discrete-time survival analyis vs interval censoring

[statlimp]

Dear All,
can anyone please explain what the difference is between discrete time survival analysis and interval censoring.
I'm wondering if discrete time is a special case of interval censoring where the intervals are regular eg once a year and assesment of event occurence occurs at the same timepoints for all subjects ???
many thanks in advance

 

[fed2]

hmmm. I think its mostly a 'bias towards the null' type situation, at least as far as non-parametrics tests go. i think your right about discrete time.

 

[statlimp]

thanks, Fed2, glad I'm not too far wrong. Sorry I'm not 100% sure what you mean by a "'bias towards the null' type situation, at least as far as non-parametrics tests go" .. do you mean that if we analyse discrete time survival by assuming the event toccurs actually at the discrete timepoint where we assess or at a midpoint (rather than by accepting the event really occured between assessment timepoints by using interval censoring) then we bias to the null and lose power ?

 

[hlsmith]

Can you describe what you mean by interval censoring?

Discrete time survival analysis doesn't have the exact time of event but records if the event transpires during a period. So say you send out a cross-sectional survey asking if the event happen during the last month. Given there are actual differences between the independent variable of interest groups - yes you obviously lose information and possibly power, since you are using a proxy for exact time to event.

 

[Miner]

Interval censoring means that the event happened between two known points in time, but the exact timing of the event is unknown. These intervals may be regular in size and not overlap, or they may be irregular in size and/or overlap.

 

[fed2]

Im not real big on interval censoring, but it is sort of 'obvious' that if you have a 'discrete' time situation, where the intervals do not overlap, then the result will be to create ties where there were none before, which would result in greater p-values for rank based tests, like log rank tests.

Here's what the SAS docs say:
The observed intervals might or might not overlap. It they do not overlap, then you can usually use conventional methods for right-censored data, with minor modifications. On the other hand, if some intervals overlap, you need special algorithms to compute an unbiased estimate of the underlying survival function.