Im analyzing the impact of different biomarkers and clinical risk scores on postoperative complications and Id like to test the importance that every factor has on the final outcome. Even if I add the same Variables, Factors and Covariates in Neural Networks (Multilayer Perceptron Network) in SPSS, I get the different result every time. What could the problem be? And what is the minimum sample size for the proper analysis in NN since it needs samples for both training and testing?
You may try Weka (https://www.cs.waikato.ac.nz/ml/weka/) to solve this issue.
Hi Danica, ANN starts with new random weights to find a solution within the accpetable error limits. Therefore it can produce different results since it is not a parametric method. You can prefer parametric/nonparametric statistical methods to get the unique result.
Sabri Erdem, yes I figured that out. But since Id like to add NN to my results and compare it to logistic regression results and C statistics, would it be OK to process NN multiple times and maybe present the mean+/- SD? Maybe in that way one would be able to see that our model doesnt have too many variations.
Weka is good option for Multilayer perception [the images certainly look better] however it does not have significance of different options as it is in SPSS!
I would recommend trying hierarchal regression!
Why NN and not decision tree ML algorithms (CART, random forest)? Random forest can handle small sample sizes. Are you comparing versus ML logistic or traditional statistics regressions?
Honestly, in this case I will choose to use principal component analysis to produce factor scores before runing alrgorithms, rather them single variables.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Sampling Design
- Sample Size