search

domenicodigangi / DynWGraphsPaper

Code for analysis of paper about dynmical sparse weighted graphs
MIT License
0 stars 1 forks source link

need to clarify the interpretation of fitnesses in misspecified (missing regressors) contexts #18

Open domenicodigangi opened 2 years ago

domenicodigangi commented 2 years ago

What do we see if we filter a dgp with a filter missing one or some (node spec) regressors? The idea is to shed some light on the interpretation of the fitness as node specific time varying "fixed" effect - fixed effect as in the regression jargon. I.e. a time varying variable capturing the propension of nodes to form links - or weights - that are not explained by the regressors used.

Practically simulate a dgp with a regressor and filter it with a missp model.

Cases of interest:

Is there a difference between the fitness' dynamics with or without the regressor? Do the fitness correlate with the regressor?

domenicodigangi commented 2 years ago

Is there a difference between the fitness' dynamics with or without the regressor? Do the fitness correlate with the regressor?

  • [ ] dgp fitness + 1 reg uniform const beta

    • [ ] filter with fitness only

The fitnesses have a high correlation with the external regression, as intuitively expected. The relation is clearer in the weighted model where non linearities play a smaller role

  • [ ] filter with fitness and regressor

In this case the average correlation between the fitness and the external regression is zero

  • [ ] dgp fitness + 1 reg 2N const beta

    • [ ] filter with fitness only

In the weighted case, Fitness correlate with external regr and the correlation is related with the beta coefficient of the dgp. In the binary case the relation between correlation and beta coeff is unclear

  • [ ] filter with fitness and regressor

Both the estimated coefficients and he fitness are in good agreement with the DGP

NEED TO ADD THE PICTURES SUPPORTING THOSE STATEMENTS (REQUIRE LABELS AND TITLES)