Trinity College Dublin Diploma in Statistics Computer Laboratory 2 (1) Measurement System Analysis and (2) Analysis of a Split Plot experiment Invitations to consider the results of Minitab analysis and their statistical and substantive interpretations are printed in italics. Take some time for this; consult your neighbour or tutor. Enter your responses in a Word document, as draft contributions to a report on the experiment and its analysis. Learning Objectives On completion of this laboratory, students should be able to use the General Linear Model command to calculate an analysis of variance for a balanced replicated fully crossed random effects layout, report on patterns of variations in one way and two way tables of means and corresponding factor plots, explain the Minitab "Expected Mean Squares" output, relate it to components of variance and the analysis of variance, explain basic quantities in measurement system capability, repeatability and reproducibility and estimate them from a relevant analysis of variance, explain the make of a split plot experiment, explain the make up of the analysis of variance table associated with a split plot experiment, provide a comprehensive report on the analysis of results of a split plot experiment with tables and graphs, analysis of variance, comparisons of treatments with control and pairwise comparisons of treatments. Exercise 1; Measurement System Analysis Basic requirements of a measurement system are that it be accurate, or unbiased, and precise. A measurement system is accurate if repeated measurements are centred in the right place and precise if the spread of the repeated measurements is small. These definitions recognise that measurement processes are subject to variation. The purpose of measurement system analysis is to classify and quantify the extent of such inaccuracy and imprecision. It is important to recognise that measurement variation may be influenced by a range of factors, including the instrument itself, the user of the instrument, materials used as part of the measurement process, the laboratory in which the measurement is made, the environment, and others. This implies that we should assess the precision of the instrument itself, repeatability, and also how that precision varies in varying circumstances, reproducibility. It is also important to recognise that the objects being measured are themselves subject to manufacturing variation. This imposes two requirements on a measurement process, that it is capable of distinguishing differing objects and that it is valid across the range of variation of the objects. In a situation where measurements are made on manufactured parts where the standard deviation of manufacturing variation is P, using an instrument whose error standard deviation is E, measurement system capability may be expressed as T (P ) E , where T(P) = 2P 2E , referred to as Total Parts standard deviation. Repeatability may be expressed simply as E. Where several operators make the measurements and variation between them has standard deviation O, reproducibility of the measurement system may be expressed in terms of 2 O2 OP 2E , referred to as Total Reproducibility standard deviation, which reflects the overall variation of the measurement process including instrument and operator variation, or as 2 O2 OP which reflects the additional variation associated with the use of different operators, additional to repeatability variation. In practice, different definitions derived from these are recommended by different standards organisations; see Measurement Notes on the course web page for more details and some references. Application When fixing floor covering made of fibrous materials to a floor using suitable glue, a key determinant of the strength of the bond is the typical angle of incidence of fibres to the floor. Varying approaches to the measurement of such angles are available. In selecting a suitable approach, knowledge of the precision of such methods is important. In a study designed to quantify the precision of one such method, each of four analysts made eight measurements of the fibre angle of each of five sections cut from a piece of floor covering that had been fixed to a sheet of base material. The results follow overleaf and may be copied from Fibre angles.xls on the course website or mstuart's get folder. Use the General Linear Model option from the ANOVA section of the Stat menu to calculate a full analysis of variance, with Analyst and Section as random effects. To facilitate study of main effects and interactions, choose to display least squares means corresponding to the term Section*Analyst and factor plots corresponding to the terms Section Analyst Section* Analyst. Also, choose to display expected mean squares and variance components. page 2 Section Analyst 1 Analyst 2 Analyst 3 Analyst 4 1 10, 20, 20, 23 20, 20, 20, 15 20, 25, 17, 22 23, 15, 23, 20 20, 19, 15, 16 20, 19, 12, 14 10, 10, 10, 5, 5, 5, 2 15, 17, 20, 20 10, 15, 15, 15 15, 13, 5, 10 8, 8, 10, 12 15, 20, 14, 16 13, 20, 15, 15 10, 10, 10, 10 10, 15, 15, 10 3 23, 20, 22, 20 25, 22, 20, 23 20, 23, 20, 20 23, 23, 22, 20 15, 20, 22, 18 15, 20, 16, 20 10, 10. 10, 15 15, 10, 10, 10 4 15, 16, 22, 15 15, 15, 22, 17 20, 22, 18, 23 23, 23, 24, 20 13, 13, 15, 20 11, 20, 13, 15 5, 10, 10, 10 10, 10, 10, 10 5 20, 20, 22, 20 27, 17, 20, 15 18, 20, 18, 23 20, 20, 18, 15 10, 14, 17, 12 11, 10, 15, 10 5, 10, 10, 10 10, 10, 10, 10 5 5 Make the table directly Copy the output on least squares means into a new worksheet and make a two way table; use Stat / Tables / Descriptive Statistics, clear summaries for "Categorical Variables" and choose to display Means of "Associated Variable" Angle. Copy your factor plots and two way table into a Word document, for coordinated examination. Report on patterns of variation in Analyst means, Section means and Section* Analyst means, as seen in factor plots, with numerical highlights from the table. Interpret the "Expected Mean Squares" output in terms of A, S and AS. Explain the basis for calculating the F-ratios in the analysis of variance table. Interpret the analysis of variance table, with reference to your earlier report, editing your report as appropriate. Calculate Repeatability standard deviation, E and Total Reproducibility standard deviation, 2A 2AS E2 . Comment on their relative magnitude. Calculate Total Parts standard deviation, T(P) = 2S E2 . Comment on its magnitude relarive to Repeatability standard deviation, E. Exercise 2; Split Plot analysis A classic split plot experiment was carried out on the Cambridge University Farm in 1931. The original plan was to investigate the effects of two novel cultivation treatments of grass-land pasture, by comparison with no treatment. The novel treatments were use of a grassland "Rejuvenator" (R) and use of a conventional harrow (H). These treatments were compared with no treatment, taken to be a control. The experiment was carried out on a field of old pasture, laid out in plots consisting of strips 4 yards wide and 45 yards long. The treatments were applied to a block of 3 adjacent plots, in a randomly chosen order. The experiment was replicated in a total of 6 independently randomised blocks, placed side by side. In addition to the randomised blocks layout for investigating treatment effects, each 45 yard plot was subdivided into 4 subplots, each of which was treated with one of four fertilisers, Farmyard manure (F), Straw (S), Artificial fertiliser, (A) and no treatment, acting as Control (C). The fertilisers were also applied in random order to the subplots within each whole plot, independently. page 3 Yield was recorded in pounds (lbs) of green produce from a single cut of each plot made on June 31, 1931. The data follow and are available in GrassLand.xls on the course website or mstuart's get folder. The main plots correspond to the columns in the table, the subplots to the entries within each column. The organisation of the data in subplots within whole plots, with the independent randomisations within whole plots, means that the comparison of treatments C, H and R is assessed by reference to the variation at whole plot level, effectively by way of a randomised blocks analysis of treatments C, H and R in blocks 1 to 6., while comparison of fertilisers, including their interactions with treatments (and, possibly, blocks) is assessed by reference to the variation at subplot level. A C F S C 266 165 198 184 H 213 127 180 127 R 208 155 200 150 C 210 150 247 188 H 222 167 203 167 R 266 163 228 157 C 220 155 190 140 H 184 118 168 128 R 184 153 174 141 C 216 159 225 174 H 178 125 149 107 R 207 135 162 113 C 202 147 184 154 H 175 118 175 112 R 184 98 144 113 C 169 132 164 116 H 142 104 145 89 R 151 69 116 101 The analysis will be reported in two stages, with the randomised blocks layout of cultivation treatments analysed initially, followed by the full split plot analysis. Reports will be in terms of tables of means and corresponding plots with significance tests and confidence intervals, as we did for the measurement system analysis. To set up the randomised blocks analysis, first calculate the whole plot yields. To do this, copy the data to Minitab, then use Data / Unstack1 on the Yields column using subscripts in the Fertiliser column, then add the unstacked columns to each other, using the Calculator in the Calc command. Since each whole plot yield is the sum of the corresponding subplot yields, this results in a column with the whole plot yields. Next, we need the appropriate Block and Treatment codes in two further columns. The first three entries in the whole plot yield column are the whole plot yields of C, H and R, respectively, in Block 1. Succesive sets of three whole plot yields are the same for successive blocks. Thus, we need a column of Block numbers, 1 to 6, in sets of 3, and a column of succesive sets of C, H, R. Do get the block numbers column, use the Make Patterned Data command in the Calc menu. To get the treatment codes column, enter C, H and R in the first three cells and copy toi the remaining cells (down to Row 18). Next, use the General Linear Model command to do the analysis of variance, with model including the new block and treatment variables as main effects; recall that the Block*Treatmentc interaction serves as the error term in a simple ransomised block experiment. Include factor plots. Use the Tables command to make a two way table, including marginal means (see Options). Provide a comprehensive report on the randomised blocks analysis of treatment effects in blocks, with table and graphs, analysis of variance, comparisons of treatments with control and pairwise comparisons of treatments. 1 To understand the Unstack command, first look at the Stack command; it "stacks" several columns into one big stack of numbers, that is, one big column. Unstack reverses this process, taking a single column and putting its numbers into several columns. The "subscript" column tells you which numbers go in which column. In this case, we unstack the Yield column using the "subscripts" in the Fertiliser column. Thus, all the numbers in the Yield column corresponding to Fertiliser code A go in one unstacked column, all those corresponding to code F go in another unstacked column, etc. The result is that we get four columns corresponding to the four fertilisers, A, C, F and S. In fact, they form the transpose of the table above. page 4 Next, calculate the split plot analysis of the original data; first list all relevant main effects and interactions to help decide on the Model terms to enter. Enter Block as a Random Effect, to force Minitab to calculate appropriate error terms. Compare the randomised blocks analysis with the top half of the split plots analysis. What is the connection? Why? Provide a comprehensive report on the split plots analysis, with tables and graphs, analysis of variance, comparisons of treatments with control and pairwise comparisons of treatments. This ends the formal laboratory. Review the Learning Objectives. Have they been achieved? page 5