It is widely believed that learning-by-doing is a major source of economic growth, human capital, and comparative advantage (Arrow 1962, Lucas 1988, Romer 1990, Yang and Borland 1991). The extent of learning is also important for the understanding of labour markets and wage dynamics. If performance increases with experience, it supports a human-capital based interpretation of upward-sloping experience-wage profiles (Becker 1964). If learning is limited, such profiles would instead have to be explained by other theories, such as contract-based theories or matching models, with important policy implications (Lazear and Moore 1984). Learning-by-doing has also been of particular interest to health economists, as medical technologies often require substantial practice to master and since learning effects have important implications for productivity growth in the health care sector.
Despite its fundamental importance, documenting learning-by-doing at the individual level has proven challenging. While a large literature has documented empirical patterns consistent with learning, there are several challenges that complicate a causal interpretation of the results. In many contexts, there is non-random assignments of workers to tasks, where more experienced workers typically take on more challenging job tasks. Another challenge is that more productive workers are more likely to stay on the job, producing a non-causal relationship between experience and performance. On top of this, high-quality data on performance are often lacking and researchers have been forced to rely on measures such as unit costs, quantity, and wages (Thompson 2001). This also makes it difficult to disentangle the specific mechanisms behind learning and what type of skills that improve.
In our recent work, we overcome these challenges in the context of heart attack treatments in Sweden (Lundborg et al. 2021). The setting allows us to break the commonly observed sorting of more experienced workers to more difficult tasks by focusing on so-called PCI heart attack treatments performed during on-call shifts (nights, holidays, and weekends). During these shifts, only one physician is present and no systematic assignment of physicians to patients can take place. We use rich data between 2004-2013 on performance – measured through physician speediness, use of medical inputs, decision-making – and on patient outcomes, which we relate to the physicians’ experience. This facilitates a causal interpretation of the effect of experience on performance.
By focusing on complex heart attack treatments, we shed light on learning-by-doing in a high-skilled setting, where the opportunities for learning are large, as the task is non-trivial and non-standardized and involves a range of decisions to be taken under time pressure.
Our main findings
Our results show that learning-by-doing occurs continually over many years. In terms of proficiency, Panel (a) of Figure 1 shows that the physicians get 21% faster in performing the heart attack treatment between their first and 1,000th case. This is a substantial productivity improvement, corresponding to a three-minute reduction in the time to identify blockages in the arteries and to perform the medical procedures. Learning is fastest over the first 600 cases, slows down thereafter, and stops after 1,000 cases. We obtain similar results for other measures of proficiency, such as the adoption of more advanced technology that requires greater manual skills (Panel b).
Figure 1 Physician experience, proficiency, decision making, and patient health
The learning process for medical decision-making follows a similar pattern, where the invasiveness of the chosen medical procedures increases over, at least, the first 1,000 cases (Panels c and d). We also show that the more invasive, and more time-consuming, treatments by experienced physicians reflect more appropriate treatments of patients. The fact that more experienced physicians are more responsive to patient characteristics when taking their decisions also point in the direction of more appropriate treatments.
We find some evidence that the learning effects translate into effects on patient health, in terms of mortality or having another heart attack, but only among high-risk patients and only over the first 150 cases (see Figure 1, panels (e) and (f) and Figure 2, panels (a) and (b)). The risk of having another heart attack or dying within one year decreases by 40% for high-risk patients over the first 150 cases. This step learning curve adds to the discussion on the amount of training needed before physicians treat patients on their own. We obtain similar results for the risk of complications during treatment (see Figure 2, panels (c) and (d)).
Figure 2 Low level of experience, patient health, and patient risk
Our results highlight the difficulties associated with studying the impacts of experience on performance. We show that it is important to account for the potential non-random sorting of physicians to patients when estimating learning curves. That is, when studying experience, it is crucial to account for sorting of workers to tasks. In our setting, we observe a strong positive correlation between physician experience and predicted patient risk during daytime shifts, suggesting that hospitals assign more experienced physicians to more complicated cases. This correlation vanishes when we use data from on-call shifts, providing us with credible variation in physician assignment needed to identify learning-by-doing effects. We also address several other potential empirical concerns, for example, showing that experience is unrelated to the number of patients treated during night shifts, assuring that our estimates are not affected by selective referral of patients during the shifts, and that early-career patient outcomes are unrelated to whether a physician stays on the job in the future.
Through which mechanisms do workers learn?
Besides documenting learning, our results also give additional insights into the mechanisms behind the learning we observe. An attractive feature of our data is that we can study how learning differs across tasks that vary in complexity. Treating high-risk patients is arguably more difficult than treating low-risk patients and our results suggest that physicians indeed learn more from treating difficult cases.
Our data also allow us to examine whether the skills of physicians depreciate over time or if the knowledge ‘sticks’. We show that experience from more recent cases is more valuable than experience from more distant ones. This suggests that fine-tuned manual skills depreciate over time, whereas more intellectual skills stick.
Our results also highlight the role of peers in the learning process: learning rates are substantially higher for physicians who have worked with more experienced colleagues. This suggests that productivity growth is enhanced by placing inexperienced workers with experienced ones in occupations where tasks are non-standardised and learning curves are long.
Does productivity growth follow wage growth?
We can also relate our experience-related performance growth to wage growth and thus shed light on the common finding that wages increase with experience. Conventional human capital theory explains upward-sloping experience-wage profiles by the accumulation of human capital, partly acquired through learning-by-doing. To distinguish such an explanation from other ones, such as a deferred compensation mechanism, we relate wage profiles to observed learning patterns. If the human capital story is correct, we would expect a tight connection between wage and productivity profiles. But if wages increase faster than productivity, this would be inconsistent with human capital theory but consistent with the theory of deferred compensation as an incentive mechanism
We find that productivity growth keeps pace with wage growth over the first four years of the physicians’ careers. Productivity growth then fades out, while wage growth continues. This suggests that a human capital mechanism may explain upward-sloping experience-wage profiles in the beginning of the physicians’ careers, while other mechanisms better explain long-run wage growth.
Our paper provides new evidence on learning-by-doing and productivity growth in a high-skill task, by documenting the presence of prolonged learning curves in the treatment of heart attacks. The cardiologists we study keep learning over the first 1,000 cases performed, both in terms of proficiency and decision-making skills. These learning effects translate into important effects on patient health, but only over the first 150 cases performed, corresponding to one year of experience.
The long learning curves contrast with those found in previous studies that deal with potential non-random assignment of workers to tasks by focusing on standardised job tasks, where performance is easy to measure and where all workers perform more or less the same task (Shaw and Lazear 2008, Haggag et al. 2017). While these studies can rule out systematic sorting of workers to tasks, it comes at the price of studying standardised and less-skilled tasks, where the learning curves are typically short and steep, such as windshield installation and taxi driving. The contrasting results likely reflects that heart attack treatments constitute more complex tasks, with higher levels of worker discretion. This is also one of the unique contributions of our paper – to document learning-by-doing in a high-skilled occupation, using an empirical strategy that handles non-random sorting of workers to tasks.
We conclude that our results support the notion that learning-by-doing can be a powerful engine for productivity growth in high-skilled occupations. It offers some insight in the process of learning in advanced tasks, where both fine-tuned manual skills and fast decision-making are needed.
Arrow, K J (1962), “The Economic Implications of Learning by Doing”, Review of Economic Studies 29(3): 155–173
Becker, G (1964), Human Capital, Columbia University Press.
Haggag, K, B McManus and G Paci (2017), “Learning by Driving: Productivity Improvements by New York City Taxi Drivers”, American Economic Journal: Applied Economics 9(1): 70–95.
Lazear, E and R Moore (1984), “Incentives, Productivity, and Labor Contracts”, Quarterly Journal of Economics 99(2): 275–296.
Lucas, R E (1988), “On the Mechanics of Economic Development”, Journal of Monetary Economics 22(1): 3–42.
Lundborg P, S James, B Lagerqvist and J Vikström (2021), “Learning-by-Doing and Productivity Growth among High-Skilled Workers: Evidence from the Treatment of Heart Attacks”, IZA Discussion Paper No. 14744.
Romer, P (1990), “Endogenous technological change”, Journal of Political Economy 98(5): 71–102.
Shaw, K and E P Lazear (2008), “Tenure and output”, Labour Economics 15: 705–724.
Thompson, P (2001), “How Much Did the Liberty Shipbuilders Learn? New Evidence for an Old Case Study”, Journal of Political Economy 109(1): 103–137.
Yang, X and J Borland (1991), “A Microeconomic Mechanism for Economic Growth”, Journal of Political Economy 99(3): 460–82.