-------- Original Message -------- Subject: Problems interpreting variable results given by different stats and landmark configurations Date: Tue, 23 Feb 2010 16:18:36 -0000 From: Richards Paul <[email protected]> To: <[email protected]> Dear List, I am new to the analyses and statistics underlying geometric morphometrics and so would appreciate your invaluable advice/comments on the methods and findings described below: 1) I am testing for associations between shape and colour in a snail species (dark, intermediate and light groups). A weak association was detected (T-test, p=0.03) using a height/width ratio derived from two linear measurements with calipers; with dark snails, on average, being higher spired and light snails flatter and more globular. With a geometric morphometric approach... Relative warp analysis using 29 landmarks (including sliders) and 16 landmarks (just fixed landmarks) described similar axes and percentages of variation with each other and the h/w ratio. So the 16LM dataset still appears to capture similar shape variation to using 29LM and the simple h/w also seems to provide a fairly good estimate. However, multivariate regression of shape against colour(in TPSregr) returned a weak significant result (p=0.035) for the 29LM dataset, but non-significant result for the 16LM dataset (p=0.28). Using 16LM is of benefit in relation to dealing with smaller sample sizes than we would like. I have also used CVA to explore shape differences between the colour groups, which found similar significant axes of variation for 16LM and 29LM. Do you have any suggestions as to why using 16LM gives such a different result to using 29LM, when the reduced number of variables should provide greater statistical power and a simple h/w ratio can capture the apparent key shape differences? My 16LM landmark coverage should capture the variation encompassed by the h/w ratio and shows ‘identical’ PCA axes to the 29LM dataset. 2) MANOVA (in SPSS) on the weight matrix variables against colour gave a slightly different result to the multivariate regression based on 29LM (MANOVA: wilks lambda: F=1.255, df=108, 716, p=0.051 versus Multivariate Regression: wilks lambda: F=1.418, df=54, 359, p=0.0350, 2000 permutations = 2.8%). I assume both tests have a similar basis, but so should I consider one more valid than the other given my question? Obviously the difference isn’t huge but could still lead to different interpretations of the biology. 3) Additionally, the Goodalls F statistic in TPSregr returned a highly non-significant result compared to wilks lambda – am I correct in understanding that this statistic is used only for dealing with comparisons of shape and size in TPSregr? In summary: why would 16LM give such a different result to 29LM, particularly when two linear measurements appear also to be able to detect the key differences? And am I using the right statistical approaches - if so what might be causing the differences observed between similar approaches (e.g. MANOVA vs multivariate regression) and which test(s) are best to stick with? Many thanks, Paul Richards ---------------- School of Biology University of Nottingham University Park NG7 2RD +44 (0)115 8213128 [email protected] This message has been checked for viruses but the contents of an attachment may still contain software viruses which could damage your computer system: you are advised to perform your own checks. Email communications with the University of Nottingham may be monitored as permitted by UK legislation. -- For now, new message AND replies should be sent to: [email protected] /* Replies will be sent to the list. */ For more information visit http://www.morphometrics.org
