How Much Did the Liberty Shipbuilders Learn?
A Graphical Summary
This is a graphical, non-technical summary of my paper, "How Much Did the Liberty Shipbuilders Learn? New Evidence for an Old Case Study." Journal of Political Economy, 109(1):103-137.
Thumbnails link to full-size pictures or to an additional page of thumbnails.
Introduction
In 1941, the U.S. Maritime Commission (USMC) embarked on a massive expansion of the merchant marine fleet under the auspices of the Emergency Shipbuilding Program.
The standard Liberty ship, an all-welded cargo ship with a displacement of 7,000 tons, was the centerpiece of this program. The original plan was to build no more than 60 such ships, contracted under a lend-lease program for the British. On March 27, 1941, Congress approved the Defense Aid Supplemental Appropriations Act, which envisaged an additional 200 ships for the lend lease program. By April, this number had risen to 306 vessels, of which 112 would be Liberty ships.
After Pearl Harbor, the US view of the shipbuilding program was fundamentally revised. Several new programs were initiated, first because of the US entry into the war, and second because of German success in the Atlantic in early 1942, during which time losses were substantially exceeding new construction. Eventually, 16 U.S. shipyards delivered a total of 2,580 Liberty ships ships, by far the largest ever production run of the a single ship design. An additional, 119 Liberty ships were produced with modified designs -- colliers, tankers and and aircraft and tank transports -- were also delivered.
A revolutionary aspect of the Liberty shipbuilding program was that a substantial portion of ship construction was undertaken off the ways (the berths in which the keel is laid and from which the ship is eventually launched).
Most yards had a linear 'conveyor belt' plan. Steel plates and shapes entered a holding area in the yard on its inland side, and passed through a large prefabrication area where major sections of the ship were constructed. The sections were then transported on rails or by moveable cranes to one the ways, and large whirley cranes lifted them onto the hull for final assembly.
Welding constituted the bulk of this work. A Liberty ship contained almost 600,000 feet of welded joints, and welding labor accounted for about one third of the direct labor employed in construction.
Once the main structures were completed, the vessel was launched and moved to the outfitting docks nearby. Another keel was typically laid on the vacant way within twenty-four hours. There, final painting, joinery and electrical work were completed, and rigging and lifeboats were added.
The same day that the final outfitting was completed, the ship was delivered to a representative of the USMC, boarded by its crew, and sent to join one of hundreds of convoys crossing the Atlantic or the Pacific.
Productivity Growth in Liberty Ship Production
The Liberty ship program attracts the attention of economists primarily because of the massive productivity gains realized in a very short space of time. Over the course of three years, labor productivity rose at an average annual rate of 40 percent. The speed of production increased even more rapidly. While it often took as much as 300 days for a yard to deliver its first ship, by 1943 delivery times were often less than one month.  One ship, the Robert E. Peary, was built in under five days. Moreover, because yards entered the program at different times, the industry average understates the rate of increase in output per worker at individual yards. Other types of vessels built by the Maritime Commission experienced similar rapid gains in productivity and production speed.
     In an article published in the 1945 issue of the Monthly Labor Review, Allen Searle showed that each doubling of cumulative output reduced labor hours per ship by between 12 and 24 percent. The phenomenal increase in labor productivity experienced during the Liberty program has long been attributed to learning from the experience gained by repeating the same production activities many times, a process known as learning by doing.Leonard Rapping's paper in the 1965 issue of the Review of Economics and Statistics is most closely associated with this learning by doing interpretation of productivity growth in the Liberty program. He estimated the parameters of a simple production function, augmented to include cumulative output as a measure of experience, using pooled data from fifteen yards. His analysis confirmed Searle's earlier work. Each doubling of cumulative output was associated with an increase in output of between 11 and 34 percent. Moreover, this apparent learning effect was robust to the inclusion of calendar time, which had no significant impact on productivity.  Finally, in 1990's Management Science, Linda Argote, Sara Beckman and Dennis Epple (using yet another variation in methodology) reconfirmed the importance of learning. In 1993, Nobel Laureate Robert Lucas described the Liberty ship data as "the best evidence I know of that bears on on-the-job productivity change. . . . What is exceptional about the Liberty ship evidence, I think, is the cleanness of the experiment, not the behavior it documents so beautifully."
Was the Liberty Ship Experiment Clean?
Numerous difficulties involved in measuring the sources of productivity growth raise the possibility that much of what has been attributed to learning by doing in many empirical studies may just be measurement error. In 1976, the economic historian Nathan Rosenberg  suggested that "cost reductions which have been attributed to learning by doing [may] have actually been due to other factors which have not been correctly identified, especially in cases where learning by doing has been defined as a residual." This paper provides empirical evidence that the Liberty ship program was not a clean learning experiment. The paper focuses on two aspects of the program. First, using new data constructed form National Archive records, it is shown that a significant part of the increase in production speed and the reduction in labor requirements can be explained by increases in the amount of physical capital installed in each yard. Second, one can link productivity growth to a deterioration in the quality of ships, measured by the probability that the ship would develop fractures. That is, the evidence suggests a tradeoff was made of quality for quantity. 
Building Capital and Building Ships
    A common feature of previous studies of learning in the Liberty ship program isthat none of them had access to data on capital investment. The solution was to use the number of  ways in each yard as a proxy for the capital stock, a measure that exhibits almost no variation over time for individual yards. It turns out that the number of ways is a very poor proxy for the amount of capital a yard had. Measures of infrastructure per way show significant variation across yards and, as one might expect, yards with more capital per way also had higher output per worker.  More important, however, is the fact that millions of dollars of capital were installed in the yards after vessel production began. In fact, as much as two thirds of all the capital that was eventually to be installed was added after the first keel had been laid. In some yards, the lack of adequate equipment and machinery in early stages demanded different labor-intensive production techniques.
    What is the effect of including the new data on capital investment in the statistical analysis of learning? The easiest way to show this is to consider a particular number that economists often try to estimate from the data. This number, which we will call  the elasticity of learning, measures the poercentage increase in the monthly rate of output that is obtained when experience (measured by cumulative output) rises by one percent. Argote, Beckman and Epple, for example, estimated that this elasticity was 0.44. My data are somwehat different (but I can obtain a very similar estimate of 0.48 if I also do not use the new capital investment data. However, on adding capital to the data set and re-estimating the elsticity of learning, I obtain a much smaller estimate of 0.26. That is, if one excludes capital investment from the analysis, the importance of learning by doing is estimated to be twice as important as it really is.
Were All Liberty Ships Created Equal?
An attractive feature of the Liberty ship as a case study of learning is that all vessels appeared to be the same. In many other industries, it is hard to separate the effects of learning on productivity from product design and specification changes. It is the combination of the strong learning effects (which turned out to be overestimated) and the unchanging nature of the Liberty ship that induced Robert Lucas to talk about the cleanness of the experiment. The final part of this study provides evidence that conventional wisdom about the second desirable feature of the Liberty program may also be mistaken. By February 1946, 362 ships, over thirteen percent of the Liberty fleet, had sufffered at least one major fracture. Of these, almost one third were "class I" fractures that threatened the structural integrity of the ship. Among the more dramatic and well publicized fracture incidents were the John P. Gaines, whose loss at sea tragically lead to the loss of ten lives; the Esso Manhattan, which fractured in calm seas at the entrance to New York harbor; and the Schenectady, which broke in two 24 hours after launching while tied up near the outfitting dock.
All this would not matter if the fractures occurred randomly. However, it turns out that one can establish a link between the speed with which a ship was built and the probability of fracturing. To show this, I matched the productivity data for every Liberty ship built with data on fractures. Using a statistical technique known as duration analysis, I then measured the effects that production speed, labor productivity had, along with various other factors, on the probabilty that an individual ship would develop fractures. The results were dramatic. If one were to holding all the other factors at their average levels, but reduce the labor hours expended in production of a ship from 1.25 million hours to 350,000 hours (these are about the range of the sample data), the probability that a ship develops fractures rises from six percent to twenty percent.  Put another way, there is strong evidence that the Maritime Commission and the yards traded quantity for quality. In fact, various contemporary documents show that this was a conscious decision on the part of top management. Even the American Bureau of Shipping, which was responsible for safety inspections instructed its inspectors not to be as rigorous as they would be in peacetime. What does this matter for studies of learning? It matters because yards were making poorer quality ships later in the war than they were at the earliest stages, and this implies that the official productivity figures, which have attracted so much attention from economists, overstate the true rate of productivity growth.
Errors and Omissions
The estimates of the contributions of capital and learning to productivity growth summarized above should be treated with caution. The study continues to omit various factors that could have contributed to productivity growth. The US Maritime Commission financed an extensive research department. There were numerous technological innovations, developed outside the industry, but which may have had important impacts on labor productivity within the industry. For example, a 1947 report to Congress documents hundreds of innovations in materials science and welding. Also important, the Maritime Commission developed and subsequently expanded labor training programs. These programs became critical as the shipyards increasingly turned to labor with no prior experience in shipbuilding. 
Conclusion
Are there general lessons to learn from this study? In terms of what on can say about learning rates generally, the answer must be no: the Liberty ship program was in many ways unique and very unlike modern private sector start-ups. But, in another sense, the answer is yes: in a case study that is widely viewed as one of the cleanest examples of learning by doing on record, the real causes of productivitity growth have turned out to be more complex and more diverse than economists have long believed to be the case.