Robotics and Manufacturing Employment

Robotics and Manufacturing Employment: A Pooled Time Series Analysis

Introduction by firms of labor-saving technology need not result in declines in manufacturing employment. Although increasing deployment of robots in the 1980s and 1990s certainly correlates with declines in manufacturing employment in some industrial countries, pinpointing a causal role for robotics across the industrialized world is quite difficult. Further research into the effects on manufacturing employment of robotics may yield more positive findings.

In this paper, I examine the influence of robots on total manufacturing employment in eight industrialized countries over a fourteen year period. While several authors and studies have predicted the effects of robots on future manufacturing employment, I have found no econometric analyses of actual effects over the previous decade. I offer such an analysis in this paper.

After briefly discussing the mixed social blessings of automation technology and reviewing the literature on the predicted effects of robots on employment, I test and discuss the results of several models for estimating these effects over a 14 year period. I find that different models yield dramatically different results.

The Lure and Curse of Labor-Displacing Technology

We can identify at least two broad arguments favoring technological innovation in automation. First, such innovation offers competitive advantages for firms. Second, over the long term, this technology benefits society as a whole. Meanwhile, traditional arguments attacking innovation emphasize automation technology's marginalizing effects on workers. Most of these arguments, pro and con, describe dynamics which would lead us to expect reductions in overall manufacturing employment.

Benefits to the firm. Market economies generate profound incentives for innovation. Competitive firms seek productivity increases--greater output levels per person-hour--and technological enhancement of a firm's production systems can foster those increases by reducing employment of human labor. Especially when the tasks to be performed are repetitive, the advantages to firms of machines are formidable indeed. In the North American automobile industry–today one of the most roboticized of all industries–automation and overall technological development was largely responsible for the tremendous increases in productivity over a thirty-year period:

In 1948, some 713,000 U.S. and Canadian auto workers produced 5.96 million new cars, trucks, and buses. In 1978, 839,000 auto workers in the two nations produced nearly 14.26 million motor vehicles. Seventeen motor vehicles were produced per automobile worker in 1978 as compared with 8.36 vehicles in 1948. These figures certainly understate the growth in output per worker since they do not take into account the increase in complexity of motor vehicles in the past generation (Ayres and Miller, 1983: 189).

Firms which automate may enjoy other advantages. Firms may deliberately employ automation to handle expected labor shortages. A contribution to a 1982 OECD report indicated that, in Japan, "(i)t is expected that industrial robots will be increasingly utilised to solve the skilled labour shortage" (Yonemoto, 1982: 239). Moreover, fewer workers means easier control and greater power of firm owners and managers over production and working conditions.

Finally, even firms which do not employ technology may benefit from the second-order effects of automation in other firms. Machines reduce the demand for and thus the wages paid to human labor. All firms may enjoy greater ease of procuring and eliminating human labor in response to changes in market conditions.

Overall, and in summary, automation technology may enhance the competitive edge of firms in market economies in various interrelated ways. These arguments tend to reinforce an intuition that increases in robotics deployment over the past decades should have caused reductions in manufacturing employment.

Benefits to Society. The foregoing would convince many that automation technology is a curse, because it tends to benefit primarily managers and owners of firms while displacing workers. But others have noted farther-ranging social benefits of technology. John Zysman writes that technological development generally, together with trade, has contributed to overall societal welfare over time:

Certainly, seen historically, trade and technology together have been part of the development motor that has sustained growing wealth in the West. Jobs may be displaced, but wealth and income are created, which in turn generate more jobs and, historically viewed, on balance better jobs (Zysman, 1994: page number not given).

If by "better jobs" he means jobs with higher wages and improved working conditions, to the extent that automation technology contributes to the process described by Zysman, we can view it as a social good. Moreover, some attempts at substituting machines for people have had strange, worker-empowering effects. An important example of this occurred in the 19th century, when the Manchester firm of Sharp, Roberts and Company created the self-activating mule, which the firm hoped and Karl Marx believed would make redundant skilled cotton spinners and destroy the spinners' union. But because the spinners did more than just spin, they managed not only to keep their jobs and union but to acquire "extensive control over use of the new technology" (Piore and Sabel, 1984: 45). These considerations suggest that, contrary to the intuitions of many, robotics deployment may not cause relentless reductions in manufacturing employment, and that it may correlate with increases in "better" jobs.

Marginalization of Workers. In spite of the foregoing optimistic observations, few would question that the dynamics of automation technology deployment and price competition among firms may seriously marginalize, possibly even eliminate altogether, a firm's workers, with concomitant increases in poverty and a host of other social ills. Although it is not precisely an example of the effects of automation technology, agriculture offers an impressive illustration of how replacing people with machines can bleed an industry of employment. In the United States, from the 1910s to the 1980s, technological innovation in agriculture reduced the percent of the work force in that sector from 30 to 4 while tripling output (Guest, 1984: 6).

A key issue concerns whether technology in the post-Fordist era can continue to create lots of high-value added, high-wage employment. One author has suggested that, in the late 20th century, the manufacturing sector will not be the locus of any job creation. Fred Block has gathered statistics on manufacturing-person ("man") hours in several industrialized countries and concludes that, along with dramatic increases in productivity, there has been a stagnation in the total number of person-hours in manufacturing. He attributes this trend to the increasing power of technology to enhance worker productivity (Block 1984).

Many of the foregoing general considerations on labor-displacing technology suggest that we should find little growth, or even decline, in manufacturing employment correlate with deployment of robotics. However, we have seen some historical material which belies these expectations. In the next section, I review predictions relating the effects specifically of robotics on manufacturing employment.

Robots and Employment Displacement

Numerous studies have predicted that the overall effects of robots on employment would be modest, but that robots would hit certain industries, particularly the automobile industry, and certain regions quite hard indeed. Close to two decades ago, Drew Lewis and Harley Shaiken predicted that automobile manufacturing would lose hundreds of thousands of jobs by the 1990s. Also, in 1980, the American Society of Manufacturing Engineers predicted that "50 percent of auto assembly operations will be performed by automated machines, principally robots" (Guest, 1984, 7).

In 1979, the Eikonix corporation published the first predictions of the employment displacement effects of robots in manufacturing. The corporation's study, and several following it, divided robots into Levels I and II. Level I robots perform automated motion but have no "sense" of the outside world, whereas Level II robots contain a camera or other device for "sensing" the world. Eikonix foresaw, by the year 2000, a 15 percent reduction in manufacturing "operative workers" by Level I robots and a 55 percent reduction by Level II robots. One study by Ayres and Miller in the 1980s predicted a 14 percent reduction in "metalworking craftworkers, operative workers, and laborers" by Level I robots and a 40 percent reduction by Level II robots over a period of one to three decades. Another study by the same authors predicted a 12 percent reduction in "all production workers in manufacturing" by Level I robots and a 33 percent reduction in this category by Level II robots over a period of one to three decades. A study by H. Allan Hunt and Timothy Hunt in 1983 predicted a 1 to 2 percent reduction in "operative workers and laborers" by 1990 (Miller, 1989, 9-10, 73). Hunt and Hunt also predicted that there would be, by 1990, 25,000 to 50,000 job tasks, mostly or totally low-skill, eliminated from the automobile industry, which would suffer a 6 to 11 percent decline "of all operatives and laborers" (Hunt and Hunt, 1983, x-xii).

Other studies have examined effects in other countries or worldwide. Kathawala and Brandyberry, writing in 1989, believe some of the results of the U.S.-focused studies apply generally, namely that, in other countries, industries such as automobiles and metalworking will experience employment reductions. However, Kathawala and Brandyberry also claim that "(i)n most companies which have moderate use of robotics, the number of workers employed did not decline and most companies either using or considering robotics do not consider decreasing labour costs as a major criterion in their decision to implement robotics." Further, they write that "(t)here seems to be no direct correlation between the use of robotics and unemployment" in industries using robotics. While there has indeed been unemployment in these industries, the authors note, this is due to low demand for the industries' products (1989, 684).

This last point illustrates an obvious, basic fact. There are several causes for unemployment and other economic trends; it is difficult, and way beyond the scope of this paper, to theoretically identify all salient causes. Nevertheless, based on the foregoing analyses, it is reasonable to expect a modest decline in manufacturing employment associated with robot deployment across countries. However, based on the forecasts of the Japanese Industrial Robot Association (quoted in the OECD report mentioned above), we may find individual countries experiencing little such decline.

In the following sections, I use and discuss several models for examining the effects of robotics on manufacturing employment over a period of 14 years.

Deriving A Basic Model

The null hypothesis is that robots have no influence on hours worked in manufacturing. The alternative hypothesis is that robots do influence hours worked. "Robots" refers to sensory and nonsensory industrial robots.
I pool cross sectional with time series data. Eight countries--the United States, Japan, Denmark, Germany, Italy, Norway, Sweden, and the United Kingdom--constitute the cross sections, and the period of 14 years from 1981 to 1994 constitutes the time series. I thus have a total of 112 cases. My dependent variable is an indexed measure of total hours worked in the manufacturing sector in these eight countries in each of the 14 years. The primary parameter I seek to estimate describes the effects on the index of total hours worked exerted by the number of robots per 10,000 workers in a nation's manufacturing sector. I include in my initial model a dummy variable for economic recession, defined as a decrease in national GDP in one year from the previous year's level. Thus the basic regression model reads:

Yi = 1 + 1X1 + 2D1 + ui

where
Yi = the indexed computation of total hours worked in the manufacturing sector
1 = intercept
X   = robots per 10,000 manufacturing workers
D   = the dummy for presence (D = 1) or absence (D = 0) of recession
ui   = error term

From this basic model, I derive the regression

Ynt = 1 + 1Xnt + 2D1 + unt

where
n = 1...N cross sections
t = 1...T time points

This particular model, the constant coefficients model, assumes that the parameters are constant across space and time, that the effect of beta on Y is the same for all countries in all years.

The error term unt takes on peculiar significance in the tests and models I run. In addition to absorbing such influences as theoretical misspecifications, random patterns of human behavior, and others, the term also accounts for data collecting and reporting errors which surely have transpired as I did this study.

Problems with Pooled Data

This constant coefficients model very likely violates some OLS assumptions. First, units of analysis which are less than fully comparable violate the assumption of homoscedasticity, that the variance of unt for all n equals 2. Second, covariance between any two disturbance terms uit and ujt for any cross sections i and j and any time point t violates the assumption of 0 covariance among error terms. Third, the assumption of 0 covariance between uit and Xit for any i and t may be violated (Sayrs, 1989: 13). Finally, we may have incorrectly specified our model, violating assumptions of linearity and inclusion of relevant variables. Let us examine these issues more closely.

Heteroscedasticity. Regarding heteroscedasticity, Stimson points out, "(o)n the reasonable assumption that variation is roughly a fixed proportion of size, analysis of units of substantially different sizes induces heteroscedasticity in any regression" (1985: 919). A look at the bar charts (attached) for the minimum and maximum numbers of robots in 1981 and 1993, respectively, for each of these countries alerts us to the possibility of such heteroscedasticity. Japan had about 20 robots per 10,000 manufacturing employees in 1981, Denmark fewer than 5; in 1993, Japan had 380 robots per 10,000 manufacturing employees, Denmark fewer than 30. (Please see also the descriptives for robots.) Scatterplots of squared residuals against predicted values, of squared residuals against robots, and in other plots (please see accompanying charts) also seem to reveal unit-induced heteroscedasticity. Applying a Breusch-Pagan-Godfrey test yields, however, a chi-square value of only 2.15, which, at two degrees of freedom, is statistically insignificant at even the .25 level. The BPG assumes a linear relationship between the independent variables and the residuals which, however, my various scatterplots do not show.

Additionally, although I do not probe for it, autoregressive conditional heteroscedasticity (ARCH), or heteroscedasticity induced by the time series, involving the dependence of the variance of uit on u2t-1, might accompany any units-induced heteroscedasticity.

Disturbance Covariance. Omission of important explanatory variables, the mere nature of certain series, and many other factors may violate the assumption of 0 covariance between any two disturbance terms uit and ujt for any cross sections i and j and any time point t. By excluding inflation, strike activity, and other influences on production and employment from the model, I force one or more systematic influences into the error terms for all cross sections and time points. The model also assumes that such influences in one time point cause no effects on hours worked in a following time point, when in fact it is quite likely that inflation or strikes at some point t-1 will, even if they are not present at point t, nevertheless exert influence on total hours worked in t. My model also does not account for economic contagion among nations. That is, to use one example, it assumes that a recession in one country will not affect total hours worked in the manufacturing sector of another country, now or in the future.

Explanatory Variable-Disturbance Covariance. There is also reason to consider that the model's right hand variables, robots and recessions, are correlated to some degree with the error terms, so that there is no longer any separate and additive influence on hours worked coming from the right hand variables and the disturbance terms. Thus inflation or strikes may be correlated with either increases or decreases in robots or with the presence or absence of recessions.

Model Misspecification. Finally, it is possible that the model is misspecified not merely with respect to omission of variables but in terms of its functional form. The "true" population function may not be a linear function. The true function may be a curvilinear relation in which hours worked decreases unevenly, such that initial worker displacement by robots is greatest in an initial period of some years, after which the decline in hours worked tapers off, as the marginal returns from adding additional robots diminish. Or, the true function may involve an autoregressive process in which one of the independent variables is a lagged value of the dependent variable. Furthermore, model misspecification can contribute to heteroscedasticity (Downs and Rocke 1979; Stimson, 1985: 919).

Sayrs notes that we can correct for autoregression by "specifying an autoregressive structure to the error" (1989: 24), and Gujarati recommends a number of remedies for autocorrelation when , the coefficient of (auto)correlation, is either known or not (1995: 426-436). I am not going to attempt such a correction, but corrections would enhance the viability of the constant coefficients approach to pooled regression.

Results from the Basic Model

Applying no corrections, and stuffing a plethora of contaminants into the error term, the results of the constant coefficients regression are statistically insignificant at the five percent level for both robots and recessions (please see output). The Durbin-Watson figure of 1.18 would alert us to the possibility of first-order autocorrelation, but this result may reflect an ARCH effect, "artifactual of nonconstant variance in the cross-sections," heteroscedasticity having been neither corrected nor controlled for (Sayrs, 1989: 20). The model seems to have no serious collinearity problems: The k value of 1.71665/.42618 is far below 100, and the condition index (defined in Gujarati as the square root of k) is barely over 2 and far less than 10 (Gujarati, 1995: 338). (The condition indexes in the printout are also considerably below 10.)

These results would imply that as robots increase, so do hours worked in the manufacturing sector. But inspection and comparison of the indexed hours worked in each country from 1981 to 1994 show, for most countries, an overall decrease over this period. Denmark is an exception, having gone from an indexed figure of 98.8 in 1981 to 99.7 in 1994 and 98.8 in 1993; Japan shows very little decrease in its 1994 level (100.1) from its 1981 level (100.3), with its 1993 level at 102.3, which reflects an increase over its 1981 level. Other countries show dramatic decreases. The United Kingdom went from 105.6 in 1981 to 77.1 in both 1993 and 1994; Sweden from 103.9 to 85.7 in 1994 and 81.4 in 1993; Norway from 104.3 to 80.0 in 1994 and 77.0 in 1993; Germany from 103.1 to 83.4 in 1994 and 87.3 in 1993; Italy from 103.4 to 83.0 in 1994 and 81.7 in 1993. The United States saw a relatively modest decline from 109.2 to 105.8 in 1994 and 103.5 in 1993 (Monthly Labor Review, 1996: 132; 1995: 102; 1990: 108).

The Monthly Labor Review statistics would seem to reflect, more or less, an extension of the trends observed in the United States by Eikonix, Ayres and Miller, Hunt and Hunt, in the 1970s and 1980s, and in both the United States and Europe, 12 years ago, by Block, who calculated only a 1.5 percent increase in manufacturing hours in the United States from 1969 to 1979 and a serious "trend in manufacturing person-hours (which) is sharply downward" in Western Europe (Block, 1984: 62-64). Yet the model I have used produces no statistically significant results, and the results it does produce imply that hours worked increase as robots increase, a phenomenon no one predicted. Whether robots are or are not bugbears for those worried about unemployment, we would like to correct the model to obtain statistically significant estimates of the influences of robots.

On the basis of the law of diminishing returns, it is reasonable to assume that the influence of robots on hours worked decreases over time. However, regressing hours worked by recessions and the natural logarithm of robots, I get similarly bewildering results. While the estimate for the intercept for recessions is significant at the 5 percent level, the estimate for ln(robots) is even less significant than the first robots estimate, although now the (negative) direction of the relation confirms the fears of luddites (please see the appropriately labeled output). Thus some other approach is required.

A Least Squares Dummy Variable Model (LSDV)

Stimson notes that "(t)he misspecification that is peculiar to pooled data, and one for which other pooled models are solutions, is the assumption of homogeneity in the dependent variable (the level of the dependent variable) across units" (1985: 919). If we introduce dummy variables for each of the units (cross sections) and years, we can "capture the effects unique to the cross-sections and those that might be unique to time" (Sayrs, 1989: 28).

In the model I will outline shortly, the cross sections dummies account for the country specific differences which theoretically result in different total hours worked in the different countries given the same numbers of robots per 10,000 manufacturing workers. The time dummies account for events occurring in time which theoretically result in different total hours worked in the different countries given the same numbers of robots per 10,000 manufacturing workers. The dummy variables take on values of either 1 or 0. The is still the same for all cross sections. From the basic model

Ynt = 1 + 1Xnt + 2D1 + unt

I specify

un t = t + µn + nt

to give

Ynt = 1 + 1Xnt + 2D1 + t + µn + nt

where
n = (1...N) cross sections
t = (1...T) time series.
t = conditional (in intercept) term for time-specific effects
µn = conditional (in intercept) term for unit-specific effects
nt = random error term

Initial results seem to contradict most of the earlier predictions. After running the regression, with dummies assigned to both countries and years, the robots beta is now a positive 0.0514, with a very high level of statistical significance. The direction of the function is such that hours worked would stay more or less the same, with a very slight positive increase. The recessions intercept is somewhat less secure at T = -1.858, significant only at the 0.0667 level. We would conclude that robots have practically no effect on total hours worked, and thus employment, in the manufacturing sector.

Rerunning the regression, this time taking the natural logarithm of robots, worsens all estimates. The standard error for ln(robots) is bigger than the estimate for ln(robots) itself. The estimate is in the positive direction. The recession intercept estimate is also statistically insignificant. What is the problem here?

By including dummy variables for both countries and years, I lose 20 degrees of freedom, reducing my initial 109 df (112 cases minus two constants and one regressor) to 89. This is not a significant reduction. There would, however, seem to be some very serious collinearity between my explanatory and my time/units dummy variables in both the unlogged and logged robots regressions. Dividing the greatest eigenvalue by the least eigenvalue yields a k of 341,477 and a condition index of 584.36 (only the eigenvalues for the unlogged robots regression are included in the outputs). A logical next step is to control for unit effects only.

However, regressing hours worked on robots and including dummy variables only for my countries, thus controlling for unit effects only, yields disappointing results. There are no estimates statistically significant at the 0.05 level, although this time the direction of the robots parameter is (a very slight) -0.010416.

Performing an analagous regression--hours worked on robots with only year dummies-- produces a highly significant 0.047511 estimate for the robots parameter. Alas, the recession intercept parameter is less significant, with a T of -1.563, corresponding to the 12 percent level.

Finally, however, one regression produces estimates which would seem to correspond to intuition as well as to most of the trends in the data. Regressing the hours worked on the logarithm of robots, holding only unit effects constant, yields parameter estimates of -3.079081 for ln(robots), significant at 0.000, and -3.394468 for the recession intercept, significant at 0.0246, all accompanied by a respectable R square of 0.65139. These are the best estimates produced. A final regression on ln(robots) using year dummies only results in parameter estimates of 1.548210, significant at the 0.0664 (two-tailed) level, for ln(robots), and -4.492972, significant at the 0.728 level, for the recession intercept.

Discussion and Conclusions

Most of the models produce results most analysts had not predicted, namely that as robots increase, so do hours worked in manufacturing. However, most of these results were not statistically significant; thus they do not invalidate our intuitions regarding the effects of robots on employment. Only one model produces statistically significant results showing that as robots increase, hours worked in manufacturing decrease.

As noted above, most countries experienced some decline in total manufacturing hours in 1994 from 1981 levels. Denmark is the only country which actually registered longer hours worked in manufacturing in 1994 than in 1981. However, its totals decreased continuously after 1986, when it registered an indexed total of 116.6 hours worked. The country increased its robots from 2 per 10,000 manufacturing workers in 1981 to 20 in 1993, and we may attribute the continuous decline in hours worked since 1986 in part to robotization.

Norway's trend resembles Denmark's: A continuous decline in hours worked from 1986 to 1993, followed by an upturn in 1994. We can probably attribute some of this to robotization, which was more intensive than in Denmark. Sweden's pattern matches Norway's exactly with respect to both hours worked and robotization, although its intensity of robotization was greater.

The Eikonix Corporation, the studies by Ayres and Miller, and the Hunts' studies focused solely on the U.S. Has the experience of the U.S. borne out these predictions? It is difficult to say, because these studies, unlike mine, distinguish between types of industrial robots and types of manufacturing employment. In 1979, Eikonix had predicted reductions in "operative" manufacturing work of 15 and 55 percent by 2000. In the 1980s, Ayres and Miller made several predictions involving different categories of manufacturing employment and robots; one was for a 12 percent reduction in "production workers" by Level I robots and a 33 percent reduction in this category by Level II robots over a period of one to three decades. The Hunts in 1983 predicted a 1 to 2 percent reduction in "operative workers and laborers" by 1990 (see above).

Nevertheless, for 1983, the indexed figure for hours worked in the U.S. is 101.0; for 1990, the figure is 107.0; for 1994, the figure is 105.8. This would cause us to consider that the Hunts' predictions were inaccurate. It would also cause us to question the forecasts of Fred Block in 1984, at least with respect to the U.S. The predictions by Eikonix and Ayres and Miller were for later time periods, and so we can't even guess at their accuracy now. Overall, the statistics do not indicate that total employment in manufacturing in the U.S. has suffered as a result of robotization and other technological change.

Japan registered a very slight decline in 1994 from 1981 hours worked, remarkable given its intensive robotization. Hours worked in manufacturing continuously increased from 1987 to 1991 amidst ever increasing robotization. This is a country in which employment in manufacturing seems particularly immune to technological change. Another is (West) Germany, which maintained very even levels of hours worked from 1983 through 1992, amidst an increase from 5 to 95 robots per 10,000 manufacturing workers from 1981 to 1993.

Great Britain and Italy present the most erratic patterns in hours worked over the 1980s and early 1990s. Hours worked in these countries, though, steadily declined from 1989, although Italy rebounded in somewhat in 1994. Italy roboticized far more intensively than Britain but consistently maintained higher hours worked in manufacturing since 1987.

Total hours worked, and thus employment, in the manufacturing industries are clearly not a function solely of robots or automation or macroeconomic trends. As I stated above, identifying all relevant variables would be a formidable task under any circumstances and an impossible one in the present circumstances.

So, in the final analysis, can we claim that robots have reduced employment in manufacturing? The answer I can offer is, "Robots have certainly displaced some workers, but overall levels of manufacturing employment since the early 1980s seem to depend little on the degree of robotization." Other than this, it is safe to conclude that countries vary in their abilities simultaneously to assimilate technology and maintain employment in manufacturing. Japan is the undisputed champion in this regard, with Germany and the United States making competitive showings. We have no evidence from our comparison that robots will affect manufacturing in the ways that agricultural technology affected employment in agriculture.

References

Ayres, Robert U., and Steven M. Miller. 1983. Robotics: Applications & Social Implications. Cambridge, Mass.: Ballinger Publishing Company.

Block, Fred. 1984. "The Myth of Reindustrialization." Socialist Review 14 (Jan/Feb), pp. 59-76.

Downs, George W., and David M. Rocke. 1979. "Interpreting Heteroscedasticity." American Journal of Political Science 23 (November 1979), pp. 816-828.

Guest, Robert H. 1984. Robotics: The Human Dimension. New York: Pergamon Press.

Gujarati, Damodar N. 1995. Basic Econometrics. New York: McGraw-Hill Incorporated.

Haller, Brandon. Date Unknown. Title of Work Unknown. Appendix, pp. 18-19.

Harrison, Bennett. 1993. Lean and Mean: The Changing Landscape of Corporate Power in the Age of Flexibility. New York: Basic Books.

Hunt, Allan, and Timothy Hunt. 1983. Human Resource Implications of Robots. Kalamazoo, Michigan: Uphohn Institute for Employment Research.

International Bank for Reconstruction and Development. 1995. World Tables 1995. New York: International Bank for Reconstruction and Development.

Lazonic, William. 1979. "Industrial Relations and Technical Change: The Case of the Self-Acting Mule." Cambridge Journal of Economics, No. 3, pp. 231-262.

Miller, Steven M. 1989. Impacts of Industrial Robots. Madison, Wisconsin: University of Wisconsin Press.

Monthly Labor Review. Annual Indexes of Manufacturing Productivity and Related Measures, 12 Countries. Volume 119 (January/February 1996), p. 132; Volume 118 (February 1995), p. 102; and Volume 113 (December 1990), p.108.

Piore, Michael J., and Charles F. Sabel. 1984. The Second Industrial Divide: Possibilities for Prosperity. New York: Basic Books.

Sayrs, Lois. 1989. Pooled Time Series Analysis, Sage University Paper No. 07-070. Newbury Park, California: Sage Publications.

Stimson, J. A. 1985. "Regression in Space and Time: A Statistical Essay." American Journal of Political Science 29 (4), pp. 914-947.

Tufte, Edward R. 1974. Data Analysis for Politics and Policy. Englewood Cliffs, New Jersey: Prentice-Hall Incorporated.

United Nations Economic Commission for Europe, (UN/ECE), Working Party on Engineering Industries and Automation, and International Federation of Robotics. 1995. World Industrial Robots 1995. Geneva: United Nations.

Yonemoto, Kanji. 1982. "Robotisation in Japanese Industries." in OECD, Micro-electronics, Robotics and Jobs. Paris: OECD, pp. 235-245.

Zysman, John. 1994. "Technology and Trade: Villains or Saviors in the Struggle for Employment." Berkeley Roundtable on the International Economy (BRIE) Working Paper No. 67.