Hi, I am wondering if anyone can explain to me if cumulative incidence (CI) is just "1 minus kaplan-Meier survival"? Under what circumstance, you should use cumulative incidence vs KM survival? If the relationship is just CI = 1-survival, then what difference it makes to use one vs. the other? And in R how I can draw a cumulative incidence plot. I know I can make a Kaplan-Meier survival plot using plot(survfit()), for example: fit<-survfit(Surv(time,status)~group,data=data) plot(fit, col=1:2) How to draw CI plot then? Thanks very much! John [[alternative HTML version deleted]]
On Jun 27, 2011, at 4:31 PM, array chip wrote:> Hi, I am wondering if anyone can explain to me if cumulative > incidence (CI) is > just "1 minus kaplan-Meier survival"?First tell us what you think CI is defined as. I suspect it is not the same. The KM estimator is cumulative product of (alive-n(dead))/alive so is the product of interval survival probabilities. I doubt that your definition of CI has a similar denominator.> Under what circumstance, you should use > cumulative incidence vs KM survival? If the relationship is just CI > 1-survival, then what difference it makes to use one vs. the other? > > And in R how I can draw a cumulative incidence plot.plot(time, cumsum(dead)) ...?> I know I can make a > Kaplan-Meier survival plot using plot(survfit()), for example: > > fit<-survfit(Surv(time,status)~group,data=data) > plot(fit, col=1:2) > > How to draw CI plot then?As above. Specify what you are seeking. There is a well-defined relationship between S(t) and the cumulative hazard. Maybe you should do a little study of those terms in texts regarding survival analysis.> Thanks very much! > > John > [[alternative HTML version deleted]]Isn't it time you learned to post in plain text? -- David Winsemius, MD West Hartford, CT
alanm (Alan Mitchell)
2011-Jun-28 16:20 UTC
[R] cumulative incidence plot vs survival plot
John, Since death precludes recurrence, censoring deaths would violate the KM estimator assumption that additional follow-up would eventually lead to an event. If your goal is to estimate the probability of recurrence, then you want CI with deaths as a competing risk. The cuminc function in the cmprsk package is a great place to start. Gooley has a great paper on the difference between CI and 1-KM (See Statistics in Medicine, 18, 695-706 (1999)). HTH, Alan Mitchell, MSc Biostatistician alanm at crab.org -----Original Message----- From: array chip [mailto:arrayprofile at yahoo.com] Sent: Monday, June 27, 2011 2:04 PM To: David Winsemius Cc: r-help at r-project.org Subject: Re: [R] cumulative incidence plot vs survival plot Hi David, Thanks for responding, and plain text ...(didn't realized I was in rich text). The endpoint is disease recurrence, I was producing a regular KM plot of recurrence-free probability. Then someone recommend using cumulative incidence is preferred because death was censored in the dataset. I did a little googling, I found CI was used often in the context of competing risk. I am totally new to competing risk and trying to understand what competing risk means and why CI is preferred than KM survival in this context. If you could share your thoughts helping me to understand, greatly appreciated. Searched archive, found people talking about cmprsk package for estimating and plotting CI. would that be the same as the code you suggested: plot(time, cumsum(dead)) Thanks very much! John ________________________________ From: David Winsemius <dwinsemius at comcast.net> Cc: r-help at r-project.org Sent: Mon, June 27, 2011 1:45:35 PM Subject: Re: [R] cumulative incidence plot vs survival plot On Jun 27, 2011, at 4:31 PM, array chip wrote:> Hi, I am wondering if anyone can explain to me if cumulative incidence> (CI) is just "1 minus kaplan-Meier survival"?First tell us what you think CI is defined as. I suspect it is not the same. The KM estimator is cumulative product of (alive-n(dead))/alive so is the product of interval survival probabilities. I doubt that your definition of CI has a similar denominator.> Under what circumstance, you should use > cumulative incidence vs KM survival? If the relationship is just CI > 1-survival, then what difference it makes to use one vs. the other? > > And in R how I can draw a cumulative incidence plot.plot(time, cumsum(dead)) ...?> I know I can make a > Kaplan-Meier survival plot using plot(survfit()), for example: > > fit<-survfit(Surv(time,status)~group,data=data) > plot(fit, col=1:2) > > How to draw CI plot then?As above. Specify what you are seeking. There is a well-defined relationship between S(t) and the cumulative hazard. Maybe you should do a little study of those terms in texts regarding survival analysis.> Thanks very much! > > John > [[alternative HTML version deleted]]Isn't it time you learned to post in plain text? -- David Winsemius, MD West Hartford, CT
Thank you Alan! Now I sort of understand what it means by competing risk! So in cuminc() function, the argument "fstatus" should be coded like: 0=censored, 1=event of interest, 2=event of competing risk. Then the function will calculate CI for each of the 2 types of events (event of interest and event of competing risk), am I correct? What about running regular Cox regression for recurrence? any problem there? for example, need to take into competing risk as well or regular Cox regression is still fine? Thanks! John ----- Original Message ---- From: alanm (Alan Mitchell) <alanm at crab.org> To: array chip <arrayprofile at yahoo.com>; David Winsemius <dwinsemius at comcast.net> Cc: r-help at r-project.org Sent: Tue, June 28, 2011 9:20:22 AM Subject: RE: [R] cumulative incidence plot vs survival plot John, Since death precludes recurrence, censoring deaths would violate the KM estimator assumption that additional follow-up would eventually lead to an event. If your goal is to estimate the probability of recurrence, then you want CI with deaths as a competing risk. The cuminc function in the cmprsk package is a great place to start. Gooley has a great paper on the difference between CI and 1-KM (See Statistics in Medicine, 18, 695-706 (1999)). HTH, Alan Mitchell, MSc Biostatistician alanm at crab.org -----Original Message----- From: array chip [mailto:arrayprofile at yahoo.com] Sent: Monday, June 27, 2011 2:04 PM To: David Winsemius Cc: r-help at r-project.org Subject: Re: [R] cumulative incidence plot vs survival plot Hi David, Thanks for responding, and plain text ...(didn't realized I was in rich text). The endpoint is disease recurrence, I was producing a regular KM plot of recurrence-free probability. Then someone recommend using cumulative incidence is preferred because death was censored in the dataset. I did a little googling, I found CI was used often in the context of competing risk. I am totally new to competing risk and trying to understand what competing risk means and why CI is preferred than KM survival in this context. If you could share your thoughts helping me to understand, greatly appreciated. Searched archive, found people talking about cmprsk package for estimating and plotting CI. would that be the same as the code you suggested: plot(time, cumsum(dead)) Thanks very much! John ________________________________ From: David Winsemius <dwinsemius at comcast.net> Cc: r-help at r-project.org Sent: Mon, June 27, 2011 1:45:35 PM Subject: Re: [R] cumulative incidence plot vs survival plot On Jun 27, 2011, at 4:31 PM, array chip wrote:> Hi, I am wondering if anyone can explain to me if cumulative incidence> (CI) is just "1 minus kaplan-Meier survival"?First tell us what you think CI is defined as. I suspect it is not the same. The KM estimator is cumulative product of (alive-n(dead))/alive so is the product of interval survival probabilities. I doubt that your definition of CI has a similar denominator.> Under what circumstance, you should use > cumulative incidence vs KM survival? If the relationship is just CI > 1-survival, then what difference it makes to use one vs. the other? > > And in R how I can draw a cumulative incidence plot.plot(time, cumsum(dead)) ...?> I know I can make a > Kaplan-Meier survival plot using plot(survfit()), for example: > > fit<-survfit(Surv(time,status)~group,data=data) > plot(fit, col=1:2) > > How to draw CI plot then?As above. Specify what you are seeking. There is a well-defined relationship between S(t) and the cumulative hazard. Maybe you should do a little study of those terms in texts regarding survival analysis.> Thanks very much! > > John > [[alternative HTML version deleted]]Isn't it time you learned to post in plain text? -- David Winsemius, MD West Hartford, CT
Alan, Let's say that I code censoring as "0", recurrence as "1" for fstat and death/competing risk as "2". If a patient did not have recurrence and lost follow-up at 2 years in terms of recurrence monitoring, but he also died at 5 years. How should I code this patient? I think I still code this patient as 0 (censoring) because lost-of-followup occurred before death, am I correct? Thanks very much! John ----- Original Message ---- From: alanm (Alan Mitchell) <alanm at crab.org> To: array chip <arrayprofile at yahoo.com>; David Winsemius <dwinsemius at comcast.net> Cc: r-help at r-project.org Sent: Tue, June 28, 2011 9:20:22 AM Subject: RE: [R] cumulative incidence plot vs survival plot John, Since death precludes recurrence, censoring deaths would violate the KM estimator assumption that additional follow-up would eventually lead to an event. If your goal is to estimate the probability of recurrence, then you want CI with deaths as a competing risk. The cuminc function in the cmprsk package is a great place to start. Gooley has a great paper on the difference between CI and 1-KM (See Statistics in Medicine, 18, 695-706 (1999)). HTH, Alan Mitchell, MSc Biostatistician alanm at crab.org -----Original Message----- From: array chip [mailto:arrayprofile at yahoo.com] Sent: Monday, June 27, 2011 2:04 PM To: David Winsemius Cc: r-help at r-project.org Subject: Re: [R] cumulative incidence plot vs survival plot Hi David, Thanks for responding, and plain text ...(didn't realized I was in rich text). The endpoint is disease recurrence, I was producing a regular KM plot of recurrence-free probability. Then someone recommend using cumulative incidence is preferred because death was censored in the dataset. I did a little googling, I found CI was used often in the context of competing risk. I am totally new to competing risk and trying to understand what competing risk means and why CI is preferred than KM survival in this context. If you could share your thoughts helping me to understand, greatly appreciated. Searched archive, found people talking about cmprsk package for estimating and plotting CI. would that be the same as the code you suggested: plot(time, cumsum(dead)) Thanks very much! John ________________________________ From: David Winsemius <dwinsemius at comcast.net> Cc: r-help at r-project.org Sent: Mon, June 27, 2011 1:45:35 PM Subject: Re: [R] cumulative incidence plot vs survival plot On Jun 27, 2011, at 4:31 PM, array chip wrote:> Hi, I am wondering if anyone can explain to me if cumulative incidence> (CI) is just "1 minus kaplan-Meier survival"?First tell us what you think CI is defined as. I suspect it is not the same. The KM estimator is cumulative product of (alive-n(dead))/alive so is the product of interval survival probabilities. I doubt that your definition of CI has a similar denominator.> Under what circumstance, you should use > cumulative incidence vs KM survival? If the relationship is just CI > 1-survival, then what difference it makes to use one vs. the other? > > And in R how I can draw a cumulative incidence plot.plot(time, cumsum(dead)) ...?> I know I can make a > Kaplan-Meier survival plot using plot(survfit()), for example: > > fit<-survfit(Surv(time,status)~group,data=data) > plot(fit, col=1:2) > > How to draw CI plot then?As above. Specify what you are seeking. There is a well-defined relationship between S(t) and the cumulative hazard. Maybe you should do a little study of those terms in texts regarding survival analysis.> Thanks very much! > > John > [[alternative HTML version deleted]]Isn't it time you learned to post in plain text? -- David Winsemius, MD West Hartford, CT
Thank you Alan again! Hope you could also share your thought on my another email about the coding of censoring before death.. Thanks again! John ----- Original Message ---- From: alanm (Alan Mitchell) <alanm at crab.org> To: array chip <arrayprofile at yahoo.com>; David Winsemius <dwinsemius at comcast.net> Cc: r-help at r-project.org Sent: Tue, June 28, 2011 11:40:52 AM Subject: RE: [R] cumulative incidence plot vs survival plot>So in cuminc() function, the argument "fstatus" should be coded like:0=censored, 1=event of interest, 2=event of competing risk. Then the function will calculate CI for each of the 2 types of events >(event of interest and event of competing risk), am I correct? Correct.>What about running regular Cox regression for recurrence? any problemthere? for example, need to take into competing risk as well or regular Cox regression is still fine? This is a little trickier. I would strongly suggest reading up on this before doing any analyses. There are a couple of different ways to do this, but the crr function in cmprsk will perform a competing risks regression. Alan Mitchell, MSc Ph: (206) 839-1708 alanm at crab.org
> Hi, I am wondering if anyone can explain to me if cumulative incidence > (CI) is ....The cumulative incidence curve and the KM are not the same, when there are multiple outcomes. See the "etype" argument to survfit, which is used to create CI curves (?survfit.formula). For testing differences between CI curves use the cmprsk library from Gray; it can also draw curves by the survfit routine has a lot more flexibility. Terry Therneau
Note that most of the nonparametric and semi-parametric competing risks analyses can be performed within the survival package. This includes nonparametric estimation of cause-specific cumulative incidence curves and the log-rank type test. It suffices to create a weighted data set as explained in Geskus, Biometrics 67, p. 39-49, 2011. Ronald Geskus Academic Medical Center Amsterdam, the Netherlands>> Hi, I am wondering if anyone can explain to me if cumulative incidence >> (CI) is ....> The cumulative incidence curve and the KM are not the same, when there > are multiple outcomes. See the "etype" argument to survfit, which is > used to create CI curves (?survfit.formula). For testing differences > between CI curves use the cmprsk library from Gray; it can also draw > curves by the survfit routine has a lot more flexibility.-- View this message in context: http://r.789695.n4.nabble.com/cumulative-incidence-plot-vs-survival-plot-tp3628772p3643659.html Sent from the R help mailing list archive at Nabble.com.