The statistical model used to generate attribution findings should be validated with in-sample as well as holdout sample, or “control“ (a sample of data not used in fitting a model) – the holdout sample is used to assess the performance of the models.