2
$\begingroup$

I want to use glmnet cox regression approach to predict survival from methylation data for cancer patients. But I couldn't find any proper reference except this one https://cran.r-project.org/web/packages/glmnet/vignettes/Coxnet.pdf, discussing its prediction values.

I am looking for answers to my basic questions.

I trained the model using real patient data, where I provided both "overall survival" as well as "vital status of the patient" from the training set. After training, When I used test data for prediction, It is giving some values with negative sign?

I assume these are number of days for which patient will survive after diagnosis, but I couldn't understant the meaning of minus sign with these values. Does these minus sign denote status of the patients.

Please correct me if I am getting it wrong.

I am new with data science and also with R. Any help and suggestions are welcome.

Thanks in advance

Improve this question
$\endgroup$
2
  • $\begingroup$ Welcome to the site. The question is kind of unclear - I think it is a translation problem. In particular, I don't know what you mean in the paragraph starting with "I was wondering...." Can you clarify? Or, do you have access to someone who can help you clarify this? $\endgroup$
    – Peter Flom
    Commented May 3, 2019 at 12:50
  • $\begingroup$ Thanks @Peter for your suggestion. I have rewritten it, I hope it is clear now. $\endgroup$
    – Researcher
    Commented May 5, 2019 at 6:36

1 Answer 1

Reset to default
2
$\begingroup$

A Cox model is an ingenious way to estimate effects on a response with censoring. However, it it not designed to predict times or status (although it is possible by making a lot of assumptions). Instead, it would predict the hazard/risk (if type = "response") which is useful e.g. to calculate a hazard ratio between two predictions/subjects. If you don't set "type" you will get the value of the linear predictor.

My suggestion is to have a minimal understanding of the used methods. In this case of Cox regression.

Improve this answer
$\endgroup$
5
  • $\begingroup$ +1 especially for your last paragraph and especially if this is real patient data and not just a course example. $\endgroup$
    – mdewey
    Commented May 3, 2019 at 12:32
  • 1
    $\begingroup$ Thanks @Michael M for your explanation. But I am a bit confused if it is predicting only hazard ratio, then what would be the significance level of this prediction. Can you please shed some light on "the value of the linear predictor", what does it mean?. I'll also appreciate if you can share some related reference for these explanations. $\endgroup$
    – Researcher
    Commented May 5, 2019 at 16:52
  • $\begingroup$ I have also gone through few other related posts like: stats.stackexchange.com/questions/33774/… where people have used glmnet cox just to select features or predictors and later used them with un-regularised coxph for predicting coefficient for HR and its significance level. So my question is will it be a better approach to predict survival of the test data than using glmnet's predict() $\endgroup$
    – Researcher
    Commented May 5, 2019 at 16:53
  • $\begingroup$ Maybe using parametric survival model (accelerated failure time model) or quantile regression for survival data could be a more natural choice than Cox regression? An alternative would be to estimate the baseline hazard function from the data and use it together with predictions of the Cox regression to find survival probabilities or median survival times. $\endgroup$
    – Michael M
    Commented May 6, 2019 at 2:13
  • $\begingroup$ Hi @MichaelM Can you please suggest how can I estimate the baseline hazard function for my data? Another question I want to ask, what is meant by linear predictor, can these values be used to check their correlation with the survival information which I already have for my test data. $\endgroup$
    – Researcher
    Commented May 8, 2019 at 15:51

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.