0
$\begingroup$

I am running high-throughput microarray data (methylation array), and after running univariate, lasso and cross-validation lasso analyses, I was able to come down to a list of 15 probes (predictors).

Now, I want to run a ROC/AUC curve in order to check whether those predictors are in fact good candidates. Problem is that the result coming out of it is a ROC curve with an AUC=1. I have been trying to twitch the fitted model (i.e. family and maxit), but the results did not change.

Here is a sample of the data (with 8 predictors) and the downstream analysis, with some explanation:

          Tumor   probe_1     probe_2     probe_3    probe_4    probe_5    probe_6    probe_7     probe_8
Benign.A4    No -5.076257 -3.18658187 -2.91627872 -3.2393655 -2.4080861 -3.9414602 -4.5844204 -2.96877633
Benign.A1    No -3.232952 -2.21518181  0.71340947 -2.1103999 -1.4563154 -4.0614544 -2.9378821 -0.90468942
Benign.C2    No -4.487701 -3.34515435 -5.35341349 -2.0355878 -2.9573763 -4.2980546 -4.3421487 -2.35597830
Benign.C8    No -3.692610 -1.24332686 -0.59115736 -3.4852858 -2.3339160 -3.1302782 -3.0943430 -1.03581249
Benign.D7    No -2.978757 -0.05097524  0.02744634 -1.4946543 -1.5593915 -2.8860660 -2.7633458 -0.99299595
Benign.D3    No -2.441925 -1.98227873 -2.13478645 -3.0265593 -2.7789079 -3.9860489 -2.8512663 -2.61804934
Tumor.A6    Yes  1.044348 -5.85637090 -4.49697162  1.5033139  0.3226736  1.5937440 -0.4881769  0.95135529
Tumor.A5    Yes  1.749187 -2.93393903 -5.54439148  2.4403760  1.6238294 -1.1699169  3.0410728  1.07437064
Tumor.A2    Yes  2.323806 -6.57693143 -5.78690184  1.7684931  2.3522317  0.3517146 -1.9972320  1.46663990
Tumor.C1    Yes  2.229316 -6.69010615 -6.22036584  0.7482678  1.3277280  0.6128029  1.3349142  1.63602050
Tumor.C6    Yes  2.888489 -5.79079519 -5.02991621  1.4605461  1.3002248  1.1498193  0.4481215  0.81473797
Tumor.C5    Yes -1.861726 -5.14400193 -5.26197761  1.0023323  0.8582683  0.5492184  0.6720438  1.73785369
Tumor.D1    Yes  2.776804 -6.78537165 -6.20280759  2.0623420  1.8291220  1.7328508  1.3667038  1.77813837
Tumor.D6    Yes  2.985209 -6.13405436 -5.92181030  1.8801728  1.1815045  2.2210693  0.1363381  2.21102559
Tumor.D8    Yes  1.670136 -6.72855542 -6.61156537  1.9847271  1.6267041 -2.8621148  0.7134887 -0.56794735
Tumor.A3    Yes  2.106628 -5.61286600 -5.75976883  2.1291475  0.5839721  1.4210874  1.2746626  1.77239233
Tumor.A8    Yes  1.798005 -5.53405698 -5.34042037  3.0262657  1.2199790  1.2448107  1.2297283  0.25649834
Tumor.A7    Yes  1.798074 -6.03775348 -5.01964376  1.2428083  2.3899569  0.6292222  0.6439477  0.92047002
Tumor.C3    Yes  1.542737 -6.54219832 -5.94287577  1.6111676  2.1889028  0.1228641  0.7950770  1.38000135
Tumor.C7    Yes  3.369420 -6.84809093 -5.88474727  2.7525838  3.2090893  1.1435739  1.2199450  0.89089956
Tumor.C4    Yes  3.179484 -6.59432541 -5.68920298  2.4093288  2.3173752 -0.3378846  1.3653768  0.66432101
Tumor.D5    Yes  2.328382 -6.41234621 -6.18003184 -0.1768171  2.1202506  2.4287615  1.7804487  0.08098025
Tumor.D4    Yes  3.051829 -7.01875245 -6.32614849  1.4200916  2.3582254  2.4981644  1.7878118  1.14826500
Tumor.D2    Yes  2.686846 -3.57625801 -6.25573666  1.6330575  0.8448418  1.4229245 -0.6461006  0.09491185

The glm analysis:

> glmcgs <- glm(Tumor ~ probe_1 + probe_2 + probe_3 + probe_4 + probe_5 + probe_6 + probe_7 + probe_8 + probe_9 + 
                  probe_10 + probe_11 + probe_12 + probe_13 + probe_14 + probe_15, data=cgshort, family = quasibinomial(link = 'logit'), maxit=100)

> summary(glmcgs)

Call:
glm(formula = Tumor ~ probe_1 + probe_2 + probe_3 + probe_4 + 
    probe_5 + probe_6 + probe_7 + probe_8 + probe_9 + probe_10 + 
    probe_11 + probe_12 + probe_13 + probe_14 + probe_15, family = quasibinomial(link = "logit"), 
    data = cgshort, maxit = 100)

Deviance Residuals: 
       Min          1Q      Median          3Q         Max  
-6.227e-06  -3.066e-07   3.076e-06   4.536e-06   6.389e-06  

Coefficients:
             Estimate Std. Error t value Pr(>|t|)    
(Intercept)  23.68423    5.23171   4.527 0.001932 ** 
probe_1       0.41584    1.03539   0.402 0.698471    
probe_2      -0.88243    0.80631  -1.094 0.305630    
probe_3      -1.14642    0.60525  -1.894 0.094819 .  
probe_4       0.08650    1.64350   0.053 0.959314    
probe_5      -1.46564    1.38381  -1.059 0.320469    
probe_6      -0.72839    1.35910  -0.536 0.606580    
probe_7       2.59539    0.48714   5.328 0.000704 ***
probe_8       2.03890    1.43339   1.422 0.192700    
probe_9       0.87683    1.52469   0.575 0.581041    
probe_10      1.79828    0.80940   2.222 0.057028 .  
probe_11      0.66033    0.93300   0.708 0.499195    
probe_12    -14.75184    2.98871  -4.936 0.001141 ** 
probe_13      3.30891    1.31239   2.521 0.035737 *  
probe_14      0.36376    0.99582   0.365 0.724368    
probe_15     -0.03516    0.91771  -0.038 0.970375    
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

(Dispersion parameter for quasibinomial family taken to be 6.771977e-11)

    Null deviance: 2.6992e+01  on 23  degrees of freedom
Residual deviance: 3.9885e-10  on  8  degrees of freedom
AIC: NA

Number of Fisher Scoring iterations: 24

PS: the reason why I am using quasibinomial here is because within the samples "Tumor", there are 2 different stages. However, there is no statistical difference in methylation levels between them (previous analysis has been performed).

And finally, the ROC curve with AUC:

> roc.final <- roc(cgshort$Tumor, fitted(glmcgs), smooth=FALSE)

Call:
roc.default(response = cgshort$Tumor, predictor = fitted(glmcgs),     smooth = FALSE)

Data: fitted(glmcgs) in 6 controls (cgshort$Tumor No) < 18 cases (cgshort$Tumor Yes).
Area under the curve: 1

My guess is because the sample size is not big enough, which would also explain the high standard error. Would that be it? And would there be any way to still evaluate the efficiency of those potential predictors in such a sample?

$\endgroup$
  • 3
    $\begingroup$ I think you should work a little on this question for making the example reproducible. I think that the main issues here are statistical and not on the programming side. This question should be moved to cross-validated. I doubt that lasso left all the 15 predictors in the model, I think that your model predicts exactly the tumor, in fact probe1 is enough to predict 100% accurately (probe1 below/above -2). $\endgroup$ – paoloeusebi Oct 15 '19 at 11:44
  • $\begingroup$ Thank you, @paoloeusebi. Yes, the question is mostly on statistics (sorry, if there is somewhere in the post that doesn't make it clear, despite the tags). And yes, the upstream analyses were not added here. Briefly, I ran (1) univariate linear regression first (in ~13,000 predictors), and got ~11,000 out of it (p<0.05) (PS: those ~13,000 were found among ~500,000 predictors initially, by limma). (2) Lasso and cross-validation, where I got to those 15 final targets. As for predicting "exactly" the tumor, I can totally see your point. But it does look very iffy, no? Thanks again. $\endgroup$ – Douglas Oct 15 '19 at 13:02
  • $\begingroup$ so you have how many observations? and 500.000 possible predictors? if your sample is little enaugh even completely random predictors could very well have a perfect association with your response variable, if you choose it from such a large poll $\endgroup$ – carlo Oct 15 '19 at 13:38
1
$\begingroup$

If the model manages to set apart successes and unsuccesses completely, AUC will be 1. you have very little data and many predictors, and some of them are very effective in predicting outcome, so no surprise here that the model appears deterministic and ROC curve is square.

Let me guess: you used maxit parameter because model estimation woudn't converge. This means that that solution is not reliable. To get one you may use generalized LASSO, or some other kind of regularization.

This is a statistics question by the way.

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ Hi, @carlo. You did get that completly right: the maxit parameter was set due to no convergence of estimation. Also, I just added a brief statistical procedures to @paoloeusebi comment above. I did do LASSO, and those 15 came out from that. Unfortunately, the sample size is quite small. Would you have more suggestions on what to try? Also, does it make sense to use the quasibinomial family here? And thanks for the spot on comment on maxit. $\endgroup$ – Douglas Oct 15 '19 at 13:06
  • $\begingroup$ I see that you used LASSO to select variables, ok. what I was suggesting is to use LASSO to estimate the final model, directly or after a first selection, in general there is no preferred behaviour, even if you can of course try to see what gets you better out-of-fold performance. $\endgroup$ – carlo Oct 15 '19 at 13:30
  • $\begingroup$ As I just pointed out in the comment above, the number of predictors you have is your biggest problem. Anyway, binomial quasi-likelihood (called quasibinomial in R) only makes sense if you have grouped data, while in the sample you posted I only see individual observations. So, no, quasi-likelihood here is pointless $\endgroup$ – carlo Oct 15 '19 at 13:42

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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