2. PREPARATIONS

Since graduating in 1933, Haavelmo had worked for four years with Ragnar Frisch at the Institute of Economics at the University of Oslo, spent one autumn term studying statistics at the University of London and another half year traveling for study purposes in Europe, and taught statistics for one year at Aarhus University in Denmark. His background as Frisch's assistant and co-worker had made him a highly experienced econometrician, although he had not published much before the war.⁸

What had he learned from Frisch? Haavelmo assisted him when Frisch was at the very peak of his scientific achievements and creativity. He was familiar with Frisch's ideas and theoretical concepts and had studied his major publications thoroughly. As his “chief computer” he had taken part in most of Frisch's numerical experiments, which covered a wide range of theoretical and applied work. This broad general background is relevant for understanding Haavelmo's motivation and also his abilities. If we were to try to pin down specific strands in the intellectual approach that Haavelmo brought with him from Frisch's laboratory, the following items would be of particular interest:

With regard to item 1, Frisch's entry into economics was marked by a combination of harsh criticism of economics as it was practiced for lack of scientificity and a demonstration of how to proceed. His maxims were well formulated, but perhaps it is fair to say that his metatheory remained on the level of slogans and catchphrases. Starting from generalities such as “For any science the object of which is the outer world there comes a moment when the logical need awakens, when the immediate, more or less emotional conception of the basic concepts must yield to an objective and exact definition,” he developed his programmatic quantification formulations that a few years later became embedded in the constitution of the Econometric Society. He repeated catchphrases such as “theoretical economics is about to enter the phase of development at which natural sciences, particularly theoretical physics, have long been, the phase in which theory gets its concepts from the observational technique.” He thus compared economics to the natural sciences but also pointed out differences, in particular that economics was by and large deprived of access to experimental data and had to cope with “passive observations.” This recognition was closely associated with his assertion of the primacy of theory, in statements such as “Economic statistics tell us nothing if not viewed against the background of a broad theoretical-economic analysis” and “The observation material is and remains a dead mass until it is animated by constructive theoretical speculation.”⁹

With regard to item 2, among Frisch's earliest work was his data analysis, set out in Frisch (1929). The core issue was simultaneity: i.e., more than one relationship holding between the same set of variables. In Frisch's terminology, the set of observations would then be “multiply collinear.” The simultaneity problem led to identification concerns in estimation. Frisch wrote two insightful essays that dealt with identification. One, the “Pitfalls” essay, had just been published when Haavelmo started to work for him in 1933; the other was the famous memorandum for the conference on the Tinbergen volumes in 1938, written just weeks before the end of Haavelmo's assistantship.¹¹

Confluence Analysis started out rather straightforwardly with a geometric consideration, assuming a great number of observations of variates between which “there exist not only one, but two independent equations…. From the distribution of points it would be absurd to try to determine the coefficients of any of the two equations that we know a priori should exist between the variates. Indeed a set of points lying in a line does not contain enough information to determine a plane.” But with data contaminated by errors of measurement or stochastic shocks, it might seem as if the plane could still be determined, an illusion Frisch called “fictitious determinateness created by random errors.” In focusing on this problem Frisch argued explicitly for disregarding “sampling theory” (probability) in order to have the problem “laid bare.”¹³

With regard to item 3, Frisch's approach to business cycle analysis was his most important project throughout the 1930s. It became known internationally mainly as a result of his “Propagation-Impulse” paper in the Cassel Festschrift (Frisch, 1933a), published just shortly after Haavelmo's assistantship began. In this widely known and discussed work, Frisch explained persistent and gradually changing cycles as an outcome of the interaction between economic structure, represented as a dynamic deterministic model, and stochastic shocks, the “rocking-horse” simile borrowed from Wicksell.¹⁵

Thus Frisch's involvement with probabilistic elements in economics was profound, making it a little puzzling that he has ended up being labeled “antiprobabilist” in the history of econometrics.¹⁶

With regard to item 4, the potential welfare implication of improved econometrics may seem somewhat remote from the probability approach, but in fact it is not. Frisch was deeply concerned about improvements in the human condition being the ultimate aim of scientific economics, although he was seldom successful when he attempted to exert direct influence upon politicians and policymakers. Haavelmo obviously took over this rationale or had adopted it already when he chose to study economics in the depths of the Depression. The importance of the concept of “autonomy” was the potential use of econometrics for policy analysis, as Haavelmo explained in 1943, contrasting the needs of “astronomer” and “planner” for information (Haavelmo, 1943a), although Jakob Marschak gave more striking formulations in distinguishing between the “engineering” and the “meteorological” types of econometric inference (See Sect. 6).¹⁷

Frisch also told Haavelmo again and again that he needed more mathematics to do well as an econometrician—between them the message was shortened to “Carthage!”¹⁸

To the extent that Haavelmo had a plan for his visit to the United States, it was strongly influenced by advice from Frisch. In 1937, after Haavelmo's first trip abroad to study with Jerzy Neyman in London, Frisch drew up a two-year study plan for Haavelmo: one year in Europe and one year in the United States. His advice was to “read a considerable amount of mathematics before leaving Norway” and then take up mathematical statistics by going back to Jerzy Neyman and staying there “as long as you think is necessary in order to get a good foundation in sampling theory.” Half a year with Neyman and a similar period with Jakob Marschak and the group he had put together in Oxford might be worthwhile. At the time of Haavelmo's departure for the United States, both Neyman and Marschak had moved across the Atlantic. Frisch also advised a visit to the U.S. Department of Agriculture. He mentioned Theodore Yntema at the University of Chicago and Charles Roos, one of the founders of the Econometric Society, and Harold Hotelling at Columbia University, adding that Haavelmo ought to go out to Colorado Springs to see the work of the Cowles Commission.²¹

3. WRITING “THE PROBABILITY APPROACH”

After arriving on the SS Oslofjord in Brooklyn, New York, on June 26, 1939, Haavelmo made his way as directly as he could to Colorado Springs, where the Cowles Commission had been located since its foundation by Alfred Cowles in 1932 as a support organization for the Econometric Society.²²

The fifth Cowles Commission Research Conference, which ran from July 3 to 28, 1939, had 50–60 out-of-town participants, most of them members of the Econometric Society. Among those present whom Haavelmo knew from Europe were Jakob Marschak, near the end of a one-year leave from his institute in Oxford as Rockefeller fellow, and Horst Mendershausen, until shortly before the conference a research associate at the commission. From Frisch's 1937 list, in addition to Marschak, there were also Charles F. Roos and Theodore Yntema.²³

The attendees thus included half of those Frisch had recommended that Haavelmo see on his study tour abroad. Jerzy Neyman, who had moved permanently from London to Berkeley in 1938, had been invited but in the end chose not to come. Instead he sent Francis Dresch, who worked with Neyman (and also with G.C. Evans). With Dresch as an intermediary, Haavelmo could arrange to visit Neyman in Berkeley as his next stop in the United States.

Harold Hotelling was not at the conference, but he had a co-worker there. Abraham Wald had fled Europe in 1938 and got a position as a research associate at the Cowles Commission.²⁴

Close to 40 lectures were given by the participants over the period of the conference. Apart from Wald's paper, nothing much was close to Haavelmo's main field of interest. There were Keynes-inspired contributions by Abba Lerner, Mordecai Ezekiel, and Jakob Marschak; the latter also discussed the use of market, budget, and income data in demand analysis with references to recent work in the field. H.T. Davis spoke on problems in the theory of business cycles, in a vein of limited appeal to Haavelmo, who might have been more interested in Edward Dobb's attempt at estimating the length of Slutsky effect cycles created by taking moving averages.²⁶

Abraham Wald's contribution to the conference, “The Fitting of Straight Lines If Both Variables Are Subject to Error,” showed that consistent estimates could be arrived at in this problem close to the heart of confluence analysis, although only two variables were dealt with.²⁷

Haavelmo presented “Statistical Testing of Dynamic Systems If the Series Observed Are Shock Cumulants.”²⁹

Haavelmo sent Frisch a letter on the last day of the conference. He felt his four weeks there had been well spent and expressed satisfaction with most of the presentations and discussions. He even seemed to have tolerated a few chatterboxes at an event that had so much to offer. He also found the conference of enormous practical value, as he met most of the individuals on the list of people he wanted to see and made arrangements for visiting Berkeley, Chicago, and Columbia.

Of those Haavelmo had met so far, he was in no doubt about ranking Wald and Marschak most highly. At this stage he had read hardly anything by Wald, who at the time of the conference had barely started his publishing in the field of theoretical statistics. Haavelmo's impression from the time they spent together was that Wald was a brilliant mathematician and statistician from whom he could learn a lot. He had also picked up at the conference that Wald would take over Hotelling's job at Columbia.³⁰

After the conference ended Haavelmo stayed on for some weeks in Colorado Springs with Wald, Marschak, Tintner, and a few others for hiking and further discussions on statistical testing and estimation. The smaller group gave Haavelmo a better opportunity to introduce his own scientific ideas, and he used it. Wald showed immediate interest in Frisch's irreducibility problem, set out in Frisch (1938). It was a mathematical problem rather than a statistical problem. Marschak had posed, perhaps provocatively, the following question with reference to Frisch's business cycle theory: Suppose an observed time series displays marked cycles of a certain length, and further that we have constructed a dynamic theory implying a differential equation for the observed variable of such a general form that it may have a cyclical solution. What then is the point of fitting this differential equation to the observed series and solving it for the time path of the variable? Won't it just give us a solution corresponding to the observed cycles?³¹

To help the reader keep track of Haavelmo's movements during the first two years he spent in the United States, the period in which he wrote “The Probability Approach,” Table 2 shows where he was, when, and with whom, though the text must be consulted for more precise information about the contacts.

Trygve Haavelmo's itinerary, June 1939–August 1941

After the Cowles Commission conference, Haavelmo left for Berkeley at the end of August 1939, arriving on the ominous day of the outbreak of World War II. One may presume that his stay in Berkeley was of great importance for his future work, as he chose in his Nobel speech in 1989 to mention his Berkeley visit 50 years later in the following passage:

From the context this is clearly significant and intriguing information about Haavelmo's progress, but presented as anecdotal evidence after 50 years it must be interpreted with caution. The importance of the quote is underlined by the fact that in this speech Haavelmo mentioned no one else as having furthered his scientific progress.

The meeting between Haavelmo and Neyman in Berkeley in 1939 figures prominently in the history of econometrics and in fact as a conversion of Haavelmo from “Frisch's thinking to Neyman's.”³³

In 1939 Haavelmo was a mature statistician and econometrician, rather than a novice. After his visit to London he had also taught statistics for one year at Aarhus University. As an econometrician he had reached the conclusion, unlike several other practitioners, that Neyman–Pearson testing was fully applicable to economic relations.³⁴

Why then was Haavelmo not more precise about the timing of events in his Nobel speech? The events were 50 and 53 years ago, respectively, but it is unlikely that Haavelmo's memory failed him. Knowing Haavelmo's habit of not wasting words needlessly (and never talking about himself), it seems more likely that he saw no point in bothering the dressed-up audience in Stockholm with itinerary details from his early years. Haavelmo's point was that Neyman had exerted a major influence on his scientific career; he had led Haavelmo onto the road of modern probability reasoning. That explained why he gave Neyman such prominence in the Nobel speech. It also explains why he was eager to meet with Neyman again in the United States, namely, because he had ideas to present and certainly would appreciate Neyman's reactions to them (cf. his seminar lecture, which is discussed later in this section).³⁵

Thus we should consider Neyman's influence as exerted in 1936 but with great effect on Haavelmo's work over the next 10 years. We might even put forward the hypothesis that without Neyman's influence Haavelmo might well have remained a theoretician and an applied and empirical economist, rather than the fully fledged econometrician he was for a period of 10 years or so.

This reinterpretation of the Haavelmo–Neyman relationship is not meant to deprecate the importance of their second meeting in Berkeley in 1939. Haavelmo attended Neyman's lectures.³⁶

Haavelmo had a chance meeting with Harold Hotelling when he happened to pass through Berkeley on his way to India. As Hotelling was the reason why Columbia University was on Haavelmo's list of places to visit in the United States, it was a fortuitous meeting. Haavelmo knew that Frisch considered Hotelling one of very few outstanding theoretical statisticians in the United States and an excellent economist too. He also went down to Palo Alto one day at the invitation of Holbrook Working of the Food Research Institute, Stanford University, who wanted to discuss his recent work with Haavelmo.

Haavelmo gave a seminar lecture at Neyman's Statistical Laboratory at the very end of his stay. He spoke on the frameworks for econometric analysis of Frisch and Koopmans, respectively, starting out with a quite general formulation of pure theory with a (theoretical) variable y as a function of number of variables x and then moving to alternative statistical formulations, the most general of which was

where the (y′′,x′′) variables were measuring errors and y_1t represented neglected variables. The A_it's were parameters; the t subscript represented the idea, he explained, that the parameters could be shifted as a result of variable changes.³⁸

The assertions reflected Haavelmo's main concern: the possibility of statistical testing of economic theories and what it required of the formulation of theoretical statements to be tested.

Errors of observation had been a general topic among econometric practitioners in the late 1920s and early 1930s. Frisch took it up in his Confluence Analysis. In the lecture Haavelmo elaborated on the implications of such errors, mainly by reference to Frisch's assumptions of measuring errors in all variables: namely, that the moment matrix of the observable variables would not tend to zero in the limit and thus minimizing the residual variance would not give unbiased estimates. He also recapitulated Koopmans's application of Fisher's maximum likelihood principle in some detail, adhering to Koopmans (1937).

Shortly after Haavelmo left Berkeley he sent Frisch a progress report on his research. His general idea of what he was after had changed since coming to the United States. As he told Frisch, he had originally thought he would go looking for a new technical method, such as a new regression approach, but had found it necessary to take up the entire question of testing economic relations on a broader basis. It is near at hand to see in this broadened perspective an influence from Neyman, but a close reading of Haavelmo's Cowles Commission contribution, not least the title he had given it, leaves little doubt that he had brought the idea of “testing economic relations” with him from home. On the other hand, the scope of the statistical testing may well have been broadened in Berkeley.

Haavelmo's approach, he intimated to Frisch, would be to write out, at least for himself, “a survey of the terrain”: namely, that economic theory had come quite a long way with regard to the construction of rational systems and in drawing correct conclusions from strictly defined assumptions but little had been done with regard to building “rational theories about the gap between theory and observations.” As soon as there was a discrepancy between theory and data—i.e., practically always—one was completely at a loss as to whether a theory was “good” or “bad,” whether the deviations were “large” or “small,” etc. One could hold a personal opinion about it and reach valuable conclusions that way, but it was useless to try to convince others on that basis. The missing “rational theory” here is obviously just a less poetic expression for the “bridge [that] was never completely built.”

Haavelmo also reported to Frisch that he had observed during his few months in the United States that the inevitable subjective element in such assessments was frequently used as an argument against Tinbergen's work. He recognized that there could be dubious aspects connected with the introduction of probability statements, but he had nevertheless come to the conclusion that this was the only possible, rational way out. When considering macroeconomic relationships, it was hardly useful to add more or less arbitrary assumptions about the nature of the random deviations. In that case one would not know which “fundamental probability set” (referring to Neyman) one was dealing with, and the probability concepts would be floating in the air.⁴⁰

Haavelmo had thus come to the conclusion, with or without Neyman's help, that testing implied more than just bringing the economic relationships formally into a probabilistic framework; prior to that probability had to be brought into the economic relationships. He elaborated to Frisch further on how to do just that:

From Berkeley Haavelmo moved on to Chicago, stopping en route for a week in Ames, Iowa, where he had been invited by Gerhard Tintner to give a seminar lecture on problems in the statistical testing of economic relations. He started out as if taking a cue from Neyman's lecture to the Econometric Society meeting in Oxford in 1936 (cf. Bjerkholt, 2005, p. 509):

Thus in Ames Haavelmo spoke out against the opinion that probability theory could not be applied to economic time series, indicating that this viewpoint was a thing of the past. Although there was widespread skepticism about bringing probability into economic theory, all that was needed was persuasive arguments.⁴³

He also expounded a somewhat pluralistic view on the foundation of probability: “The kind of probability statements wanted may be a matter of taste, but once these questions are settled the probability calculus leads in point of principle to definite testing and estimation proceedings.”⁴⁴

In Chicago Haavelmo spent much of his time in the library of the Cowles Commission, which was located in the Social Science Building of the University of Chicago. Theodore Yntema had become research director. There were ten people on the research staff of the commission at the time, including Oskar Lange, who was professor in the Department of Economics, and Dickson H. Leavens, the managing editor of Econometrica.⁴⁵

Marschak also showed up in Chicago. On one occasion Haavelmo took part in a discussion of the Tinbergen–Keynes debate in Yntema's office with Marschak and Lange. As an outcome of this discussion, Marschak and Lange wrote a rather strongly worded response to Keynes's highly controversial review of Tinbergen (1939) and submitted it to the Economic Journal at the beginning of 1940.⁴⁷

From Chicago Haavelmo traveled via Philadelphia to Columbia University, New York, where Wald was teaching. Haavelmo and Wald had respected each other from their first meeting in Colorado Springs. Haavelmo would have felt that he had much to learn from Wald, who also took a great interest in Haavelmo's concerns and ideas. At Columbia Haavelmo attended Wald's lectures on statistics. Wald had not written much on modern statistics when he took up his post at Columbia. His first work in the field was Wald (1939b), held by Wolfowitz (1952) to be his most important paper. The lectures he gave at Columbia in 1939–1940, some of which Haavelmo attended, were “noted for their lucidity and rigor.” Lecture notes taken by students later introduced many of the new generation of American statisticians to the theory of the analysis of variance.⁴⁸

Early on in the term Haavelmo started to write sections of the treatise he wanted to produce during his stay. (It cannot be ruled out that he had sketched plans for the treatise and drafted sections already in Berkeley.) A little later he gave it the title “Theory and Measurement of Economic Relations.” He planned to finish it by the end of 1940. The outline of the treatise did not change much over time, the overall structure and sections remaining very much as drawn up at Columbia that spring. To make his argument, Haavelmo found it necessary to include a brief introduction to modern statistics for fellow econometricians.⁴⁹

In addition, Haavelmo worked with Wald on the “determination of coefficients”—i.e., identification—in simultaneous structural relations with reference to the “irreducibility” problem in Frisch (1938). Clearly, this was a continuation of the discussions at the 1939 Cowles Commission conference. Haavelmo reported to Frisch that their discussions so far had led to the conclusion that it was possible to come up with rules that were sufficiently general for practical application. Wald found the mathematical problem far from trivial and seemed quite convinced that there were no theorems in the general mathematical literature that would be of any help. The problem could be stated as one of ascertaining the rank of certain matrices once the unknown coefficients were entered. The rank could thus be insufficient for “critical” values of the coefficients. It seemed, according to Haavelmo, “not always possible to give rules—in the form of a finite number of steps—wherewith one can verify if just these critical values of the coefficients are possible values with regard to a set of observations constrained by the simultaneously fulfilled relations.”⁵⁰

Haavelmo also reported to Frisch on his attempt to introduce probability distributions into theoretical relations to represent the different behavior among individuals in similar situations. He applied this to the consumption–saving decisions of individuals and nicknamed the project “Distribution-Dynamics.” The general idea was to assume an initially given joint probability distribution of consumption, income, and wealth and a stochastic relation between growth in income and growth in consumption for each individual. It would then be possible to study the change in the distributions of consumption, income, and wealth resulting from the saving process (for given wages and interest rates). The focus was on the consequences of random differences in individual behavior. When an individual one year “by random” saves more than usual it will “systematically” influence his income next year, etc. At any point in time the market relation between consumption, income, and wealth could be found as the expectation of consumption for given income and wealth in the current simultaneous distribution of consumption, income, and wealth. Haavelmo was insistent about including wealth in the explanation of the propensity to consume, on the grounds that it would influence the stability of the consumption under year-to-year variations in income. During the spring Haavelmo had contact with Frederick C. Mills at Columbia and the National Bureau of Economic Research and was invited to spend some time at NBER to explore data that could be used for his Distribution-Dynamics project.⁵¹

From around the end of February Haavelmo and Marschak were in touch again in New York. Marschak, whose fellowship had expired, had decided to stay instead of returning to his Institute of Statistics in Oxford when offered a position at the New School for Social Research in New York. Marschak had sent Haavelmo the paper he and Lange had written in response to Keynes's review of Tinbergen and looked forward to discussing it with him. He intimated to Haavelmo that he and Lange had found one or two further difficulties with Tinbergen's work, but had chosen not to elaborate upon them in the paper as they were not touched upon by Keynes.⁵²

At the time Marschak was working on his contribution to a memorial volume for Henry Schultz. This volume was not published until 1942, but Marschak dated his contribution very precisely to March 1940.⁵³

That the simple two-equation model is used for discussion of identification in both Marschak (1942) and Haavelmo (1943a) can, however, be traced back to Frisch's “Pitfalls” essay, which Haavelmo knew well from his early years with Frisch but which was also well known to Marschak, who had intervened to calm the “acrimonious debate” between Frisch and Leontief in 1933–1934.⁵⁶

Haavelmo had by that time done some work on the “Theory and Measurement” treatise but had also spent time drafting two papers inspired by discussions in Colorado Springs the previous summer. One was on the question Marschak had raised. Haavelmo showed in the paper “The Inadequacy of Testing Dynamic Theory by Comparing Theoretical Solutions and Observed Cycles,” eventually published as Haavelmo (1940a), that a “shock-cumulant” in the Frisch–Slutsky sense—i.e., a dynamic equation with stochastic shocks added to it, with a weight system consisting of two exponentials—could well result in cycles. Fitting a second-order linear difference equation to the shock-disturbed constructed series without trend elimination would recover the exponential functions with exponents close to the true values. Fitting after trend elimination resulted in a cyclical solution. Therefore letting the apparent properties of the observed time series—i.e., cycles—influence the ex post specification of the dynamic properties of the theory would give little insight by fitting to data. As Marschak had quipped, one would just get a description of what one already had observed.⁵⁷

The other paper inspired by discussions in Colorado Springs the previous year was completed by Haavelmo in the early spring of 1940 and titled “Some Generalizations of the ‘Cobweb’ Theory.” The basic model looked like a simplified version of Frisch's propagation model and was simply

In the paper Haavelmo first studied the effect of autonomous cyclical shifts in demand and supply functions to reflect business conditions. Then he went on to study “stock regulation policy” to stabilize prices within the model.⁵⁸

Haavelmo left New York late in June 1940 to attend the sixth (and last) Cowles Commission Research Conference, which ran from July 1–26, 1940. Although the commission had moved to Chicago the previous year, the conference was still held in Colorado Springs. Haavelmo and Wald had decided early on that they would both attend and also this time stay on for awhile under the shadow of Pike's Peak.

Among those who took part again were Davis, Dresch, Mendershausen, Roos, Wald, Holbrook Working, and Yntema. Among the newcomers were Oskar Lange, Wassily Leontief, and Paul Samuelson. Haavelmo had even arranged for a couple of Norwegian economists to be invited too. Irving Fisher, who was in his seventies, showed up to speak on the velocity of circulation of money and to give a public lecture, “Which Way to Peace?” On his first sabbatical Leontief was en route to California and Mexico and spoke with reference to his forthcoming book, The Structure of American Economy 1919–1929, which launched input-output economics. On the same day Samuelson spoke on the correspondence principle; his paper on the subject was soon to appear in Econometrica (Samuelson 1941).

By 1940 Wald had gone deeply into statistical theory. In his paper “A New Foundation of the Method of Maximum Likelihood in Statistical Theory” he had succeeded in showing how, under quite general assumptions, all maximum likelihood estimators tended toward a normal distribution.⁵⁹

Haavelmo's paper “The Problem of Testing Economic Theories by Means of Passive Observations” dealt with what would become a core issue in “Theory and Measurement.” He posed the key question in the lecture: “Can we measure economic structure relations by means of data which satisfy simultaneously a whole network of such relations—i.e. data obtained by a passive watching of the game and not by planned experiments?” This was clearly an outcome of the discussions he had had with Wald about Frisch's “irreducibility” problem.⁶⁰

Haavelmo was again satisfied with the time he spent in Colorado Springs, as there had been several good lectures, if fewer than the year before. The discussions in 1940 had been more “vacation-tainted,” and there had been no afternoon sessions as in 1939. The most valuable aspect of his entire stay in Colorado Springs that summer was probably the chance to discuss various topics at length with Wald. After the conference, which turned out to be the last Cowles Commission summer conference, Haavelmo and Wald went hiking in Colorado together with Arne Skaug, Haavelmo's colleague from Oslo.⁶¹

Haavelmo left Colorado and went back to Columbia University with Wald at the end of August and remained in New York during September 1940. From October he settled in at Harvard. Among the teachers he had the most contact with were Joseph Schumpeter, Gottfried Haberler, and Alvin Hansen, with all three of whom he had “innumerable interesting discussions and informal talks.”⁶²

At Harvard Haavelmo got his treatise typed up. This was also an opportunity to go over the argument and improve the presentation. Time became less pressing when he was notified that his fellowship would be extended to cover 1941. He wanted one or two qualified readers to look at his manuscript, not so much the technical sections as his entire argument. He gave a carbon copy to Schumpeter.

Also at Harvard was the German-born Hans Staehle, whom Haavelmo had met a few times in Europe.⁶³

In 1941 Haavelmo received his first formal appointment at the University of Oslo. He was given “universitetsstipend,” a kind of research fellowship comprising limited teaching duties, which he each year dutifully asked permission to be relieved from.⁶⁵

Haavelmo remained in Chicago until after the middle of February. In the meantime he had been invited by Arthur Smithies to come to the University of Michigan, Ann Arbor, to give lectures on confluence analysis. Smithies indicated to Haavelmo that he had heard from others that “your missionary zeal will lead you to accept such a suggestion.” This turned out to be perfectly correct, but Haavelmo tried to limit the visit to “a few days” on his way from Chicago to Harvard. Smithies pushed for him to stay “as long as you can and at least long enough to do justice to your subject,” and in the end they settled on one week, during which Haavelmo offered to “take part in as many group meetings as you could arrange.” He came to the university and lectured each day for one week on “statistical regression techniques,” presumably giving the students a taste of his own ideas in addition to the Frischian confluence analysis.

Haavelmo obviously liked being at Harvard, not least because of the presence of the many econometricians there and at nearby MIT.⁶⁶

In writing to Frisch about his missionary activity, Haavelmo commented that it was strange how few people really understood the importance of “multi-collinearity” for the question of numerical determination of structural equations. He suggested to Frisch that even Koopmans, who gave a lecture as part of Staehle's course, was confused on this point.⁶⁸

Haavelmo could tell Frisch in the middle of the term that he was nearing completion of his manuscript, which he described as being on “statistical verification of economic relations.” Schumpeter was still reading it, and Haavelmo intended to let Wald look at the more technical parts. Haavelmo may have considered the manuscript practically completed when he dated the preface April 1941.

Shortly afterward he also let Leonid Hurwicz read the manuscript. Hurwicz had come to the United States in 1940 and had been hired as a research assistant by Paul Samuelson at MIT. It seems likely that Hurwicz and Haavelmo were brought into contact with Samuelson as an intermediary. Hurwicz had frequent contact with Haavelmo at the end of the spring term, and Haavelmo received comments on his manuscript in June 1941.

Haavelmo was keen to spend another summer with Wald. The two previous summers they had spent together in Colorado Springs had been very productive periods for Haavelmo. He had used them to air a lot of ideas with Wald, who had offered much constructive criticism. Haavelmo told Wald in May 1941: “I have learnt more during our previous summers together than during all the time I have spent elsewhere in this country.”

Wald, who had spent every summer he had been in the United States at Colorado Springs, wanted to go there again. Haavelmo, however, much preferred the forests and lakes of Maine, which were closer to his favorite landscape. He checked out a number of possible locations, Millinocket on Pemadumcook Lake and Greenville near Moosehead Lake, among others. He knew he would have to put pressure on Wald to tempt him away from Colorado Springs and went to some lengths to achieve this: “Maine has got everything that Colorado Springs has, and it has got more … to choose Colorado Springs instead of Maine would definitely be to give up something for something inferior”!⁷¹

For Haavelmo it was a valuable opportunity to go through the complete treatise with Wald. Their summer discussions resulted in lots of changes to the manuscript, including the correction of a number of errors in the statistical formulas, but not in any major rewriting of the treatise. The time together gave Haavelmo a chance to scrutinize the presentation of the more mathematical parts, and it may have inspired Wald to do work in a direction that would complement and support Haavelmo's argument, which was perhaps one motivation behind Mann and Wald (1943).

While in Maine Haavelmo wrote to Olav Reiersøl, his old friend in Oslo, congratulating him on a piece on confluence analysis he had published as the lead paper in the January 1941 issue of Econometrica. It was a major breakthrough in an attempt to develop econometrics further within Frisch's confluence analysis paradigm and introduced the ideas of instrumental variables (without using the term) that Reiersøl would develop in his doctoral dissertation.⁷³

Haavelmo also reminded Reiersøl of their many conversations about Neyman's and Pearson's disregard for Fisher's maximum likelihood method as having no foundation with regard to the most powerful tests, shortest confidence intervals, etc., and gave him advance notice that Wald had shown in a series of papers that the maximum likelihood method could be firmly grounded on the basis of Neyman–Pearson's scheme and thus Fisher was right after all!

Haavelmo came back to Harvard with a revised manuscript toward the end of August 1941. In the preface, he acknowledged debts to two people: Ragnar Frisch (“his influence can be traced at almost every point”) and Abraham Wald, whose suggestions and help had clearly been comprehensive, particularly in the formulations of the “statistical sections.”⁷⁵

One of the latest changes to the manuscript was that Haavelmo dropped the idea of placing a quote at the start of each part of the treatise. For the first three parts he carefully selected quotes by Pareto, Poincaré, and Chuprov but for unknown reasons struck them out again.⁷⁶

In a letter he sent to Frisch that did get through, Haavelmo explained that the more he had worked on the questions of statistical verification of economic laws, of structural versus confluent relationships, etc., the more convinced he had become about the necessity and value of studying such problems from the viewpoint of probability and random variables. In the treatise he had decided to go the whole way and put all the problems on a probabilistic basis. It turned out that they could all be formulated so that they fell under the Neyman–Pearson scheme for testing statistical hypotheses.

In the treatise he had tried to give the general principles for such a formulation. At the beginning he had thought that this would cover only a special group of econometric research problems and that other kinds of apparatus would be needed for a whole group of problems “where probability considerations do not apply.” But he decided to abandon that way of looking at things. Problems might arise for which another technique was better, but with regard to the problems then being discussed in econometric research, they could, as far as Haavelmo could see, be formulated—precisely and also generally—as questions about testing statistical hypotheses in the Neyman–Pearson sense.⁷⁹

4. TREATISE COMPLETED

Let us try to determine which elements of the treatise Haavelmo brought with him from Europe and also make some observations on the process of writing the treatise.⁸⁰

Despite Haavelmo's statement in the preface that the idea of undertaking the study that resulted in the treatise “developed during my work as an assistant to Professor Ragnar Frisch” no evidence has been found that he brought with him any kind of outline of what he was going to do or even a working title for it. On the other hand we cannot rule out that he brought with him from Europe notes that later were destroyed. In fact, this seems likely, as he had had ample time to prepare during his year in Aarhus prior to departure. Also notes related to his Copenhagen lecture and the 1939 Cowles Commission conference are missing. Hence, we are somewhat at a loss in discerning pieces of his treatise brought from Europe.⁸¹

The wording in his application of “connecting economic theory and statistical observation” was decidedly vague. When he was offered a fellowship by the Rockefeller Foundation in late 1939 he gave as his research program “problems of statistical testing of economic relations.” As in his original application, there was also a second part of the program: namely, “problems of individual economic behavior in relation to saving.”

Surely, Haavelmo must have started thinking about the report he would put together on his scientific results early on during his stay. He had studied modern statistics for several years, adding to the training in premodern statistics he had obtained from Frisch's lectures and seminars. He had not read Kolmogorov, but, judging from his references, he had studied a considerable number of books on modern statistics, particularly Jerzy Neyman's works. He had in fact studied in three phases: first in London in 1936; then in the last year or two before leaving for the United States, when he was teaching in Denmark; and, finally, he dug even deeper into probability fundamentals after he got to the United States, thanks to the direct contact he had with Neyman but, perhaps more importantly, with Abraham Wald.⁸²

While still in Europe he seemed to have reached the firm conclusion that the only way of testing economic theory and hypotheses was to formulate them in such a way that the Neyman–Pearson theory of statistical testing could be applied. That the hypotheses of economic theory were of a statistical nature was his main theme in the Copenhagen lecture he gave just before he left (Bjerkholt, 2005, pp. 518–520). On the same occasion he also seemed fully convinced and argued forcefully that the widespread opinion among economists that hypotheses about economic time series did not lend themselves to statistical testing had no merit. He repeated this position in Ames, in a lecture perhaps partly based on the Copenhagen paper.

With regard to his reliance on the Neyman–Pearson theory of testing he did not seem to have considered alternatives, supposedly for the simple reason that he did not know of any alternative and thus did not reflect on it either. Hence, we cannot predict how he would have reacted to the criticism of the Neyman–Pearson paradigm that came long after Haavelmo had left the econometric field for good. We just note here that a leading antagonist, James J. Heckman, found it “unfortunate that the Neyman-Pearson paradigm continues to be so influential in econometrics” and that “the Neyman-Pearson theory espoused by Haavelmo and the Cowles group takes a narrow view of science.”⁸³

We may note here also his somewhat pluralistic conception of probability. In Ames he had talked about the kind of probability statements one preferred being a matter of taste. This pluralism is also very evident in “The Probability Approach.” He said there that to interpret probability of “frequency of occurrence” will in many cases seem “rather artificial.” He mentioned, alternatively, probability as a measure of “a priori confidence,” apparently liking it better(?). He then moved on to deplore the “futile discussion” of what probabilities really are and wound up by arguing that the common “logical consistency requirements” of the various probability concepts were all we needed.⁸⁴

Haavelmo kept three manuscript versions of the treatise in his files. (For the discussion that follows, we will use TM as an abbreviation for Haavelmo (1941c) and number the three manuscripts chronologically.) TM-1 is handwritten and comprises all sections except those of part 1. The bulk of it was very likely written at Columbia between January and June 1940, although some sections may have been written later. TM-2 was typewritten (by a typist at Harvard?) and comprised all sections. It was basically just a typewritten version of TM-1 with the addition of the six sections of part 1 and various typewritten insertions, including the preface. On the first page Haavelmo wrote his name and “Harvard 1940,” but that was the only date in the manuscript. It seems reasonable to see TM-2 as the manuscript Haavelmo planned to complete within 1940 but then got a chance to improve when the time constraints were relaxed. TM-3 is clearly the manuscript from which TM was typed at Harvard in August 1941. It is based on the carbon copy of TM-2 with insertions of handwritten and typewritten pieces and represents a substantially revised version of TM-2. The improvements from TM-2 to TM-3 are of interest for the progress of Haavelmo's thinking but cannot be dealt with in more detail here.

An argument can be made for thinking that Haavelmo's investigation was limited at the outset to testing and verification or, more generally, to the scientification of economics, whereas the problem of estimation of simultaneous equations was not included. Simultaneity, confluence, or anything in that direction was not mentioned in his brief study-purpose formulations. Neither was the topic taken up in the two lectures Haavelmo gave in 1939 (see the preceding discussion). Haavelmo's report to Frisch just after leaving Berkeley, when he said that he had widened his horizon from looking for a new regression method to taking up testing of economic relations on a broader basis, with no mention of confluence problems, also supports this argument. This is consistent with Haavelmo's letter to Frisch when the treatise was just finished, saying that he had expected to work out principles for the testing of economic hypotheses for only a special group of problems but found that he could encompass in his approach all those problems currently being discussed—i.e., also problems with confluent relations. Finally, the introductory paragraph of the preface, which characterizes his treatise as “an attempt to supply a theoretical foundation for the analysis of interrelations between economic variables,” was not in the original draft but was inserted in the TM-3 version. There is no other, even remote, reference to simultaneity in the general introductory part of the preface.

On the other hand, the confluence problem of simultaneous equations and the ensuing identification problem were included in the TM manuscript from the beginning. From the preceding narrative we can easily see why. Haavelmo had chosen at the Colorado Springs conference in 1939 to speak about Frisch's inversion problem as a single-equation problem, as a highly relevant case for discussing estimation and testing. After the conference he took up the irreducibility problem in Frisch (1938), which resulted in Wald and Haavelmo discussing and working on it early in 1940. It is quite likely that the incorporation of simultaneity in the treatise started here, with Wald's suggestion that it was a nontrivial but still solvable problem, at least to some extent. Wald pointed to the rank condition as one approach to resolving the coefficient indeterminacy and may also have guided Haavelmo toward the use of the Gramian criterion. Haavelmo wrote all of this into TM-1 and most likely drew considerably on Wald also when he went over and rewrote these sections in Maine in 1941, making the argument more general and also more elegantly formulated. Thus after telling Frisch that he would present his “distribution-dynamics” idea at the 1940 Cowles conference, Haavelmo instead changed his mind and discussed for the first time the identification problems arising in a simultaneous model under passive observations and the need for a constructive approach to cope with them.

Why then did Haavelmo continue as a missionary for confluence analysis and bunch maps long after he had drafted his treatise with all the ideas that went into “The Probability Approach”? He lectured on confluence analysis on a number of occasions, well into 1941. His booklet with Staehle was hectographed only two or three months before “Theory and Measurement” was completed. His two-page paper, written in the autumn of 1940 to explain Tinbergen's findings that the rate of interest had little influence on investment (Haavelmo, 1941b), had a two-page appendix not only explaining but also exalting Frisch's confluence analysis as “a really powerful tool.” The explanation is simple, despite the view taken in much of the literature on the history of econometrics that confluence analysis and the probability approach represented alternative and irreconcilable paradigms. To Haavelmo, as indeed to Frisch, confluence analysis—cum bunch maps—provided an exploratory and diagnostic tool in data analysis whose (main) function was to discover “multi-collinearity,” for which purpose Haavelmo had used it to advantage on a great number of occasions. That this was the case follows directly from a remark Haavelmo made in his lecture notes with Staehle:

The first two chapters of “The Probability Approach” deal very little with probability but can rather be said to represent Haavelmo's philosophy of science. These parts of “The Probability Approach” have been more or less neglected in the literature on the history of econometrics. In the second chapter Haavelmo discusses what is meant by an economic law, and in the first chapter he establishes concepts and terminology for discussing “how the bridge is to be built.” An interesting assessment of Haavelmo's approach to the philosophy of science is given by Davis (2000), who finds it very modern, easy to fit into the “semantic approach to the philosophy of science” that has emerged in the last two to three decades. There is no obvious source or influence one can point to as having inspired Haavelmo in this regard. He could have given his main results from chapters III and IV without the philosophy of science framework (but the work would have been much poorer for it). One can read into Haavelmo's approach an attempt to take Frisch's quantification banner beyond the level of slogans. Apart from Davis (2000), this part of Haavelmo's work has been little studied. There are a number of significant changes in the relevant parts between the 1941 and 1944 versions of the treatise, suggesting that he gave them considerable attention when he reworked the treatise in 1943, perhaps because he was not fully satisfied with the original version.⁸⁷

5. ADJUSTMENT TO EXILE AND PROMOTION OF “THE PROBABILITY APPROACH”

At the outbreak of World War II Haavelmo may well have shared the opinion of many Norwegians that Norway could not stay out of a war between Germany and Great Britain for long. On April 9, 1940, Norway was attacked by the German war machine. Resistance lasted only a few weeks, and the Norwegian government left Norway on 7 June 1940 on the British cruiser H.M.S. Devonshire and became a government in exile in London. Haavelmo had enrolled in the Norwegian army in 1932 and was part of a field artillery unit. He had got permission from his military unit to study in America until January 1, 1941. If he had been in Norway on April 9, 1940, he would most likely have taken part in military skirmishes against the Blitzkrieg. People he met in the United States said that he was lucky to be out of the country, but Haavelmo may have looked at it differently.

A few months after the German invasion of Norway the Norwegian government in exile had its consulates in the United States draw up lists of Norwegians who were willing to do wartime work. Haavelmo signed on, but nothing happened. He kept well informed about developments in Oslo, though, mainly through the grapevine of Norwegians in the United States, among whom were a couple of fellow economists from Oslo. There was more bad news than good, and as 1940 turned into 1941, beyond the time he had ever considered staying in the United States, Haavelmo despaired.

This must have been obvious to others, and Paul Samuelson put it succinctly many years later: “[Haavelmo] was prepared to row back to Europe but that proved to be impractical.”⁸⁸

Postal communication with Norway was possible until the attack on Pearl Harbor. In his last letter to Frisch before this interruption, Haavelmo reflected on the wartime situation, concluding that theoretical work was more important than ever before:

But the letter to Frisch did not really reveal how upset Haavelmo was about the war situation or rather about the situation that the war had put him in. The autumn term of 1941 may in retrospect be viewed as a period of adjustment for Haavelmo—adjusting to exile.

Unlike Wald, Haavelmo chose not go to the Econometric Society meeting held in Chicago on September 2–4, 1941. It was a special meeting in connection with the fiftieth anniversary of the University of Chicago, with papers that would certainly have been of interest to Haavelmo. Koopmans spoke on the applicability of sampling theory to time series and Wald on large-scale distribution of the likelihood ratio. Among other participants were Marschak, Oskar Lange, Jacob Mosak, and R.L. Anderson. Instead Haavelmo prepared to go to Washington, DC, to meet a diplomat in the Norwegian Legation in Washington, a man he did not know personally but who was an acquaintance of Frisch. In a letter he sent in advance, Haavelmo explained the situation and expressed his belief that what he was working on was something that could be of some value after the war. He was confident that it would not be particularly difficult for him to get a teaching position but said that he was willing to put everything aside for meaningful work in one of the institutions or organizations working for the Norwegian government.⁹¹

But the trip to Washington in mid-September was a disappointment. In the Norwegian Legation Haavelmo was politely but firmly rebuffed—for the time being there was no work suitable for him, either in the United States or in London. He was advised that it would be better if he kept on with whatever he was doing and learned as much as possible in the United States that could be of postwar use but also avoided becoming a financial burden and refrained from causing trouble. On top of that he heard through the grapevine that people in London were saying that he and others in similar situations should really be doing something “for the Norwegian cause.” Haavelmo reacted with anger and despair, but the official rebuff also meant a kind of release from such worries.⁹²

Haavelmo also had another unpleasant experience in Washington, to say the least. He visited the immigration authorities about his visa, necessitated by new regulations from July 1, 1941. The new regulations meant that the State Department had to approve his visa, and this required affidavits from two independent persons, who, according to Haavelmo, had to guarantee his maintenance for the rest of their life. And not only that, for each of the two persons letters of reference from others were needed to vouch for them. The affidavits, furthermore, were to be given on forms “that I personally would feel very ashamed to ask anyone to fill out,” because of the questions about communist activity, etc. Haavelmo thus was confronted with the anticommunism rampant within the FBI at this time. He and the circle he moved within were hardly pursued and wiretapped to the same extent as, say, J. Robert Oppenheimer and his circle of friends in the Bay Area.⁹³

In the autumn at Harvard, Haavelmo devoted himself to studies not least to Keynesian theories of savings and investment. He had completed his treatise and thus fulfilled his main study purpose. He could hope for some feedback from the people who chose to read his treatise. In the meantime, he attempted to “dynamize” Keynes and took part in lectures and discussions on economic and monetary policy. He increased his contact with Harvard faculty members further in this third term in Cambridge. His lecture notes on confluence analysis were used as teaching materials in Staehle's course, and Haavelmo would drop by occasionally.

He was, however, facing a practical problem. His fellowship expired at the end of 1941, and he needed a job. Toward the end of October 1941 he decided to contact Marschak in New York to inquire about a temporary teaching or research post at the New School. Marschak was, as always, able to conjure up some ideas. One was to get Haavelmo onto the teaching staff at the New School, another that his Rockefeller Foundation–financed research project on demand studies could be broadened in scope to include Haavelmo as a researcher. In both cases his ideas presupposed that the Rockefeller Foundation would foot the bill, which made sense as the foundation had, under the circumstances, at least a moral obligation to look after Haavelmo beyond the current fellowship term.⁹⁵

Marschak was excited about the prospects of getting Haavelmo to the New School and went into detail about his teaching program for the forthcoming term. A considerable part of the teaching staff of the New School had been hired under a special “Refugee Scholar” program, financed by the Rockefeller Foundation. This program was not meant for people like Haavelmo but rather for scholars arriving as refugees from Europe. Marschak figured, however, that the statutes could be bent a little to accommodate Haavelmo. He raised the issue with the Rockefeller Foundation and with the management of the New School. He had no trouble convincing the New School that hiring Haavelmo would be a good idea. In the middle of November, Haavelmo received a letter from the dean, Hans Simons, offering him an appointment as lecturer in economics, which Haavelmo accepted immediately.⁹⁶

Marschak pursued his other idea of research cooperation with equal enthusiasm. He suggested that Haavelmo should come down to New York to discuss the matter fully and meet with faculty members. Before he left, Haavelmo sent a two-page note outlining under five headings research ideas he would like to pursue, one of which was demand functions, thus matching what Marschak had to offer.⁹⁷

This distinction had become a key element in Haavelmo's thinking, not least when he considered postwar applications of his work, but it had not been much accentuated in his treatise. After meeting and discussing with Haavelmo in New York, Marschak immediately sent the Rockefeller Foundation a three-page note, the bottom line of which was a proposal for the foundation to finance Haavelmo to work on the project for the first six months of 1942. Marschak elaborated on Haavelmo's qualifications, emphasizing his considerable expertise in practical econometric work with Frisch and his more recent acquisition of “statistical and logical fundamentals of econometrics (application of probabilities to economic time series),” and stated that

In short, this was the first application for funding research applying the methods that Haavelmo had introduced.

The distinction between the “meteorological” and “engineering” aspects of econometric inference had become an equally important idea with Marschak, although his understanding of it was not yet as fundamental as Haavelmo's. For both Marschak and Haavelmo, economic and statistical research had little meaning unless it was for the betterment of society, perhaps something that was so obvious it was taken for granted. But how was one to get at the autonomous relations required for (social) “engineering”? As Haavelmo could elaborate, passive observations of changing “economic weather” might, using the most appropriate methods, reveal information about parameters of autonomous relations.¹⁰⁰

Haavelmo left Cambridge for good early in December 1941 to go to New York, where he started to cooperate with Marschak immediately on the statistical demand functions project, drawing up a plan for the work ahead. Again he could also take part in the annual Econometric Society meeting, which was held in New York between Christmas and the new year. He expected to start teaching at the New School in the spring term. Everything was, however, dependent upon Rockefeller Foundation approval.

Haavelmo could still not stop thinking about his role—or lack of role—in the war effort, but the despair was gone. He thought about the situation in more constructive ways, influenced perhaps by his involvement with economic policy issues for postwar use at Harvard the previous term. He put it to a friend in London as follows:

Two weeks later Haavelmo was asked to take up a position as statistician at the Norwegian Shipping and Trade Mission (Nortraship). He told his friends: “I have been drafted for civil service at the Norwegian Government's Shipping Ministry in New York.” The nature of the “draft” was not quite clear, as it was hardly something he had to accept, but it was what Haavelmo really wanted, what he had been begging for.¹⁰²

In addition to compiling and analyzing statistics, Haavelmo worked in a small research division, following developments of the international shipping market and trying to get some idea of what the postwar shipping situation would be. The workload was not too heavy, and the location gave him an opportunity to keep in touch with Marschak and his fellow econometricians.¹⁰³

In 1940 Jakob Marschak had initiated a seminar group on economic and econometric topics that met once a month at the National Bureau of Economic Research. In addition to Marschak, originally the participants were Koopmans, Mendershausen, Wald, and J.J. Polak, plus some of Marschak's students and occasional guests. Haavelmo might well have been there as a guest in 1940. After a long break, Marschak reconvened the group with a seminar on December 6, 1941, which Haavelmo attended. From then on it met regularly on Saturdays every second month, usually at the NBER's Hillside retreat in Riverdale, where people could stay overnight. Haavelmo, Paul Samuelson, Hans Staehle, and Wassily Leontief were among those invited to join, as was Joseph Schumpeter. Abraham Wald arranged for graduate student Kenneth Arrow to attend some of the seminars.

In February 1942 Haavelmo addressed Marschak's seminar group with a lecture entitled “The Nature and Logic of Econometric Inference” at Hillside. According to the invitation that went out the meeting would address

1. Is there any danger in considering economic data as stochastic variables? Is there any use in so doing?

2. Is it always objectionable to use free-hand methods of curve-fitting?

3.

4. Is it more dangerous, in general, to draw inference from economic time series than from other kinds of statistical information?¹⁰⁴

Haavelmo's opening paragraphs put these questions from current discussions in perspective:

He then went on in characteristic philosophical manner to convince the audience to look at the economic data as produced by “observation-producing mechanisms” about which little was known but which could be exhaustively classified by the notions of probability distributions and stochastic variables. He gave questions 1 and 4 the most attention before he focused on “the nature of the relationships we operate with in economic theory,” particularly behavioristic relations, which are approached “imagining a certain kind of experiment.” Then came the key question: “Can we get at these relations, if they are really true, by a passive observation of the game?” He demonstrated, using a two-equation example, that there were indeed limits to the possibility of our rejecting false hypotheses imposed by passive observations of simultaneous equations. He made passing remarks on questions 2 and 3(a) and seemed to throw up his hands at question 3(b), stating, “No matter how refined our analytic tools become, there will always be a difference between good economists and bad ones.” He ended with a joke:

One of those present, Kenneth Arrow, wrote many years later:

Later in April 1942 Haavelmo became acutely ill with appendicitis and had to go to the hospital for an operation. Lying in bed gave him an unexpected opportunity to rethink some of the problems he had dealt with in his treatise. Recovering from the operation, he drafted the “Statistical Implications” paper (Haavelmo, 1943a), which brought out a core part of his argument in a more concise form and also discussed how to make predictions, which had been only touched upon in the 1941 treatise. He finished a draft in June 1942 and gave it to Marschak for comments. A little later he sent it to Schumpeter, who in August 1942 passed it on to the managing editor of Econometrica with his own and Marschak's recommendations. For the third year in a row there were only three issues of Econometrica, so it had to wait to be published until the first issue of 1943.¹⁰⁷

Around this time Haavelmo heard from Marschak that Koopmans was doing similar work for the British Merchant Shipping Mission in Washington. He contacted Koopmans soon after, suggesting that there was too little contact between these units and complaining that meetings between shipping missions of various Allied countries “never get down to a cold discussion of specific technical research problems.” Koopmans made it plain that he would welcome a visit, and Haavelmo went to Washington for a few days in July 1942. They were both bound by secrecy rules not to discuss facts and figures, but to Haavelmo that was irrelevant for the contact he wanted, namely, to discuss with Koopmans what kind of research related to their work would be most important. As he had just recently read Koopmans (1942), the visit also gave him an opportunity to discuss econometric problems.¹⁰⁸

Haavelmo addressed Marschak's seminar group a second time in 1942. His lecture took place on December 12 at the New School and was entitled “Problems of Estimation and Prediction in Economic Dynamics.”¹⁰⁹

In March 1943 Haavelmo was invited to speak “on econometrics or another statistical problem” to a group of students of economic statistics and mathematics at MIT. The weekly seminars in mathematical statistics were organized by two graduate students, Joseph Ullman and Lawrence Klein.¹¹¹

What “should be” was, however, far from what actually was the case. Much energy was spent on constructing rational models, involving exact relationships that were much too rigid to identify the theoretical variables involved with some actually observable economic quantities. The relationship between such exact economic models and economic reality was similar to the relationship between rational mechanics and the bodies and motions observed in the real physical world. But the difference in the degree of precision was (and would always be) tremendous. In economics it was not sufficient to set up a system of exact relationships first and then allow for certain small deviations in the applications to facts. One had to start from scratch with a probabilistic formulation of models; otherwise one would have to either declare all theories wrong or call almost any theory right by allowing for sufficiently large discrepancies (in a subjective manner).

Economists had been reluctant to give rigorous stochastic formulations of their theories. Many argued that economic quantities showed variations of a more general and complex type than those of which stochastic variables were capable. Comparing their theoretical relationships with facts, they thought they had a better case if they spoke about “errors,” “unexplained residuals,” etc., in a more vague and general manner instead of introducing the idea of probability distributions. This was a grave misunderstanding. Haavelmo attempted to show

He then elaborated on the implications of simultaneous equations as set out in Haavelmo (1943a).

6. THE HAAVELMO–MARSCHAK NEXUS AND THE PUBLICATION OF “THE PROBABILITY APPROACH”

Over a period from the end of 1942 and into the first half of 1943 Haavelmo reworked his 1941 treatise in his free time.¹¹⁴

In the revised treatise he added a chapter VI, “Problems of Prediction,” drawing upon the paper for the canceled Econometric Society meeting and the “Statistical Implications” paper. He then set out to rearrange the sections in “a more logical order.” This resulted in a considerably modified, improved, and shortened chapter I, “Abstract Models and Reality,” where he set out his scientification approach, the metalanguage terms he introduced, etc. Chapter II, “The Degree of Permanence of Economic Laws,” was left virtually unchanged, and a new chapter III, “Stochastical Schemes as a Basis for Econometrics,” was put together from sections cut out of other chapters, with the addition of a completely new section, “The Practical Meaning of Probability Statements.” The key chapter in the treatise changed its title from “The Verification of Economic Theories Treated as a Problem of Testing Statistical Hypotheses” to just “The Testing of Hypotheses.” Was there in this perhaps a slight withdrawing from the aim of verifying theories?

The preface was virtually unchanged. He added an acknowledgment of indebtedness to Jakob Marschak for “stimulating conversations,” and, in addition to thanking Leonid Hurwicz, he also thanked Edith Elbogen of the NBER for reading and commenting on the manuscript.¹¹⁶

The revision was complete by the beginning of June. Haavelmo gave the only version of the manuscript to Oskar Lange, the acting editor of Econometrica, when he met him in New York. After reading it, Lange returned the manuscript with the remark that “the book should be published by all means.” Before returning it, Lange had discussed the question of publication with Cowles and Leavens. As the manuscript was much too long for an ordinary paper, a separate issue of Econometrica was one possibility (although paper shortage was an obstacle), a Cowles Commission monograph another (although the monograph series was exclusively for Cowles Commission staff, which Haavelmo was not). Lange had also mentioned other possibilities.¹¹⁷

Marschak had moved to Chicago at the beginning of the year to become research director of the Cowles Commission. Haavelmo (1943a) had shortly afterward appeared in the January issue of Econometrica. After a couple of months of adjusting to his new environment and new tasks, Marschak sat down with Hurwicz, who had been recruited shortly before by Yntema as a research associate, to study Haavelmo's article. Unfortunately, they found mistakes in Haavelmo's formulas! Marschak somewhat rashly instructed Leavens to contact Haavelmo to ask for an erratum slip to be inserted in the next issue of Econometrica. Haavelmo saw at a glance that Marschak and Hurwicz had acted a little too quickly.¹¹⁸

After further studies of Haavelmo (1943a), in May 1943 Marschak entered into correspondence with Haavelmo to clarify the meaning of various sections of the paper. Assisted by Hurwicz, Marschak must at the time have represented the Cowles Commission's total pool of econometric expertise. He was apparently primarily motivated not so much by thinking about the Cowles Commission's future research program as by short-term concerns about applying Haavelmo's new ideas to his practical research work. He was involved in demand studies with George Garvy and was about to start work with William Andrews on the estimation of production functions. In these applications Marschak's concern was again the “meteorological versus engineering” types of econometric inference. Marschak wanted policy conclusions, and he wanted to be sure to get the “engineering” answers.¹¹⁹

Much of the discussion in the exchange of letters revolved around the merits of least squares estimation versus maximum likelihood in problem examples that Marschak provided and when and why “indeterminacy” (= unidentifiability) occurred. The problems turned out to be a little more difficult to resolve than Marschak had originally thought. Marschak and Haavelmo wrote four letters in May–July 1943.¹²⁰

When the dialogue started, both Marschak and Haavelmo were eagerly waiting for Mann and Wald (1943) to appear in the next issue of Econometrica. Mann had presented their results at the Marschak seminar (not attended by Marschak) earlier in the year. Haavelmo, who had been there, got the impression that Wald and Mann “had cleared up everything, practically speaking, at least for large samples, and that the maximum likelihood method seemed to be a good method in all cases.”

By the time the letter exchange ended, Marschak had still not seen the revised manuscript. He had no qualms about its importance and conveyed to Haavelmo that the best solution would have been to print and distribute it as a free supplement to Econometrica, but that was, alas, precluded by war regulations. Marschak also suggested that it would be easier to publish the manuscript as a Cowles Commission monograph if it was a little longer.¹²¹

Haavelmo did not like this suggestion much, partly for the reason that he did not think he could find time for a major rewrite, not even the addition of an extensive summary. He had been through a revision and put in all he had to say on the matter, adding that it seemed to him that “the whole thing now hangs rather nicely together as one compact argument, and such that it can—and must—be read in one stretch, which I would consider an advantage.”¹²²

In mid-July Marschak came up with the neatly thought out idea that Haavelmo's treatise and Mann and Wald (1943), which had still not appeared, could be published together as a Cowles Commission monograph. Given the importance of Mann and Wald (1943) for the procedures Haavelmo recommended, this was a very sensible proposal. Marschak had already got around the obstacle of Haavelmo not being a Cowles Commission staff member by getting approval from members of the commission to offer Haavelmo a research associateship. There had never been a research associate not in residence before, but Marschak used his recent correspondence with Haavelmo to convince the commission that there was de facto intensive cooperation between the commission and Haavelmo. The appointment did not involve any salary and was, as Marschak said, “a purely moral tie.”¹²³

Haavelmo liked Marschak's idea for publishing the treatise and accepted the appointment as research associate. He soon found out from Wald that he had no objection either. It would perhaps have been better if the Mann and Wald paper had been rewritten for the monograph, but Wald had no time for that and suggested a direct reprint with a Cowles Commission foreword explaining that the paper had been added because it dealt “directly and more explicitly with a group of fundamental statistical problems that have been raised” in Haavelmo's study.¹²⁴

After some further discussion of practical issues, such as the insertion of more cross-references and compiling a common list of relevant literature, etc., Marschak suggested that the book title should be The Theory of Economic Measurements, with all three names listed as authors. Haavelmo suggested that the overall title be amended to Contributions to the Theory of Economic Measurements. For Haavelmo's part Marschak's suggestion was “The Stochastic Analysis of Economic Relations.” Haavelmo turned this down, arguing, “I would like the title of my part to express the fact that my analysis deals not only with methods of statistical analysis but also with the foundations for this, namely the probabilistic formulation of economic models.” He suggested either “The Probability Approach in the Theory and Measurement of Economic Relations” or merely “The Probability Approach in Econometrics.” The latter appealed immediately to Marschak.¹²⁵

In mid-October Marschak received and read Haavelmo's revised manuscript and sent editorial comments. He suggested several minor changes and the addition of a conclusion “instead of stopping in the midst of details.” Haavelmo accepted the idea, but what he wrote was more like an afterthought. He started it out rather apologetically:

But he ended it by raising the flag on behalf of economic research:

Unfortunately, the entire book-publishing plan collapsed because of the wartime rationing of paper.¹²⁶

At the beginning of 1944 a new idea about publishing the book came from Chicago: namely, as a supplement in Econometrica, which had published less than 300 pages in 1943 and thus had a quota that could be carried over to 1944. But Mann and Wald (1943) would have to be dropped, partly for paper quota reasons and partly because it would hardly be proper under the circumstances to reprint in Econometrica what had been published the year before. Instead a two- to three-page summary of Mann and Wald (1943) could be added (but it was not). Haavelmo's treatise would be listed also as an issue in the monograph series of Cowles Commission monographs.¹²⁷

On March 15, 1944, Haavelmo was transferred to the Norwegian Legation in Washington, DC, with an assignment to work for the Royal Norwegian Purchasing Company, an agency directly under the Norwegian government in exile. The new assignment gave him the title of commercial secretary and diplomatic status. Haavelmo was not particularly happy about these frills of his new position or about the access it gave to the cocktail party circuit. He also knew that the transfer meant there was little chance of being released until half a year after the war had ended. In an interim period of a couple of months Haavelmo traveled repeatedly between New York and Washington.

An editorial note by the acting editor of Econometrica in the April 1944 issue announced that “[i]n view of the deficiency of material, it has been decided to publish, instead of the July issue, a special supplement containing an extensive monograph by Trygve Haavelmo, ‘The Probability Approach in Econometrics.’ A preliminary version of this was privately circulated by the author.”¹²⁹

The last discussion about the contents of “The Probability Approach” was another brief exchange with Chicago. Marschak was still studying Haavelmo (1941c) and had difficulties reassuring himself that the argument in Section 21, “An Illustration of the Problems of Estimation,” was “mathematically watertight.” He had been to New York in April and discussed this with Haavelmo. After that he had studied the chapter with Koopmans, who in the meantime had become a research associate at the Cowles Commission, and was now convinced that it was correct after all. But he found the presentation unfortunate, as in Haavelmo's example the least squares estimation of the reduced form gave the same results as maximum likelihood estimation, as the number of parameters in the original and in the reduced form was exactly the same. Marschak was afraid that the reader would assign great importance to the apparent conclusion and not notice that it was not valid generally.¹³⁰

Koopmans added on the same occasion that one of the footnotes seemed to imply that before Wald's work there were no established reasons to prefer the maximum likelihood estimates over other estimates.¹³¹

Haavelmo brushed off the idea of making changes at this late stage. Instead he pointed out that the exchanges between Marschak and himself almost made a whole little book by themselves. Thus instead of making some explanatory notes in Haavelmo's treatise, it would be better to write a separate monograph on regressions and least squares estimates, dealing in particular with

1. The general theory of least squares estimation and least squares prediction formulae.
2. The relation between maximum likelihood estimates and least squares estimates.
3. The questions regarding determinacy of coefficients.
4. The meaning of dependent and independent variables.

It read like a proposal to produce the famous Monograph 10 (Koopmans, 1950)!

Soon after, the special July issue of Econometrica with “The Probability Approach” appeared. At the end of July Haavelmo received a double set of reprints, one with an Econometrica cover and one with a Cowles Commission cover.¹³³

Later in the year, in December 1944, Haavelmo was notified that he had been elected a fellow of the Econometric Society.¹³⁴

8. INCIDENTAL DETAILS ON HAAVELMO'S STAY IN THE UNITED STATES