You’re likely aware of the story. Having developed General Relativity, a theory of gravity that improved on Newton’s account, Einstein concluded that the fabric of space is warped by the presence of mass and thus that light rays will travel on distorted paths, following the warped space. He then predicted that this could be observed during a solar eclipse, when the apparent position of stars near the sun would be distorted by the mass of the sun. Britain’s leading astronomer, Arthur Eddington, set out to observe the 1919 solar eclipse, and triumpantly confirmed Einstein’s prediction. The story then made the front pages of the newspapers, and Einstein became a household name.
You’re also likely aware of the revisionist account. That the observations acquired by Eddington were ambiguous and inconclusive, and that he picked out the subset of measurements that agreed with Einstein’s prediction. Thus Eddington’s vindication of Einstein was not warranted on the data, but was more a “social construction”, arrived at because Eddington wanted Einstein’s theory to be true. Thus Einstein’s fame resulted, not from having developed a superior theory, but from the approval of the high-status Eddington.
The story is often quoted in support of the thesis that science — far from giving an objective model of reality — is just another form of socially-constructed knowledge, with little claim to be superior to other “ways of knowing”. Even those who may grant that science can attain some degree of objectivity can point to such accounts and conclude that the acceptance of scientific ideas is far more about the social status of their advocates than is commonly acknowleged.
A new paper by Gerry Gilmore and Gudrun Tausch-Pebody reports a re-analysis of the data and a re-evaluation of the whole story. Their conclusion, in short, is that Eddington’s analysis was defendable and correct. Where he placed more credence on some observations than others he was right to do so, and the measurements really did favour Einstein’s value for the deflection of the stars’ positions.
Thus, the later revisionist account by philosophers John Earman and Clark Glymour, taken up in accounts of science such as The Golem by Harry Collins and Trevor Pinch, are unfair to Eddington.
Gilmore and Tausch-Pebody say in their article:
Earman and Glymour conclude: “Now the eclipse expeditions confirmed the theory only if part of the observations were thrown out and the discrepancies in the remainder ignored; Dyson and Eddington, who presented the results to the scientific world, threw out a good part of the data and ignored the discrepancies. This curious sequence of reasons might be cause enough for despair on the part of those who see in science a model of objectivity and rationality.”
Our re-analysis shows that these strong claims are based entirely on methodological error. Earman and Glymour failed to understand the difference between the dispersion of a set of measurements and an uncertainty, random plus systematic, on the value of the parameter being measured. They speculated but did not calculate, and their conclusions are not supported by evidence.
Their error was left unchallenged and the strong conclusions and accusations they derived from it were used not only to question the scientific method then applied, but also to undermine the scientific integrity and reputation of an eminent scientist.
The crucial observations came from two different telescopes, a 4-inch telescope at Sobral, in Brazil, and an astrograph sent to Principe Island, off West Africa. Einstein’s theory of gravity predicted a deflection (for a star at the sun’s limb) of 1.75 arcsecs, while a calculation based on Newtonian gravity predicted half that value, 0.87 arcsecs.
Gilmore and Tausch-Pebody present the table below, listing the measured deflection, and how much it differed from the Einsteinian, Newtonian and zero-deflection models. The z value is the difference, in units of the measurement’s error bar, and P(z) is the probability of obtaining that measurement, were the model correct. The data clearly prefer Einstein’s value for the deflection.
Observations were also made with a third instrument, an astrograph taken to Sobral. However, the resulting images were “diffused and apparently out of focus”, resulting in a systematic error that was large and unquantifiable. Crucially, being unable to evaluate the systematic distortion, the observers could not arrive at a proper uncertainty estimate for these data points, without which they could not be combined with the measurements from the other two telescopes.
Gilmore and Tausch-Pebody conclude:
The original 1919 analysis is statistically robust, with conclusions validly derived, supporting Einstein’s prediction. The rejected third data set is indeed of such low weight that its suppression or inclusion has no effect on the final result for the light deflection, though the very large and poorly quantified systematic errors justify its rejection.
Scientists, being human, are of course fallible and prone to bias. To a large extent they are aware of that, which is why techniques such as double-blinded controlled trials are routinely adopted. And in some areas, such as the replication crisis in psychology, scientists have certainly not been careful enough. But, overall, it does seem that science succeeds in overcoming human fallibility, and that the consensus findings arrived at are more robust than critics sometimes allow.