Running the Analysis

To run a Multilayer Perceptron analysis, from the menus choose:
Analyze > Neural Networks > Multilayer Perceptron...

Figure 1. Multilayer Perceptron: Variables tab and context menu for Length of stay

Length of stay [los] has an ordinal measurement level, but you want the network to treat it as scale.
Right-click on Length of stay [los] and select Scale on the context menu.
Figure 2. Multilayer Perceptron: Variables tab with dependent variables and factors selected
Select Length of stay [los] and Treatment costs [cost] as dependent variables.
Select Age category [agecat] through Taking anti-clotting drugs [anticlot] and Time to hospital [time] through Surgical complications [comp] as factors. To ensure exact replication of the model results below, be sure to maintain the order of the variables in the factor list. To this end, you may find it helpful to select each set of predictors and use the button to move them to the factor list, rather than use drap-and-drop. Alternatively, changing the order of variables helps you to assess the stability of the solution.
Click the Partitions tab.
Figure 3. Multilayer Perceptron: Partitions tab
Type 2 as the relative number of cases to assign to the testing sample.
Type 1 as the relative number of cases to assign to the holdout sample.
Click the Architecture tab.
Figure 4. Multilayer Perceptron: Architecture tab
Select Custom architecture.
Select Two as the number of hidden layers.
Select Hyperbolic tangent as the output layer activation function. Note that this automatically sets the rescaling method for the dependent variables to Adjusted Normalized.
Click the Training tab.
Figure 5. Multilayer Perceptron: Training tab
Select Online as the type of training. Online training is supposed to perform well on "larger" datasets with correlated predictors. Note that this automatically sets Gradient descent as the optimization algorithm with the corresponding default options.
Click the Output tab.
Figure 6. Multilayer Perceptron: Output tab
Deselect Diagram; there are a lot of inputs, and the resulting diagram will be unwieldy.
Select Predicted by observed chart and Residual by predicted chart in the Network Performance group. Classification results, the ROC curve, cumulative gains chart, and lift chart are not available because neither dependent variable is treated as categorical (nominal or ordinal).
Select Independent variable importance analysis.
Click the Options tab.
Figure 7. Options tab
Choose to Include user-missing variables. Patients who did not have a surgical procedure have user-missing values on the Surgical complications variable. This ensures that those patients are included in the analysis.
Click OK.