In-depth look at income and wealth data (pt. 2.5 of 3): A small note on wealth inequality from the archaeologist's perspective

I've been working on fellowship applications these past few weeks so naturally I began my research in a germane area of literature and ended up somewhere completely random. And by completely random I mean not even within the field of economics anymore and at best tangential to my original topic of investigation, but fascinating. I stumbled upon Ten Thousand Years of Inequality: The Archaeology of Wealth Differences, a volume on the archaeological studies of wealth inequality, and given its relevance to the posts I've been writing on inequality I figured I would make a small note on how inequality is being measured for a society that lived nearly two thousand years ago.

I haven't written a formal post on the Gini coefficient but given that this article uses it extensively in the archaeological context I preface by stating a few things: (1) the Gini coefficient is notably a simple measure of inequalities (most commonly income inequality) therefore it has its limitations that are well summarized in this Wikipedia post; (2) it has also been subject to revision and extensive debate as well as the creation of alternative measures of inequality including the Atkinson Index that may be more informative if certain contexts; (3) given my lack of knowledge of archaeology (and my lack of knowledge more broadly on the range of applications that the Gini coefficient has had in diverse fields within social science) I don't assess whether or how the Gini coefficient was applied and rather introduce it as a thought-provoking application outside of the realm of economics.

Feinman, Faulseit, and Nicholas (2018) provide estimates of wealth inequality for the Classic period in the history of the pre-Hispanic Valley of Oaxaca, Mexico based on archaeological house excavations at six pre-Hispanic settlements. They rely principally on architectural constructions and space to proxy for wealth and apply the Gini coefficient to three architectural variables: terrace area, house size, and patio area. They also utilize distribution of artifacts such as obsidian and other rare items.

The Lorenz curve in the figure below shows the Gini coefficient constructed based on the house sizes for all of the houses in the sample (a total of 36 excavated houses across all six sites in the Valley of Oaxaca) with a coefficient of 0.35 and a 95 percent confidence interval between 0.31 and 0.39. A Gini coefficient of 0 represents perfect equality whereas 1 represents perfect inequality.

The Gini coefficients from all samples (houses, patios, and terraces) and excavation sites are indicated in the figure below. They range from 0.35 to 0.43.

The authors find based on their analysis that wealth inequality during this time was low compared to other urbanized and preindustrial settings (confirming extant evidence).

While there was notable variation between the periods that the authors link to changes in the socio-political structures of the time, they specifically note that "[t]he consistently low Gini values... are informative, especially as indicators of wealth inequality, because they challenge the long-term notion that archaic states were always starkly divisible into the rulers and the ruled, with dramatic differences in resources and quality of life between the two. This coercive/despotic vantage on archaic states is well ensconced in the historical/social sciences for preindustrial times (e.g. Mann 1977; Wittfogel 1957) but now is being challenged as not uniformly applicable (Blanton 2016; Blanton and Fargher 2008), with some historical polities seen as having had a more collective institutional orientations and lower degrees of wealth inequity (e.g. Mann 2016, for a change from his earlier perspective)."

A few comments and questions came to mind about the application of Gini coefficient in this context:

1. The sets of data used in these analyses of the Classic period in particular are considered by the authors to be large and representative (likely given the difficulties involved in excavations and the number of houses, patios, and terraces that are still intact after thousands of years) but they are still subject to the well-described small sample size bias associated with the Gini coefficient. Smaller samples are biased towards having smaller Gini coefficients which may make it difficult to compare excavated sites in the graphic below (from the Smithsonian Magazine article about this book) with the United States today, which is not based on excavated evidence and has a much, much larger sample size. 
2. Comparisons across past civilizations, though, if they are based on similar sample sizes may be more informative than comparisons between past civilizations that were excavated and modern day societies. Similarly, I would find variation within a given civilization over time (as evidenced in the article) to be informative especially in conjunction with changes in socio-political structures. This is hinted at by the authors when they quote from Piketty (2015) on the impacts of these structures on inequality: "[O]ne should be wary of any economic determinism in regard to inequalities of wealth and income... The history of the distribution of wealth has always been deeply political... How this history plays out depends on how societies view inequalities and what kinds of policies and institutions they adopt." 

In-depth look at income and wealth data (pt. 2 of 3): Wealth

First, to preface with why wealth as distinct from income is relevant to economists and to policymakers at large. Kopczuk (2014) discusses the importance of understanding the wealth distribution: "the extent to which the well-off are going to rely on work vs. return to their wealth in the future is clearly important for assessing the extent to which a society will view itself in some way a meritocracy." Wealth is an important determinant of labor force participation and therefore impacts productivity and economic growth. It also has important implications for inequality, intergenerational mobility, and, consequently, implications for democratic institutions whose stability is reliant on a meritocratic society or at least the verisimilitude of a meritocratic society.

It should be noted that estimates of wealth inequality and the top wealth shares are not as widely agreed upon as estimates of income inequality and labor income shares. There are a few main data sources for estimating wealth inequality that are aptly summarized in Alvaredo, Atkinson, and Morelli (2018):

1. Household surveys including the U.K. Wealth and Assets Survey and the U.S. Survey of Consumer Finances;
2. Administrative data on individual estates at death; 
3. Administrative data on wealth of living from annual wealth taxes; 
4. Administrative data on investment income that are capitalized; and 
5. Lists of large wealth-holders (e.g. Forbes).

These data sources are discussed in great detail in Kopczuk (2014)'s "What Do We Know About the Evolution of Top Wealth Shares in the United States?" which specifically discusses the U.S. Survey of Consumer Finances (1), the mortality multiplier method with individual estate data (2), and investment income data (4). Each of these data sources is subject to different concerns. Household surveys and list of the wealthiest individuals are recent phenomena and cannot be used for estimates prior to the 1950s when the household surveys on wealth were first implemented. Administrative data on wealth of the living based on wealth taxes cannot be recouped in most developed countries because only a few developed countries, most notably France and Norway, have a wealth tax to begin with. Therefore, most researchers rely on estate tax records on individual estates at death or on reported taxable capital income.

The primary concern with estate taxes is that the distribution of estates of the deceased must be projected to the population at large: i.e. a multiplier method must be used in order to answer the question, how does the distribution of wealth among the deceased reflect the distribution of wealth among the living? Mortality multipliers are inverses of mortality rates based on various criteria, for example, wealthy individuals tend to have lower mortality rates and increased longevity compared to less wealthy individuals and therefore a higher mortality multiplier would be applied to the upper estate ranges meaning there are relatively more individuals living within those ranges than lower ones. For more on recent discussions of the relative longevity of the wealthy see Saez and Zucman (2016) and Chetty et al. (2016).

Kopczuk presents a few interesting stylized facts about wealth that provide a good introduction to the wealth distribution and methods of estimating it:

• Wealth is highly concentrated (top 10 percent holds between 65 and 85 percent of the total wealth, top 1 percent holds between 20 and 45 percent of total wealth based on time period); 
• While the methods of estimating the wealth distribution disagree on the timing it is clear that wealth concentration hit its apex prior to the Great Depression and declined after that; 
• Different methods lead to varying estimates for the top 1% for several reasons: one is that the estate tax multiplier method uses the individual as the unit of observation, surveys use the household, and the capitalization method uses tax units; another is that tax evasion impacts the administrative tax-based methods (estate tax and capitalization) but not the survey-based methods. Some capture debt (estate tax returns) whereas others do not (capitalization). 

In a recent issue of the Journal of Public Economics commemorating Tony Atkinson's work, Alvaredo, Atkinson, and Morelli (2018) provides new evidence on the evolution of top wealth shares in the U.K. To choose one of the most interesting facets of the discussion of wealth that they present in the article, it is enlightening to view the top wealth shares compared to the wealth shares excluding housing.



The top 1%'s total wealth share and wealth share excluding housing tracked each other for much of the late 20th century but the authors note the divergence between the two trends in the 21st century, wherein the share of the top 1% of wealth holders of total wealth increased much more rapidly than its share of wealth excluding housing. In other words, the growth of wealth excluding housing is likely to be a more significant contributor to rising inequality than is the growth of housing wealth. In fact, they even mention that increases in housing prices serve an equalizing effect for the top 1%:

"It appears that housing wealth has moderated a definite tendency for there to be a rise in recent years in top shares in total wealth apart from housing. When people talk about rising wealth concentration in the U.K., then it is probably the latter that they have in mind... The results show how the impact of a general rise in house prices has changed over the period but it is always equalizing for the top 1%. At the beginning of the period a rise of 25% led to a reduction of some 1 percentage point in the share of the top 1% but the effect became smaller over time."

It should be noted, however, that trends in the housing market - particularly the resurgence in the private landlord and "buy to let" over the past three decades - likely have impacts on other areas of the wealth distribution apart from the top 1% of wealth owners (though these impacts are not addressed in this paper). This New York Times article from last year, for example, is a news feature that discusses the role that homeownership plays in propagating existing wealth and income inequalities. These topics and the lower rungs of the wealth distribution more broadly are areas for further investigation, but for the time being Alvaredo, Atkinson, and Morelli (2018) highlight how granularity in wealth data can be used to better identify the causes of growing wealth inequality over the past few decades and, while they utilize estate data and the mortality multiplier method in their analysis, can also be triangulated with other methods and data sources to form a more comprehensive understanding of the wealth distribution.

Sources

1. Alvaredo, F., Atkinson, A., Morelli, S. (2018). Top wealth shares in the UK over more than a century. Journal of Public Economics.
2. Kopczuk, W. (2014). What do we know about the evolution of top wealth shares in the United States? NBER Working Paper 20734.
3. Chetty, R., Stepner, M., Abraham, S., Lin, S., Scuderi, B., Turner, N., Bergeron, A., Cutler, D. (2016) The association between income and life expectancy in the United States 2001-2014. Journal of American Medical Association.
4. Saez, E., Zucman, G. (2016) The distribution of US wealth, capital income, and returns since 1913. Quarterly Journal of Economics. 

In-depth look at income and wealth data (pt. 1.5 of 3): Non-traditional data and machine learning approaches

While this was originally meant to be a three-part series on income and wealth data, it would have been an oversight to not include some discussion of the non-traditional data and machine learning approaches to collecting information on poverty. These data are particularly relevant in developing countries where traditional sources of data - administrative data and survey data - are not collected as widely, regularly, or thoroughly. This can be for several reasons: nationally representative surveys are expensive and the costs of data collection too high, challenges associated with data collection in conflict-affected areas (discussed in greater detail in a previous publication I worked on), and large proportions of the population are employed in the informal economy meaning there is little by way of administrative tax records at the lower end of the income distribution.

Yet, information on poverty is still needed in these countries to inform evidence-based policymaking by governments, international organizations, and non-profits. A brief article by researcher Joshua Blumenstock published a few years ago in Science, "Fighting Poverty with Data", discusses the frontier of research in this area that aims to supplement the traditional sources of data on wealth and inequality with machine learning approaches. Blumenstock discusses, for example, the rise in use of nightlight data to track economic productivity and growth citing one paper which utilizes nightlight based measures to study the impact of sanctions on North Korea. In fact, a paper that I reviewed earlier in the year on the impact of Chinese aid projects on local corruption used nightlight data to proxy for local economic activity in areas around active and inactive Chinese aid sites.

More novel and more interestingly, the author cites research in machine learning that uses satellite imagery in conjunction with nightlight data to identify the visual features of relatively wealthier areas (which have brighter nightlight) that would allow researchers to leverage daytime satellite images to better track poverty in developing countries. There are limitations to this approach for example that nightlight is not an ideal measure of activity at the lower end of the income distribution - where all is dark - but with further research these approaches could be very useful in the developing country context.

Mobile phone data - which was discussed in part in the above article and in greater detail in this other Science piece also by Blumenstock - is also promising. Using mobile phone logs, researchers extract statistics including volume, intensity, and timing of phone calls, the structure of the individual's network of contacts, and mobility and migration information based on geospatial markers and whittle down to the statistics that can be used to predict socioeconomic status. In the case cited in this article, the researchers paired consenting individuals' mobile phone data with survey data that they collected on individual income and wealth in order to train the model. It should be noted that mobile phone data is subject to greater ethical and privacy concerns than publicly available data. While the research cited here aimed to obtain macro level statistics to inform policymaking it is clear that attempting to obtain a more granular understanding for specific demographics will be challenging. ICT access and use is far from universal and, often, those who are excluded from its access are the most vulnerable. This is similar to the challenges with using conflict data wherein the data on those who are the most vulnerable and impacted by conflict is the data that is the most challenging to collect and to collect accurately. This is not, however, meant to generalize, given that some of the poorest regions of the world have reasonably high mobile phone penetration but rather a cautionary note when assessing whether data are representative with respect to specific populations.

For example, with respect to a recent project that I've worked on, there is high mobile phone penetration in sub-Saharan Africa despite low income. Yet, while its neighbors in East Africa have experienced fast growth in mobile phone ownership and usage in the past five years, Ethiopia has fallen behind largely due to government ownership of the nation's telecom monopoly which has limited expansion and service. Further, analyzing the distributional data on mobile phone usage indicates that women are far less likely to own and use mobile phones than men - consistent with the findings in many developing countries - and that any data collected from these devices in a hypothetical scenario would only be representative of a specific demographic.

And yet, despite the challenges, non-traditional sources of data offer promise particularly in geographic areas where recent, traditional data on wealth and poverty are not available. Research in this interdisciplinary area will be interesting to watch in the near future.

In-depth look at income and wealth data (pt. 1 of 3): Background

For some time now I have been interested in writing an in-depth post on income and wealth data in order to discuss how the study of inequality - in conjunction with the data and methods that enable this study - has progressed over time. While this was initially intended to be a single post, it quickly became evident that there was too much to discuss within too short a space. In this first post of a three-part series, then, I focus on providing the background for a more granular discussion of wealth and income in the next two parts.

Given the topic at hand, it is noteworthy that several articles on inequality and tax and redistribution policy were published in a recent special issue of the Journal of Public Economics honoring the late Tony Atkinson. For an introduction to that series of papers see here. My previous post on individual and household level inequality is based on a paper within this special issue. Additionally, a recent issue of the Quarterly Journal of Economics features an article that combines national accounts data with micro data to produce estimates of inequality in the U.S. that are consistent at the macro level.

In light of expanding research on inequality, its growing presence in policy debates in developed countries, and the evolution of both data and methods that enable its rigorous study, it is useful to take stock of the existing data sets and methods used by researchers to answer some of the most pressing questions in public economics today: those that deal with the distribution of wealth and income in our societies and the reasons for widening or stagnant inequality levels. We can also assess what types of questions we are now able to answer and how our answers to these and other - yet unasked - questions can become more accurate through improved data collection and methods and how future data collection can fill existing gaps in our knowledge.

To begin, the World Inequality Database - a database of global wealth and income inequality data co-founded by Tony Atkinson - provides a concise description of data and research in this field over the past twenty years. Two important trends:

1. Most studies on inequality have until very recently focused on income rather than wealth. The key reason is the greater availability of micro data to study income, which is taxed and therefore observable in administrative data, as opposed to wealth, which in most developed countries is not taxed apart from an estate tax upon death. A secondary reason is that it has not been made evident until recently - likely for similar data reasons - that wealth concentration plays a large role in the inequality we see within developed countries. Piketty (2014)'s Capital in the Twenty-First Century was not the first but perhaps the most prominent description of the growing role of capital in widening divisions between haves and have-nots.
2. Current efforts are aimed at producing distributed national accounts that combine administrative micro data with national accounts macro data - ledgers of assets and liabilities at the national level - in order to reconcile inequality estimates that are created based on micro data with the national accounting. This publication from the founders of the WID discusses the motivation and methodology for the creation of these "distributed national accounts." It notes the historical background, "[by] combining the macro and micro dimensions of economic measurement, we are of course following a very long tradition. In particular, it is worth recalling that Kuznets was both of the founders of the U.S. national accounts and the author of the first national income series and also the first scholar to combine national income series and income tax data in order to estimate the evolution of the share of total income going to top fractiles in the U.S. over the 1913-1948 period (see Kuznets, 1953)." The article cited above from the QJE, Piketty, Saez, and Zucman (2018), presents "distributed national accounts" for the U.S., which they note is distinct from government statistical agencies' work in this area.

Discussion of the main data types and their roles in inequality studies

Administrative micro data

To preface a discussion on administrative tax data for wealth and income studies, I provide context for use of this data for social sciences research more broadly. Administrative data are collected for the purposes of registration, transaction and record keeping, and are often linked to public service delivery. They are typically collected by public sector agencies and can be used in administration systems in education, health, and taxation, among other departments of the public sector. It should be noted that these data are "found" data and are not collected for the purposes of research as survey data are. The social sciences, and economics in particular, have shifted to using administrative data over survey data sources in recent years for several reasons.

Specifically as noted in this white paper to the National Science Foundation: "Administrative data are highly preferable to survey data along three key dimensions. First, since full population files are generally available, administrative records offer much larger sample sizes... Second, administrative files have an inherent longitudinal structure that enables researchers to follow individuals over time and address many critical policy questions, such as the long term effects of job loss (von Wachter, Song, and Manchester, 2009) or the degree of earnings mobility over the life cycle (Kopczuk, Saez, and Song, 2010). Third administrative data provide much higher quality information than is typically available for survey sources, which suffer from high and rising rates of non-response, attrition, and under-reporting."

Access to this data is not without its challenges in many developed countries. Nordic countries have been leaders in enabling researchers to access de-identified administrative or "register" data but other countries, such as the U.S., have been relatively slow to follow. Given the central role that administrative data has come to play in social sciences and economics research in particular (see the two charts on the number of publications in leading economics journals that employed administrative data in this presentation from researcher Raj Chetty, who also co-authored the white paper cited above), it is clear that access to these data has important implications that are outlined in an article published in the Economist last month on the topic.

Administrative tax data are widely used in income and wealth inequality studies. For example, wealth inequality is largely studied through either estate tax records - in order to create wealth distributions of wealth at death and to extrapolate from those records the distribution of wealth among the living using the mortality multiplier method - or through taxable capital income (it should be noted that only one-third of total capital income is reported on tax returns which is why it is challenging to estimate wealth based on this quantity). Similarly, income inequality is studied through income tax records. Given the socioeconomic and demographic data contained in these records we are able to answer (or attempt to answer) a wide range of social science research questions based on micro data. Yet, the missing piece is information on movements in the economy at large over time (e.g. increase in fraction of retired individuals or declines in household size) which could have implications for inequality.

As noted by Piketty, Saez, and Zucman (2018), studies that use micro data exclusively are unable to answer questions such as: (1) what fraction of economic growth accrues to different parts of the income distribution, (2) what fraction of the increase in income inequality is due to changes in share of labor vs. capital in national income as opposed to changes in the distribution within labor or capital earnings, (3) how does government redistribution impact inequality (i.e. we are only able to observe pretax income using micro data series which does not allow us to observe the changes in the income distribution between pre- and posttax). To answer these questions, they argue, merging micro data with national accounts data at the macro level is valuable.

National accounts macro data

On the macro side side, national accounts data aggregate output, expenditure, and income activities of each sector of the economy. While income and consumption measures are important for evaluating standards of living they offer only a static picture of well-being. Specifically, income and consumption reflect current well-being: how much a household or an economy is producing and consuming at present, but they do not provide much insight into a household or economy's long-term or future well-being (beyond making assumptions that current well-being and consumption are highly correlated over time). This is where national accounts data can be useful: data on a household or economy's ownership of marketable assets and contraction of debts can provide insight into long-term or future well-being though it may be cross-sectional rather than longitudinal.

For a valuable introduction to balance sheets and the national accounts data see here for a discussion from the French National Institute of Statistics and Economic Studies (INSEE). It should be noted that the definitions of "assets" and whether or not they provide "economic advantages" refer specifically to those items that have market values. This would exclude, as stated by INSEE, "items that one might expect to see in the accounts (human capital, natural heritage, natural State property, household durables, pension entitlements linked to the allocation system, etc.)" They note as a rule of thumb that only items that are featured in the capital and financial accounts are included as assets in order to maintain internal consistency. The capital account and financial account link the opening and closing balance sheets to one another: they specify what happened to the accumulation of capital based on capital consumption, assets sold and acquired, discoveries and inventions, and nominal holding gains as a result of price fluctuations.

These data, and specifically the national income measures in these data, may be relied upon to fill the gaps in our knowledge from the tax data. Specifically, there are gaps between the reported income and the national income that are not captured in micro studies: imputed rents of homeowners and taxes on top of unreported and untaxed labor income in the form of tax-exempt fringe benefit. Piketty, Saez, and Zucman (2018) estimate that the fraction of national income reported on tax returns in the U.S. has declined from 70 percent in the late 1970s to roughly 60 percent today which indicates that micro data alone may underestimate the level and growth of income in this country and perhaps more so for certain parts of the income distribution than others depending on what exactly is being excluded from the tax data that is present in the national income.

For a more in-depth description of the methods and the process by which these two data are being combined, I would look to the article. The authors effectively illustrate both the motivation and the methods for incorporating national income macro data into inequality studies. In the next part of this three-part series, I will discuss the data and research on wealth inequality specifically to provide greater detail on wealth estimates using estate tax data compared to those using capital income.

Sources

1. Piketty, T., Saez, E., Zucman, G. (2018). Distributional National Accounts: Methods and Estimates for the United States. Quarterly Journal of Economics.
2. Kleven, H., Luttmer, E. (2018). A Special Issue of the Journal of Public Economics: Honoring the Work of Sir Anthony B. Atkinson (1944-2017). Journal of Public Economics. 
3. Blundell, R., Joyce, R., Keiller, A.N., Ziliak, J.P. (2017). Income inequality and the labour market in Britain and the US. Journal of Public Economics. 

Selected articles on AI and job displacement

Given my hiatus from posting and recent time constraints (multidimensional vector spaces occupy most of my time now that summer session has started), I'm discussing here a few interesting papers rather than providing an in-depth review of a single topic or piece of literature as I usually do. But worry not, I will be back to discuss the econometric details in another post soon enough.

And, a thank you to Intelligent Economist for mentioning this blog in his "Top 100 Economics Blogs of 2018" and a thank you to everyone who is visiting as a result of that post. I've tried to add some value with this blog (both for myself and for my readers) and I hope that you've found it valuable. Thanks very much for reading.

Automation and labor
In an earlier post from January, I discussed Acemoglu and Restrepo (2017)'s paper that models the relationship between automation and labor force displacement.

In the earlier post I identified key policy items to focus on in a future defined by automation:

1. Identify market failures that contribute to "excessive" automation or the adoption of technologies that are only marginally more cost effective than labor and lead to little productivity gain or job creation
2. Determine whether and to what degree jobs will be created at all in the process of automation if the adoption of new technologies leads to marginal but limited productivity gains
3. Address the inequality implications inherent in the displacement of jobs that require particular skill sets and the creation of jobs that require another
4. Identify the type of jobs that are created and the quality of those jobs
5. Prepare the labor market for "new skills" and a culture of lifelong learning

A recent paper from Jason Furman and Robert Seamans discusses a lot of these key items and more tangible policy proposals, including universal basic income and guaranteed employment, that would address the labor market implications of a future that will come to depend heavily on artificial intelligence. One challenge associated with automation that they mention in the paper is the decline in the male labor force participation rate, which I discussed in my previous post on rising inequality in male labor market outcomes.

The decline in male labor force participation is a signal that, at least in part, existing policies have had limited positive impact on (3) addressing the inequality implications inherent in job displacement and creation - the disproportionate impact on low-skilled labor - and (5) preparing the labor market for "new skills" and a culture of lifelong learning - failure to re-integrate workers that have been displaced by the system into new jobs requiring new skill sets. The paper highlights that addressing labor market transitions for individuals who have been displaced from their jobs is more challenging than it appears. For a discussion on skills in the context of automation see this new report from McKinsey Global Institute.

The paper also discusses non-labor related policy issues with rising automation including a need for new approaches to antitrust regulation. In particular, they draw attention to the fact that large datasets can serve as a barrier to entry in the AI field. I mentioned the role of big data in competition in an earlier post in the context of Amazon's edge in entering the grocery market: Amazon's access to high quality data on consumer preferences through its dominance of e-commerce retail is non-negligible given that its competitors in the grocery store market that it entered will have much less of that type of data. Even more so in the case of AI, data could serve as a crucial factor for entrants meaning there need to be novel ways of thinking about competition (or lack thereof) in these markets due to this new barrier to entry. It is also interesting to think about how institutions and laws such as the recent European Union General Data Protection Regulation can play a role in this area by limiting data retention.

Another point of further reading is the European Commission's "Analysis of the impact of robotic systems on employment in the EU". It adds value to existing literature because it is one of the first studies to use firm-level data to assess the impact of robotics on productivity (finding a significant and positive effect on labor productivity but not identifying an effect on employment levels which is an interesting finding that will have to delve into further in another post). The common alternative - using macro level data on productivity - has a more limited scope in terms of understanding what happens at a granular level. This is perhaps not as relevant for isolating a causal impact as for using descriptive statistics to explore the topic in greater detail and to ask more refined questions about automation's effects on firm behavior. 

A third paper in the current issue of Labour Economics also merits a mention given that it adds another layer of complexity to the relationship between automation and job displacement. Lordan and Neumark (2018) find that increases in the minimum wage lead to significant decreases in automatable employment held by low-skilled workers and that, while there is significant heterogeneity across industries and demographics, well-intentioned minimum wage laws interact with rising automation to have adverse impacts on a vulnerable population. One important question for this and economic research on AI more broadly: to what extent can new technologies be grouped together to analyze the impact of their adoption on the labor force? This paper relies on the U.S. Consumer Population Survey from 1980-2015, during which time a range of new technologies were adopted with potentially different implications and benefits of adoption for firms.

One of the key takeaways from Acemoglu and Restrepo (2017) is that new technologies have varying effects on firm productivity and the creation of new jobs. They discuss the "so-so technologies" that are only marginally more cost-effective and lead to little job creation. This implies that nuances in the type of technology and to what extent they increase firm productivity are extremely important. The nuances are, however, more important in determining job creation rather than job displacement based on their model.

A new report debunking myths on agriculture in Africa

The World Bank came out with a notable publication discussing common myths and truths on agriculture in Africa for policymakers and practitioners. The publication discusses the common myths in the table below and identifies whether or not they are true based on detailed data from the World Bank's Living Standards Measurement Survey.

It is notable both as a primer for researchers to get a more accurate, big picture sense of small holder agriculture and because its format effectively marries the technical details with concise policy takeaways without eliminating the relevant nuances across countries and settings.

Trends in household income inequality: comparing the U.S. and Britain on labor, marriage, and government transfers

A recent paper highlighted by the Institute for Fiscal Studies and forthcoming in Journal of Public Economics presents an intriguing look at the relationships between individual labor market outcomes, household composition/spousal labor market outcomes, government tax and transfer systems, and household income inequality in the U.S. vs. Britain over the past four decades.

Blundell et al. (2017)'s descriptive analysis compares individual labor market outcomes in the two countries' by education level, income level, and gender and compares spousal labor market outcomes and government tax and transfer systems to provide a comprehensive look at the components of household income inequality. Their analysis enables us to connect each country's experience of/response to key shared events - including the rise in female labor force participation, the decline in low-skilled labor, and the 2007-08 financial crisis and recession - to household income inequality.

Many of the findings presented confirm existing ideas about the interaction between the tax and transfer system and labor market outcomes. The paper adds value in its use of micro data through 2015 and its use of data that has been standardized to facilitate comparison between the two countries. It adds to a literature on the role of the welfare state in exacerbating or alleviating individual-level labor market inequalities. This literature, and the inequality literature more broadly, formed the crux of the late economist Tony Atkinson's life's work and his numerous contributions in the field built the foundation for modern studies of inequality.

In fact, in his work on redistributive preferences and the welfare state, Atkinson (2000) discusses the responses of various countries to the universal shift in labor markets in industrialized countries away from low-skilled labor: "We are concerned not only with policy before and after a shift in the external circumstances, but also with how different societies respond to the same shift. It is striking that a number of OECD countries have in common a rise in the inequality of market incomes (incomes from earnings and investments) between 1980 and the mid-1990s, but that the outcomes in terms of disposable outcomes (after direct taxes and social transfers) differed."

So as not to diminish the rich nuances in Blundell et al. (2017), I only mention that one of the paper's findings is exactly this: inequalities in market incomes between the two countries are very similar but this is not the case for the disposable incomes at the household level. I discuss the findings below.

Inequality in male labor market outcomes has increased significantly in both countries

• The two countries' experiences in the Great Recession were different: in the U.S., real incomes for the most part kept pace with inflation whereas Britain experienced a sharp drop in real incomes, particularly for the top income percentiles and the educated. Adjustment in the U.S. came in the form of declines in employment rather than in real wages whereas in Britain employment on both intensive and extensive margins was relatively robust. 
• I suspect there are a couple reasons for the disparity:

1. High subsidization of capital vs. labor may bias the U.S. economy towards lower but more productive employment levels (due to higher investments in capital).
2. Stricter labor market regulation in Britain may imply that adjustments to shocks take the form of real wage declines rather than layoffs and declines in employment.
3. Differences in the composition of employment between the two countries post-recession (numbers of full-time, part-time, and self-employed workers) may impact the aggregate figures on wage growth and hours worked.

• This discussion is connected to the authors' second finding: while both countries have experienced a significant rise in male income inequality in the past four decades, in the U.S. this increase is largely driven by male hourly wage inequality whereas in Britain it is driven by fewer hours worked by men at the bottom of the distribution. Therefore while employment was relatively robust in the Britain it is because any response to the recession was part of a longer term trend in decreasing hours for male low-skilled workers on the intensive margin.
• This is clearly illustrated by the green dotted lines in the two graphs below on hours worked for G.B. Men who left education at or below 16 years of age and U.S. Men with less than high school education. The line on the left for G.B. indicates a steep downward trend beginning 1995.

• Finally, the authors provide evidence of the wage stagnation in the U.S. continues to make headlines. They state that the only group of male workers that has a higher median real wage today compared to 1979 is those with a college education (compared to those without high school, those with high school, and those with some college who have not seen any improvement to their real wages in the past four decades). 

Lower marriage rates among the bottom half of the income distribution indicate inequalities in household composition and spousal income

Assortativeness of marriage - the tendency for people to marry others who are found in roughly the same area of the wage distribution - is a trend that has only increased in the U.S. in the past twenty years (in Britain it has remained constant). A second, commonly discussed finding is the decline in marriage rates - among the entire wage distribution but more sharply among the lower half of the income distribution including low-skilled and unemployed men. Together the findings indicate that, rather than alleviate male earnings inequalities, the marriage market has likely amplified those inequalities.

The tax and transfer system in Britain has done a much better job of ensuring that the inequality in male labor market outcomes has not translated into large household income inequalities

The following image presented by the authors is the most striking:  while male earnings inequality (red dotted line) has increased steadily in both countries and perhaps even more sharply in Britain, household net income inequality (labor earnings plus government transfers minus taxes) in Britain has not grown in the past twenty years. This is not the case in the U.S. indicating that tax and transfer systems in Britain have done a much better job at ensuring that disparities in male labor market outcomes have not translated into as large disparities in net household income. 

The authors identify the following when discussing the disparity in tax and transfer outcomes between the U.S. and Britain:

• Much more generous social welfare programs in Britain vs. the U.S. in particular due to successive Labour governments from 1997-2010. 
• The one and very important exception being the recessionary period: in the U.S., average transfer generosity increased greatly in response to the recession and were in place through the six-year recessionary period whereas in Britain, fiscal consolidation policies beginning in 2011 indicated a reduction in social programs. 
• Welfare policy in Britain that does not link transfers to work status indicating net income growth of non-workers whereas this is not the case in the U.S. where the generosity of welfare for non-working families declined greatly in the past two decades.

Follow-up questions 

The study raises several further points of research/questions to be answered by existing research - 

What are the differences in structural factors that led the decline in low-skilled labor to manifest itself as a decline in hours worked in Britain vs. a stagnation of real wages in the U.S.? In Britain, real wages experienced a sharp decline only during the recession and prior to the recession were even on the uptick for most men. Yet their hours of work had been declining for decades. To research this further we would need to look at the labor force participation rates of low-skilled men to determine whether the U.S. experienced similar decreases in employment (though on the extensive rather than intensive margin) that are masked by lower labor force participation rates among low-skilled men.

Through what channels did monetary and fiscal policy in the U.S. in the recessionary period contribute to the stabilization of real wages? The decline in real wages in the recession and post-recession Britain has been attributed to a number of factors - high inflation due to high energy prices, expansion of lower-paid, self-employment or part-time jobs rather than full-time jobs, limited investment in capital and as a result low levels of productivity - a number of which should also be issues in the U.S.

Though the study doesn't delve deeply into female labor market outcomes it paints an interesting picture of stability in women's employment and wages over the past forty years and particularly during the recession. This is likely due to higher relative attrition of women from the labor force at times when jobs are hard to be found given that men remain the main earners in most households, but again we would need to see the labor force participation rates to be sure. 

Sources

1. Blundell, R., Joyce, R., Keiller, A.N., Ziliak, J.P. (2017). Income inequality and the labour market in Britain and the US. Journal of Public Economics. 
2. Atkinson, A. (2000). The welfare state, budgetary pressure and labour market shifts. Scandinavian Journal of Economics.
3. Atkinson, A. (1992). Towards a European social safety net. Fiscal Studies.

"Intelligent evolution of humanity"

First of all, I'm sorry for the long hiatus in posts: midterms, a bout of the flu, and starting a new work project have all taken away from the time that I normally spend on this blog.

Given that the project that I've started working on focuses on the gender gap in agricultural productivity levels in sub-Saharan Africa, I visited a number of seminal papers in agricultural economics and in the study of firm and individual-level productivity and efficiency. In this survey of the literature, I stumbled on Theodore Schultz's Nobel Prize lecture, "The Economics of Being Poor", after having read his work on smallholder farmer efficiency ("poor-but-efficient" hypothesis which states that farmers are calculating economic agents who are highly efficient in a traditional agricultural environment (Schultz, 1964)) and his consequent focus on human capital and the gains to labor productivity and entrepreneurial ability as the key to improving the well-being of the agricultural population.

In his lecture he discussed the entrepreneurship inherent in agriculture and criticized the government price distortions in developing countries: "experts fail to recognize how efficient they [small farmers] are... This allocative ability is supplied by millions of men and women on smallscale producing units; agriculture is in general a highly decentralized sector of the economy.... The allocative roles of farmers and of farm women are important and their economic opportunities really matter (Schultz, 1978b)." Yet he maintained that despite their efficiency in a traditional agricultural setting, small farmers needed further investments in human capital and skills to be just as entrepreneurial and efficient in a dynamic setting (i.e. one with constant technological and economic change made more dynamic by increased globalization in past decades).

He also gave the following hopeful exposition on humanity addressing natural resource constraints (economic growth models to come such as Nordhaus (1992) viewed natural resources including land and energy as lags on economic growth): "It is ironic that economics, long labelled the dismal science, is capable of showing that the bleak natural earth view for food is not compatible with economic history; that history demonstrates that we can augment resources by advances in knowledge. I agree with Margaret Mead: 'The future of mankind is open ended.' Mankind's future is not foreordained by space, energy, and cropland. It will be determined by the intelligent evolution of humanity."

In other words, that land is a fixed resource and that traditional sources of energy are a depleting resource do not preclude the fact that investments in human capital can modify the production function and the relationship between the traditional inputs and output.

Look no further than the growth and takeoff of alternative, renewable energy sources and the expansive yields from sustainable farming as examples. As much as Schultz critiqued government for distorting agricultural systems that he viewed as otherwise efficient, he may not have appropriately appreciated government's unique role in incentivizing the resource allocation decisions of firms towards that "intelligent evolution of humanity" that he spoke of. For example, we've seen in the last decade the role of government in incentivizing alternative, renewable energy sources through large-scale investments and subsidies for the research and development and market viability of those energy sources.

The interesting and challenging piece of the lecture is connecting those two goals: improving the human capital and skills of much of the population (key to improving well-being) and moving towards a more sustainable future that is not tied to traditional, fixed or depleting natural resources. While positing these ideas, he also provided advice to economists that is still relevant today:

"We all know that most of the world's people are poor, that they earn a pittance for their labor, that half and more of their meager income is spent on food, that they reside predominantly in low income countries and that most of them are earning their livelihood in agriculture. What many economists fail to understand is that poor people are no less concerned about improving their lot and that of their children than rich people are."

If you want a more humorous speech, try this one that he gave at the Nobel banquet. 

China and the future of development aid

At the end of last year, research organization Aid Data published the first extensive data set documenting Chinese aid flows around the world. The Chinese government is notoriously secretive about its development aid outflows: it does not publish any project-level or country-specific data on its own nor does it work with international organizations that attempt to quantify and release this information. To create the data set, researchers at Aid Data scoured publicly available news reports, official embassy documents, and aid/debt information from other countries for the past five years. The Tracking Underreported Financial Flows methodology that they rely on is detailed here.

China's lack of transparency has been cited as a growing issue given its increasing role on the international stage. In the past few years, China surpassed the U.S. in terms of annual spend on development aid and has established itself as one of the key development players in Africa. Critics - such as Moises Naim in this opinion post years ago in the New York Times - have raised concerns that in competing with Western donors and international organizations such as the World Bank, China is seen as the "no strings attached" donor likely to give to undemocratic regimes and countries with poor institutions that would be subject to higher scrutiny under traditional Western giving and lending practices. 

It is unsurprising, then, that researchers have already taken to the new Aid Data data set to answer a myriad of questions about the impact of Chinese aid on local economies. In this post I look specifically at the paper, "Chinese aid and local corruption" published last month in the Journal of Public Economics. Authors Isakkson and Kotsadam (2018) employed Aid Data's Chinese Official Finance to Africa data set to identify locations with: (1) ongoing Chinese aid projects, and (2) those selected for future Chinese aid projects. They then connected this data with Afrobarometer survey data eliciting survey respondents' experiences with corruption (whether they "had to pay a bribe, give a gift, or do a favor to government officials") in order to estimate the effect of an ongoing Chinese aid project in a given location on the level of local corruption. Because they geocode both data sets and restrict their sample to only those aid projects for which they can identify a granular location, the authors are able to identify respondents within 50 or 25 km of aid project locations to analyze their experiences with localized corruption. 

The authors find that Chinese aid projects have a statistically significant effect on local corruption with point estimates of a 3.5% (bribes given to "avoid a problem with the police") or 2.7% (bribes given to "get a document or permit") increase in bribery in locations with ongoing Chinese projects relative to locations selected for future Chinese projects. They speculate that Chinese aid increases local corruption through two potential mechanisms: first, that presence of the donor changes the cost-benefit structure of engaging in corruption (i.e. if the donor is indifferent as to the "means" by which a project is completed and is willing to reward for the "ends" of completing it then this raises the benefits associated with corruption). Second, they posit that the donor is in a position of power to influence social norms and create institutional change. A donor's acceptance or propagation of corrupt activity could worsen norms (noting that norms are easier to change for the worse than the better). Finally, they employ the same strategy around World Bank aid project locations and do not find any effect of these projects on local corruption. 

Estimation strategy

The paper employs a model similar to a difference-in-differences model wherein the responses of individuals who live near a site that is currently developed by the Chinese are compared to the responses of individuals who live near a site that will be developed by the Chinese in the future. In the following regression model, the authors employ the difference between the coefficients on "active" and "inactive" as the key parameter of interest. Individuals located within the radius of an ongoing Chinese aid project are "active" ("active" = 1). Those located within the radius of a future Chinese aid project are "inactive" ("inactive" = 1). And those that are outside the radii of any current or future Chinese aid projects are neither active nor inactive ("active" = 0; "inactive" = 0). 

(1) Yit = β1 activeit + β2 inactiveit + αs + δt +y Xit +εivt 

Isakkson and Kotsadam employ this method rather than interpreting the coefficient on the "active" dummy to avoid the ex-ante assumption that "there is no relationship between project localization and the pre-existing institutional characteristics of project sites." In other words, if the locations for Chinese aid projects were selected based on certain institutional characteristics it is very possible that those institutional characteristics are correlated with corruption levels and that, as a result, interpreting the coefficient on "active" alone erroneously captures pre-existing differences in corruption levels between locations with Chinese aid projects and those without. 

To control for the geographic and time-based variation in the data set - which includes data from across the African continent and spanning 2000-2013 - the authors include spatial fixed effects, year fixed effects, and a set of individual controls. While the baseline results indicate that Chinese aid projects led to an increase in local corruption, the various iterations do lead to questions:

1. The authors point to two statistics from the regression output to determine whether the parameter of interest is significant: coefficient on the "active" dummy variable (if there is an effect this should be positive and statistically significant) and the statistic for an F-test testing the hypothesis "active - inactive = 0" (if the effect on corruption is a result of a Chinese aid project this hypothesis should be rejected). The baseline results indicate that both with respect to police bribes and permit bribes the coefficient on "active" is positive and highly statistically significant and the F-test hypothesis can be rejected at the 5 percent level. See Table 1 for details.
2. However, the sensitivities indicate that the coefficient on "active" is statistically significant across most but not all iterations and the F-test cannot be rejected at the 5 percent level in at least one of the iterations. The results indicate that the effects are stronger for police bribes than they are for permit bribes which leads to questions about the mechanism that leads to increased corruption and why it impacts police bribes more so than permit bribes. See Table 2 for details.
3. Furthermore, the authors conclude that World Bank aid projects do not similarly lead to an increase in local corruption not because the coefficient on "active" in that sample set is not positive or statistically significant (in fact it is significant in several iterations) but because the F-test results indicate that it cannot be rejected that the "active" and "inactive" coefficients are equal. But why is it that both "active" and "inactive" locations with World Bank projects see a higher level of local corruption than those without World Bank projects (where Chinese aid projects don't, because the "inactive" coefficient is not significant in most Chinese projects)? It's not a question that this paper seeks to answer but there should be a reasonable hypothesis for why locations selected for World Bank vs. Chinese aid projects differ in this way.
4. My main question reading this paper was whether the implementation of an aid project leads to a change in the demographic population of a locality. Given that the Afrobarometer survey is not a panel data set, the same individuals are not necessarily interviewed pre- and post-implementation of an aid project. 
1. While it is possible that the implementation of a project leads to the corruption of existing actors it also seemed possible that it led to inflows of new actors into the locality due to a possible increase in local economic growth and activity. If the increase in local corruption is due to the influx and changing composition of the locality this is distinct from an increase due to corruption of the existing population. 
2. The authors attempt to address this question by analyzing whether there are more police stations in active aid areas vs. inactive aid areas (to address the claim that more bribery is a result of more police stations rather than more corruption), stating: "Neither do we find any evidence that the results are driven by increased resource flows making the project areas into 'honey pots' attracting corrupt actors." However, the empirical investigation does not seem to answer the original question of whether aid projects lead to an influx of corrupt actors. 
3. It boils down to how the effect is interpreted: in the case of this paper, the parameter of interest does not distinguish or isolate the two effects presumably because both lead to an increase in local corruption whether by migration or by impacting existing populations. 

Implications for the future of development aid

Overall, the implications of the paper that Chinese aid projects lead to local corruption are the first step in understanding how different forms of aid (and specifically "no strings attached" aid) can create institutional change and impact social norms. While the quantification of this impact is important, the paper does not explicate the mechanism by which these projects increase corruption and without that linkage it is difficult to prescribe appropriate policy solutions to improve local governance and reduce corruption. But the paper reaffirms questions about China's development strategy to work within existing entrenched systems to create economic growth vs. the traditional Western approach to attempt to improve governance and create institutional change at the same time. The broader takeaways for the field of development aid:

• It is clear that conceptualizing Chinese vs. Western aid as a competition is not the most effective way to improve growth and development in these economies, rather, assuming Chinese aid will continue at its current rate how can each set of aid practices complement and supplement one another? Transparency and data availability make it easier to answer these questions. 
• Given the importance of international coordination in aid, this new availability of data on Chinese aid offers a novel opportunity for other donors to provide a value-add to these economies in sectors and projects that the Chinese are not investing in and to advocate more strongly for better governance and effective democratic institutions given the apparent worsening of certain aspects of local governance as a result of Chinese aid projects. 
• China's increasing role in development and aid places places a need for further introspection on the part of international organizations such as the World Bank, specifically: how should the organization continue to advocate for good governance and effective democratic institutions while simultaneously recognizing the need to work with one of the largest unilateral donors that may not be interested in propagating those norms? Will the World Bank's role and priorities change as funds from China and from private investors play an increasing role in the growth of developing economies? How can it position itself most effectively in this rapidly changing space and provide a distinct value-add?

Sources

1. Isaksson, A. and Kostadam, A. (2018). Chinese aid and local corruption. Journal of Public Economics.