Technological uncertainty and earnings dispersioneconterms.net/pbmeyer/research/micro/inequality_20080331.doc · Web view2008/03/31  · Using decennial U.S. Census data from 1960

Superstardom and technological turbulence: job-linked sources of earnings inequality

Peter B. Meyer1

Office of Productivity and Technology, Bureau of Labor StatisticsMarch 31, 2008

Preliminary and incompleteFeedback welcome, to [email protected]

Abstract. The paper analyzes trends in the dispersion of earnings within occupations in the Current Population Survey since 1968 and the decennial U.S. Census since 1960. New media technologies make it easier to transmit certain kinds of work, such as athletic performances, to wider audiences around the world, enhancing the relative payoffs to the most-favored performers. Earnings inequality rose within these occupations, consistent with the superstars effect described by Rosen (1981). Earnings inequality rose within occupations which call for working closely with new semiconductor and information technologies, such as electrical engineers and computer programmers. It is argued that these occupations experienced technological uncertainty, which leads to extraordinary opportunities, obsolescence, and therefore turbulence. The uncertainty and superstars effects would naturally occur to some extent in many occupations. Therefore we examine also occupations in which these effects are likely to be the weakest – those that call for personal interaction with other individuals. On average inequality within occupations at this other extreme has not risen.

1 The author thanks Christopher Taber, Joel Mokyr, Leo Sveikauskas, Sabrina Pabilonia, and many others for their comments and advice. Thanks to Anastasiya Osborne for much research assistance. Views expressed are those of the author, not the U.S. Bureau of Labor Statistics.

1

1. Introduction

New information technology is one cause of the rise in income inequality in the U.S.

since 1970. Autor, Katz, and Krueger (1998) showed that wage inequality tended to rise

more in those industries which had more computers, which invested more into computers as a

fraction of overall investment, and whose employees used computers more. By their

accounting, 30 to 50 percent of the increase in income inequality since 1970 could be

attributed to the use of computer technology.

Particular capabilities of computer technology produced some of this rise in inequality.

For example, much of the computer-related rise in inequality could be explained by

reorganizations of tasks to use computers to do routine work.2 The broader category of

semiconductor technology also enabled the expansion of media markets through quicker and

cheaper communication, for example on CDs and through cable television systems. This

increased the scale of distribution of the most-preferred performers and therefore their

competitive advantage in revenue terms. This superstars effect, discussed by Rosen (1981),

will be measured here.

Influential new technologies arrive along with technological uncertainty which can cause

economic turbulence and a temporary increase in earnings inequality. It is difficult to predict

the future of an immature technology with great potential. Organizations may innovate,

reorganize, and change their products and processes to capture the benefits of reduced costs,

higher quality output, and to avoid obsolescence. The new technology can therefore produce

a wave of experimentation, new engineering standards, and entrant firms. Characteristically

many of the entrants fail and a few become big successes. This turbulence among

organizations can widen the distribution of individual earnings in affected occupations by (1)

temporarily opening up valuable opportunities (such as starting a firm, or receiving incentive

stock options); (2) making opportunities depreciate rapidly, especially for those using older

2 Autor, Levy, and Murnane (2003) demonstrates this.

2

technologies and methods; and (3) expanding the range of activities and knowledge in the

affected occupation.

Subsequent sections of this paper discuss the theories of the superstars effect and

technological uncertainty, and the institutional mechanisms through which they could raise

earnings inequality. Using decennial U.S. Census data from 1960 to 2000 and March Current

Population Survey data from 1968 to 2000, occupations that experienced amplification

through the media and technological turbulence will be shown to have had rising inequality

of earnings compared to other occupations.

2. Media amplification and the superstars effectRosen (1981) modeled an effect that would occur in certain labor markets as they grow

in size. In the model, the services of sellers vary in quality, and the sellers can deliver

services to many buyers simultaneously. That is, joint consumption is possible, in the sense

that many listeners can hear a single musician and many readers can read a book, without

imposing significant costs on one another’s experience. In this environment, an expansion of

the market (quantities demanded and supplied) for the service leads to more revenue for the

top-quality sellers, but less revenue for the least-preferred sellers (who now have more

competition) and therefore there is a rise in revenue inequality among the sellers. As a

market expands, revenue inequality would increase in certain labor markets.

Several standard examples illustrate the theory. An athlete or musician before the age of

mass media performed only for those who were present. Spectators might have been almost

as likely to come to see the tenth-most famous one as the very most famous one, since it was

hard to rank them and the opportunity to see either one was rare. But once all musicians had

recordings for sale, buyers could more easily buy any of them, and the best or most famous

one may get most of sales and unknown ones no sales. Similarly, before broadcasts, if there

were only one basketball game available in town, it would not face direct competition for

basketball fans, but when there are several games on television the most appealing one may

3

get all the viewers. It follows that inequality at the top of these performance occupations

would increase as more and more recorded and packaged versions of their performances

became more widely available. Availability increases with technology (e.g., the invention

and standardization of compact disks, computer networks, and cable television), and also with

expanded trade and globalization. The work of famous musicians and authors is available

around the world, and successful performers are international celebrities. These are

superstars, in Rosen’s memorable language. Labor markets with superstars have two

distinctive characteristics: (a) the outputs of different sellers are not perfect substitutes for one

another in the minds of buyers, and (b) there are economies of distribution, meaning that the

costs of production rise more slowly than the number of buyers.

Distinctive niche sellers of such services such as athletes, dancers, or reporters, can

benefit directly from expanded media outlets and indirectly from being interviewed or

discussed on cable TV, through global broadcasts of American channels, and Internet

connections to homes and workplaces. A famous athlete, musician or author can have a

larger audience now than ever before. Top earners in these professions thus benefit

disproportionately from improvements in information and communication technology, and

globalization.

The empirical definition here is made up of occupations Rosen (1981) used as examples,

adjusted for what is available in the data and also by the other occupations in which the effect

seems to be visible. The media-amplified occupation groups are defined to be: actors,

directors, or producers; artists (artistic painters, sculptors, craft-artists, and print-makers);

athletes; authors; dancers, dance teachers, and choreographers; designers; editors and

reporters; musician or composers; and photographers. Table 6 shows three definitions: this

list, Rosen’s examples, and the examples in Frank and Cook (1995). The mean earnings in

performance jobs have not risen much on average (as shown in table 3) perhaps because

performers enter these tournament-type professions and reestablish a kind of Malthusian

equilibrium.

4

Rosen (1981) discussed a related effect, of more intense competition among

professionals that can occur in an environment where communication and transportation are

easier. It has become easier over time to compare surgeons or attorneys through phone

recommendations or online information, and also easier to travel to such specialists.

Therefore the value of a one percent increase in expected performance to a customer could be

increasingly valuable in these professions, and competition could have increased at the top of

the professions. Indeed there could be a tendency for superstars with international audiences

to appear. This argument is plausible, but the evidence on earnings does not show this.

The main superstars hypothesis to be taken to the data is that in professions whose output

can be reproduced or amplified by computer or television communications, earnings

inequality rose over the recent decades. This is not because the technologies are new per se,

but because they have been used increasingly to transmit work content and performances.

3. Technological uncertainty and turbulenceHere technological uncertainty means a lack of common knowledge and agreement about

what production technology will be relevant in the future. “It involves not only lack of

knowledge of the precise cost and outcomes of the different alternatives, but often also lack

of knowledge of what the alternatives are.” 3 Uncertainty in markets associated with a new

technology takes several forms such as uncertainty over prices, tools and materials, products

and customers, financing, and the work force. A core assumption in the Greenwood-

Yorukoglu (1997) theory is that people vary tremendously in how well they can apply the

“skill” of learning in response to new situations, and this leads to a rise in income inequality

during the adoption of a radically new technology. In that model there is a productivity

slowdown at the same time, as employers try to usefully adapt the new invention.

3 Dosi (1988) p.1134. The subject is also well discussed in Rosenberg (1996). No source seems to give a direct definition but this one seems to be approximately what they mean. Tushman and Anderson (1986) measured uncertainty by the magnitudes of the errors in forecasts of demand growth made by financial analysts. This was much higher in semiconductors than in other industries.

5

In the theory of Greenwood and Yorukoglu (1997) and Greenwood (1997), workers have

the opportunity to improve the technology of the firm at times of radically new technology.

Workers differ in their ability to make such improvements, therefore there is an increase in

the dispersion of worker productivities and in the dispersion in the distribution of earnings

when innovations are most desired by the employers. This hypothesis, that workers differ

greatly in the ability to adapt (the adaptability hypothesis) has been made in other models of

technological change such as those of Caselli (1997), Rubinstein and Tsiddon (1999) and

Galor and Moav (2000). Offering evidence for this interpretation, Bartel and Sicherman

(1999) found wages and the wage premium to education were higher in industries which had

more technological change by several measures, including research and development as a

fraction of sales. Using a specification with fixed effects on individuals they controlled for

unmeasured abilities of workers who switched industries over time. They found that

technologically changing industries tended to employ workers who had more of this

unmeasured ability, and more years of formal education. On the employer’s side, Hunter,

Kobelsky, and Richardson (2003) and Chun, Kim, Lee, and Morck (2004) have found that

greater investment by firms in information technology was correlated with more volatile

earnings, sales growth, and stock returns subsequently. Indeed information technology

projects have for decades had high rates of finishing later than planned or with an unexpected

outcome.4

Semiconductor chips that are the physical basis of most of the new information

technology could have caused this effect before microcomputers even existed.5 The transistor

was invented in 1948, the integrated circuit in 1959, the microprocessor in about 1971, and

microcomputers soon after that. Throughout the history of the semiconductor industry there

has been plenty of evidence of uncertainty such as asset price fluctuations, retreats by firms

4 Brooks (1975) attributes this to complexity and the interrelationships of those on the project.5 Meyer (2005) showed that the adoption of mass-production steel technology in the U.S. in the 1870s was associated with a rise in earnings inequality in the iron- and steel-making industries but not in other industries.

6

from product markets they had just entered, and poor understanding of production processes.

These also characterized the software and e-commerce markets as they developed.

Uncertainty about the future means today’s choices are gambles, with noisy payoffs, and this

could induce noise into the semiconductor-related occupation wages.

Below we will compare statistics measuring inequality of salaries within occupations tied

to semiconductors and information technology to inequality measures from other

occupations. This functional approach is qualitatively different from an approach of

decomposing inequality into worker attributes like demographics or education, which has

been done well elsewhere. That work leaves open the possibility that measures of education

have a signaling role or a credentialing role in a tournament game for high incomes, not a

skill role. Here we look for a more narrowly functional effect of the new technologies on a

different dimension where signaling is less likely. The hypothesis about turbulence is

different from the Greenwood-Yorukoglu hypothesis about adaptability, and also different

from the more common hypotheses about skill bias, though these are not strongly

distinguished here. Evidence for the uncertainty paradigm takes two forms here: anecdotes of

uncertainty from the players themselves, statistics about earnings inequality in some groups

changing differently from others. In principle uncertainty could also be measured by

volatility or variation in company profits, or errors in profit forecasts, as in the classic

Tushman and Anderson (1986).

Since 1968 semiconductor chips have improved dramatically and fallen in price while

the quantities produced have skyrocketed. The resulting products (such as those shown in

Table 1) and changes in work process have redefined white collar work around the world.

Semconductor performance improvements have followed an exponential pace since 1959

known as Moore’s Law. They result from the efforts of a variety of specialists including a

class of electrical engineers and other specialists, and these improvements then reverberate to

buffet the population of electrical engineers and computer specialists. Electronic design and

software design changed dramatically. Electrical engineering became less about continuous

7

flows of electricity and more about digital encoding. The dramatic technological changes put

them in a state of technological turbulence that is more intense than that felt by denizens of

other occupational categories.

Table 1. Examples of new semiconductor-related technologies

New kinds of hardware

New kinds of software

Disk drives (1960s) Word processors (circa 1976) Semiconductor memory (1971) Electronic spreadsheets (circa 1978) Microprocessors (1971), pocket calculators, about 1973

Graphical user interfaces (with mice, icons, and drop-down menus)

Bit-mapped video output (beginning late 1970s) Object-oriented computer languages Internet hardware (1970s and on) Web (1990) Microcomputers (beginning in late 1970s) Client / server distinctions Mobile phones Streaming transmission of content Handheld music and game devices E-commerce Handhelds PDAs (personal digital assistants) Web search (late 1990s and on)

Making and selling a new device is involves risky predictions. Jerry Kaplan, a founder

of the company that sold the first pen-input computers, wrote, “We are building an unproven

product for an unproven market. And the key to success is to reduce risk whenever and

wherever we can.” Many of the companies that pioneered new devices were themselves

startups, adding another layer of uncertainty: “Anyone who has managed a startup knows that

predictability is an illusion.” (Kaplan, 1994). Thompson’s (1967) theory of organizations is

constructed around the idea that “the central problem for complex organizations is one of

coping with uncertainty. . . . technologies and environments are major sources of uncertainty

for organizations, . . . differences in those dimensions will result in differences in

organizations” (p. 13). Relatedly, Lindberg (1995) offers advice specifically to managers

who confront technological uncertainty.6

6 Among those recommendations: select for skills and intellectual capacity in potential employees; train existing workers; install new technology in stages, to delay some risks and to enable learning-by-doing; establish trust with potential vendors; be flexible organizationally.

8

In this study, electrical engineers and software development employees are taken to be

the high technology employees who can change the employer’s products or processes in

response to technological opportunities. Therefore technological uncertainty predicts a

widening of the earnings distribution in these occupations. The next sections discuss several

dimensions of this.

3.1 Changing prices

An evolving technology causes changes in the market prices for its inputs and outputs.

Intel cofounder Gordon Moore described price trends and fluctuations for semiconductor

devices since the late 1950s this way:

A 20 to 30 percent price decrease per year is about average, although this average consists of periods of time when prices fall very rapidly and when they might even increase if supplies are tight. Not only does the price fall for a given integrated circuit, but as the complexity of the chip increases, the price per electronics function decreases from product generation to generation as more and more functions are integrated into a single structure. Today a complete circuit containing several million transistors costs less to the user than did a single transistor thirty-five years ago. 7

Microcomputer designers faced extraordinary declines and fluctuations in the prices of

components. Price change and volatility were permanent parts of the environment, as

performance of the best semiconductor devices, disk drives, and computer networks

improved so much. For example, one read-only memory chip declined in retail price from

$110 in 1983 to $10 in 1984, and to $3 in 1985. (Morris, 1990, p. 78). Kaplan (1994, p. 35)

described the engineer’s situation this way:

7 Moore, 1996, p. 56. R. J. Gordon (1990) estimated that computer hardware prices declined on average 19% per year from 1954 to 1984 period.

9

Hardware people are tinkerers and gamblers. Their challenge is to assemble things out of standard manufactured parts, as cheaply and reliably as possible. The problem is that these parts are constantly changing -- in configuration, price, and availability. Hardware people must spend their leisure time poring over catalogs, price lists, and specification sheets ... A short supply of parts instantly creates a black market, with skyrocketing prices. Brilliantly designed circuits become doomed products if a single component is unavailable.

This disequilibrium affected sophisticated computer makers such as the Digital

Equipment Corporation (DEC). DEC had surprised IBM in the late 1960s and early 1970s by

delivering minicomputers that competed with IBM’s mainframes. IBM then introduced its

own minicomputers. Observing the expanding class of microcomputers in the late 1970s,

IBM released its PC in 1981 to unexpected success. DEC never effectively confronted the

low-end competition from personal computers. DEC did try a series of half-measures to

confront the challenge but relative paralysis seemed to overcome the company. By the 1990s

it was losing money, permanently. These surprising shakeups illustrate the economics of

technological uncertainty. In more predictable markets, the top firms are less likely to be

overturned.

3.2 Novelty and uncertainty over tools, materials, and products

New technologies may have “properties and characteristics whose usefulness cannot be

immediately appreciated.”8 Semiconductor work in the 1960s was characterized by failures

that were not well understood. It is now thought that the failures were results of uncontrolled

impurities in the silicon. “Sometimes the problems would disappear for no apparent reason,

only suddenly to reappear. Solutions that worked one time might not work the next.

8 Rosenberg, 1996, pp. 340-349. As an example, Rosenberg discusses early fiber optic technology: “It took a number of years for some of the attractive properties of fiber optic technology to become apparent: the lack of electromagnetic interference, the conservation of heat and electricity, and the enormous expansion in bandwidth that fiber optics can provide -- the last feature a consequence of the fact that the light spectrum is approximately 1000 times wider than the radio spectrum.” (p. 342). Analogously, it was not generally understood when integrated circuits were first made how compact and reliable they could be.

10

Semiconductor manufacturing was so poorly understood that some problems were given

colorful names, such as ‘Purple Plague’ and ‘Red Death.’ Scientists routinely referred to

‘black magic’ and ‘witches' brew’ in describing their process techniques.” (Berlin, 2001, p.

83) Such language implies uncertainty even as the companies were creating and selling

products. Relatedly, the scientists were often doing engineering work. Several analysts have

observed that after the invention of the first transistors, most of the work in this area has been

fundamentally cumulative9, depending more on empirical discoveries than on scientific ones.

Rising productivity in this area depended on experimentation and imitation.

Manufacture of each generation of smaller integrated circuits requires unproven

technology which if successful gives the leader a temporary monopoly. When Intel was a

startup, its first product was immediately imitated so profits from it did not fund the company

for long, but its next product, a silicon gate metal oxide semiconductor memory chip, was not

successfully imitated for seven years. That lead-time advantage allowed the company to fund

other operations at length. The second product is an example where the technological

uncertainty worked to the advantage of the player who was first to market. Handwriting-

input computer maker GO had the opposite experience when standard manufactured chips

turned out not to behave according to their specifications outside the bounds of what regular

PCs would put them through. GO also had trouble writing on flat displays for which the

existing technology had not been used in a production environment (Kaplan, pp. 56 and 108).

GO had not planned on great difficulty with those devices -- the gamble was a surprise.

Similar problems occurred in software development, where project schedules were

dubious. Among the problems were unreliable software development tools, ambiguities in

engineering standards, and unexpected and intricate external product design problems in the

user interface. Some engineering managers believed that expanding a project development

9 The distinction between “cumulative” and “science-based” technologies has been attributed to Nelson and Winter (1982). See also Bessen and Maskin (2006).

11

team could proportionately reduce the project’s duration but often this was dramatically not

the case. (Maguire, 1994, and the classic Brooks, 1975).

3.3 Uncertainty over potential customers and what they want

Rosenberg (1996) lists issues of technological uncertainty, among them:

the impact of an innovation depends on later complementary inventions; one does not

at the beginning see the whole technological system built around the original invention

the original invention is targeted at some particular problem to begin with, and its

useful scope may expand and evolve in a way that is hard to predict

This paper, like Rosenberg’s, conflates uncertainty about making a product with the

market uncertainty associated with selling a product under the general heading of

technological uncertainty.

There may be no common agreement about the form in which a new technology is best

delivered to customers. The technology is complex and comes in many designs, possibly

from competing vendors. In the management literature this is sometimes described as the

period before a dominant design. Thomas Edison, for example, was making awkward

phonographs that could record on and play from cylinders made of wax and cardboard for

fifteen years before it became clear that the main mass market use for the phonograph was to

play pre-recorded music. (Norman, 1998, c. p 30.) Edison lost a similar battle when his

company committed to using direct current electricity when alternating current became the

standard. In the semiconductor context, an electronics magazine wrote, “In the 1962-64

period nearly every semiconductor producer got into the [field-effect transistor] business –

and out again just as quickly when optimistic predictions failed to materialize.”10

After the microprocessor had been commercialized in 1971 it was not obvious how to

make money from it. From Freiberger and Swaine, p. 14:

10 Morris (1990, p. 44) quoting from Electronics magazine.

12

Intel’s marketing department was cool to the idea of releasing the chips to the general engineering public. Intel had been formed to produce memory chips, which were easy to put to use and were sold in volume like razor blades. Microprocessors presented enormous customer support problems for the young company. [Inventor Ted] Hoff countered with ideas for applications. For instance, one could build an elevator controller around a chip. Moreover, the processor offered cost-reduction to an electronic design engineer, and the engineer would thus make the effort to design it into products. Hoff knew he would.

Even after the microprocessor was an established product, Intel did not venture into the

business of selling applications for it, although in retrospect many of these would have been

valuable businesses. In the early 1970s at Intel “talk had come up about getting into end

products, designing machines around the microprocessors, even about using a microprocessor

as the main component in a small computer. Microprocessor-controlled computers, however,

seemed to have a marginal sales potential at best. Noyce felt that microprocessors would find

their chief market in watches.”11

Starting in January, 1975, a little known store in Albuquerque offered the first extensible

microprocessor-based computer kit for sale. Customers could order it by mail and receive a

kit and build an Altair, then potentially integrate other hardware and software into it. The

company that produced the kit was operating near the edge of bankruptcy. Owner Ed Roberts

had to borrow to his limits to offer the computer kit. He was worried that hardly anyone

would buy it. In fact, electronic hobbyists placed hundreds of thousands of dollars in orders

within the first two months. The same problem occurred with the IBM PC for which IBM

under-predicted sales in its first year by a factor of six. Such mis-estimates also occurred

with the 1978 release of the first electronic spreadsheet program, Visicalc. Its publishers

believed small businesses were the natural first market but in fact it was hard to convince

them, whereas sales to corporate middle managers took off. These examples of unexpectedly

11 Freiberger and Swaine, pp. 15-16. Integrated circuit co-inventor and Intel co-founder Robert Noyce was a legend beforehand and long afterward. He was not just ill-informed.

13

rapid success are legendary.12 They occurred simultaneously with others (like Intel’s Noyce)

who could have sold products into these markets but did not, and others who did try, but

failed. It is intrinsic to the uncertainty of the environment, and the brief opportunities it

presents, that there were surprising successes and numerous failures or missed opportunities.

These pioneers were the experts, but made poor predictions. Perhaps this is a kind of

failure, and they could have done better. The “designers and managers involved in each new

generation of computers consistently failed to anticipate the uses that would be found for

their machines” (Ruttan, pp. 90-91). And “The microcomputer is a product that came out of

nowhere, at least in the sense that established firms initially misunderstood its uses and

underappreciated its importance” Langlois (1992, p.5). The uncertainty perspective is that

such errors are not idiosyncratic, but are likely when the situation is novel.

In a new-technology environment, customers may specifically ask for a feature, design or

product that is inferior to another approach, or harder to deliver. “The customer is always

right” is particularly false when the customer does not know what is possible (Southwick, p.

vi and p. 161). The usual approaches to figuring out what customers want are inappropriate,

according to a psychology professor turned Apple vice president:

But focus groups can be very misleading. They tend to reveal what is relevant at the moment, not about what might happen in the future. Users have great trouble imagining how they might use new products, and when it comes to entirely new product categories – forget it. (Norman, 1998, p. 192).

Christensen (1999) and Utterback (1996) documented cases of established firms that

were ruined because they repeatedly followed the advice of customers and made only minor

updates to their product technology, when fundamental technological changes were called for

instead. Customers naturally did not know this, could not evaluate it, or had fundamentally

different interests from the vendor. Possibly this is what happened to the vacuum tube

12 On the Altair, this account is drawn from Freiberger and Swaine, p. 37; on the PC, from (Langlois 1992, p. 23); on Visicalc, from (Cringely, 1992, circa p. 64).

14

producers in the 1950s. They had the financial resources to branch out into transistors, but

few did (Morris, p. 80).

Henderson (1993) offers theories of why after a radical or drastic invention, entrant

producers often overcome incumbent producers. One hypothesis is that the incumbent’s

incentives to cannibalize an existing product line are weak, whereas the entrant’s incentives

to survive and prosper are strong. Another theory is that the incumbent’s organization “falls

prey to inertia and complacency” (p. 248), partly because the incumbent is organized to

ignore certain kinds of information. For example, an incumbent firm’s research and

development department may naturally respond to product improvements by competitors by

investing in improving its own technology, even when the better strategy is to abandon its

existing technology. It can be hard for organizations to arrive at that consensus, however, so

Christensen (1999) recommends that established firms form subsidiaries to specialize in

potentially disruptive new technologies.

These issues are sources of productivity dispersion among individuals since sometimes it

is possible for individuals to avoid investing in the wrong technology, or to adapt in some

other way. Some people will wisely evaluate whether they must abandon a technology. IBM

executives did this when a manager stated that in order to compete in the new microcomputer

market it would have to abandon its risk-avoidance tactics and proprietary designs in order to

get a PC product out fast enough to be relevant. Large competitors such as Xerox, HP, and

DEC did not.

We conclude that while the technologies are new and changing and have an

undetermined future, there is not a common agreement about who might be potential

customers and what they want.

3.4 Uncertainty over industry entrants, financing, and survival

15

New firms are often the pioneers of discontinuous technological innovations, and new

firms often face great difficulty obtaining financing. Much of the effort of GO executives

was devoted to finding sources of financing (Kaplan, 1994, pp. 60-80). Venture capitalists

specializing in computer technology startups have had high but variable returns. “Industry

benchmarks have it that out of ten venture capital deals, one is going to turn out to be a super

winner, two if you are really lucky; and the rest will turn out to be bankruptcies or barely

break even.” (Perez, 1986, p. 106). In this quote the break-even firms were lumped in with

the bankruptcies. For a venture capital firm, a quickly bankrupt client can actually be better

than a slowly bankrupt one or a firm that barely survives since long-lived firms that are not

acquired and do not go public take much more management time. To reduce risks across

clients, venture capital firms syndicate (share) deals.

Waves of initial public offerings of stocks represent another kind of financial uncertainty

for new firms. Payoffs to early investors are much higher if the firm can go public early

(Perez, 1986, p. 140), and this depends on a kind of cycle in venture capital financing and

public investor interest. Financing for a startup is therefore more easily available when a flow

of initial public offerings of stock is anticipated. Such waves have peaked in 1959, 1961,

1969, 1983, and 1999-2000. In 1983, for example, four times as many firms went public as

in 1982, and they raised more money than all new issues raised in the ten previous years or in

the next two years (Perez, 1986).

There was an influx of new semiconductor makers after a practical integrated circuit was

demonstrated in 1959. Tilton (1971, p 53) counted a leap in industry size from about 25

firms selling transistors in 1955-58 to about 50 in 1960-63. A number of new firms,

including Intel, spun off from the innovating firm Fairchild Semiconductor (Malone, 1985;

Berlin, 2001, pp. 84-85). There was not a great wave of entry of microprocessor producers

after its invention. But with the microprocessor, semiconductor memory, and floppy disk

drive – all commercialized about 1971 -- there was a rise of downstream industries, especially

makers of personal computers.

16

Intel, Hewlett-Packard, and DEC all declined to mass-manufacture personal computers,

in the early stages when one of them could have led that industry.13 In later years when they

did try, the experienced computer makers usually failed financially in the personal computer

market. IBM was the main exception. This is a kind of productivity dispersion among the

firms illustrating the uncertainty in the market and technology.

Tushman and Anderson (1986, p. 448-9 and 455-6) measured uncertainty by the errors in

forecasts of industry sales by financial analysts summarized by the publication Predicasts

Forecasts. Forecast errors were high in the minicomputer business both before and after

1971 (when semiconductor memory was adopted into minicomputers, the floppy disk drive

was first used, and the microprocessor was invented.) Uncertainty by this measure was much

greater throughout the 1967-1976 period in the minicomputer market than in their other cases

of technological discontinuity – the appearance of the DC-3 airplane in 1959 and the

widebody jets in 1969 on the airline industry, and the first feasible huge computerized kilns

to make cement on cement makers.14

Of the top ten U.S. manufacturers of vacuum tubes in 1955, only one (RCA) was a

significant producer of integrated circuits in 1978. (Morris, 1990, circa p. 80) One study

measured a rate of 2.8 new semiconductor firms per year in the 1950s (the transistor period)

13 The founders of Apple had hoped to make computers for Hewlett-Packard but HP turned them down and they instead started their own company. DEC’s founder and president did not foresee uses for home computers in 1977 (Norman, p. 234). Xerox’s Alto office computer had a user interface something like a Macintosh in 1974, and could be networked. However, the Alto was expensive and unpleasantly slow to use. Xerox executives were ambivalent about the attempt and did not rush to try again. But if Xerox had revised the product quickly and seriously they might then have led the industry. Similar technology later emerged from Apple Corporation partly because Apple cofounder Steve Jobs took a now-infamous tour of Xerox’s laboratory, and hired his tour guide to work at Apple, where they imitated parts of the Alto’s design.14 Their measure was average error in forecasts of sales, by industry analysts I believe, which was 146% in minicomputers from 1967 to 1971 and 136% from 1972 to 1976. Their other cases were the cement industry (38% from 1963-1967 and 80% from 1968 to 1972) and the airlines (16% from 1955-1959, 78% from 1960-64, 19% from 1965-1969, and 49% from 1970-74. (These figures are from their Table 4. Tushman and Anderson drew the conclusion that uncertainty by this measure was generally higher in the period after a major technological discontinuity. The uncertainty in the industry affected by semiconductors was twice as great as in the other examples.

17

and a rise to 4.7 new firms per year from 1960 to 1972 (the integrated circuit period), but then

a decline to .7 new firms per year from 1973-78. (Levin, 1982, p. 44; Wilson, Ashton, and

Egan, 1980. Berlin (2001), p. 61, says waves of spinoffs and startup firms appeared in 1961-

3 and 1968-9. These waves were centered around specific technologies – the planar

manufacturing process and the integrated circuit, first, then large-scale-integration density

techniques in the second. Often startups had specific niche markets to which they thought

they could offer a product and get established before larger companies stepped in.) In a

sample of 35 semiconductor companies started before 1975 only one-fifth remained

independent in 1980 (Morris, 1990, p. 85). Thus there were substantial flows of entrants and

exits from semiconductor manufacture.

The new developments in semiconductors made new downstream products and industries

possible, such as digital watches, microwave ovens, mobile phones, personal computers, and

personal computer software. The early years of personal computers and software were

turbulent. The first industry sales leader in microcomputers was MITS (the maker of the

Altair) in 1975, then IMSAI in 1975-78 (Langlois, 1992, p.12), then Apple (1978-81), then

IBM, then Compaq. IBM’s attempted to regain control from the clone-makers with its

proprietary standards for the PS/2 hardware design and OS/2 operating system, but these

were overturned in 1987 by revised industry standards -- the EISA hardware standard and the

Windows operating environment which evolved into a complete graphical operating system.

Downstream from the personal computer, in the personal computer software industry,

there was dramatic evidence of uncertainty and turbulence. New companies took the field

almost entirely. Few minicomputer or mainframe software developers made software for

microcomputers. Fundamental principles of software development remained the same, but

techniques and culture did not translate well (Langlois and Mowery, 1996, p. 74).

18

Consider personal computer operating systems. Before making its PC, IBM expected

DRI, the established leader in personal computer operating systems,15 to supply the new

computer’s operating system. DRI unexpectedly declined to discuss a contract with IBM.16

IBM offered the contract instead to Microsoft which seized the opportunity and agreed to

provide an operating system. Microsoft quickly acquired one, adapted it to look like the DRI

product, and came up with PC-DOS. This turned out to be an astonishingly important

product, changing the whole industry in a path-dependent way as Microsoft could define

standards, exercise leadership, and undermine alternatives. Microsoft had not been in the

operating systems business in 1979, but by 1987 had locked down a kind of dominant

position. Even so, it had to slowly develop a radically different technology, a graphical user

interface, following the Macintosh.

Applications software categories were more turbulent yet. Industry-leading word

processors repeatedly lost their position: Electric Pencil (1976), Wordstar (1978),

WordPerfect, and finally Microsoft’s Word. In the 1980s the leading database program was

Ashton-Tate’s dBase II, which ran aground on technical complexity and was replaced as the

leader by Borland’s Paradox, until Borland self-destructed. The inventors of electronic

spreadsheets made millions in the 1970s selling their Visicalc program, but it was then beaten

by the focused effort of startup Lotus to make its 1-2-3 program on the then-new IBM PC.

This program made billions in revenue but then lost out to Microsoft’s Excel.

So survival in the personal computer industry’s first ten years was hard to predict even

by experts in the technology and market. Partly because of uncertainty in the product design

and potential customer base, a shakeout often follows the industry boom of new-technology

entrants. Companies take varied approaches to meeting customer needs. Some work, but

15 DRI stood for Digital Research International, which was a toned-down version of its original name, Intergalactic Digital Research. Such hobbyist producers can exist partly for the joy of it and may not be profit maximizing. In the early pre-profitable phase of a technology, such hobbyists may have the best available technology, before conventional microeconomics has kicked in.16

? Probably this was because IBM required signing an obtrusive nondisclosure agreement A nondisclosure agreement was a routine prerequisite to seeing or hearing about secret technology or marketing plans. It stated that the signer would not tell others of what he or she would see or hear in the private discussions. According to Cringely’s account, IBM’s nondisclosure agreement required both that the IBM technology be kept secret, and also that IBM would not guarantee to keep secret the technology of the other firm. This last unusual clause may be the reason Digital Research refused to sign.

19

others are not implemented well enough, or do not become well enough known, or for some

other reason do not become established.

3.5 Uncertainty over possible colleagues

Electrical engineers design semiconductor devices. Engineers with experience with a

very new technology are by definition not widely available, and their experiences are

idiosyncratic. In labor market terms, the supply of workers is thin and heterogenous. So the

right person for a job may not exist, or may not be available. A key engineer designing a new

technology product may be irreplaceable since the market may disappear before a substitute

could be competent (Perez, 1986, p. 104-5). Hiring is chancy too since the talents and skills

needed may not be clearly identified and agreed upon. Turnover in the semiconductor and

software industries tends to be high,17 so there is uncertainty about who will be in next year’s

workforce. Almeida and Kogut (1999) show that Silicon Valley has a distinctive pattern of

job mobility compared to other areas where semiconductor patents come from, and that when

comparing regions, high mobility predicts more patent citations holding other factors

constant.

3.6 Uncertainty hypothesis for the data

If answers existed to the many questions above about the markets, methods, and timing

of the new technology, they were not common knowledge. Workers may differ dramatically

in their ability to adapt to this situation, and employers are willing to pay more for workers

within the high-tech occupations who can respond to the uncertainty in a useful way and

improve the employer’s technology. Aside from the worker’s capabilities, different

employers and different circumstances can determine whether successful adaptation is

17 This is shown in Fallick, Fleischman, and Rebitzer (2005), and is widely believed in the relevant industries.

20

possible. Very new technologies produce a kind of disequilibrium situation in which workers

and employers make various gambles on the basis of partial information. In principle these

are different forces but the general proposition can be tested indirectly. The hypothesis to be

taken to the data is that earnings dispersion rose within occupations which involved designing

semiconductor products or using novel, incomplete or malfunctioning computer systems.

The empirical definition used here includes five occupations. In the combined occupation

scheme these are titled electrical engineers, electrical engineering technicians, computer

software developers (usually called computer programmers in the original Census

definitions), systems analysts, and data processing repair persons.

4. Data sources and definitions of variablesThe data come from the annual March CPS (Current Population Survey) and from the

ten-year population Census. These are repeated cross section samples of the U.S. population.

The 1960, 1970, 1980, and 1990 Census samples have results from 1% of the U.S.

population, and the 2000 sample has results from 5% of the population. The Census data

came from the IPUMS project site (Ruggles and Sobek, 1997), and the CPS data came partly

from Unicon and partly from the IPUMS project site (Current Population Surveys, Unicon,

1999; and King, Ruggles, and Sobek, 2003).

The data set includes respondents between 16 and 75 years old who reported a positive

income and for whom an occupation was recorded. Occupations were assigned by Census

specialists, and mapped into one of several hundred codes. The category systems for these

21

occupation codes changed each decade.18 A standardized occupation category system defined

by Meyer and Osborne (2005) is used here.19

The regressions to follow exclude respondents with zero or negative earnings, by the

definition of earnings relevant to each regression. Earnings were usually measured by wages

and salaries, and in some regression also include self-employment (or “business”) income.

Data on capital gains income and stock options was not consistently available most years, and

is not used in any regression here.

Income variables are said to be top-coded if the reporting agency does not report exact

values of high incomes. The Census and CPS report only top-coded incomes to protect the

privacy of respondents with distinctive levels of income. Similarly, negative values for self-

employment income are bottom-coded. The average of top-coded incomes each year is

known, and that value was imputed for each top-coded income as if it were every such

person’s true income.20 The IPUMS data source filled in estimates for the some top-coded

incomes in the decennial Census data. Not every year has top-coding taken into account

however. Fixed effects on years may help adjust for this somewhat in the regressions to

follow.

 5.0 Findings about earnings dispersion

18 The 1960 Census definitions were used in the 1968-1970 CPS, the 1970 Census definitions were used in the 1971-1982 CPS, the 1980 definitions were used in the 1983-1990 CPS, the 1990 Census definitions were used for the 1991-2002 CPS, and the 2000 Census has been used in the CPS starting in 2003. Apart from this remapping, occupations are reported with error for a variety of reasons. For example, the respondent may have had several jobs during the year, and only one will be recorded. There has also been no adjustment for workers who worked only part of the year; their incomes and occupations are taken literally as describing their experience in the year.19 It was also used in Autor, Katz, and Kearney (2006) and was adopted by IPUMS.org as the definition of their occ1990 variable. 20 I am indebted to Marcela Perticala, Finis Welch, Unicon Corp., and Larry Rosenblum for their advice and methods of handling topcoding and estimation outside the reported range.

22

Average salaries for electrical engineers and software developers rose at roughly the

same pace as earnings in other occupations, as shown in Table 3. Electrical engineers and

electrical engineering technicians were a stable proportion of the workforce over the entire

period, as shown in Table 4. The software categories, however, did not exist in the first

sample period. They appeared first in the 1970 Census and the 1971 CPS for the first time.

They grew quickly. As fractions of the workforce, programmers doubled and systems

analysts quadrupled from the 1970s to 2000, by which time these software workers made up

over 1% of the U.S. workforce.

The graphs then show earnings inequality within groups representing an occupation for

each year from 1968 to 2001. One measure of within-group inequality used here is the

standard deviation of log-earnings. Another measure used here is the coefficient of variation

of incomes within an occupation or industry. The coefficient of variation of a distribution is

its standard deviation divided by its mean. Both measures of income inequality are robust to

inflation and inflation measures.21 Observations of inequality were dropped from the graphs

if they were estimated from fewer than 10 respondents.

The population within each of these occupations evolved slowly and did not change

much in terms of years of education, age distribution, or other measured attributes. Let us

assume that the populations did not change very much relative to the population. Then these

choices of groups fit a structural explanation on the labor demand side. If electrical engineers

were dramatically affected by new technology, the hypothesis that people vary greatly in their

ability to adapt to a new technology suggests that great differences will appear in the

productivity of engineer A versus engineer B. In the Greenwood and Yorukoglu (1997)

environment this would lead to increased dispersion in what employers would pay them. The

same logic holds for any group that is segmented from a broader labor market -- it should

21 Meaning: if all wages were changed by the same proportion, inequality by this measure would be unchanged. This makes it possible for inequality measures to sidestep any complicated issues about comparing prices over time. All the measures of inequality in this paper are relative to other incomes in the same year.

23

show increased earnings dispersion. There could also be substitution in and out to the broader

population, but fewer than 5% of engineers per year in this data set transition to another

occupation, even to another engineering specialty. We assume here that these transitions

between groups are caused by uncorrelated factors and do not affect the distribution of the

abilities-to-adapt within the group.

Figures 1-5 show two measures of inequality for each class of occupations listed in

Tables 5-7. The occupations shown in Figure 1 have a face-to-face component and are

providers of services which do not have economies of scale in distribution. In principle there

should not be much of a superstars effect spreading their earnings distributions over time.

These professions do not face the technological uncertainty of the semiconductor, computer,

or software industries. Earnings dispersion in these occupations trends slightly downward

since 1968.

Figure 2 shows earnings dispersion for electrical engineers, electrical engineering

technicians, computer programmers, systems analysts, and data processing repair persons.

These groups confronted technological change and uncertainty directly, from declining

semiconductor prices, quality improvements, and continuing novelty in products, processes,

and markets. They created and experienced Moore’s Law most directly. There is indeed a

trend toward rising inequality in these occupations. One might think that all technical

occupations experience this effect but Figure 3 shows that other engineering categories did

not.

Figures 4 and 5 show the closest available occupations to those mentioned in Rosen’s

article, where he forecast a superstars effect. Rosen’s examples (listed in Table 6) combine

the joint-consumption performers with specialists like surgeons and lawyers whose services

might be bid up in price by the expanded sources of demand in a market with wider or easier

communication and transportation. It seems that the joint-consumption effect is visible

(Figure 4) but the bidding-up effect is not (Figure 5). The relevance of the joint-consumption

assumption is that if consumers can simultaneously benefit from the output of the worker

24

without causing negative effects on one another, and economies of distribution improve with

better communications technologies, there will be a time trend toward greater inequality.

Figure 3 shows those occupations which seem to show this media-amplified effect. In these,

some kind of fame is possible, different sellers are not perfect substitutes for one another, and

there are extreme economies in distribution. Rosen’s superstars effect is visible, in the sense

that there is a clear rise in inequality of wages and salaries.

Some media-amplification and technological uncertainty occurs in many occupations.

For example, most academics and researchers operate in environments where some kind of

economies of distribution occur (e.g. through publication or product manufacture), where

reknown is possible, and where suppliers are imperfect substitutes for one another.

However there do exist occupations at the other extreme from the high-tech and media-

amplified groups. England, Budig, and Folbre (2002) distinguished a set of occupations in

the 1980 Census that do care work, meaning work involving face-to-face service to a

recipient which increases the recipient’s capabilities. Most of these occupations were

providers of medical services, teachers at any level, and social and religious workers. These

occupations tend to allow few economies of distribution, or and in some of them different

suppliers are near-perfect substitutes for one another. In the data we shall see these

occupations do not have widening earnings distributions over the period studied. England et

al (2002) have a broader category of interactive service work, including care work but also

attendants and sales workers of various kinds. The results to be shown are slightly weaker if

this larger interactive category is used in place of just the care workers.22

The excluded category, separate from media-amplified occupations, technologically

uncertain occupations, and care work occupations, includes most managers, sales workers,

researchers, analysts and most clerical workers. The working hypothesis about them is that

they can experience some of these effects, more than care workers do.

22 Blinder (2006) referred to a similar category of personally delivered service jobs, motivated by the principle that they would be hard to relocate in another country and keep the same customers.

25

5.1. Key regressions

Tables 5-7 define several groups of occupations. One group are the five occupations

which have the main groups of people who work with novel and uncertain semiconductor

computer technology, based on my understanding of this kind of work. Designers and testers

of new semiconductor chips, for example, are in the categories of electrical engineers and

electrical engineering technicians. The design and test of digital semiconductor chips was a

new activity which joined this field whose previous canonical activity was the design of

physical circuits. The new work looked increasingly like programming.

The category of computer programmers include any software developers who are

creating new software tools. Many of these are using recently made hardware and software

to do their work. Data processing repairers do work that is not familiar to me but involves

directly interacting with recent computer technology precisely when it is not working, that is,

under unfavorable and perhaps uncertain circumstances.

A second category of occupations includes those whose work can be jointly consumed by

many customers through amplification by some kind of communication or production

technology. Rosen (1981) discussed joint consumption, and mentioned also the increasing

effect of bidding for specialists whose work was not perfectly substitutable for one another.

Based on the evidence in the graphs it appears that the key issue is whether joint consumption

through communications and transportation media are possible. See Table 4 for the empirical

definition of these media-amplified occupations.

The dependent variables in the regressions in tables 8 and 9 are measures of dispersion of

incomes in an occupation-year in one of the two data sets. We do not have hypotheses about

the levels of dispersion but rather their changes over time. Therefore we include fixed effects

on the occupations and measure only the trends. Year fixed effects are included to help

screen out a number of possible problems, such as any effects of business cycles and any6

26

effects from adjusting for top-coding in some years but not others. Perhaps most important,

the year effects remove some effects artificially created by the sharp changes in occupational

category systems which occur in Census years. The standardized occupation system can only

imperfectly compensate for changes in the way observations were originally categorized.

The test which is most illuminating is a test of these hypotheses: that the high tech

category and the media-amplified category will exhibit rising inequality over time in the

regressions, and that the care work category will not. Results from the OLS regression on the

CPS data are in Table 8, and for the Census in Table 9. With a fixed effect on occupations

and years, we regress the measure of dispersion on a year trend for each of the (a) media-

amplified occupations, (b) occupations facing the most semiconductor-related technological

uncertainty by the earlier definition, and (c) care work occupations. Observations were

weighted by the size of the sub-samples from which the dependent variable was estimated.

Here are the central results from the first regression in the first panel of each table:

Table 2: Key predictors of trends in earnings dispersion within occupations

Predictor

Dependent variable is standard-deviation of ln(wage and salary income) within each occupation-years

in annual CPS in decennial CensusCoefficient p-value coefficient p-value

Annual trend for media-amplified occupations

0.023 0.000 0.025 0.000

Trend if high tech uncertain / turbulent occupation

0.020 0.000 0.014 0.000

Trend if care work occupation

-0.002 0.268 -0.005 0.221

Regressions include fixed effects for occupation and year. Coefficients are in bold if the p-value is smaller than .05.

27

By the definitions of the groups and the inequality measure used here, the media-

amplified (superstars) professions and the technologically turbulent (high tech) professions

experienced growing earnings inequality within them. It is detected at a high level of

statistical significance in both data sets. Furthermore the hypothesis that care work jobs do

not experience rising inequality based on the forces of technological turbulence or media-

amplification is supported. In these occupations the trend in inequality is not statistically

significantly different from zero and is more likely to be negative than positive.

It does not seem that these results are artifacts of growing or shrinking occupations, or

rising or falling wages generally in the occupations. For example, the number of systems

analysts more than quadrupled in the period, as shown in Table 4, but the number of electrical

engineers did not rise much, yet we see a rise in earnings dispersion within both groups.

Tables 8 and 9 show other inequality measures. One of them is the coefficient of

variation, which is the standard deviation of the population of wages divided by the mean

wage. This measure (like the standard deviation measure above) is fairly robust to inflation

and measures of it -- that is, by both measures, an inflation that affects all wages would not

change the inequality measure. Another measure used in these tables is the interquartile

range, meaning the difference between the 75th percentile wage and the 25th percentile wage.

A third measure includes not only the wage and salary income, and also self-employment

income, and takes the standard deviation of the log of this measure of income. Using any of

these inequality measures, the media-amplified professions are increasingly unequal to a

statistically significant degree, and in most of the regressions the high tech ones are too.

The remaining regressions test variants of these hypotheses. Consider first this

alternative. One might think that the high tech professions are not so narrowly determined.

Perhaps the turbulence hypothesis would apply to larger categories of engineers, technicians,

or mathematically sophisticated workers. To test this, regression 2 in panel A of table 8

shows the same regression as above with groups of engineering occupations and technician

occupations added to the regression. The coefficients are opposite in sign to those for the

28

turbulent semiconductor-influenced occupations. In fact, the earnings distributions of

engineering and technician professions seems to have become more compressed over time.

Any force of technological turbulence has been overwhelmed by something else -- perhaps,

greater standardization in training, certification, and tools available to these other categories

of work.

A second alternative comes from Rosen (1981), which defined two kinds of superstars

occupations -- not only those which have electronically transmissible work content, but also

those which might draw greater and greater competition from buyers. Rosen gave the

example of doctors and lawyers (and others, listed in table 6). In regressions 3, 4, and 5 of

Table 8, the distributions of earnings seem to be compressing in the medical and legal

professions. More information technology may induce greater standardization in these

professions which overwhelms any superstars properties. For example, access to Westlaw

and Lexis-Nexis may put lawyers at different employers on a more even footing than they

were decades ago. In the medical area, standardized data about pharmaceutical drugs and

other treatments may be better verified and certified, or more widely available, than it was in

1970. Or, insurance companies may face more perfect competition than they did in 1970.

There is not firm confirmation of earnings compression from the regressions in table 9, whose

coefficients for doctors and lawyers jump around and are not statistically significant in most

cases.

Frank and Cook, in their 1995 book The Winner-Take-All Society, have a similar, yet

more expansive category of professions in which competition could become more intense and

incomes more unequal. Interpretations are required to test their propositions in this data,

since there is not a perfect mapping between the occupations they used as examples and the

ones identified in the data, but Frank and Cook do indicate that researchers, faculty,

managers, and sales professions should experience some winner-take-all phenomena, which

is roughly analogous to the superstars effect. In the regressions in tables 8 and 9, these

professions do not seem to have rising inequality. Some of these jobs are also defined as care

29

work occupations, and seem to look more like care professions according to the inequality

trends than like superstars professions.

A kind of direct test of uncertainty, analogous to the one in Meyer (2005) is possible.

Predictors of wages ought to lose some traction, that is, predictive power, in a period of

uncertainty. The regression in Table 10 covers CPS wages from 1992 to 2002 (the time

restriction has to do with the data available temporarily) using occupation, weeks working in

the year, age, and education as predictors. Subtracting the predicted wages from the

observed wages generates a series of wage residuals. Regressing the square of these residuals

on various predictors tells us how the magnitudes of the errors in the predictions are a

function of time, occupation, or other regressors. (The procedure is potentially analogous to

an ARCH specification of volatility in finance, but not precise enough here for the analogy to

be strong.) In Table 11, we see that high tech occupations had slightly improving

predictability over this period. Perhaps this is because the education variable includes more

relevant computer experience in the population at the end of the period than it did at the

beginning. More usefully, perhaps, we see that for the media-amplied professions, the wages

regressions has lost a lot of traction – for performers, formal education may matter less and

less. If this finding is reliable in various studies, it could be used to define the superstars

professions in a statistically reliable way, rather than depending on the ad hoc beliefs or

examples of particular researchers.

Generally speaking, the results for the high tech effect were weaker in the regressions

from the Census than from the CPS. One possible reason for this comes from the Census

1960 measures of inequality for electrical engineers and electrical engineering technicians,

which are higher than in 1970. In 1960, there were almost no semiconductor engineers.

Perhaps inequality among electrical engineers (such as experts in telephone systems and

power systems) was declining at that time, and the population of semiconductor engineers

was too few to make a difference. If so, inequality among electrical engineers would have hit

a low some time between 1960 and 1980, and only then would the effect to be estimated have

30

dominated. A re-specification which would allow this could strengthen the results,

statistically. A second reason that there would be noise in the 1960 data is that the Census

had so many fewer classifications of occupations then. Careful work (outlined in Meyer and

Osborne, 2005) could make it possible to impute 1970 occupations to 1960 data on the basis

of dual-coded data sets with other variables taken into account besides the 1960 occupation

classification. These are avenues for further improvement.

6. Unmeasured income or wealthSome effects of technological uncertainty could be under-measured using the measure of

wage and salary income that has been used above. First, wage and salary measures do not

include all income coming from high technology employers. Employees of startups in the

new microcomputer-related industries often received stock options, whose returns were

dramatically unequal. In one survey of high technology employees, 10% of executives, 85%

of managers, and 42% of other employees participated in stock option plans in 1997

(Southwick, 1999, p. 165). 24.1% of electrical engineers in one survey were offered

employee stock option plans (IEEE, 1995, p. 5-2). The CPS has a measure of capital gains

income but it is not incorporated into the dependent variables in this paper. The Census does

not have a measure of capital gains.

Second, technologically turbulent occupations could well have more self-employment or

consulting income than other occupations do. However, a comparison of regression 10 in

table 8 to regression 1 suggests the opposite; the high tech effect is weaker once self-

employment income is taken into account. This is a surprise and calls for further

investigation.

Third, some expansion in payoffs in an industry or occupation could take the form of

expanded opportunities for promotion or demotion and unemployment. For example, if some

electrical engineers became managers, or founders of startups, then there was prospective

diversity in the population that the displayed measures of current-year income did not

31

include. The founders of Apple, Microsoft, Cisco, Google, and other technical companies

would disappear from the technologically-uncertain category at the time they became

managers, and yet their extraordinary wealth afterward is partly attributable to the

opportunities that came about because of their special knowledge of the relevant technologies

and the great, fast-depreciating opportunities. And on the other side, large fractions of high

tech startups go under. So the present value of a high tech employee’s career path may be

volatile compared to others even when taking present-day income into account. These

differences have not been measured here. CPS respondents report their occupation and

industry in both the previous year and the current one. By this measure (not shown),

transitions into and out of these occupations do not seem to have changed over time for these

occupations. It could be that the later state – manager, or founder, for example – has a more

volatile income or wealth prospect in the later period than the earlier period. Particularly

undercounted are the volatile effects of engineers leaving stable firms to start their own or

those who join one and receive stock in the new venture. If in the future it were possible to

use income tax data for this study the effects might be stronger.

Relatedly, high income values for wage-and-salary, have been censored (“top-coded”) to

protect the privacy of respondents. When possible, estimates of the true values were used in

this study (as explained earlier), based on averages over high-income groups. Using these,

measures of the mean incomes of a group are unbiased, but most measures of inequality like

standard deviation are reduced by the substitution.23 This bias probably hides some of the

uncertainty and superstars effects, since huge fortunes were made through these processes.24 23 The standard deviation will usually decline, and cannot increase, if one replaces observations by their average. To see this, consider a set {1, 2, 3} and its standard deviation, and what the standard deviation would be if we replaced any two of those observations by their average. This would not change the mean of the population, but it reduces the influence by observations far from the mean on the sum of squared differences from the mean. 24 Consider the top incomes of star athletes, or of high tech company founders who are now the richest Americans. (The great fortunes of Bill Gates and other Microsoft executives, for example, arose through a highly path-dependent, technologically uncertain process.) Replacing these incomes by averages introduces a bias across occupations. Further research may be possible in data sets with high-income Americans such as the Survey of Consumer Finances, or Census and CPS data before the

32

7. ConclusionThere is evidence here of a rise in earnings dispersion within the media-amplified

occupations since 1968, supporting Rosen’s superstars hypothesis about joint consumption

and trends in economies of distribution. The data here did not support the hypothesis of

Rosen (1981) and Frank and Cook (1995) that expanded markets would produce substantially

more unequal incomes in the larger class of professions where the output of different workers

is not substitutable.

There is evidence also of a long term rise in earnings dispersion within the high tech

occupations. Turbulence or uncertainty in information technology seems to have been high

throughout the period, presumably because it is closely related to the rapid improvements in

the capabilities of semiconductor chips. This would produce in the population several

effects: obsolescence of previous skills; big opportunities; and qualitative differentiation in

the tasks of people in these jobs.

Hopefully methods like these can enable observers to detect long-term technological

change and turbulence in a statistical way. The U.S. “productivity slowdown” after 1973

may be related to the chaotic economic processes of inventing new devices, competing in

disequilibrium markets, and setting new standards. It also could help discussion of inequality

move beyond a generic, one-dimensional “skill bias”, and toward a substantive understanding

of the effects of particular technological changes on particular jobs. Another long term

benefit could be a clarification of the kinds of labor market regulation that restrict adaptations

to technological change. To gain the benefits of what is technologically possible, it may be

necessary for a subset of workers to operate in disequilibrium environments for long periods,

and some economic and regulatory environments may support this better than others.

top-coding in public use samples.

33

Std dev of ln(weekly earnings) for care work occupationsyear

Std dev of log-oc incs in Censu Std dev of log-oc incs in CPSPhysicians

0

.5

1

1.5

Dentists Optometrists Podiatrists Other health and therapy jobs Registered nurses Respiratory therapists

Occupational therapists

0

.5

1

1.5

Physical therapists Speech therapists Therapists, n.e.c. Physicians' assistants Earth, environmental, and marine Biology instructors

Chemistry instructors

0

.5

1

1.5

Physics instructors Psychology instructors History instructors Sociology instructors Math instructors Education instructors

Law instructors

0

.5

1

1.5

Theology instructors Home economics instructors Humanities instructors Other academic subject instructo Kindergarten and earlier school Primary school teachers

Secondary school teachers

0

.5

1

1.5

Special education teachers Other teachers, pre-college Vocational and educational couns Librarians Social workers Recreation workers

Clergy and religious workers

1960 20030

.5

1

1.5

Dental hygenists

1960 2003

Licensed practical nurses

1960 2003

387

1960 2003

Dental assistants

1960 2003

Health aides, except nursing

1960 2003

Child care workers

1960 2003

Figure 1a. Inequality within care work occupations for years 1960-2000 (Census) and 1968-2003 (CPS).

Within these occupations, earnings dispersion generally drifted down over this period.

A core activity of these jobs is face-to-face contact with the recipient of the service. The care work category was defined by England, Budig, and Folbre (2002). The hypothesis in the text is that these occupations are not affected by changes in technology and globalization that produce economies of scale in distribution for other occupations.

The coefficient of variation (defined to be the standard deviation divided by sample mean) is a measure of earnings dispersion. It is unaffected by a general inflation changing all wages by the same percentage.

34

Each observation displayed represents a sample size of at least ten salaries.

35

CV of ln(weekly earnings) for care work occupationsCV of ln(weekly earnings) for care work occupationsyearyear

in decennial Census in decennial Census in CPS in CPSPhysiciansPhysicians

00.5.511

1.51.5

DentistsDentists OptometristsOptometrists PodiatristsPodiatrists Other health and therapy jobsOther health and therapy jobs Registered nursesRegistered nurses Respiratory therapistsRespiratory therapists

Occupational therapistsOccupational therapists

00.5.511

1.51.5

Physical therapistsPhysical therapists Speech therapistsSpeech therapists Therapists, n.e.c.Therapists, n.e.c. Physicians' assistantsPhysicians' assistants Earth, environmental, and marineEarth, environmental, and marine Biology instructorsBiology instructors

Chemistry instructorsChemistry instructors

00.5.511

1.51.5

Physics instructorsPhysics instructors Psychology instructorsPsychology instructors History instructorsHistory instructors Sociology instructorsSociology instructors Math instructorsMath instructors Education instructorsEducation instructors

Law instructorsLaw instructors

00.5.511

1.51.5

Theology instructorsTheology instructors Home economics instructorsHome economics instructors Humanities instructorsHumanities instructors Other academic subject instructoOther academic subject instructo Kindergarten and earlier school Kindergarten and earlier school Primary school teachersPrimary school teachers

Secondary school teachersSecondary school teachers

00.5.511

1.51.5

Special education teachersSpecial education teachers Other teachers, pre-collegeOther teachers, pre-college Vocational and educational counsVocational and educational couns LibrariansLibrarians Social workersSocial workers Recreation workersRecreation workers

Clergy and religious workersClergy and religious workers

19601960 2003200300.5.511

1.51.5

Dental hygenistsDental hygenists

19601960 20032003

Licensed practical nursesLicensed practical nurses

19601960 20032003

387387

19601960 20032003

Dental assistantsDental assistants

19601960 20032003

Health aides, except nursingHealth aides, except nursing

19601960 20032003

Child care workersChild care workers

19601960 20032003

Figure 1b. Coefficients of variation (standard deviation divided by sample mean) for years 1960-2000 (Census) and 1968-2003 (CPS).

Within these occupations, earnings dispersion generally drifted down over this period.

A core activity of these jobs is face-to-face contact with the recipient of the service. The care work category was defined by England, Budig, and Folbre (2002). The hypothesis in the text is that these occupations are not affected by changes in technology and globalization that produce economies of scale in distribution for other occupations.

36

The coefficient of variation (defined to be the standard deviation divided by sample mean) is a measure of earnings dispersion. It is unaffected by a general inflation changing all wages by the same percentage.

Each observation displayed represents a sample size of at least ten salaries.

37

Std dev of ln(weekly earnings) for high-tech/uncertain occupationsyear

Std dev of log-oc incs in Censu Std dev of log-oc incs in CPS

Electrical engineers

0

.5

1

1.5

Computer systems analysts, admin Electrical engineering technicia

1960 2003Computer software developers

1960 20030

.5

1

1.5

Data processing equipment repair

1960 2003

Figure 2a. Integrated circuit chips doubled in capacity each 18 months over this period, and the work content of many holders of these occupations changed dramatically in response. These induced great uncertainty about future technologies which the text argues generated an increase in earnings dispersion. We see above a rise in the dispersion of salaries by this inequality measure.

The standard deviation of log-incomes is a measure of earnings inequality that is unaffected by a general inflation changing all wages by the same percentage. Each observation displayed represents a sample size of at least ten salaries.

38

CV of ln(weekly earnings) for high-tech/uncertain occupationsyear

in decennial Census in CPS

Electrical engineers

0.51

1.5

Computer systems analysts, admin Electrical engineering technicia

1960 2003Computer software developers

1960 20030

.51

1.5

Data processing equipment repair

1960 2003

Figure 2b. The coefficient of variation of a sample of wages is the standard deviation divided by the mean. This is a measure of earnings inequality that is unaffected by a general inflation changing all wages by the same percentage. Each observation displayed represents a sample size of at least ten salaries.

39

Std dev of ln(weekly earnings) for other engineers and techniciansyear

Std dev of log-oc incs in Censu Std dev of log-oc incs in CPS

Aerospace engineers

0

.5

1

1.5

Materials and metallurgial engin Petroleum, mining, geo engineers Chemical engineers

Civil engineers

0

.5

1

1.5

Industrial engineers Mechanical engineers

1960 2003

Engineers not elsewhere classifi

1960 2003Engineering technicians, n.e.c.

1960 20030

.5

1

1.5

Mechanical engineering technicia

1960 2003

Figure 3a. Other engineering and technical jobs, apart from those closely associated with semiconductor improvements, do not show a rising trend in the standard deviation measure.

40

CV of ln(weekly earnings) for other engineers and techniciansyear

in decennial Census in CPS

Aerospace engineers

0.51

1.5

Materials and metallurgial engin Petroleum, mining, geo engineers Chemical engineers

Civil engineers

0.51

1.5

Industrial engineers Mechanical engineers

1960 2003

Engineers not elsewhere classifi

1960 2003Engineering technicians, n.e.c.

1960 20030

.51

1.5

Mechanical engineering technicia

1960 2003

Figure 3b. By the coefficient of variation measure.

41

Std dev of ln(weekly earnings) for media-amplified occupationsyear

Std dev of log-oc incs in Censu Std dev of log-oc incs in CPS

Writers and authors

0

.5

1

1.5

Designers Musicians and composers Actors, directors, and producers

Art and craft makers

0

.5

1

1.5

Photographers Dancers

1960 2003

Art and entertainment performers

1960 2003Editors and reporters

1960 20030

.5

1

1.5

Athletes, sports instructors, an

1960 2003

Figure 4a. In these jobs different producers are imperfect substitutes for one another, and they have great economies of distribution as large scale computer and other information networks grew. Following Rosen (1981), a superstars effect is possible at the top. Indeed there is a rise in the dispersion of earnings above.

Within most of these occupations, earnings inequality rose substantially. The text argues that this was partly because of the growth in markets and technological aspects of distribution which made it easier for the demand for their services to be satisfied by a few “superstars”.

42

CV of ln(weekly earnings) for media-amplified occupationsCV of ln(weekly earnings) for media-amplified occupationsyearyear

in decennial Census in decennial Census in CPS in CPS

Writers and authorsWriters and authors

00.5.511

1.51.5

DesignersDesigners Musicians and composersMusicians and composers Actors, directors, and producersActors, directors, and producers

Art and craft makersArt and craft makers

00.5.511

1.51.5

PhotographersPhotographers DancersDancers

19601960 20032003

Art and entertainment performersArt and entertainment performers

19601960 20032003Editors and reportersEditors and reporters

19601960 2003200300.5.511

1.51.5

Athletes, sports instructors, anAthletes, sports instructors, an

19601960 20032003

Figure 4b. In these professions, different producers are imperfect substitutes for one another, and they have great economies of distribution as large scale computer and other information networks grew. Following Rosen (1981) we may expect a rise in earnings dispersion within these professions, where a superstars effect is possible at the top. We see above a rise in the dispersion of salaries by this inequality measure.

Within most of these occupations, earnings inequality rose substantially. The text argues that this was partly because of the growth in markets and technological aspects of distribution which made it easier for the demand for their services to be satisfied by a few “superstars”.

43

44

Std dev of ln(weekly earnings) for doctors and lawyersyear

Std dev of log-oc incs in Censu Std dev of log-oc incs in CPS

Physicians

0

.5

1

1.5

Dentists Veterinarians

Optometrists

1960 20030

.5

1

1.5

Podiatrists

1960 2003

Lawyers

1960 2003

Figure 5a. Inequality within groups of doctors and lawyers

45

CV of ln(weekly earnings) for doctors and lawyersyear

in decennial Census in CPS

Physicians

0

.5

1

1.5

Dentists Veterinarians

Optometrists

1960 20030

.5

1

1.5

Podiatrists

1960 2003

Lawyers

1960 2003

Figure 5b. In these professions, different producers are imperfect substitutes for one another, but they do not have strongly growing economies of distribution as a result of new technologies. Because they are competing in larger markets in which comparison and travel are easier, Rosen (1981) predicted there would be a rise in earnings dispersion within these professions, where a superstars effect is possible at the top. However this is not supported in the evidence, above. It seems to be necessary for there to be joint consumption of the output for the changes in communication and transportation to have this effect as seen in Figure 4. The predominant effect seen here seems to be that these are face-to-face occupations in which technology uncertainty and superstars effects actually have the least effect on the earnings distribution. What we see looks more like increasingly perfect competition as forecast by Stigler (1960).

46

Table 3. Illustrative rises in nominal wage-and-salary change by occupation

Job titleChange in average

nominal wage-and-salary income, 1970 Census-

2000 CensusPhotographers 200%Judges 266%Computer systems analysts 290%Economists 307%Industrial engineers 326%Metallurgical and materials engineers 333%Mechanical engineers 338%Aerospace engineers 341%Editors and reporters 348%Chemical engineers 362%Civil engineers 369%Electrical and electronic engineers 378%Architects 387%Accountants and auditors 390%Secretaries 414%Librarians 439%Computer software developers 534%Nurses 667%Lawyers 899%Podiatrists 1059%

These are ratios of the average salaries in selected occupations based on unweighted observations from the 1970 Census and 2000 Census.

They illustrate (in a rough way) that the high tech occupations are not receiving disproportionately more pay per capita than thirty years ago.

47

Table 4. Frequency of selected occupations

Data come from the U.S. Census data downloaded from IPUMS. NA stands for “not available.” The occupation classification is that of Meyer and Osborne (2005).

Percentages of US population aged 16-75, weighted by Census weights

Occupation 1960 1970 1980 1990 2000Accountants and auditors 0.46% 0.61% 0.68% 0.98% 0.98%Architects 0.03% 0.04% 0.07% 0.09% 0.11%Aerospace engineer 0.05% 0.05% 0.06% 0.09% 0.06%Metallurgical and materials engineers 0.02% 0.01% 0.02% 0.01% 0.02%

Chemical engineers 0.04% 0.04% 0.04% 0.04% 0.04%Civil engineers 0.15% 0.14% 0.14% 0.16% 0.15%Electrical engineer 0.16% 0.23% 0.21% 0.29% 0.20%Industrial engineers 0.09% 0.14% 0.13% 0.11% 0.11%Mechanical engineers 0.14% 0.15% 0.13% 0.12% 0.16%Engineers not elsewhere classified 0.09% 0.16% 0.20% 0.24% 0.01%

Computer systems analysts and computer scientists NA 0.07% 0.13% 0.28% 0.38%

Registered nurses 0.78% 0.88% 0.95% 1.19% 1.31%Librarians 0.10% 0.12% 0.14% 0.13% 0.11%Economists 0.02% 0.05% 0.07% 0.09% 0.01%Lawyers and judges 0.18% 0.22% 0.34% 0.46% 0.49%Photographers 0.05% 0.06% 0.07% 0.10% 0.08%Editors and reporters 0.11% 0.14% 0.15% 0.17% 0.15%Licensed practical nurses 0.29% 0.25% 0.33% 0.28% 0.36%Electrical and electronic engineering technicians 0.09% 0.13% 0.18% 0.25% NA

Computer software developers NA 0.13% 0.21% 0.40% 0.76%

Secretaries 2.13% 3.08% 3.08% 2.72% 2.32%

The software occupations have quadrupled as a fraction of the work force since 1970. Programmers and systems analysts were not separately counted in the earliest classification. Electrical engineers did not grow much as a fraction of the population. We have not matched particular occupations to the electrical engineering technician category from the 2000 Census.

48

Lawyer and judges, and accountants and auditors were also growing occupations over this time period.

Table 5. Conjectured Moore’s Law occupations

Jobs in these categories were especially sensitive to change and turbulence in information and communications technologies.

Electrical engineersElectrical engineering technologistsComputer programmersSystems analystsData processing equipment repairers

49

Table 6. Media-amplified jobs, experiencing superstars effect

Panel A. Rosen (1981) illustrated the superstars discussion with these examples. The phrasings do not conform perfectly to Census occupation classifications. These were examples of occupations providing services which were imperfectly substitutable with one another:

artistsauthors (3 times)authors of textbookscomediansdoctors (3 times)economic theorists and methodologistslawyers (2 times)musical soloistsnetwork news broadcastersnews reportersperformer on televisionperformers (theater, TV, and movies)performers (2 times)

pro athletes (3 times)scholars (as writers)singerssurgeonswriters

Then, quoting Marshall (1947), who described this phenomenon too:business menbarristersjockeyspaintersmusicians

Doctors, lawyers and other experts could experience more intense bidding for their services in larger markets. This could generate a superstars effect, though it seems to be overwhelmed by other factors in the data; see Figure 5. The other occupations on the list also have the key property of joint consumption on the part of the client or consumer, which is to say they are the ones subject to media amplification, and this does seem to support a superstars effect in the data.

Panel B. The definition used in this paper, and tested in the data. Services by these occupations can be amplified through larger markets, which are generated by more advanced communications and transportation technologies. These support joint consumption of output as hypothesized by Rosen (1981). In the context of imperfect competition, these economies of scale in distribution leads to superstars effect – growing inequality of earnings. The occupations examined here for this effect in Figure 4 are:

Actors, directors, or producersArtists (artistic painters, sculptors, craft-artists, and print-makers)AthletesAuthorsDancers, dance teachers, and choreographersDesignersEditors and reporters

50

Musician or composersPhotographers

Panel C. Frank and Cook (1995) extend this kind of discussion substantially. Their book refers to an even wider category of professions within which increasingly intense competition would be possible. Their book refers to these occupations at various points as experiencing winner-take-all phenomena:

fashion modelsscreenwritersactors directorscomposersbusiness consultantsfinanciersjournalistsaccountantschiropractorsdentistssalespeople

painterswritersmusiciansathletesbusiness managersauthorslawyersacademic facultyscientistsresearchers news reportersperformers on television

For this paper, media-amplified jobs are the ones shown in listed in the first part of table 4. Engineers have occupations between code 44 and 60, inclusive. Technicians have jobs from 203 to 225, inclusive. Scientists are those with jobs between 68 and 83, inclusive. Academic faculty are those with jobs between 113 and 154, inclusive. Managers are those with jobs between 4 and 22, inclusive. Doctors have jobs 84 through 88, thus including dentists and veterinarians. Lawyers have occupation 178.

51

Table 7. Care work jobs

These are the care work occupations, as defined by England, Budig, and Folbre, 2002, as mapped into the occupation categories used here. These kinds of work involve face-to-face interactions with clients or customers and involve increasing the recipient’s capabilities.

PhysiciansDentistsOther health and therapyOptometristsPodiatristsNursesPhysical therapistsSpeech therapistsPhysicians' assistantsBiological science instructorsChemistry instructorsPhysics instructors History postsecondary teachersPostsecondary teachers of sociologyMath teachers, postsecondaryPostsecondary teachers of educationTeachers of law, generally postsecondaryPostsecondary theology teachersHome economics postsecondary teachersAcademic subject instructors, n.e.c.Teachers (secondary, primary, and earlier)LibrariansClergy and religious workersDental hygienistsLicensed practical nursesDental assistants Child care

52

Table 8. Predictors of earnings dispersion in occupation-years in CPS

Panel A. Dependent variable is standard deviation of log-wage-and-salary within occupation-year

PredictorsRegression 1 Regression 2 Regression 3 Regression 4 Regression 5 Regression 6

coeff p-value coeff p-

value coeff p-value coeff p-

value coeff p-value coeff p-

valueTrend if media-amplified job 0.0023 0.016 0.0022 0.019 0.0022 0.018 0.023 0.000 0.018 0.000 0.018 0.000

Trend if high tech job 0.0069 0.000 0.0068 0.000 0.023 0.000 0.020 0.000 0.016 0.000 0.020 0.000

Trend if care work -.0003 0.549 -.0004 0.386 -0.002 0.254 -0.002 0.382 -0.005 0.018 -0.005 0.017

Trend if engineer         -0.007 0.034         -0.010 0.001

Trend if technician         -0.006 0.016         -0.009 0.000

Trend if doctor             -0.017 0.011 -0.018 0.007 -0.019 0.006

Trend if lawyer             -0.007 0.062 -0.011 0.002 -0.011 0.001

Trend if scientist                 0.008 0.159 0.008 0.179Trend if college faculty                 -0.001 0.574 -0.002 0.489

Trend if manager                 -0.031 0.000 -0.031 0.000

Trend if sales job                 -0.004 0.122 -0.004 0.092

Annual trend overall -.0022 0.000                    

35 year fixed effects no yes yes yes yes yes

386 occupation fixed effects Yes yes yes yes yes yes

sample size 11187 11187 11187 11140 11140 11138

Adjusted R-squared 0.68 0.71 0.90 0.90 0.91 0.91

Figures in bold are statistically significant at the .05 level (that is, the p-value<.05). In the regressions, each observation has been weighted by its sample size. Robust standard errors underly the p-values. In the underlying calculations of inequality, observations of income were equal-weighted, that is, did not use the CPS-assigned person-weights which adjust for demographic characteristics. Year fixed effects help compensate for the changing number of occupations -- fewer occupation categories mechanically tends to mean more variation within them, even if there were no substantive change.

Table 8.

Panel B. Dependent variable is named measure of dispersion within occupation-years in CPS

Predictor

Dep var: coefficient of variation of wage-and-salary

income within occupation-years

Dep var: interquartile range (75th percentile wage income

minus 25th percentile of wage income) within occupation-

years

Dep var: standard deviation of wage-and-salary income plus

self-employment income within occupation-years

Regression 7 Regression 8 Regression 9 Regression 10 Regression 11 Regression 12

Coeff p-value coeff p-

value coeff p-value coeff p-

value coeff p-value coeff p-

valueTrend if media-amplified 0.008 0.000 0.008 0.000 0.072 0.000 0.051 0.000 0.007 0.000 0.007 0.000

Trend if high tech 0.004 0.000 0.004 0.000 0.029 0.000 0.009 0.000 0.009 0.000 0.009 0.000

Trend if care work -.0002 0.85 -0.001 0.129 0.029 0.000 0.018 0.001 0.000 0.913 -.0002 0.801

Trend if engineer     -0.0001 0.905     0.011 0.496     -0.001 0.302

Trend if technician     -0.001 0.214     -0.008 0.002     -0.003 0.001

Trend if doctor     -0.024 0.000     -0.178 0.000     -0.002 0.427

Trend if lawyer     -0.013 0.000     -0.293 0.000     0.003 0.364

Trend if scientist     -0.001 0.411     -0.007 0.091     -0.001 0.615

Trend if college faculty     0.008 0.000     -0.011 0.083     0.004 0.623

Trend if manager     -0.002 0.242     -0.138 0.000     -0.003 0.010

Trend if sales job     0.005 0.001     -0.013 0.063     -0.001 0.015

35 year fixed effects yes yes yes yes yes yes381 occupation fixed effects yes yes yes yes yes yes

sample size 11138 11138 11140 11140 11140 11140

Adjusted R-squared 0.81 0.81 0.67 0.73 0.75 0.75

CV stands for coefficient of variation, which is the standard deviation of the sample divided by the mean of the sample. Coefficient on regression constant is not shown.

Figures in bold are statistically significant at the .05 level. In computing the standard errors each observation has been weighted by its sample size. Robust standard errors underly the p-values. In the underlying calculations of inequality, observations of income were equal-weighted, that is, did not use the person-weights which adjust for demographic characteristics. Year fixed effects help compensate for the changing number of occupations defined by the Census -- fewer occupation categories mechanically means more variation within them, even if there is no substantive change.

The hypotheses of interest were that over this time period (a) the trend in the media-amplified occupations and the technologically uncertain occupations have been increasingly dispersed over time, and that (b) the face-to-face service occupations have not.

Table 9. Predictors of earnings dispersion in occupation-years in Census

Panel A. Dependent variable is standard deviation of log-wage-and-salary within occupation-years

Figures in bold are statistically significant at the .05 level. In computing the standard errors each observation has been weighted by its sample size. Robust standard errors underly the p-values. In the underlying calculations of inequality, observations of income were equal-weighted, that is, did not use the person-weights which adjust for demographic characteristics. Year fixed effects help compensate for the changing number of occupations defined by the Census -- fewer occupation categories mechanically means more variation within them, even if there is no substantive change.

Predictors of std dev of ln(wage) in Census

Regression 0 Regression 1 Regression 2

coeffp-

value coeffp-

value coeffp-

valueTrend if media-amplified 0.026 0.000 0.025 0.000 0.023 0.000

Trend if high tech 0.025 0.001 0.014 0.000 0.013 0.000Trend if care work -0.005 0.474 -0.005 0.221 -0.009 0.077Trend if engineer -.0003 0.940

Trend if technician -0.005 0.361Trend if doctor 0.012 0.434Trend if lawyer 0.015 0.278

Trend if scientist -0.004 0.445Trend if college faculty 0.004 0.370

Trend if manager 0.017 0.005Trend if sales job 0.007 0.240

Annual trend overall -.010 0.0055 year fixed effects no yes yes

387 occupation fixed effects yes yes yessample size 1635 1635 1635

Adjusted R-squared 0.77 0.88 0.88

Table 9.

Panel B. Dependent variable is named measure of dispersion within occupation-years in Census

Predictors

Dep var: coefficient of variation of wage-and-salary

income within occupation-years

Dep var: interquartile range (75%ile to 25 %ile) of wage-

and-salary income within occupation-years

Dep var: coefficient of variation of wage-and-salary income plus self-employment

income within occupation-yearsRegression 1 Regression 2 Regression 3 Regression 4 Regression 5 Regression 6

coeffp-

value coeffp-

value coeffp-

value coeffp-

value coeffp-

value coeffp-

valueTrend if media-amplified 0.006 0.009 0.01 0.03 0.07 0.00 0.06 0.001 0.01 0.001 0.01 0.001

Trend if high tech 0.001 0.593 0.00 0.41 0.03 0.13 0.01 0.712 0.012 0.006 0.008 0.031

Trend if care work -0.004 0.011 -0.01 0.00 -0.02 0.23 -0.02 0.142 -0.004 0.026 -0.006 0.007

Trend if engineer     0.00 0.00     0.03 0.025     0.009 0.004

Trend if technician     0.01 0.03     -0.02 0.459     -0.002 0.547

Trend if doctor     -0.02 0.00     -0.03 0.714     0.016 0.001

Trend if lawyer     -0.01 0.00     -0.13 0.053     0.012 0.145

Trend if scientist     0.00 0.00     0.03 0.132     0.001 0.702

Trend if college faculty     0.00 0.15     0.03 0.197     0.006 0.007

Trend if manager     -0.01 0.00     -0.08 0.013     0.002 0.264

Trend if sales job     0.00 0.22     -0.02 0.558     0.003 0.199

Constant yes yes yes yes yes yes

35 year fixed effects yes yes yes yes yes yes387 occupation fixed effects yes yes yes yes yes yes

sample size 1635 1635 1635 1635 1635 1635

Adjusted R-squared 0.91 0.91 0.69 0.71 0.91 0.91

CV stands for coefficient of variation, which is the standard deviation of the sample divided by the mean of the sample. Coefficient on regression constant is not shown.

Figures in bold are statistically significant at the .05 level. In computing the standard errors each observation has been weighted by its sample size. Robust standard errors underly the p-values. In the underlying calculations of inequality, observations of income were equal-weighted, that is, did not use the person-weights which adjust for demographic characteristics. Year fixed effects help compensate for the changing number of occupations defined by the Census -- fewer occupation categories mechanically means more variation within them, even if there is no substantive change.

Table 10. Earnings regression from 1968-2003 CPS(last updated aug 18, 2005)

The dependent variable is the log of the weekly earnings (defined as wage and salary plus self-employment income) for individuals with over $40 in weekly earnings.

Predictor coefficient p-valueyear trend 0.040 0.000age .242 0.000age squared -.007 0.000age cubed .00008 0.000age to the fourth -.0000004 0.000years of education -.060 0.000Year of educ squared .004 0.000age * educ .0007 0.000

occupation fixed effects (387 categories) included

sample size 2,508,091Adjusted R-squared 0.57

10b from census

The dependent variable is the log of the weekly earnings (defined as wage and salary plus self-employment income) for individuals with over $40 in weekly earnings.Sample size 9,739,077R-squared is .51

Predictor coefficient p-value

Age .075 0.000age squared -.0007 0.000years of education -.024 0.000Year of educ squared .003 0.000Year 2000 1.772 0.000Year 1990 1.456 0.000Year 1980 .978 0.000Year 1970 .401 0.000occupation fixed effects (387 categories) included

Table 11. Residuals from earnings regression from 1968-2003 CPS

The dependent variable is the square of the residual from the wage regression in Table 10. Very little of the change in residual magnitudes is explained by the regression – a third of one percent, using the R-squared measure. The purpose of this regression is to see if it possible to detect technological uncertainty or media-amplification from the data, or to distinguish between them statistically. It does seem that the media-amplification attribute raises the magnitude of the residuals over this period, but not for the high tech turbulent professions.

Predictor 

Regression 1 Regression 2

coefficient p-value coefficient p-value

trend if media-amplified 0.449 0.000 0.473 0.000trend if high tech turbulent -0.026 0.001 0.022 0.004trend if care work 0.139 0.000 0.114 0.000trend if engineer 0.024 0.006trend if technician -0.219 0.000trend if doctor 0.682 0.000trend if lawyer 0.468 0.000trend if scientist -0.149 0.000trend if manager  0.016 0.344

trend if college faculty 0.128 0.000trend if sales job 0.116 0.000Year effects  yes yes

sample size 2,508,091R-squared 0.02

Table 12. Computer use by occupation in 1984

An October 1984 CPS survey supplement recorded answers to the question, “Does [respondent] directly use a computer at work?” This illustrates that the use of a computer is not very closely linked to technological uncertainty as discussed in the text. Some closer involvement with technology change is needed.

Occupation group % of work force % of these who use computer at work

Public officials and administrators 0.46% 33%Other managerial and administrative 6.98% 34%Management-related 2.76% 51%Engineers 1.35% 58%Mathematical and computer scientists 0.39% 83%Natural scientists 0.35% 54%Health diagnosing 0.61% 21%Health management and treatment 1.72% 24%Teachers, college and university 0.64% 39%Teachers outside college and university 3.34% 27%Lawyers and judges 0.58% 27%Other professional specialty 3.09% 23%Health technologists and technicians 1.00% 26%Engineering and science technicians 0.94% 41%Other technicians 0.86% 70%Supervisors and proprietors 2.90% 23%Sales representatives, finance and business 1.69% 40%Sales representatives, commodities 1.24% 25%Sales workers, retail and personal 5.94% 9%Sales related 0.05% 5%Administrative support supervisors 0.55% 59%Computer equipment operators 0.58% 90%

Secretaries, stenographers, and typists 4.55% 40%Financial records processing 2.29% 38%Mail and message distributor 0.73% 8%Other administrative, including clerical 6.62% 40%Private household service 1.34% 1%Protective service 1.53% 17%Food service 5.42% 2%Health service 1.69% 6%Cleaning and building 2.91% 2%Personal service 2.03% 3%Mechanics and repair workers 3.84% 12%Construction trades 4.48% 3%Other precision production, craft, and repair 3.65% 12%Machine operators and tenders, except precision 5.07% 5%Fabricators, assemblers, inspectors, samplers 2.67% 7%Motor vehicle operators 3.11% 2%Other transportation and material-moving 1.33% 4%Construction laborers 0.79% 0%Freight, stock, and materials handlers 1.45% 3%Other handlers, equipment helpers, laborers 2.24% 3%Farm manager and operators 1.60% 4%Farm workers and related 2.37% 1%Forestry and fishing 0.22% 4%Armed forces 0.06% 0%Overall 100%, 77452 observations 20%

References

Abernathy W. J. and Utterback, James M. 1978. Patterns of industrial innovation. Technology Review, Vol. 80 (June-July).Almeida, Paul, and Bruce Kogut. 1999. Management Science.Autor, David H., Lawrence F. Katz, and Alan B. Krueger. 1997. Computing Inequality: Have Computers Changed the Labor Market? NBER

Working Paper 5956.Autor, David H., Lawrence F. Katz, and Alan B. Krueger. 1998. Computing Inequality: Have Computers Changed the Labor Market?

Quarterly Journal of Economics CXIII:4 (November), pp 1169-1213.Autor, David H., Lawrence F. Katz, and Melissa S. Kearney. 2006. The Polarization of the U.S. Labor Market. American Economic Review,

96:2 (May, 2006). pp 189-194.Autor, David H., Richard Murnane, and Frank Levy. 2003. Quarterly Journal of Economics. Bartel, Ann P., and Frank R. Lichtenberg. 1987. The comparative advantage of educated workers in implementing new technology. Review

of Economics and Statistics LXIX:1 (Feb 1987), pp. 1-11.Bartel, Ann P. and Nachum Sicherman. Technological Change and Wages: An Interindustry Analysis. Journal of Political Economy 107:2

(April, 1999): 285-325.Berlin, Leslie R. Robert Noyce and Fairchild Semiconductor, 1957-1968. Business History Review 75 (Spring 2001): 63-101. Berman, Eli, John Bound, and Zvi Griliches. Changes in the Demand for Skilled Labor within U.S. Manufacturing: Evidence from the Annual

Survey of Manufacturers. Quarterly Journal of Economics, Vol. 109, No. 2. (May, 1994), pp. 367-397.Bessen, James and Eric Maskin. 2006. Sequential Innovation, Patents, and Imitation. Institute of Advanced Study School of Social Science

Working Paper 25.Blinder, Alan S. Offshoring: the next industrial revolution? Foreign Affairs, March/April 2006.Bound, John, and George Johnson. Changes in the Structure of Wages During the 1980s: an Evaluation of Alternative Explanations. American

Economic Review, LXXDII (1992).Braun, E. and S. MacDonald. 1978. Revolution in Miniature: The history and impact of semiconductor electronics. Cambridge University Press.Brooks, Frederick P., Jr. 1975. The Mythical Man-month: Essays on Software Engineering. Reading, MA: Addison-Wesley. Campbell-Kelly, Martin. Not Only Microsoft: The Maturing of the Personal Computer Software Industry. Business History Review 75 (Spring

2001): 103-45.Campbell-Kelly, Martin, and William Aspray. Computer: A History of the Information Machine. New York: Basic Books, 1996.Caselli, Francesco. 1997. Technological Revolutions. Draft paper, Graduate School of Business, University of Chicago.Christensen, Clayton M. 1999. The Innovator's Dilemma. Harvard Business School Press.Chun, Hyunbae, Jung-Wook Kim, Jason Lee, and Randall Morck. Patterns of comovement: the role of information technology in the U.S.

economy. NBER working paper 10937. http:/www.nber.org/papers/w10937. November 2004.Cringely, Robert X. 1992. Accidental Empires. Addison-Wesley.Current Population Surveys, March 1962-98 [machine-readable datafiles]/conducted by the Bureau of the Census for the Bureau of Labor

Statistics. Washington: Bureau of the Census [producer and distributor], 1984-1998. Santa Monica, CA: Unicon Research Corporation [producer and distributor of CPS Utilities], 1999.

Dosi, Giovanni. Sources, Procedures, and Microeconomic Effects of Innovation. Journal of Economic Literature 26:3 (Sept. 1988), 1120-1171.

England, Paula, Michelle Budig, and Nancy Folbre. 2002. Wages of Work: The Relative Pay of Care Work. Social Problems 49:4, pp. 455-473.

Fallick, Bruce, Charles A. Fleischman, and James B. Rebitzer. 2005. Job Hopping in Silicon Valley: Some Evidence Concerning the Micro-Foundations of a High Technology Cluster. Federal Reserve Working Paper 2005-11, available at www.federalreserve.gov/ pubs/feds/2005/200511/200511pap.pdf

Freiberger, Paul, and Michael Swaine. 1984. Fire in the Valley: The Making of the Personal Computer. Osborne/McGraw-Hill.Galor, Oded, and Omer Moav. 2000. Ability Biased Technological Transition, Wage Inequality, and Economic Growth. Quarterly Journal of

Economics 115, 469-498 (May 2000).Gordon, Robert J. 1990. The Measurement of Durable Goods Prices. University of Chicago Press.Greene, William H. 1997. Econometric Analysis, third edition. Prentice-Hall.Greenwood, Jeremy. 1997. The Third Industrial Revolution: Technology, Productivity, and Income Inequality. Washington, D.C.: The AEI

Press.Greenwood, Jeremy, and Yorukoglu, Mehmet. 1997. 1974. Carnegie Rochester Conference Series on Public Policy, 46 (June): 49-95.

http://www.econ.rochester.edu/Faculty/GreenwoodPapers/1974.pdfHecker, Daniel. 1999. High-technology employment: a broader view. Monthly Labor Review, June 1999, pp 18-28.Henderson, Rebecca. Underinvestment and incompetence as responses to radical innovation: evidence from the photolithographic alignment

equipment industry. RAND Journal of Economics 24:2 (Summer, 1993), 248-70.Hirsch, Paul M. From Ambushes to Golden Parachutes: Corporate Takeovers as an Instance of Cultural Framing and Institutional Integration.

American Journal of Sociology 91:4 (January 1986): 800-837.

Hunter, Starling, Kevin Kobelsky, and Vernon J. Richardson. Information Technology and the Volatility of Firm Performance. MIT Sloan School of Management Working Paper 4449-03, November 2003. http://ssrn.com/abstract=475641

Institute of Electrical and Electronics Engineers, Inc. 1995. IEEE U.S. Membership Salary & Fringe Benefit Survey 1995. Institute of Electrical and Electronics Engineers, Inc.

Jovanovic, Boyan, and Yaw Nyarko. A Bayesian Learning Model Fitted to a Variety of Empirical Learning Curves, in Brookings Papers on Economic Activity, Microeconomics, 1995. Brookings Institution. 247-299.

Juhn, Chinhui, Kevin M. Murphy, and Brooks Pierce. 1993. Wage Inequality and the Rise in Returns to Skill. Journal of Political Economy 101 (June 1993): 410-442.

Kaplan, Jerry. 1994. Startup: A Silicon Valley Adventure. Penguin Books.Katz, Lawrence F. and Kevin Murphy. 1992. Changes in Relative Wages, 1963-1987: Supply and Demand Factors. Quarterly Journal of

Economics 107 (February 1992): 35-78.King, Miriam, Steven Ruggles, and Matthew Sobek. Integrated Public Use Microdata Series, Current Population Survey: Preliminary

Version 0.1. Minneapolis: Minnesota Population Center, University of Minnesota, 2003. Klepper, Steven, and Elizabeth Graddy. The Evolution of New Industries and the Determinants of Market Structure. The RAND Journal of

Economics, Vol. 21, No. 1 (Spring, 1990), pp. 27-44.Kuznets, Simon. 1955. Economic Growth and Income Inequality. American Economic Review 45:1-28.Langlois, Richard N. External Economics and Economic Progress: The Case of the Microcomputer Industry. Business History Review 66

(Spring 1992): 1-50.Langlois, Richard N., and David C. Mowery. 1996. The Federal Government Role in the Development of the U.S. Software Industry.

Chapter 3 of The International Computer Software Industry: A Comparative Study of Industry Evolution and Structure, edited by David C. Mowery. Oxford University Press.

Levin, Richard C. 1982. R&D Productivity in the Semiconductor Industry: Is a Slowdown Imminent? In Understanding R&D Productivity, edited by Herbert I. Fusfeld and Richard N. Langlois. Pergamon Press.

Levy, Frank, and Richard. J. Murnane. 1992. U.S. earnings levels and earnings inequality. A review of recent trends and proposed explanations. Journal of Economic Literature 30 (Sept. 1992):1333-1381.

Lindberg, Per. 1995. Managing and Organizing for Advanced Manufacturing Technology. In New Technology and the Workers’ Response: Microelectronics, Labour, and Society, edited by Amiya Kumar Bagchi.

Maguire, Steve. 1994. Debugging the Development Process. Microsoft Press.Malerba, Franco. The Semiconductor Business. 1985. U of Wisconsin Press.Malone, Michael S. 1985. The Big Score: The Billion-Dollar Story of Silicon Valley. Doubleday & Co.Meyer, Peter B. 2005. Turbulence, inequality, and cheap steel. US Bureau of Labor Statistics working paper 375, available online at

http://www.bls.gov/ore/abstract/ec/ec050010.htmMeyer, Peter B. 2006, in progress. Updated unified occupation categories for 1960-2000 Census occupations. Available at

http://econterms.net/pbmeyer/research/occs/WEAI2006/occ_defns.doc.Meyer, Peter B. and Anastasiya Osborne. 2005. Proposed category system for 1960-2000 Census occupations. US Bureau of Labor Statistics

working paper 383, available online at http://www.bls.gov/ore/abstract/ec/ec050090.htmMoore, Gordon E. Intel -- memories and the microprocessor. Daedalus. Spring 1996.Morris, P.R. 1990. A History of the World Semiconductor Industry. Peter Peregrinus, London.Murphy, Kevin M., and Finis Welch. The Structure of Wages. Quarterly Journal of Economics 107:1 (Feb., 1992), 285-326.Norman, Don. 1998. The Invisible Computer. The MIT Press.Perez, Carlotta. 1997. Technological revolutions and financial.Perez, Robert C. 1986. Inside venture capital: past, present, and future. New York: Praeger.Rock, Arthur. 2001. Interview by Intel Corporation. Visible on the Web on Aug 23, 2001.

http://www.intel.com/intel/museum/25anniv/int/rock.htm. Rosen, Sherwin. 1981. The Economics of Superstars. American Economic Review 71:5 (Dec., 1981), 845-858.Rosenberg, Nathan. 1996. Uncertainty and Technological Change. In Mosaic of Economic Growth, edited by Ralph Landau, Timothy Taylor,

and Gavin Wright. Stanford University Press.Rubinstein, Yona, and Daniel Tsiddon. 1999. Coping with technological progress: the role of ability in making inequality so persistent.

Unpublished paper, NBER summer institute and Tel Aviv University.Ruggles, Steven and Matthew Sobek et. al. Integrated Public Use Microdata Series: Version 2.0. Minneapolis: Historical Census Projects,

University of Minnesota, 1997. http://www.ipums.umn.edu.Ruttan, Vern. Technology, growth, and development: an induced innovation perspective. Oxford University Press, 2001. Solon, Gary, Robert Barsky, and Jonathan A. Parker. 1994. Measuring the Cyclicality of Wages: How Important is Composition Bias?

Quarterly Journal of Economics 109:1 (February), pp. 1-25.Southwick, Karen. 1999. The Silicon Gold Rush. John Wiley and Sons.Schultz, Theodore W. 1975. The value of the ability to deal with disequilibria. Journal of Economic Literature 13:3 (Sept 1975), pp 827-846.Thompson, James D. 1967. Organizations in Action: Social Science Bases of Administrative Theory. McGraw-Hill.Tilton, John E. 1971. International Diffusion of Technology: The Case of Semiconductors. Brookings Institution.

Tushman, Michael L., and Philip Anderson. 1986. Technological Discontinuities and Organizational Environments. Administrative Science Quarterly 31: 439-465.

U.S. Dept. of Commerce, Bureau of the Census. CURRENT POPULATION SURVEY, OCTOBER 1984: SCHOOL ENROLLMENT [Computer file]. Washington, DC: U.S. Dept. of Commerce, Bureau of the Census [producer], 1986. Ann Arbor, MI: Inter-university Consortium for Political and Social Research [distributor], 1986.

Utterback, James. 1996. Mastering the Dynamics of Innovation. Harvard Business School Press.Wilson, Robert W., Peter K. Ashton, and Thomas P. Egan. 1980. Innovation, Competition, and Government Policy in the Semiconductor

Industry. D.C. Heath and Company.