| | 8 | * some further parallax tests : I extracted objects with measured parallaxes (not necessarily significant) for |b| > 10.0, with a minimum number of measurements, time range, and parallax factor range, but with no limit on colors or magnitudes. I then looked at the number of positive and negative parallaxes. For the full sample, I (30M objects), there were roughly equal numbers of positive and negative values, saying the false positive rate was ~100% (n.b.: false positive measurements of parallax -- the objects themselves are very likely real because I required at least 10 measurements). restricting the sample based on the quality of the parallax measurements (chisq values, delta-chisq w.r.t. proper motion, parallax signal-to-noise) did not substantially improve the contamination rate, nor did restricting to certain magnitude ranges. However, limiting to significant proper motions and/or high galactic latitude (|b| > 45) increased the real parallax fraction [(pi_pos - pi_neg) / pi_pos] to roughly 40%. My interpretation is that the rate of (non-Gaussian) outlier errors is relatively high (probably ~10%) so that we are very sensitive to an effect equivalent to the Lutz-Kelker bias : background objects are being scattered to (apparently) significant parallaxes and swamping the real objects (since there are so many more background objects). The implication is that, if we want to do blind parallax-based searches, we need to do a better job removing the non-Gaussian outliers. This can probably be done by (a) correcting the Koppenhoffer effect and (b) outlier rejection in the parallax fitting itself (currently not used). |
