I have a gene CFAP97 with expression values and clinical information like FollowupDays and patient_vitalstatus.

Using the gene expression data first I classified the samples into CFAP97_low and CFAP97_high groups. Using the surfit function with all the information available I wanted to know which group is good for survival and which group is worse.

fit <- survfit(Surv(FollowUpDays, patient_vital_status) ~ CFAP97,
                       data = data)


low_high    records n.max   n.start events  *rmean      *se.rmean.  median  0.95LCL 0.95UCL
CFAP97=high  176     176     176     25   2746.767649   261.3634514 2475    1910     NA
CFAP97=low   420     420     420     99   2534.369762   144.8695301 2532    1849     NA

From the above summary I see that, median survival is approximately 2532 days for low group and 2475 days for high group, suggesting a good survival for CFAP97_low group compared to CFAP97_high group. And 0.95LCL is other way around, big number for high group. Which one should I choose to say worse or good survival?

But when I made a plot out of this, it looks like this enter image description here

So, from the plot it looks like CFAP97_low group is worse for survival compared to high. But in the summary results I see it different.

Am I wrong in interpreting the plot? Can anyone clear my confusion. thanq


1 Answer 1


Although the median survival estimates are 2532 and 2475 for the low and high groups respectively, those values are not distinguishable statistically based on what's shown in the summary table. The lower 95% confidence level for CFAP97=low is 1849 days, which is below the point estimate of 2475 days for the CFAP97=high group. So the median survival values are not significantly different by the usual statistical criterion of p < 0.05. You should make no claims of difference based on median survival values.

Looking at the survival curves indicates that median survival is not capturing what's going on here; the final displayed values of survival are close and barely below 50%, so any estimates of median survival in this case are necessarily imprecise (explaining the NA values for the 95% upper confidence limits for median survival). Until a bit past 4 years (~1500 days), however, survival in the CFAP97=low group is consistently below that of the CFAP97=high group.

It looks like there is no overall simple "better/worse" distinction between the groups with respect to survival. The data suggest an initial survival advantage for the CFAP97=high group that diminishes over time.

Furthermore, this type of single-predictor survival modeling can be difficult to interpret, as it ignores all of the clinical information for the patients and the expression of all the other genes for which you might have information (up to 20,000 or so in microarray or RNAseq studies). Even in linear regression this poses a problem, as omitting predictors correlated both to outcome and to the variable you are examining in your model can affect the coefficients for that variable. It's even worse in survival analysis, as omitting predictors related to outcome can affect the coefficients of variables that aren't even correlated to them. This inherent omitted variable bias should make you think very carefully about what you are going to learn reliably from this one-variable-at-a-time approach.


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.