Is there a way to test if the distributions of the two samples of censored data? As the data is not defined exactly, Kolmogorov-Smirnov test does not seem to be directly applicable.

Generally speaking, how do you call the department of statistics that works with imprecise (we only know the upper and lower bounds) data? How big the samples should be, as a function of uncertainty on data, to allow sensible analysis? Using the middle of the interval as the precise value seems to lead to bias.

In my case, it is the casualties of war - e.g. in Iran-Iraq war, it is 300,000 - 1,100,00, https://en.wikipedia.org/wiki/Iran%E2%80%93Iraq_War, in WWII it was 8,668,000 - 11,400,000 (Soviet side, https://en.wikipedia.org/wiki/World_War_II_casualties).

  • 1
    $\begingroup$ "Interval analysis" lies at the intersection of computer science and statistics. It concerns working with intervals of data by making no assumptions about how the data might be distributed within those intervals. $\endgroup$ – whuber Feb 25 '20 at 17:50

If only the lower and upper bounds are known for response data, this is referred to as interval censored data. A large number of statistical tools for right censored data have been generalized to interval censored data. Some of the fundamental tools include:

NPMLE (generalization of Kaplan Meier curves, which themselves are a generalization of the Empirical Distribution Function): this is the most basic method for looking the CDF/survival function of a univariate distribution. This can be fit a number of ways in R, I'm going to recommend icenReg's ic_np, since I'm the author.

Log-rank statistic: similar to the log-rank statistic for right censored data, this is one of the basic tools for examining statistical significant differences between two univariate distributions. Implementations of this are available in R's interval package. From the original question, it sounds like this might be the test that most directly answers the OP's question.

Regression models: several of the classic right censored regression models are available. The canonical model is the proportional hazards model, although it's this author's believe that this is due more to historic development of methodology. Similar to the right censored case, the baseline distribution is not necessary to specify. Alternatively, there exist the proportional odds model and accelerated failure time (AFT) model. Personally, I believe the AFT model should be the default unless it fits the data poorly, as the AFT model is easier to interpret. All the models listed can be fit with icenReg (either ic_sp, ic_par or ic_bayes). The AFT model for interval censored data can also be fit with survival's survreg function.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.