APPENDIX B 



TEST PROGRAM - FIBERGLASS POLYESTER LAMINATES 

 STATISTICAL ANALYSIS OF TEST DATA 



By Mr. C. Daniel 

 Consulting Engineering Statistician 



This Appendix is presented for statisticians and gives an explanation of the procedures 

 used in obtaining the results, and a discussion of some of the problems encountered in 

 analyzing the test data. 



The following types of laminates were evaluated: 



Ml - Mat - 2 ounces per square foot 



M2 - Woven Roving - 25-27 ounces per square yard 



M3 - Cloth - 10 ounces per square yard 



M4 - Mat with 1 ply of 10 ounce cloth on each face 



M5 - Woven Roving with 1 ply of 10 ounce cloth on each face 



Original Design and its Modifications 



The plan first proposed was a Graeco-Latin Cube, sometimes also called a quarter 

 replicate of a 4 4 factorial experiment. The four factors were: Material, Fabricator, 

 Thickness and Testing Laboratory. 



To this plan were later added: 



Four thicknesses of material Ml to form a sequence of five thicknesses 

 for this material. 



16 duplicate panels arranged in a Graeco-Latin square. 



Duplicate specimens in every panel for every property. 



Discussion of Original Design 



The general plan used (quarter replicate of a 4^) was well adapted to estimate differences 

 among materials, among fabricators, etc. It was not well adapted to measuring the differing 

 variabilities of the several materials. 



The testing of duplicate panels was not properly randomized. Thus the use of the panel- 

 to-panel observed variability as a measure of error in the analysis of variance is excluded. 



The use of duplicate specimens was largely unnecessary for the estimation of the smallest 

 component of variance. The interactions with angle (within a panel) generally check the speci- 

 men standard deviation quite closely with ample degrees of freedom. When the two do not 

 check, the angle interaction estimate is preferable. 



B-l 



