To control for this possibility, we regress our different measures of accrual management on the partitioning variable Sanc_Event and a set of control variables, including firm-fixed effects.
In the UDA specification, we expect the coefficient (beta _{1}) to be positive if Sanc_Aud is used as the partitioning variable and negative if Event_Sanc is included in the regression model.
Panels A and B in Table 7 summarize the results concerning our analysis of loss avoidance when applying a conditional logistic regression, in which the dichotomous output Loss depends on the partitioning variable, Sanc_Event, which captures the pre- and post-sanction period and a set of control variables, including firm fixed effects.
The terms of the model were chosen by observing properties of the simulated z-scores with respect to their corresponding partitioning variable, mean and depth in the tree.
Quantile regression was used on the simulation data to create starting values for the coefficients for τ for each partitioning variable.
We then use the data generated from each partition in each of these simulations (z-score, partitioning variable, expected count and depth of the winning partitions) to train the coefficients of a threshold model that gives the desired false alarm rate.
Figure 3 shows running times for random graphs using the fastest ILP version (using partition variables and cutting planes).
However, neither version is as effective as the partition variable formulation.
For the partition variable formulation (Section ' Partition variable ILP'), we initially add all independent set constraints (5) and those P3 constraints ((6), (7)) for which the vertices u, v, w induce a P3 in the input graph.
Possibly the independent set, which interacts with all clusters, makes local search approaches less effective here compared to SCE. Figure 5 shows the running times for the fastest ILP approaches, that is, the partition variable ILPs with cuts, and the heuristics for both problems.
