2
$\begingroup$

I have count data (basically histone modification data).

The counts represents number of reads falling into each genomic region.

I need to divide my data into $\text{low|medium|high}$ based on the counts.

What method should I use in order to get the most statistically sound result?

Improve this question
$\endgroup$
2
  • $\begingroup$ Would you be able to elaborate on why you need to categorize your data and why exactly three categories? What do you understand the phrase "statistically sound result" to mean? What do you hope that will do for you? $\endgroup$
    – whuber ♦
    Commented May 14, 2015 at 21:53
  • $\begingroup$ My aim is to use these results in association analysis (www-users.cs.umn.edu/~kumar/dmbook/ch6.pdf) of histone data wrt other genomic features. I could just categorize the data into two (present or absent) but I thought to divide into 3 kinds "low" ,"medium" and "high". That way I can have 3 categories for each mutation e.g H3K4me1_low (0 or 1),H3K4me1_medium(0 or 1),H3K4me1_high(0 or 1) and use them in conjunction with other genomic features. $\endgroup$
    – saad khan
    Commented May 14, 2015 at 22:20

1 Answer 1

Reset to default
1
$\begingroup$

If you're not doing any formal statistics on the categories themselves consider doing k-means clustering with 3 clusters. In this case, we are clustering in only one dimension which is the count of reads for each genomic region. After the clustering is done, you can manually determine which cluster is the high/medium/low count cluster.

K-means clustering isn't guaranteed to find the globally best clustering and people usually run it multiple times with different starting points. But I think with only 3 clusters in one dimension you'll find very similar clusters each time.

Improve this answer
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.