Webinars on Data Analysis with Nondetects

Videos on Statistics for Data with Nondetects (Censored Data)

All videos here are free to view. They 'give yu a taste' of methods covered in our Nondetects and Data Analysis course. You'll now find that course on this website under the Training menu. Taking the course is now free, but doesn't come with any support by email with the insructor. But there's a lot to learn, and its free.

Two changes have occurred since these videos were recorded:
1. Our online training center is now closed. Most of its content for nondetects has been moved to this website.
2. Scripts or software "that come with the course" mentioned in the videos are now in the NADA2 package for R, freely available on the CRAN site https://cran.rstudio.com . There is no need for an extra scripts file.

1. The Cost of Complacency

There is a cost to methods like substituting one-half the detection limit for nondetects. When this is done for multiple observations within the dataset, it affects the standard deviation. The result can be confidence intervals that are too short, trends determined that are not really there, failure to find differences between contaminated and control sites that are really there, and inaccurate no-effect levels. Among other problems. There is a cost to thinking that substitution 'can't be too bad'. This is an intro to why using methods for censored data is important for environmental sciences.

approx. 30 mins

pdf of Powerpoint slides

2. The Mystery of Nondetects: How Censored Data Methods Work

The Mystery of Nondetects attempts to take the mystery out of the methods used for data analysis with nondetects. It demonstrates how methods that do not substitute values for nondetects actually work. This is not widely understood by the environmental science community -- one of the most frequent questions I am asked is "But what number do I put in for the nondetects when I use them?" The answer is "you don't". The reasons why this is so, and how these methods work, will be presented in this webinar.

approx. 60 mins

pdf of Powerpoint slides
Q & A from the webinar

3. Testing Groups of Data with Multiple Detection Limits

Testing differences between groups of data is familiar to scientists -- ANOVA for means and Kruskal-Wallis for cdfs (percentiles). If differences occur, multiple comparison methods determine which groups differ from others. Tests with the same objectives exist that are designed for censored data, data subject to (one or multiple) detection limits. In this webinar you will learn how these tests work and how you can compute them. These tests from the survival analysis discipline avoid substitution of values for the nondetects. All of these methods are available in the NADA2 package for R.

approx. 50 mins

pdf of Powerpoint slides
Q & A from the webinar

4. Fitting Distributions to Data with Nondetects

Which distribution best fits your data? How do you know? Finding the correct distribution can enable better estimates of high percentiles and tolerance intervals, especially with small datasets. Finding the best fitting distribution is very important prior to running parametric regression or group tests. When deciding which distribution fits best, you must always include, and never delete, the nondetects. This video shows you how.

approx. 45 mins

pdf of Powerpoint slides
Q&A from the webinar

5. Correlation and Regression for Data with Nondetects

Test whether one (regression) or multiple explanatory variables significantly affect the response variable. Use numerical criteria to choose which of the multiple variables to use. Perform a nonparametric straight line model, the Akritas-Theil-Sen line. Compute several types of correlation coefficients. You can do it all, without substituting fabricated values.

approx. 56 mins

pdf of Powerpoint slides
Q & A for this webinar

6. Trend Analysis for Data with Nondetects

Are concentrations, some of which are nondetects, changing over time? Can I tell even when there are multiple detection limits used? Here parametric and nonparametric trend methods for data with nondetects are demonstrated, including seasonal regression and the Seasonal Kendall test for trend. Substitution with 1/2 DL or similar numbers should be avoided -- it can obscure trends that are present, or put in a trend that is not there in the field.

approx. 60 mins

pdf of Powerpoint slides
Q & A for the webinar

7. Incorporating Greater Than and Less Than Values in Data Analysis

One way of representing censored data in a database is the "interval endpoints" format. Two columns are used with the first being the low end of possible values for the variable (often 0 for censored chemical data) and the second column holding the highest possible values (the detection or quantitation limits). One benefit of storing data this way is that it allows 'greater thans' to also be stored in the same two columns. Most censored methods for data analysis can incorporate both 'less thans' and 'greater thans' as interval-censored data and compute everything from means to hypothesis tests and regression. This webinar will give you examples of how to do these types of analyses.

approx. 50 mins

pdf of Powerpoint slides
Q & A for the webinar

8. Matched Pair Tests for Data with Nondetects

Paired differences such as concentrations Before/After a treatment at a series of sites or Upstream/Downstream of an outfall at several points in time are usually analyzed for change with a paired t-test (parametric) or signed-rank test (nonparametric). But what methods can be used when data include values below detection or quantitation limits without substitution of numbers for nondetects? This video will present what methods are available and demonstrate their use with R software.

approx. 35 mins

pdf of Powerpoint slides

9. NADA2: Everything You Can Do With Nondetects

What statistical analyses can you do today for data with nondetects without substituting numbers like ½ the detection limit? It is essentially every analysis you do when there are no nondetects. Our NADA online course teaches you to use the NADA2 package for R, including routines for drawing boxplots, scatterplots with fitted models, and probability plots to determine how well a standard distribution such as the normal, lognormal or gamma fit the data. You can compute prediction and tolerance intervals, and perform hypothesis tests (parametric, nonparametric and permutation varieties). Follow that up with multiple comparison tests to determine which groups differ from others. You can compute correlation coefficients, build and evaluate regression models to find the best model. You can perform trend analysis such as the seasonal Kendall test while adjusting for the effect of exogenous variables that are not time. You can even compute multivariate procedures such as cluster analysis, NMDS plots, PCA (Principal Components Analysis) and multivariate group and trend tests. All without substitution for nondetects. Take a tour of what this R package can do.

approx. 53 mins

pdf of Powerpoint slides
Q & A for the webinar