Timeline for Fitting multiple linear regression models to select molecules for which a feature of interest significantly alters concentration
Current License: CC BY-SA 4.0
11 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Jul 2 at 13:44 | comment | added | Frank Harrell | Example: of the 92 proteins suppose that many are co-expressed so that there are 7 clusters of interrelated proteins, with very low correlation between a protein in one cluster and a protein in any other cluster. Score each cluster with the first principal component. Use these component scores to predict Y. A related but sometimes better alternative: sparse principal components analysis. All of this is discussed in RMS. | |
| Jul 2 at 8:29 | comment | added | maglorismyspiritanimal | @FrankHarrell Could you give me an example of a less specific question to ask? | |
| Jul 1 at 22:15 | comment | added | Frank Harrell | The information content in the data is rarely sufficient for you to select "winners" especially if the candidate features are correlated. Data reduction reduces the feature space to what the sample size will allow you to analyze. N too small? Ask a less specific question. Ask questions about patterns rather than individual features. | |
| Jul 1 at 14:28 | comment | added | maglorismyspiritanimal | @FrankHarrell I just don't understand how that will let me find proteins that are different between groups. How can I know what's a non-trivial effect? How can I interpret the results of unsupervised learning in this case? Wouldn't that make it very likely that I'll find separations that have nothing to do with the samples being patients or controls? | |
| Jul 1 at 13:44 | comment | added | Frank Harrell | Testing against zero means that you are interested in even trivial effects. Assessing evidence for effects that are more than trivial would be an improvement. Feature selection is almost doomed even without collinearities. More often than not, data reduction (unsupervised learning; analysis masked to Y) allows you to better live within sample size limitations. Don’t separate hard-to-separate predictors. Predict Y using collapsed dimensions. More here. | |
| Jul 1 at 13:05 | comment | added | maglorismyspiritanimal | @FrankHarrell Ok, so I should simply not try feature selection at all because it's impossible? Then why can't I do the boring hypothesis testing against 0 instead? | |
| Jul 1 at 11:44 | comment | added | Frank Harrell | Importance = 0 is not an interesting value. You need to estimate the amount of importance, and hypothesis testing against zero is very boring. More to the original point, importance measures are a basis for selecting “winning” features if you must, and getting uncertainty intervals for them shows the difficulty (and usually the impossibility) of doing reliable feature selection. | |
| Jul 1 at 8:54 | comment | added | maglorismyspiritanimal | @FrankHarrell What will the importance measure show? You mean that the CIs will include 0 for all the models? | |
| Jun 30 at 11:17 | comment | added | Frank Harrell | Sorry a link was incorrect. Use hbiostat.org/rmsc/validate | |
| Jun 30 at 11:12 | comment | added | Frank Harrell | The feature selection part of the analysis is doomed. This will be exposed by computing an importance measure (absolute $z$ statistic, likelihood ratio $\chi^2$ statistic, etc.) for each feature and bootstrap the process to get confidence intervals on I the importance measures. Examples are here and here. | |
| Jun 30 at 10:16 | history | answered | Robert Long | CC BY-SA 4.0 |