Africa

Why Economists Get It Wrong

Zed Books Ltd

Contents

Misunderstanding economic growth in Africa

For two decades, mainstream economists studying African economic growth have been trying to explain a chronic failure of growth – something that never occurred (Mkandawire and Soludo 1999). I will explain how that could happen. I will also show how that attempt affects our ability to understand economic growth in Africa today. The work of economists is always evolving over time due to changes in methodology, theory and technology (Morgan 2012). In the 1980s, some important innovations took place that had significant implications for the scholarly work of economists. The first step was a gradual tendency away from theorizing about economic growth toward conducting empirical tests of growth theory. As we will see, it is one thing to think about what the factors that determine economic growth might be, but it is quite another to use real-world data from all the countries of the globe in order to test whether the theory of what drives economic growth holds.

Theoretically, the basics of economic growth are straightforward and self-explanatory. The factors that drive economic growth – called the factors of production – are labor, land and capital. Together, these three factors produce goods and services. Economic growth is therefore a result of the increased use of the factors of production. Output increases when more people, more capital or more land is used for production.

Although this is called economic growth, it is a particular kind of economic growth known as extensive growth. In extensive growth, you only get more for more: that is, the increases in economic growth are proportional, or less than proportional, to increases in the use of resources. Intensive growth is what happens when growth is more than the proportional increases in the factors of production. Any growth that is not attributable to the increased use of resources is often referred to as 'total factor productivity' in the growth literature. The total factor productivity increase can be a result of improving the quality of the inputs. And this is often interpreted as technology – which has a broad definition and ranges from things such as smart gadgets to general changes in how production is organized to legal frameworks and structures in society at large (Jones 1988).

For an intuitive way of understanding why output may be increasing more than would be proportional to the increase in inputs, I use the example of moving a sofa. If you are one person trying to move your own sofa, productivity is very low. It is hard to get a grip on the sofa. If you get help from your friend (or you add one more unit of labor), total productivity more than doubles. If you add another friend, there is no increase in productivity; there may even be a decrease in output, as your second friend might distract you, order pizza or otherwise get in your way when you are trying to move the sofa. Productivity could also benefit from capital investment technology (such as using a strap, putting the sofa on a trolley, or using a forklift or a van). Such increases in output would require both capital and technology, and they would only be justified if you were in the business of moving many sofas regularly. Thus, there is more than one way of increasing output – some comes through adding more input/s and some through improved methods and technology – and the relationship between outputs and inputs is not linear, nor is it always rational to maximize output.

The bottom line is that economic growth goes far beyond mere increments of labor and capital. The sources and determinants of economic growth are not limited to basic processes of accumulation. How production is organized matters, and this extends to the very fabric of society – the rules that govern human behavior, or what economists have become accustomed to referring to as 'institutions' (see Bardhan 2005). We can understand this, but to go ahead and use data on labor, capital and GDP growth and then empirically measure it is quite another matter.

Initially, economists were quite happy to leave total factor productivity as an unexplained residual. In a framework called 'growth accounting', you enter increments in capital or labor as explanatory or independent variables in a system of equations that explain changes in economic growth. By doing this, you can, with a few assumptions, calculate how much of the economic growth is accounted for by labor and capital and how much of that growth is total factor productivity (Crafts 2002).

While early growth models did not seek to explain different rates of total factor productivity, this changed with a move toward models of economic growth in which total factor productivity was explained within the model, or endogenously. The turning point is best illustrated by what is called the Lucas paradox. In a famous journal article, Lucas posed the question: 'Why doesn't capital flow from rich to poor countries?' (Lucas 1990). The article presented the basic tenets of 'new' growth theory. According to the standard assumptions of neoclassical growth theory, India should have had 50 times more investment than it had received and it should have joined the richest economies in the world. Since this had not happened, Lucas suggested that maybe something was missing from the old exogenous growth models and proposed a new endogenous model. In the new model, variables that captured policies and other features of the economy were added to the basic factors of production as determinants of economic growth.

Thus, the theoretical growth literature went from focusing solely on rates of capital accumulation to emphasizing human capital and other country-specific characteristics that determine differences in country-level growth. This development has continued to the present; economists today think that growth is determined by institutions and historical events. We will see that this change in thinking about economic growth was not driven by theoretical considerations alone. The literature evolved through a process of empirical trial and error, technological and methodological innovation, and the availability of new datasets. In the 1980s and 1990s, more global datasets on economic growth, education levels and other variables became available, and, equipped with greater computing power and new models of economic growth to fit the new tools, the work of growth economists became concerned with running regressions and looking for correlations between datasets. Thus, we went from heroically making assumptions about the uniformity of economic activity and stability of returns across time to hoping that we could say something about the relative importance of increments in labor and capital in accounting for increases in output, and then to a much braver approach of hoping that we could say something about the causes of economic growth. Wording shifted from talking about accounting for growth and correlates of growth toward the more ambitious determinants of growth – and finally to the causes of growth. It is clear to any economists that this is a messy job, but just how messy is not communicated clearly. Before we proceed to lay bare this process of scholarly innovation to solve the African growth puzzle, we should go through some of the basic pitfalls of statistical testing – particularly that of looking at correlates in datasets.

Correlation is not causation

It would be difficult to think of a truism that is more often repeated and is so frequently violated in the social sciences. The contribution of mainstream economists to the study of poor economies and their politics is fundamentally based on observing country characteristics on the one hand and indicators that describe levels of development on the other. In a regression analysis, one is assumed to explain the other. This is the leap of faith from correlation to causation, and, for various reasons, mistakenly taking this leap has led economists astray in the elusive quest to explain growth in poor economies.

Since the 1990s, the bread and butter work of growth economists has consisted of running different datasets against each other in cross-country regressions and searching for a publishable result that proves or tests a hypothesis of economic development. There is nothing wrong with empirical testing. In fact, it should be encouraged. The problem arises when the testing is not guided by theory.

As an example, I tend to use the myth that storks bring babies. It is obvious to those of us who have an understanding of how babies are made that these little wonders do not make it into this world thanks to the ministrations of storks. This runs contrary to the explanation given to young children in folklore and in modern cartoons (such as Disney's movie Dumbo). So where does this hypothesis of storks bringing babies come from? The stork and baby myth is a very convenient explanation based on a deliberately misinterpreted correlation.

For the origin of the myth we need to go back to the time before electric lights and industrialized societies emerged in northern Europe. When it got darker and darker in the months following the equinox, and as the harvest had been completed and animals brought home from summer pastures, there was less work to do outside and more leisure time inside. People spent most of their leisure time in dark houses covered by blankets to keep warm. Enough said. For many, this transition meant that little babies emerged nine months later, about the time when spring came – and this coincided with the storks returning from southern countries. And there you have it: a misguided theory was born to explain how babies arrived.

The lesson of this example is that correlation is not causation. If you run the datasets of stork arrivals against baby births, there is a significant positive relationship. This abstract manner of conducting empirical research would indeed find a relationship between storks and babies. However, because such an investigation would not be grounded in a basic theoretical understanding of how babies come into this world, the researcher would still be a long way from unearthing the causal mechanisms. It is more than likely that he or she would draw the wrong type of policy advice from the research findings. Intriguingly, the arrival of babies and storks are linked. The researcher has found a real-world pattern but has failed to understand the relationship. Based on the investigation, the researcher might construct a model that suggests that, in order to reduce fertility in Denmark in the nineteenth century, policy makers should have focused their efforts on reducing the migration of storks. The scheme would have been expensive and misguided and, quite obviously, would not have had its intended effect. Conducting blind empirical tests by running different datasets against each other will lead researchers and policy makers astray.

Garbage in, garbage out

It is equally important to understand that the quality of the observations in the dataset matters. If you ask an economist about the evidence supporting their conclusions, they will direct you to the inferential statistical results and tell you about coefficients of determination, statistical significance and robustness tests. By contrast, if you ask a historian about evidence, he or she will respond by telling you about the quality of the primary observations.

This is a key difference. In my previous book on African development statistics, I pointed out a gap in our knowledge (Jerven 2013b). Since the arrival of global datasets on economic growth, poverty, politics and other variables, there has been surprisingly little questioning of the quality of the underlying observations. In some fields, such as the quantitative study of the correlates of war, there has been more probing of the quality of the different datasets (Cramer 2006); however, in the study of economics, there has been a tendency to take these datasets as 'facts'. The problem is that these data are not data in the literal meaning of the word – something that is given. Instead, they are often actively created and coded observations.

The standard defense of inferential statistics is that garbage in the dataset will bias the results toward zero. In statistics, one talks of this as false negatives and false positives. The easiest example is pregnancy. If you are declared not pregnant when you really are pregnant, that is a false negative. The false positive would be to state that you are pregnant when you are not. In the case of messy data, if one finds no relationship, it is often interpreted as finding that 'a' does not cause 'b'. Many zero results, which could be false negatives – such as finding that there is no relationship between development aid and economic growth – have been dominating the debates on economic development for more than a decade now (Easterly 2006). But that might be because data on both aid and growth are bad. The right answer is that we do not know because we cannot trust the data. So garbage data do matter, and they determine how we understand central questions in economic development.

Even more problems appear if there are systematic errors (as opposed to random errors) in the data. This happens all the time. In datasets with economic variables, there is often a tendency to understate production in certain sectors or to overstate poverty in particular areas. This problem becomes even more important with what are often called subjective datasets. These kinds of datasets are collections of opinions and responses to standardized questions. A classic example can be found in datasets on corruption. There are few direct observations and therefore very few systematic data on corruption (Aidt 2009), so instead the most commonly used datasets on corruption are a selection of opinion surveys: someone calls up businesspeople and asks them how corrupt country Y is on a scale from one to ten. Do you think that Swedes might understate the level of corruption in Sweden and that foreign businessmen might overstate corruption in Nigeria? This is how these subjective datasets are collected and it means that a large amount of literature built on correlations between corruption and growth is standing on very shaky foundations.

It is not just potential systematic errors that result from using subjective metrics in the datasets. It might also be that the datasets are not suitable for testing because the measures for corruption do not assess actual levels of corruption but rather they give an overall opinion of the efficiency of the economic and political system at a very specific point in time. That means that the endogeneity problem – or reversal causality – is built into the datasets from the outset.

Let us say that you conducted a survey of the quality of governance in Malawi in the fall of 2010. At that time, the country was experiencing several interruptions to its petroleum supply (and therefore also to the provision of electricity and the transportation of goods and people). This was because of temporary balance-of-payment problems. In a survey of visitors to Malawi that month, you would get results that ranked Malawi very poorly with regard to the quality of institutions compared with the evaluations you might collect in a month or year when there were no such interruptions. Yet the leadership and the set of policies in place during the alternative time frame may be the same. In the hands of the economist, this dataset enters the field of growth regressions, and what was really just a snapshot of a complex picture turns into a very specific causal story where 'poor economic governance' caused 'slow growth'. Meanwhile, we have learned nothing at all about the role played by the quality of institutions during this temporary shortfall in imports, about how the quality of institutions related to slow growth, or about what the actual physical and political causes of the shortfall were.

Errors attached to the numbers used in the economic growth literature are indeed often systematic, and thus we may have serious measurement bias, particularly when approaching policy questions in poor countries. Whether the end result is that the analysis finds no relationship or that it finds the wrong relationship, the outcome is unsatisfactory. Poor numbers will mislead us.

One of the most innovative areas of economic research is the one that comes up with 'proxies' to measure phenomena that are not easily quantified or observed. A particularly telling example comes from India, in a paper by Besley and Burgess (2000) investigating the effect of land reform on poverty. A priori, there are good reasons to think that redistributive land reform (taking property from the rich and giving it to the poor) might reduce poverty. But can it be measured? Poverty might be poorly measured, but with land reform it is even harder. A lot of land legislation was passed in India, but, as any purveyor of Indian rural agrarian history knows, those who had land were very resistant to distributing it. A good historian or a good ethnographer could answer this question by sifting through archives for documents proving that land shifted hands, or by using interviews to map land change over time. For the macroeconomist, this is unnecessary. The authors instead used the total number of land reforms undertaken as a proxy for land distribution and checked whether the number of land reforms in a state was correlated with poverty decline. They found some evidence of a link, but you would have to be willing to take a leap of faith to accept the results. Arguably, the authors provided a mathematically precise answer to the wrong question. We already think that there is a relationship between pro-poor land reform and reduced poverty. What we might have wanted to know is what kind of land reform actually causes land to change hands.