BINARYDELAY = β<sup>0</sup> + β1MAJORITY + β2BOARDOWNER + β3MANAGERAGE + β4WOMAN + β5TENURE + β6TIES + β7BOARDSIZE

We will also run three additional BLRs to check how: (a) non-violators differ from mild violators, (b) non-violators differ from severe violators, (c) mild violators differ from severe violators. The latter BLRs help to disclose, how the results vary when the severity of the violation is incorporated into the analysis.

Moreover, in further analysis, we divide the firm population into two subpopulations based on either the median size or median age, in order to check the robustness of the base results with respect to firm size and age differences. Additional BLRs are run in the subpopulations, which enable us to outline how smaller/larger or younger/older firms differ from the base results. The usage of more categories (e.g., breaking the firm population based on size or age quartiles) is not reasoned, as the ranges of size and age variables are not wide enough to justify the usage of a large number of subpopulations. We do not apply size and/or age as control variables due to (serious) multicollinearity issues, which can emerge from applying them with the chosen independent variables (e.g., with variables MANAGERAGE or TENURE).

It is not rational to use different types of logistic regressions (e.g., multinomial or ordered) herewith, as by keeping BLR as the only method, we can exactly compare the coefficients in different models, and by doing that, outline whether the independent variables behave differently when various contexts (i.e., the severity of delay, firm size or age) are altered. Finally, we run bootstrapping with 100 subsamples in order to study, how the coefficients of independent variables vary in the subpopulations of the whole population.
