I was doing some exploratory data analysis, and I found a couple of outliers. In what circumstance is it correct to remove them before fitting the linear models? They are not misrecorded but they are definitely affected by abnormal factors which are not measured in the data (for example natural disasters and wars).
As you've already identified that your outliers are not data errors, then what you're telling me is that these anomalous points don't fit your probability model, and you've identified that they are probably outliers because the lack of particular measurements concerning them does not allow them to be incorporated into your probability model.
Excluding such data points would be a valid way forward, provided that their exclusion was carefully documented in the write-up!