As the title informs, I'm currently working with a high dimensionality DNA methylation data. After performing batch correction, normalization and missing value imputation. I'm left with a significantly high dimensionality data, so I was looking for dimensionality reduction techniques and came across this literature. So as stated in this work I tried to perform univariate cox analysis to reduce the feature count but I'm not sure what cut-off value should I use for q-value so that I can eliminate features which are false positive and at the mean time not increasing false negatives.
Thank you for reading through my thread, if you need any additional detail please feel free to ask me. I'll respond within few minutes.
•
written
by
Yugan Gogul Muthukumar •
1 answer
There is no "correct" answer to this. No threhsold will be give you all the true positives with no false positives. The best threshold is going to depend on your personal judgement of the dataset, and the problem, and the biology. When we want to be fairly confident in our hits, and don't care about missing false positives, we usually use a q-value threshold of 0.05. This means that on average, we expect on average 5% of our hits to be false positives. But is seems that if your goal is to reduce your feature space for further analysis, this might not be the optimal threshold.
Consider the following dataset
Threshold (A) | Positives (B) | E(FP) = AxB | E(FN)
--------------+---------------+-------------+-------
0.3 | 1000 | 300 | X
0.05 | 100 | 5 |X + 605
You have to decide on the basis of what you want to do next, whether the trade of 605 extra false negatives, for 295 fewer false negatives is worth it.
Log in to answer this question.
