Skip to content

Multiply censored observations

2 messages · is@beii@ m@iii@g oii gheme@t@c@, Cade, Brian

Original
is@beii@ m@iii@g oii gheme@t@c@
# 
BODY { font-family:Arial, Helvetica, sans-serif;font-size:12px; }
 Hi everyone, 
 I am working on a project where I need to conduct non-seasonal and
seasonal trend tests of multiply censored data (i.e., data with
multiple detection 
 limits). 
 An old thread on this mailing list
(http://r-sig-ecology.471788.n2.nabble.com/Seasonal-Mann-Kendall-with-multiple-detection-limits-td7578991.html
[1]) 
 suggests that it is possible to perform these types of trend tests in
R using functions from USGS packages: 
        1.  USGSwsQW package has a function called kendallATS.test
that performs a non-seasonal Mann-Kendall trend test;
        2.  restrend package has a function called censSeaken that
performs a seasonal Kendall trend test. 
 Now, I looked for the USGSwsQW package and couldn't find it. 
Instead, I found the USGS package smwrQW
(https://rdrr.io/github/USGS-R/smwrQW/), [2] which 
 includes the function kendallATS.test.  The restrend package I could
find and it does include the censSeaken function. 
 The syntax of the kendallATS.test function is kendallATS.test(x, y,
na.rm = TRUE), where in my case x will stand for year and y will be a
concentration 
 subjected to multiple detection limits.  According to the help file,
y needs to be 'any data that can be converted to a left-censored data
object'. 
 The USGS packages seem to have an as.lcens() function which would
enable me to specify y as a left-censored vector of concentration
values subjected to 
 a single censoring limit.  But they also have an as.mcens() function
which would accept multiple censoring/detection limits.  
 Initially, I thought that kendallATS.test would accept both lcens and
mcens data values.  But it only accepts lcens data values, so I am not
sure 
 whether we are supposed to somehow convert mcens to lcens data values
prior to feeding them to kendallATS.test?  If we are supposed to do
this 
 conversion, I am also not sure how the conversion should work as far
as as.lcens is concerned. 
 As an example, if y consists of the values "

Links:
------
[1]
http://r-sig-ecology.471788.n2.nabble.com/Seasonal-Mann-Kendall-with-multiple-detection-limits-td7578991.html
[2] https://rdrr.io/github/USGS-R/smwrQW/),
[3] mailto:isabella at ghement.ca
[4] http://sitemail.wvthosting.com/<a href=
Cade, Brian
# 
An alternative to consider would be to use the censored quantile regression
option available in the quantreg package.  This handles left or right
censoring with or without multiple censoring values.  You could just
estimate the censored conditional median (0.50 quantile) to provide an
estimate similar to Mann-Kendall trend test, but have the option of
estimating censored estimates for other quantiles that may provide
additional insight to relationships if there is substantial heterogeneity.
You would have to include the necessary predictors and functions to handle
the seasonal adjustments but that should be quite doable.  The censored
quantile regression approach has much greater modeling flexibility
(inclusion of other predictors, splines on predictors, etc) than the
Mann-Kendall testing approach.

Brian

Brian S. Cade, PhD

U. S. Geological Survey
Fort Collins Science Center
2150 Centre Ave., Bldg. C
Fort Collins, CO  80526-8818

email:  cadeb at usgs.gov <brian_cade at usgs.gov>
tel:  970 226-9326
On Mon, Jan 28, 2019 at 1:00 PM <isabella at ghement.ca> wrote: