Our work is motivated by a problem that arises in measuring low concentrations of organic pollutants with gas chromatography - mass spectroscopy (GC-MS). Typically, an environmental sample contains many compounds, most of which are not of interest. The gas chromatograph first isolates what is hoped to be the pollutant of interest from all other compounds in the environmental sample. The mass spectrometer then ionizes the isolated compound, applies tests to the ions to verify that the isolated compound is the pollutant of interest, and converts the amount of the most abundant ion into a concentration. The measured concentration and the pass/fail result from the verification test are recorded by the instrumentation. If the verification test is failed, the measurement is often suppressed and only the word nondetect is reported, because the GC--MS verification tests are imperfect at low concentrations. Chemists believe that all the verified GC--MS measurements belong to the pollutant of interest, but unverified GC--MS measurements may or may not belong to the pollutant. The policy of reporting only ``nondetect'' avoids reporting irrelevant measurements that do not pertain to the pollutant. An obvious question is whether some of the unverified measurements have useful information about the distribution of measurements on the pollutant.
Such data are represented by the following statistical model. Some measurements come from a distribution F that is of interest while others come from another, irrelevant distribution G. Some measurements from F are verified and known to be from F. The other, unverified, measurements may be from F or from G. Not all measurements from F are equally likely to be verified, and no measurement from G is ever verified.
How well a feature T(F) of F can be estimated when there are unverified data depends on what can be assumed about F, G, and the conditional probability v(x) of verifying a measurement of x from F. If F, G and v are unrestricted, then more than one choice of (F, G, v) gives the same distribution p of the observable x, and thus T(F) cannot be uniquely estimated from data. But for some choices of T(F), such as the mean, the set of possible values is a finite interval that is easily estimated. An example using data from a Love Canal study shows that the interval of mle's of T(F) can be sufficiently small to be useful. When v is unrestricted, the mean of just the verified data and the mean of all the data belong to the interval of mle's.
Paper: Postscript