2.4. Statistical analysis
We used a nested general linear model (GLM) to reflect the natural
clusters within the data set at three different hierarchical levels: country,
province and village; and to determine which factors are important in
determining the magnitude of the dependant variable (Crawley, 1993).
The data were analysed using the SPSS package for windows, version 12.
The distributions of total surface area for each system and of the
household catch were positively skewed, therefore data were logtransformed
before being analysed. The association between the villages
and system-categories was explored for each country separately using
Chi-square tests. To fulfil the test assumptions (less than 20% of the
expected count b5) we had to combine the different types of ponds.
Three villages responsible for low counts were removed from this
particular analysis. Yangnoi (Thailand) and Svay Cheak (Cambodia)
contained only farmers with ponds, whereas Hoang Nguyen (Vietnam)
contained only farmers without ponds. To test whether removing the
three villages influences the results, we ran a Fisher's exact test. All tests
were regarded as statistically significant at pb0.05.