 |
 |
Dear list, Thanks to everyone who responded to my query, it was really helpful. A summary of responses: Everyone who responded said that it was very much valid to statistically test for interactions if one of the main effects was non-significant (this was the basis of my initial query - just the naive idea that it might not be ok for some reason) Several cited the case (which is also the case in my data), where the main effect of var1 is significant in the presence of sex only. This is because the effect of var1 is positive for girls and negative for boys and so when the sex-adjusted main effect is tested with no interaction these correlations cancel each other out. A second suggestion which also came up referred to the possibility that an adjusted main effect could be nonsignificant, but the interaction would highlight that the association was significant for one group but not the other. However, perhaps the most important responses for me highlighted that implicit in my question was that I was basing a decision of what to test in my analysis based on the statistical significance or not of a coefficient. All three respondents who raised this pointed out the dangers of this and emphasised the importance of making analytical decisions based on logical and / or theoretical concerns. So thanks to everyone for your thoughts, I hope I've done justice to your responses. Jamie -- Dept. of Geography, Queen Mary, University of London Mile End Rd E1 4NS Tel: 020 7882 5400 -- Dept. of Geography, Queen Mary, University of London Mile End Rd E1 4NS Tel: 020 7882 5400