predicting duration
Locked
 
|
Given time to event data, I think of survival analysis (use a hammer if all you know is "hammer"). I need to compute a predicted time to event given a set of predictors. My reading of Singer and Willett (ALDA) is that group level hazard or survival curves can be computed and median survival points computed.
Is there another alternative to analyzing to time to event data? Thanks, Gene Maguin ===================== To manage your subscription to SPSSX-L, send a message to [hidden email] (not to SPSSX-L), with no body text except the command. To leave the list, send the command SIGNOFF SPSSX-L For a list of commands to manage subscriptions, send the command INFO REFCARD |
|
|
While my background training involved courses using the same book you reference (including a course taught by singer herself), it's the only good way I know of. I mean, anything I can think of would just be another hazard model.
The warning I got was, once you get a new hammer, everything's a nail. Still, the brad hammer we have sure seems right for the brads you want to use, so to speak. This certainly isn't a screw problem. Matthew J Poes Research Data Specialist Center for Prevention Research and Development University of Illinois 510 Devonshire Dr. Champaign, IL 61820 Phone: 217-265-4576 email: [hidden email] -----Original Message----- From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Maguin, Eugene Sent: Monday, October 01, 2012 10:55 AM To: [hidden email] Subject: predicting duration Given time to event data, I think of survival analysis (use a hammer if all you know is "hammer"). I need to compute a predicted time to event given a set of predictors. My reading of Singer and Willett (ALDA) is that group level hazard or survival curves can be computed and median survival points computed. Is there another alternative to analyzing to time to event data? Thanks, Gene Maguin ===================== To manage your subscription to SPSSX-L, send a message to [hidden email] (not to SPSSX-L), with no body text except the command. To leave the list, send the command SIGNOFF SPSSX-L For a list of commands to manage subscriptions, send the command INFO REFCARD ===================== To manage your subscription to SPSSX-L, send a message to [hidden email] (not to SPSSX-L), with no body text except the command. To leave the list, send the command SIGNOFF SPSSX-L For a list of commands to manage subscriptions, send the command INFO REFCARD |
|
|
In reply to this post by Maguin, Eugene
I don't have any analysis in mind, but here is a consideration --
I always feel a bit itchy when someone says, "time to an event", without being clear that it is a statistical expectation following a constant risk. Or some other risk-function. There are several distinct models out there, and other models that are mixtures. 1. Infection: It takes several days for the virus/bacteria to thoroughly colonize to cause disease. The disease-event is expected within a certain window of time. The risk-function is high within the window, otherwise low. 2. Relapse: There can be a "constant-risk" for several years, where there is (say) a 2% attrition at every month. There will be fewer new relapses in year 3 or year 5 *only* because the sample N is a fraction of the original N. 3. Aging: There can be (3a) an increasing risk. For human aging, the daily mortality rate doubles for every nine years or so after the age of 30. The risk function can be described as U-shaped since it is very high and (3b) declining in the first weeks of life. Survivorship curves can be useful for all these, but you should not think entirely in the same way about their times-to-event. -- Rich Ulrich > Date: Mon, 1 Oct 2012 11:54:55 -0400 > From: [hidden email] > Subject: predicting duration > To: [hidden email] > > Given time to event data, I think of survival analysis (use a hammer if all you know is "hammer"). I need to compute a predicted time to event given a set of predictors. My reading of Singer and Willett (ALDA) is that group level hazard or survival curves can be computed and median survival points computed. > > Is there another alternative to analyzing to time to event data? > >... |
Re: predicting duration
|
|
In reply to this post by Maguin, Eugene
On 10/1/2012 10:54 AM, Maguin, Eugene wrote:
> Given time to event data, I think of survival analysis (use a hammer > if all you know is "hammer"). I need to compute a predicted time to > event given a set of predictors. My reading of Singer and Willett > (ALDA) is that group level hazard or survival curves can be computed > and median survival points computed. > > Is there another alternative to analyzing to time to event data? If there is no censoring in your data (everyone experiences the event and you know the exact time of the event for everyone), you can compute a mean or median time without Kaplan-Meier or Cox proportional hazards regression or any of that other fun stuff. If there is censoring, then there are also some parametric alternatives. You need to choose a time to event distribution (Weibull is a common choice). Usually these parametric approaches are not too much different, though. Steve Simon, [hidden email], Standard Disclaimer. Sign up for the Monthly Mean, the newsletter that dares to call itself average at www.pmean.com/news ===================== To manage your subscription to SPSSX-L, send a message to [hidden email] (not to SPSSX-L), with no body text except the command. To leave the list, send the command SIGNOFF SPSSX-L For a list of commands to manage subscriptions, send the command INFO REFCARD |
|
|
In reply to this post by Rich Ulrich
Rich: I was using ‘time to event’ both as a casual shorthand but also to try to elicit something other than either a discrete time or continuous time survival model. For instance, it occurred to me, perhaps incorrectly, that a survival analysis is ‘like’ a poisson limited to a count of one rather than 1 to n. What I’m looking for is the ability to get a predicted time as one would get a predicted value from a multiple regression. I imagine that predicting time to the focus event is of great interest in a number of subject areas in business and science. Steve Simon: I suppose my data are unusual. The data are discharges from a residential treatment facility. Every kid admitted gets discharged and the admit and discharge dates are known to the day. So, in that respect time is just a continuous DV, probably not normally distributed but transformable. The complication is that kids are discharged to one of three levels of care, which also matters. So, I think the true model is a competing risks survival model. You mentioned regression and I’ve thought about this problem as a normal regression and treating care level as a predictor rather than an outcome. The practical question is how closely would the two models (normal regression and CR survival) track each other and I don’t know that. Gene Maguin From: Rich Ulrich [mailto:[hidden email]] I don't have any analysis in mind, but here is a consideration -- > Date: Mon, 1 Oct 2012 11:54:55 -0400 |
|
|
Gene, Doesn’t the scenario you describe involve censored data? I have to imagine that people enter treatment at different times and exit treatment
at different times. In addition, some won’t have left treatment at the conclusion of your data collection. These would be censored cases. I don’t see how anything but a hazard model would work here. In addition, at one point would some of these other suggested
approaches be either completely or somewhat mathematically equivalent to the cox model? I.e. you get so creative you reinvent the wheel? I did an analysis looking at retention in a home visiting program for at risk young mothers, in which we tried to see if program implementation
factors could predict a change in a mother’s risk for leaving (Mothers could leave early, be discharged, have a service extension if they were young enough). When I first ran these results we had the same continuous time variable you describe, and a log transformation
normalized it. I ignored censored cases and ran a linear regression (at the time it was because I didn’t know how to run a cox regression model). The peer review ripped it apart. A second analysis then involved a cox regression, and a comparison of results
yielded similar generalized conclusions, but obviously with hazard ratio’s around implementation, and probabilities of leaving early for individuals added the outcomes. When I referenced in the paper the earlier regression, the reviewer response essentially
stated the earlier analysis was of no merit to reference and to focus only on the new analysis. I think I’m just too young in this field to argue if this was a fair assessment or not, just my experience. I actually am evaluating a much larger set of home
visiting programs now and have the ability to do this again and do it right (ten years later my knowledge is much greater, my ability is better, and my sample will be sufficient to run such an analysis). Hopefully this time it will lead to a publication.
Matthew J Poes Research Data Specialist Center for Prevention Research and Development University of Illinois 510 Devonshire Dr. Champaign, IL 61820 Phone: 217-265-4576 email:
[hidden email] From: SPSSX(r) Discussion [mailto:[hidden email]]
On Behalf Of Maguin, Eugene Rich: I was using ‘time to event’ both as a casual shorthand but also to try to elicit something other than either a discrete time or continuous time survival model.
For instance, it occurred to me, perhaps incorrectly, that a survival analysis is ‘like’ a poisson limited to a count of one rather than 1 to n. What I’m looking for is the ability to get a predicted time as one would get a predicted value from a multiple
regression. I imagine that predicting time to the focus event is of great interest in a number of subject areas in business and science.
Steve Simon: I suppose my data are unusual. The data are discharges from a residential treatment facility. Every kid admitted gets discharged and the admit and discharge
dates are known to the day. So, in that respect time is just a continuous DV, probably not normally distributed but transformable. The complication is that kids are discharged to one of three levels of care, which also matters. So, I think the true model is
a competing risks survival model. You mentioned regression and I’ve thought about this problem as a normal regression and treating care level as a predictor rather than an outcome. The practical question is how closely would the two models (normal regression
and CR survival) track each other and I don’t know that. Gene Maguin From: Rich Ulrich
[hidden email] I don't have any analysis in mind, but here is a consideration -- > Date: Mon, 1 Oct 2012 11:54:55 -0400 |
|
|
Normally, it would. I agree. However, the dataset is kids admitted and discharged over a 6 year period. Kids who were in treatment at the start date were deleted and kids who had not been discharged by the end date were also deleted. In the conventional NIH study, this would never be done but here, the data is simply a segment of time from an ongoing process. It is believed by the agency that during this time the process was stable. If it wasn’t, the results won’t be very useful; if it was, the results may be useful. Also unlike a conventional NIH (or any other funded study), the predictors, which are all intake characteristics, will be interesting but that is not the reason for the analysis. Gene Maguin From: Poes, Matthew Joseph [mailto:[hidden email]] Gene, Doesn’t the scenario you describe involve censored data? I have to imagine that people enter treatment at different times and exit treatment at different times. In addition, some won’t have left treatment at the conclusion of your data collection. These would be censored cases. I don’t see how anything but a hazard model would work here. In addition, at one point would some of these other suggested approaches be either completely or somewhat mathematically equivalent to the cox model? I.e. you get so creative you reinvent the wheel? I did an analysis looking at retention in a home visiting program for at risk young mothers, in which we tried to see if program implementation factors could predict a change in a mother’s risk for leaving (Mothers could leave early, be discharged, have a service extension if they were young enough). When I first ran these results we had the same continuous time variable you describe, and a log transformation normalized it. I ignored censored cases and ran a linear regression (at the time it was because I didn’t know how to run a cox regression model). The peer review ripped it apart. A second analysis then involved a cox regression, and a comparison of results yielded similar generalized conclusions, but obviously with hazard ratio’s around implementation, and probabilities of leaving early for individuals added the outcomes. When I referenced in the paper the earlier regression, the reviewer response essentially stated the earlier analysis was of no merit to reference and to focus only on the new analysis. I think I’m just too young in this field to argue if this was a fair assessment or not, just my experience. I actually am evaluating a much larger set of home visiting programs now and have the ability to do this again and do it right (ten years later my knowledge is much greater, my ability is better, and my sample will be sufficient to run such an analysis). Hopefully this time it will lead to a publication. Matthew J Poes Research Data Specialist Center for Prevention Research and Development University of Illinois 510 Devonshire Dr. Champaign, IL 61820 Phone: 217-265-4576 email: [hidden email] From: SPSSX(r) Discussion [hidden email] On Behalf Of Maguin, Eugene Rich: I was using ‘time to event’ both as a casual shorthand but also to try to elicit something other than either a discrete time or continuous time survival model. For instance, it occurred to me, perhaps incorrectly, that a survival analysis is ‘like’ a poisson limited to a count of one rather than 1 to n. What I’m looking for is the ability to get a predicted time as one would get a predicted value from a multiple regression. I imagine that predicting time to the focus event is of great interest in a number of subject areas in business and science. Steve Simon: I suppose my data are unusual. The data are discharges from a residential treatment facility. Every kid admitted gets discharged and the admit and discharge dates are known to the day. So, in that respect time is just a continuous DV, probably not normally distributed but transformable. The complication is that kids are discharged to one of three levels of care, which also matters. So, I think the true model is a competing risks survival model. You mentioned regression and I’ve thought about this problem as a normal regression and treating care level as a predictor rather than an outcome. The practical question is how closely would the two models (normal regression and CR survival) track each other and I don’t know that. Gene Maguin From: Rich Ulrich [hidden email] I don't have any analysis in mind, but here is a consideration -- > Date: Mon, 1 Oct 2012 11:54:55 -0400 |
|
|
In reply to this post by Maguin, Eugene
"Predicting time to a focus event" may be of less interest than
you imagine, since it is mainly useful when there is an increasing risk. Otherwise, for any sort of "constant risk" model, even if you spot factors that elevate the risks, the predictions for individuals are often so poor that they not very useful. Perhaps, you may hope to predict well enough that it would be fair to use the predictions to dictate more intensive treatment. As to the question of discharge, to 3 levels of care -- The best predictor of "discharge" in my own experience, from several decades ago, was the existence of somewhere for this patient to go, and someone willing to accept them. I think, with your data, I might first examine whether there were systematic differences between the three levels, regardless of time, in order to decide whether these levels should be collapsed. -- Rich Ulrich Date: Mon, 1 Oct 2012 14:31:35 -0400 From: [hidden email] Subject: Re: predicting duration To: [hidden email] Rich: I was using ‘time to event’ both as a casual shorthand but also to try to elicit something other than either a discrete time or continuous time survival model. For instance, it occurred to me, perhaps incorrectly, that a survival analysis is ‘like’ a poisson limited to a count of one rather than 1 to n. What I’m looking for is the ability to get a predicted time as one would get a predicted value from a multiple regression. I imagine that predicting time to the focus event is of great interest in a number of subject areas in business and science.
Steve Simon: I suppose my data are unusual. The data are discharges from a residential treatment facility. Every kid admitted gets discharged and the admit and discharge dates are known to the day. So, in that respect time is just a continuous DV, probably not normally distributed but transformable. The complication is that kids are discharged to one of three levels of care, which also matters. So, I think the true model is a competing risks survival model. You mentioned regression and I’ve thought about this problem as a normal regression and treating care level as a predictor rather than an outcome. The practical question is how closely would the two models (normal regression and CR survival) track each other and I don’t know that.
Gene Maguin ...[snip, previous] |
|
In reply to this post by Maguin, Eugene
One, possibly peripheral, issue:
At 02:31 PM 10/1/2012, Maguin, Eugene wrote, >The data are discharges from a residential treatment facility. Every >kid admitted gets discharged and the admit and discharge dates are >known to the day. So, in that respect time is just a continuous DV, >probably not normally distributed but transformable. I wouldn't expect the lengths of stay to be anything like normally distributed. The common way to get a normally distributed quantity is as a sum or mean of a number of (reasonably) independent factors with (reasonably) similar variances. (The factors themselves may well not be observable. For example, if how tall people are is approximately normally distributed, it's fair to guess that it's the result of a large number of genetic factors affecting height.) I wouldn't think LOS would be determined by anything of the kind. But, why do you care? Very few analytic methods require a normally-distributed independent variable. And most of all, why would you consider transforming to make the distribution more normal? Unless there's a theoretical basis for the transformation you use, the transformed variable is an artificial construct, and any results from analyzing it are very hard to interpret. Apologies if I'm making too much of this, raising a question about a remark you may have tossed off casually and never considered important. -Best wishes, Richard ===================== To manage your subscription to SPSSX-L, send a message to [hidden email] (not to SPSSX-L), with no body text except the command. To leave the list, send the command SIGNOFF SPSSX-L For a list of commands to manage subscriptions, send the command INFO REFCARD |
| Free forum by Nabble | Edit this page |