Should I have "normal" sampled data in my dataset?

Question

I am busy working on a project to find the reasons why kids in normal households are doing badly in school.

I have a dataset of which consists of kids that live in environments where the family is middle class, has access to necessary facilities and the kid is not suffering from any disorders but is failing grades in school.

It is understandable for kids that has a poor living condition to have problems at school but for kids that has all the necessities in life to do at least average in school, needs a bit more research.

Now that I have this dataset, does it make sense to add kids with the same living environments that are doing ok at school to the dataset? I am planning to use SOM for data mining if that helps.

Answer 1

In short, yes.

If your goal is to understand the drivers behind poor school performance (or even model + predict school performance), you will need both "positive" and "negative" outcomes in your dataset.

Of course, you need to be careful that the positive outcomes are "similar" enough to the negative outcomes to avoid introducing distortions into your data - i.e. ensure that the data you are adding are from middle-class families with access to facilities etc.

You should also consider the ratio of positive:negative outcomes in your final data set, as you will ideally like to avoid imbalanced data. However, there are techniques and approaches to handle imbalanced data if necessary.