Skip to main content

Timeline for Detecting initial trend or outliers

Current License: CC BY-SA 3.0

11 events
when toggle format what by license comment
Oct 3, 2012 at 7:18 vote accept Toby Kelsey
Oct 1, 2012 at 23:59 comment added Michael R. Chernick @IrishStat I would only do that because I believe any effective formal statistical method would not do better than a heuristic when the sample size is only ten and the first 2 to 4 observations are very likely to be outliers.
Oct 1, 2012 at 22:19 comment added IrishStat @Michael, It is always possible to supress the deteftion of trends and variance changes thus effectively applying Bayes Theorem. I believe that you had your comments reversed regarding my two ridiculous examples. Also the way I read your original comment was to throw away the first 4 values of the 10 observations in the (vain ?) hope that would be a good rule of thumb. THis could be thrpwing out the baby with the bathwater.
Oct 1, 2012 at 21:10 comment added Wayne @MichaelChernick: Would reversing the time series and doing CUSUM be helpful? We'd know that the first six values are good, so it's a matter of whether things go bad "after" that.
Oct 1, 2012 at 21:10 comment added Michael R. Chernick @IrishStat I hope that I didn't take your comment too seriously. It was hard for me to discern how much of it was tongue in cheek.
Oct 1, 2012 at 21:09 comment added Michael R. Chernick @IrishStat It is not the first ten it is the only 10! Both of you examples indicate a linear trend. The second has some noise (maybe) while the first does not. But a trend is a problem for him. It may have a trend or it may have a high variance before it stablizes. All those points are suspect according to the OP. This is a problem that defies statistical analysis because the sample size is far too small to detect one or both of these anomalies and identify the point of stability. A good physical model with time series analysis might work if there were just a little more data available.
Oct 1, 2012 at 20:45 comment added IrishStat @michael No I don't concede that. If you have two points 5 and 10 and the next point is 20, I am willing to question the "20". If i have three points 5,10,15 , I have mo problem. Time series methods are the answer to a maiden's prayer regarding the "general term" i.e. given a sequence , how to generalize (forecast ) the sequence. Intervention DEtection is simply a ploy to separate the regular (values descibable by the general term) and thos that are not. To suggest that one throw away 4 of the first 10 is to me a minor blasphemy ( jsu kidding ! ).
Oct 1, 2012 at 20:18 comment added Wayne @MichaelChernick: I, too, wonder if he shouldn't just drop the first four measurements all the time. Obviously, if things "settled down" after two measurements for a particular test, he'd be throwing away 25% of his real data, but trying to fit a model to 10 points (of which up to 40% are outliers) seems impossible. Perhaps he needs to expand and formalize how it is he decides "by eye" what's "settled down" and what's not.
Oct 1, 2012 at 19:58 comment added Michael R. Chernick @IrishStat I will give you that but do you concede that the series may be too short for time series analysis to work very well?
Oct 1, 2012 at 19:57 comment added IrishStat If it is time series data then it is a time series opportunity ....But that's just my opinion. Since I own a hammer , everything looks like a nail !
Oct 1, 2012 at 18:31 history answered Michael R. Chernick CC BY-SA 3.0