**6. Discussion**

The first approximation to the regression model was to include the variables one by one to determine which variables could explain the answer.

It was interesting that the categorical variables did not explain the model. However, the data analysis could suggest that, had the sample been more prominent, they could have influenced the result. This study aimed not to distort the sample by performing a bootstrapping, not only because of the possible distortion of the sample but also because of the complication implied when the variables do not follow a normal distribution. The sample size is considered to be representative of the period of study.

With the first approach, the variables that could explain the probability of an excursion are determined to be: water-depth, percentage of generators online, Wind force and Wave height. The first two variables belong to the DP system configuration, and the last two are related to meteorological conditions.

However, when all of them are entered into the model, we obtained that only the first two explain the answer, while the other two could not improve the model already created. Although, this could give an idea of the less importance of the meteorological variables when explaining the excursion, it should not be

forgotten that the meteorology, and especially the wind force (which creates waves with a height that is proportionally correlated to the force in knots), can also influence the probability of a unit having a loss of position while performing DP drilling operations.

The two selected independent variables, water-depth and percentage of generators, can explain the probability of losing position. This possibility will increase when the water depth has small values and the percentage of generators has significant values. These results are very interesting from the operator's point of view, as the lower values of water depths have traditionally been a common drilling ground where DP was not necessary, and other methods were used to achieve the position keeping. This could partly explain the problems of DP station keeping incidents when the drilling operations take place in shallow waters.

Studying the mean relative ratios for each main cause group, it is interesting to note that the model can explain environmental-, computer- and human-caused incidents more precisely than other causes. Within the group of human causes, the incidents without a human cause have a better prediction using the proposed model.

In general, this model correctly classifies 79% of the incidents, which is considered to be a very good prediction overall.

When the data is split into subgroups defined by the existence or not of a human cause, it can be seen how the mean percentage of thrusters is significantly more prominent for the cases where there is no human cause and smaller when there is a human cause. However, it does not explain whether the probability of an excursion is bigger or smaller for any of the subgroups. However, it can suggest that it would be a significant variable when the dependent variable is used to determine the possibility of a human error.

Always taking the general regression model from above into account, the regression model for the subgroup without human cause proposes the percentage of generators and wind force as variables that could explain the model. Wave height is not significant (although with a p-value of 0.104, it could be said that it is at the edge of being significant), and the water depth does not even enter the equation.

The fact that water depth was not even entering the equation when considered individually suggests it does not influence the probability of excursion when the cause is not human.

When studying the cases in the subgroup human cause, we obtain that the only variable that could explain the model is water depth. However, its p-value is bigger than 0.1, after entering the equation and because of this, it is rejected. In the iteration of the model, it can be seen how when this variable is included, the percentage of cases that are correctly classified decreases for the first group (no excursion). At the same time, the percentage of correctly-classified cases improves for group 2 (excursion). Probably the comparatively small number of cases (only 9 out of the 42 cases in total) contribute to the decision of rejecting this variable from the equation. Nonetheless, it suggests that this variable can be expected to be added to the model when a bigger sample is studied.
