Abstract

Return to:

Subject Index
Author Index
Keyword Search

Simple Statistics for Laboratory Data Analysis

James August, CQA
Director of Quality Management
American Biltrite, Inc. Tape Products Division
Moorestown, NJ

The use of statistics in the daily operation of test laboratories tends to be very formal. Six Sigma for Design, Design of Experiments, Gauge R&R, and ANOVA may be used to help design and analyze product and process characteristics in research labs. The use of statistical approaches can provide important preventive actions since many of the common SPC (Statistical Process Control) tools can detect small changes to your process and thus predict changes to your future results. As well, statistical tools are excellent for monitoring the effectiveness of corrective actions.

In traditional production test analysis, one set of test results is reported. The set may include some replications and all data and/or data averages are reported. In a statistical analysis the same data set is subjected to an analysis for validity; trends and significance are reported along with the data. The statistical approach requires a little more effort, but provides considerably more information.

Only a statistical interpretation will yield consistently valid evaluations of test data. But often, the rapid turnaround required in the production quality environment prohibits the use of lengthy, repetitive test processes and analyses. The concept of statistical thinking is embodied in the question: "How do these results compare with the results that were expected from this process?" And, there are several approaches that can help answer this question including the use of simple techniques that don't require major investment in time and energy. Two such tools for analyzing laboratory test data are Chauvenet's Criterion for one-time tests and the X-mR Chart for variable trend analysis.

In the production of pressure sensitive tapes, laboratory evaluation of the coating quality is part of the production process. Performed in real time, it is necessary for technicians to be able to make rapid evaluations of test data. Testing frequently encompasses peel adhesions, thicknesses and coat weights. Because we are dealing with a semi-continuous web process, it is not uncommon to periodically evaluate three or more samples from across the web in order to determine process behavior. Unfortunately, there are two problems with this: the multiple samples are not independent and the sample size may be too small to interpret unambiguously. The inherent statistical problems with small, multiple specimen, samples from a web have been addressed by Frost (1), Roisum (2) and others.