The devil is in the footnote

Recent claims on inequality trends in India are based on a set of questionable assumptions. They warrant closer scrutiny

SELDOM DOES ONE come across footnotes that discredit an entire academic paper, particularly one written by distinguished economists like Thomas Piketty. The footnote reproduced below is from a recent paper issued by the World Inequality Lab, ‘Income and Wealth Inequality in India, 1922-2023: The Rise of the Billionaire Raj’ (Nitin Kumar Bharti, Lucas Chancel, Thomas Piketty and Anmol Somanchi, hereafter BCPS).

“Our results are tentative given we do not observe both income and wealth for the same set of individuals and instead draw inferences based on the full distributions of income and wealth we estimate.”(footnote 36, page 24, emphasis added).

Estimating the distribution of income is a tricky business, especially for those countries that do not have an official income survey. One of us (Bhalla) was a member of India’s first statistical commission (2006-2009) under the chairmanship of the late Suresh Tendulkar. He pleaded with the commission that India should conduct an income distribution survey. So far, India has not conducted such a survey, though it is hoped that after BCPS’S shocking analysis (documented below), India will do so and thus diminish the need for imaginary declarations as contained in the footnote above.

It is well recognised that even official income distribution surveys can be an inaccurate reflection of the “actual true unobserved” income distribution. Not all households are covered by survey schedules, and the billionaire rich are almost never covered in any survey that we know of, anywhere in the world. Hence, the original important contribution by Piketty and his colleagues at the World Inequality Lab (WIL) was to merge tax and non-survey information with survey data in order that a better guesstimate of income distribution could be arrived at.

There are many indices of inequality but what Piketty and the WIL have popularised is the share in income of the top X per cent of the population — and their preferred X is the top 1 per cent (for the technically inclined, the share of the 100th percentile). But, making an original contribution does not give one a license to literally kill the “original” survey data.

Let us explain. In a popular article in a refereed journal World Bank Economic Review (2005), Nobel laureate Abhijit Banerjee and Thomas Piketty diagnosed the Indian distribution for the years 1922-2000 and reached the conclusion that in 1999 the top 1 per cent’s share in the Indian income was just 9 per cent.

Thirteen years later in 2018, Chancel and Piketty updated their analysis of inequality in India to 2015 and took the opportunity to increase the share of the top 1 per cent in 1999 to 14.7 per cent. That is a 63 per cent increase in the share for the same year 19 years earlier. All data go through revisions, but survey data does not and neither does tax data show such large revisions — maybe a minuscule one to two percentage points.

Remember, in a democracy, the ruling party is answerable to Parliament and the press for any estimates. Sadly, economists making conjectural revisions of such high magnitudes are answerable to none. As documented and discussed in two articles (‘Piketty has got it wrong’, IE, January 20, 2018, and ‘Inequality, myth, and reality’, August 11, 2018) Bhalla had asked as to the source of the data and the data itself for the basis of these upward revisions. No response was forthcoming from the authors.

Now, six years later and coincidentally around election time, the Chancel-piketty team has been joined by two additional authors, Nitin Bharti and Anmol Somanchi, and the analysis has been extended by seven more years to 2022. Now, 25 years later, the 1999 estimate of the share of the top 1 per cent has been inflated to read 21 per cent. Hence, the original Banerjee-piketty 1999 estimate of 9 per cent has been revised to 21 per cent. It is an interesting question as to whether at any time in world economic history has any past estimate been revised upward by 133 per cent in 19 years (2005 to 2024) or at a compound annual growth rate of 4.5 per cent a year.

As the introductory footnote explicitly states, the authors estimate a distribution on unobserved data and get a ladder increase in inequality estimates for many years earlier with each update of new information for later years. If these estimated distributions are incorrect, then so are the results. A conclusion derived from a set of questionable assumptions must be robust enough to hold steady even as some of the assumptions are violated. This is not the case with the Piketty et al method. For example, Piketty led WIL’S prior work on inequality in the US and claimed a large increase in post-tax income shares of the top 1 per cent between 1960 (magic 9 per cent again) to 15 per cent in 2019. Over the last decade, these results have begun to be questioned. Gerald Auten and David Splinter argue in a recent article ‘Income Inequality in the United States: Using Tax Data to Measure Long-term Trends’ (forthcoming, Journal of Political Economy) that there has hardly been any change in post-tax share of the top 1 per cent for the last 60 years — the share has stayed constant at approximately 8-9 per cent even though the pre-tax share has substantially increased. Remember, that Piketty et al concentrate on inequality post-tax.

Geloso, Magnes, Moore and Schlosser (2022) conclude that Piketty “overstates inequality levels”. In a slightly older paper, Magnes and Murphy (2014) found evidence of “pervasive errors”, “opaque methodological choices” and “cherry-picking of sources to construct favourable patterns from ambiguous data”.

The US represents a much richer information environment with a plethora of data on income and wealth, especially relative to an emerging market economy like India. If the WIL makes errors in an information-rich environment, then how accurate can their estimates be in a more data-constrained environment? Speculating and estimating a distribution conforming to your priors is not economic analysis, robust or otherwise. More dangerous and unfortunate is to make policy recommendations based on dodgy statistics derived from questionable methodological assumptions.

For example, the authors continue with the same 75-year-old foreign conclusion of the impending breakup of India and piously conclude: “The ‘Billionaire Raj’ headed by India’s modern bourgeoisie is now more unequal than the British Raj headed by the colonialist forces... It is unclear how long such inequality levels can sustain without major social and political upheaval.” They dig further and deeper into their self-constructed hole by recommending “implementing a super tax on Indian billionaires and multimillionaires, along with restructuring the tax schedule to include both income and wealth, so as to finance major investments in education, health and other public infrastructure”. Major investments are based on real data, not on upgraded and inflated estimates of distributions.

It is surprising that not many in India have tried to look under the hood as they take the BCPS conclusions at face value. Perhaps, it is easy to accept conclusions when they conform to one’s prior beliefs. But this is precisely when a good analyst must question the path taken to arrive at such conclusions. The world would be a better place, and so would policy discourse if more attention was devoted to important questions rather than to engagement with assumed distributions and motivated conclusions.

Bhalla is an independent researcher and Bhasin is a New York-based researcher. Views are personal

As the introductory footnote explicitly states, the authors estimate a distribution on unobserved data and get a ladder increase in inequality estimates for many years earlier with each update of new information for later years. If these estimated distributions are incorrect, then so are the results. A conclusion derived from a set of questionable assumptions must be robust enough to hold steady even as some of the assumptions are violated. This is not the case with the Piketty et al method.

C R Sasikumar