Hi, along with others, I have ran a study to determine if the disclosure of genotype (risk/non-risk) impacts adherence to dietary recommendations for the reduction of salt. We have collected a lot of data on subjects including: anthropometric, inter-arm difference and average blood pressure, genotype, assessed genetic understanding and attitudes, and assessed dietary salt intake by dietary recalls and 24-h urine collections pre-disclosure and 4-weeks after. We determined ‘adherence’ as a greater reduction in dietary salt after 4-weeks.
I was convinced the statistical tests we would use would be a manova, however the majority of the variables mentioned above are abnormally distributed (Shapiro-Wilk, univariate test of normality).
Although, another question on here indicated that univariate tests of normality may not be applicable when using them in a manova.
Secondly, I came across a question where someone suggested if the abnormal distribution did impact the manova a Munzel-brunner method in comparison to the Friedman’s two-way anova would be the better option.
I am looking for opinions/recommendations as to my next steps as I have very limited statistical knowledge.
Thanks in advance.
Your interest is about adherence (to dietary recommendations). How do you measure adherence? This should be a single variable. The rest you describe seem to be (possible) explanatory variables. Their distribution is not relevant to the probelm. What is of some importance is then only the conditional distribution of adherence (conditional on the explanatory variables). The assumption that this conditional distribution should be normal can simply be cheked by diagnostic plots of residuals from a model fitting adherence by the explanatory variables (normal-quantile-quantile, residual-vs-fitted).
If adherance is a binary measure (like yes/no), you don't need to check the conditional distribution, because you should anyway use a binomial model ("logistic regression") then. If adherance is a percentage value (bounded in (0, 1)), you should use a beta-regression. If adherance is a time (until an event), you would use a survival model (e.g. Cox proportional hazards, or a Weibull model etc).
You can use SPSS or ANOVA to test your proposed hypothesis whether it is in conformity with your previous assertion
Jochen Wilhelm Thanks for your response. We were using the hypothesised ‘greater reduction of salt intake after 4-weeks’ in those that were in the ‘risk’ category compared to the non-risk subjects.
The "reduction of salt intake after 4 weeks" could be measured quantitatively (like in g/day) , as a metric variable. Under the null hypothesis (no change), it should have a unimodal symmetric distribution, for which the normal model should be appropriate. So fit a linear model, including all the suspected covariables (explanatory variables) and relevant interactions, and then have a look at the residuals. If the diagnostic plots don't screem at you that something is strange, the model should be ok.
- Badges
- Science topic
- Similar topics
- Filing