that large (e.g., Johnson 1993, Mc- ness press indicates that only the manufacturers with Cartney 1994, Clark 1995). full product lines are able to maintain shelf space in Because the empirical setting for this study is limited the mass distribution outlets (e.g., Armstrong 1996, to one technologically dynamic industry, future re- Choi 1996). search should consider other markets and time periods For firms, these findings suggest that there may be no before the findings discussed here can be generalized. need to continually increase the level of R&D expendi- In addition, a careful analysis of the product line deci- tures. Instead, the immediate concern for firms is to sions by firms over time would add to our understand- make sufficient returns from increasing product line ing of product lifetimes. It is expected that factors such ‘‘complexity’’ (e.g., Child et al. 1991). Some firms in the as competitive pressure (e.g., number of firms, number personal computer industry are finding that the man- of products, industry concentration), market opportu- ufacturing and marketing costs associated with broad nity (e.g., installed base, industry growth rate, price product lines are greater than the revenue associated trends), internal pressure (e.g., market share, firm age, with incremental demand from numerous product al- time since last product introduction or withdrawal), ternatives, and they are thus pruning the weaker prod- and internal opportunities (e.g., firm sales growth, ucts from their lines (e.g., Carlton 1993a). In addition, number of product models in line, available technology) firms with products that have relatively short life-cycles will significantly influence a firm’s decisions regarding continue to face strong competition on product features ‘‘new’’ product introductions and ‘‘old’’ product with- (e.g., Ramstad 1995) and still have problems drawals. These effects may also differ over time as the demand (e.g., Carlton 1993b, Schneidawind 1993, King industry evolves. 1995). It is hoped that the empirical analysis presented in For consumers, these conditions imply an increase in this paper shows the potential dangers associated with product options, configurations, and technologies. Un- blindly accepting conventional wisdom espoused by the fortunately, this expanded choice also comes with popular press. The lesson in this case is: do not rush

MANAGEMENT SCIENCE/Vol. 44, No. 6, June 1998 773

3b2a 0005 Mp 773 Wednesday Jun 03 01:11 PM Man Sci (June) 0005 BAYUS Analysis of Product Lifetimes new products to market unless a careful analysis of the Hays, L., ‘‘Too Many Computer Names Confuse Too Many Buyers,’’ benefits and costs indicates that this is indeed the best Wall Street Journal, June 29, 1994, B1, B6. Helsen, K. and D. Schmittlein, ‘‘Analyzing Duration Times in Market- strategy.5 ing: Evidence for the Effectiveness of Hazard Rate Models,’’ Mar- keting Sci., 11 (1993), 395–414. 5 The financial support of UNC’s CATO Center for Applied Business Hof, R., ‘‘Inside Intel: It’s Moving at Double-Time to Head Off Com- Research and the Roy and Alice H. Richards Bicentennial Professor- petitors,’’ Business Week, June 1, 1992, 86–94. ship Fund is gratefully acknowledged. Thanks are also extended to , ‘‘Intel: Far Beyond the Pentium,’’ Business Week, February 20, Howard Susskind for his assistance in researching the personal com- 1995, 88–90. puter industry, and to the anonymous reviewers. Johnson, B., ‘‘Intel Puts Pentium Chip on a Pedestal,’’ Advertising Age, References 3 (1993), 46. Kerin, R., P. R. Varandarajan, and R. Peterson, ‘‘First-Mover Advan- Alsop, R., ‘‘Companies Get on Fast Track to Roll Hot New Brands,’’ tage: A Synthesis, Conceptual Framework, and Research Propo- Wall Street Journal, July 10, 1986. sitions,’’ J. Marketing, October (1992), 33–52. Armstrong, L., ‘‘Savior From Down Under?’’ Business Week, January King, R., ‘‘Packard Bell’s Woes Reflect a Bad Bet, Not a Marketwide 15, 1996, 85–86. Drop in PC Demand,’’ Wall Street Journal, November 29, 1995, B10. Bayus, B., ‘‘Have Diffusion Rates Been Accelerating Over Time?,’’ Mar- Lambert, D., ‘‘First, Fast, and On-Time. The Path to Success; Or is It?’’ keting Letters, 3, 3 (1992), 215–226. Presented at Hewlett Packard’s Conference, , ‘‘Are Product Life Cycles Really Getting Shorter?’’ J. Product In- 1996. novation Management, 11 (1994), 300–308. Langlois, R., ‘‘External Economies and Economic Progress: The Case Brandt, R., ‘‘Intel: What a Tease—and What a Strategy,’’ Business Week, of the Microcomputer Industry,’’ Business History Rev., 66 (1992), February 22, 1993, 40. Carlton, J., ‘‘Apple to Peel Away Confusion Over Its Existing Product 1–50. Lines,’’ Wall Street Journal, October 18, 1993a, B1, B3. Levitt, T., ‘‘Exploit the Product Life Cycle,’’ Harvard Business Rev., , ‘‘Popularity of Some Computers Means Buyers Must Wait,’’ Wall November–December (1965), 81–96. Street Journal, October 21, 1993b, B1, B12. , ‘‘Innovative Imitation,’’ Harvard Business Rev., September-October , ‘‘Befuddled PC Users Flood Help Lines, and No Question Seems (1966), 63–70. to be Too Basic,’’ Wall Street Journal, March 1, 1994, B1, B6. McCartney, S., ‘‘Computer Buyers Are Smart to be Suspicious,’’ Wall Child, P., R. Diederichs, F. Sanders, and S. Wisniowski, ‘‘The Manage- Street Journal, September 12, 1994, B1, B8. ment of Complexity,’’ McKinsey Quarterly, 4 (1991), 52–68. Meldrum, M., ‘‘Marketing High-Tech Products,’’ European J. Market- Choi, A., ‘‘Digital to Exit Home-Computer Sector, Focus on PCs for ing, 29, 10 (1995), 45–58. Business Customers,’’ Wall Street Journal, January 30, 1996, A3, Meyer, M. and J. Utterback, ‘‘The Product Family and the Dynamics A8. of Core Capability,’’ Sloan Management Review, Spring (1993), 29– Cimento, A., J. Kluge, and L. Stein, ‘‘Excellence in Electronics,’’ Mc- 47. Kinsey Quarterly, 3 (1993), 29–40. Patterson, M., Accelerating , Van Nostrand Reinhold, New Clark, D., ‘‘Intel Puts Chip Under Pentium Brand, Meaning Some Won’t York, 1993. Reap Its Benefits,’’ Wall Street Journal, September 20, 1995, B8. Pierz, K., ‘‘Benchmarking New Product Development Funding,’’ J. Cox, W., ‘‘Product Life Cycles as Marketing Models,’’ J. Business,40 Product , 12 (1995), 43–53. (1967), 375–384. Polli, R. and V. Cook, ‘‘Validity of the Product Life Cycle,’’ J. Business, Cox, D. and D. Oakes, Analysis of Survival Data, Chapman and Hall, 42 (1969), 385–400. New York, 1984. Qualls, W., R. Olshavsky, and R. Michaels, ‘‘Shortening of the PLC— Dumaine, B., ‘‘How Managers Can Succeed Through Speed,’’ Fortune, An Empirical Test,’’ J. Marketing, 45 (1981), 76–80. February 13, 1989, 54–59. Ramstad, E., ‘‘PC Makers Looking for New Ways to Stand Out,’’ News Eisenhardt, K., ‘‘Making Fast Strategic Decisions in High Velocity En- & Observer, September 17, 1995, 6F. vironments,’’ Acad. Management J., 32 (1989), 543–576. Rosegger, G. and R. Baird, ‘‘Entry and Exit of Makes in the Automobile Foster, R., Innovation: The Attacker’s Advantage, Summit Books, New Industry, 1895–1960,’’ OMEGA, 15, 2 (1987), 93–102. York, 1986. Rosenau, M., ‘‘Speeding Your New Product to Market,’’ J. Consumer Fraker, S., ‘‘High Speed Management for the High Tech Age,’’ Fortune, Marketing, 5 (1988), 23–36. March 5, 1984, 62. SAS Institute, SAS/STAT User’s Guide 6.03, SAS Institute, Cary, NC, Goldman, A., ‘‘Short Product Life Cycles: Implications for the Mar- 1988. keting Activities of Small High-Technology Companies,’’ R&D Schnaars, S., Managing Imitation Strategies, Free Press, New York, Management, 12, 2 (1982), 81–89. 1994. Grubler, A., ‘‘Long Term Patterns and Discontinuities,’’ Technological Schneidawind, J., ‘‘Hottest PCs Can’t Keep up with Demand,’’ USA Forecasting and Social Change, 39 (1991), 159–180. Today, May 25, 1993, 1B. Hardy, Q., ‘‘Cirrus Logic’s Poor Outlook Sends Stock Tumbling, Stalk, G., ‘‘Time—The Next Source of Competitive Advantage,’’ Har- Clouds Industry,’’ Wall Street Journal, December 29, 1995, B3. vard Business Rev., 66 (1988), 41–51.

774 MANAGEMENT SCIENCE/Vol. 44, No. 6, June 1998

3b2a 0005 Mp 774 Wednesday Jun 03 01:11 PM Man Sci (June) 0005 BAYUS Analysis of Product Lifetimes

Stalk, G. and A. Webber, ‘‘Japan’s Dark Side of Time,’’ Harvard Business von Braun, C., ‘‘The Acceleration Trap,’’ Sloan Management Review, Fall Rev., 71 (1993), 93–102. (1990), 49–58. Steffens, J., New Games: Strategic Competition in the PC Revolution, Per- , ‘‘The Acceleration Trap in the Real World,’’ Sloan Management gamon Press, New York, 1994. Rev., Summer (1991), 43–50. Utterback, J., Mastering the Dynamics of Innovation, Harvard Business Wildstrom, S., ‘‘Wanted: A PC for the Masses,’’ Business Week, January School Press, Boston, MA, 1994. 15, 1996, 18. Verity, J., ‘‘Deconstructing the Computer Industry,’’ Business Week, Ziegler, B., ‘‘PC Makers’ Big Push Into the Home Market Comes at November 23, 1992, 90–100. Risky Time,’’ Wall Street Journal, November 1, 1995, A1, A9.

Accepted by Ralph Katz; received May 1996. This paper has been with the author 5 months for 2 revisions.

MANAGEMENT SCIENCE/Vol. 44, No. 6, June 1998 775

3b2a 0005 Mp 775 Wednesday Jun 03 01:11 PM Man Sci (June) 0005