Read on Twitter
“Understanding Persistence” by Morgan Kelly re-examines the historical persistence literature in economics from a spatial statistics perspective. I think the results are MASSIVELY important for social science more broadly so here is a thread on the matter (1/13) #EconTwitter
Firstly, if you are doing any empirical analysis where the units of observations are distributed in space it is VITAL that you perform a series of spatial robustness exercises (2/13)
These robustness exercises include adding “World Bank Region” fixed effects in cross-county studies. Most studies use continent fixed effects despite there being vastly different within continent geographical attributes (compare Sub-Saharan and North Africa for example) (3/13)
You can see the effect of long-range spatial autocorrelation in cross-country regression in these lovely looking maps (4/13)
For local studies Kelly recommends the inclusion of latitude and longitude directions as control variables to capture spatial trends. For example, there is a striking relationship between household consumption and latitude in Dell’s Mita paper (5/13)
Other recommended spatial robustness checks are excluding observations that both have extreme values and are at the edge of maps or spatial grids and introducing Malaria prevalence as a control variable (6/13)
The effect of these robustness checks is typically, but not always, large. For instance, the coefficient on Dell’s Mita paper shrinks to less than a third of the amount reported in the paper (7/13)
Another related worry we should have is that we are just fitting “spatial noise”. Kelly demonstrates this via a Monte Carlo experiment where a spatially correlated Y is regressed on a spatially correlated X and the implied t-ratios and p-values are calculated (8/13)
The image below shows the extent to which this problem occurs in data sets that we commonly use in economics. Failure to account for spatial autocorrelation can lead to a rejection of the null in 40pc of cases with the nominal 5pc cut-off! (9/13)
Is there something we can do about this? Yes, thankfully there is! Kelly’s recommendation is to estimate the Variance-Covariance using a sandwich method with the “meat” adjusted to take account the residual spatial autocorrelation (10/13)
This approach is, in essence, the Conley (1999) except Kelly recommends that the weighing kernel K(.) be estimated from the data, not imposed. Doing this is relatively straightforward in #Rstats as you can estimate the Matérn function via ML (11/13)
The final plot is the most striking. It shows how the t-ratios change once basic spatial robustness checks and the st. errs. are spatially adjusted. Overall, it appears that the reported t-ratios in the persistence literature are inflated by spatial nature of the data (12/13)
Lastly, it is important to note that this paper is not an attack on the quality of scholarship of any study. These papers have their own robustness exercises and multiple specifications, so the Kelly findings are not evidence that any specific paper is discredited (13/13)
