Significance

Hi all,

I am using the following for estimating cross and own elasticities for products within molecules on a panel database:

est = LinearDML(model_y= LassoCV(cv=[(fold00, fold11), (fold11, fold00)], tol=0.0001,n_alphas = 1000,max_iter=100000), 
                                       model_t= MultiTaskElasticNetCV(cv=[(fold00, fold11), (fold11, fold00)],n_alphas = 1000,max_iter=100000),
                                       cv = [(fold0, fold1), (fold1, fold0)],
                                       linear_first_stages=True, fit_cate_intercept=False)

Now, focusing on just the cross elasticities, the problem here is that out of the 2500 cross estimates that I compute, only 290 are significant. Moreover more than a half of such estimates are negative, which looks strange to me.

Observations on the data I have selected: since the algorithm imposes (at least on panels) that the number of periods for which products within groups (molecules) are observed is the same for all, I am deleting the "excess time periods" with respect to the product having a lower number of time periods. This means for instance, that if a molecule has 3 products, one with 14 quarters, another with 18, and the latter with 40, I am basically randomly deleting 4 periods for prod.2 and 26 periods to prod.3 in a way that, however, intersections between periods remains. So, it may be the case that I have too few products to compute cross elasticity, but, still, I am not convinced

No repository file or test is identified; start with the shown LinearDML configuration and the documentation for panel-data requirements and cross elasticities. Done would require a reproducible example and a confirmed expected result before a code change can be scoped.