By Vidhu Shekhar
Not long ago came the claim that India had overstated its GDP growth by roughly 2.5 percentage points a year. Subsequent scrutiny exposed a curious pattern: indicators changed, specifications changed, and several tests used to infer a structural break had too little statistical power to distinguish a genuine change from ordinary noise. As the evidence crumbled under scrutiny, the conclusion that India’s growth was less impressive than reported stood untouched, as if it had never needed the evidence at all.
Now comes another statistical exercise reaching a familiar destination. Remarkably, it cites that contested claim to call its own estimate conservative.
Kevin and Robin Grier of Texas Tech University claim that by 2023, Indian incomes were roughly 10% below where they would have been without Narendra Modi. The instrument this time is synthetic control. Construct an artificial India by mixing other countries until the cocktail closely tracks actual India before 2014, then treat whatever that cocktail does afterwards as “India without Modi.”
Their synthetic India is 38% Ethiopia, 28% China, 25% Bangladesh, 7% Pakistan and 2% Philippines. Actual India, the paper concludes, finished about 10% behind this statistical creation.
Before auditing the recipe, back-test the machine. Take the paper’s own five countries, fit the weights only on data through 2004, and ask the method to predict the decade that follows. By 2014, it “discovers” an Indian shortfall of 10%, under Manmohan Singh, with Modi nowhere in sight.
The pre-2004 fit is exactly as tight, in relative terms, as the fit the paper celebrates for its Modi model. A machine that finds the Modi growth catastrophe before Modi arrives deserves deep scepticism when it finds one afterwards.
Why does the machine misfire? Because there is no unique synthetic India. My replication finds many, and they disagree.
To be clear, within the paper’s chosen PWT 11.0 dataset and chosen 14-country pool, its result appears robust. The problem begins one level earlier: why that data vintage, that GDP concept and that donor pool?
Change the data vintage and the verdict changes. Run the same method on the same 14 countries using World Bank data, and about six of every ten well-fitting cocktails put actual India ahead. Use the previous Penn World Table release, and roughly three-quarters of the 200 best-fitting models put India ahead by 2019; with PWT 11.0, only 6% do, and by 2023 none do. India’s economic history did not change when a statistical database was revised, but the estimated “Modi effect” did.
Change the GDP concept and the verdict changes again. The paper uses PWT’s expenditure-side measure, intended principally for comparing living standards. Use instead the PWT measure recommended for comparing growth rates, keeping the authors’ donor pool, and the alleged 10% deficit disappears. India finishes between roughly parity and 6% ahead, with a better pre-2014 fit.
Then change the donor pool. Apply the authors’ stated selection logic more broadly and 29 economies qualify, not 14, creating 118,755 possible five-country baskets. Very different mixes reproduce India’s pre-2014 path closely and then head in opposite directions. Set aside the war-torn and sanctioned states, and in the remaining pool a better-fitting model has India finishing 3.7% ahead; and across the 50 best fits India leads in 54% of cases. The median gap is effectively zero.
This is the central problem. Matching the past does not identify one India-without-Modi. It identifies many synthetic Indias that agree beautifully about the past and disagree sharply about the future. The paper selects one and then writes as if history selected it.
Its defence is that 30 years of close fit lets us assume actual and synthetic India are similar even in unobserved respects. Modern econometric theory grants no such licence: a model can match history superbly and still get the aftermath wrong.
The paper rightly asks that donor economies not be distorted by idiosyncratic shocks during the treatment window. Applied even-handedly as a sensitivity test, screening out not only economies wrecked by war, default or sanctions but equally those at the opposite extreme of the growth distribution. The surviving counterfactuals fit as well as or better than the published model while placing India within roughly 3% either way.
Then come the “p = 0.00” results. With such small donor pools, the standard exact test cannot produce a p-value below about 0.07 for GDP or 0.08 for governance. The paper’s own footnote concedes that its zeros mean only that India ranked most extreme in a small placebo race. Those zeros cannot establish conventional 1-in-20 significance: the race simply has too few horses.
And so goes another attempt to belittle India’s growth story. An imaginary India assembled from other countries’ data is elevated into the benchmark, against which the lived economic experience of 1.4 billion people is judged.
The recipe is familiar. Build an elaborate counterfactual. Produce a graph where the historical lines overlap beautifully. Assign a causal label to the later gap. Report a certainty the design cannot supply.
Then tell 1.4 billion Indians that the economy they actually lived through should have been much richer, on the authority of a country that has never existed.
(Vidhu Shekhar is associate professor, finance, at SPJIMR. Views expressed are personal.)
Disclaimer: The views expressed are the author’s own and do not reflect the official policy or position of Financial Express.