Can statistics do without artefacts?

This companion paper to Chatelain and Ralf (2012), “Spurious regressions with near-multicollinearity” put their results into the contexts of the history of statistics, of the current publication bias in applied sciences and of the substantive versus statistical significance debate. This article presents a particular case of spurious regression, when a dependent variable has a coefficient of simple correlation close to zero with two other variables, which are, on the contrary, highly correlated with each other. In these spurious regressions, the parameters measuring the size of the effect on the dependent variable are very large. They can be “statistically significant”. The tendency of scientific journals to favour the publication of statistically significant results is one reason why spurious regressions are so numerous, especially since it is easy to build them with variables that are lagged, squared or interacting with another variable. Such regressions can enhance the reputation of researchers by stimulating the appearance of strong effects between variables. These often surprising effects are not robust and often depend on a limited number of observations, fuelling scientific controversies. The resulting meta-analyses, based on statistical synthesis of the literature evaluating this effect between two variables, confirm the absence of any effect. This article provides an example of this phenomenon in the empirical literature, with the aim of evaluating the impact of development aid on economic growth.