Hello everyone,
I apologize for maybe banal question but I am confused. I am trying to write up results of GEE analysis I performed in R (geeglm). I am viewing various articles and I am puzzled. In articles there are usually tables that contain all the factors and their estimates but the final model contains only some of teh factors. Therefore my question is what analysis was used to obtain this full factor table? Is it the model with all of the factors? But then the estimates and the significances of the factors differ from a final model...
Thank you for any advice!
Can you attach an example of a manuscript that causes confusion?
Thank you.
for example here: http://www.sciencedirect.com/science/article/pii/S002209811500177X
I am not sure if this is the best example. They put to the result table all the variables even non-significant ones. But what I am trying to grasp is whether it is enough to show significant factors and simply mention the non-significant or whether t is necessary to state all the estimates, SE and P values for those non significant factors as well.
Thank you for help :-)
Dear,
In this table, we added all variables that were included in the final model (also some non-significant factors were included, because the interaction effect with that factor and another factor was significant!). In my opinion,you should list all variables of your final model and give the F-statistics and p-values here.
Best regards,
Nele
thank you for responce I did not want to offence you I am just trying to understand what is really appropriate to show in the publication, because those things are not usually in any textbook on statistic methods.
In my case when I compose several factors the significance is different then in the final model chosen by backwards selection.
No problem,
I think you should add the details of all the factors in your final model.
Nele
You want to state all factors that you measured in the methods section.
In MANOVA, MANCOVA, or regression models you want to make the simplest hierarchical model possible. Simplest has the fewest factors with the greatest predictive power. There are many ways to evaluate models. Adjusted R-square, or AIC are some options. Hierarchical means that if independent variables A and B are not significant, but their interaction is significant then you report the model with A + B + A*B. Likewise if the A*B*C*D*E*F interaction term is significant then all of the lower interaction terms need to be included.
Things get less certain if you have 85 dependent variables and you are now reporting the effect of some treatment on each variable. The problem is that by convention we only report significant outcomes. So I now have:
1) Significant outcomes that are significant because there really is a relationship between dependent and independent variables.
2) I have significant outcomes that are significant by chance alone (Type 1 error)
3) Significant outcomes by correlation: If A causes B and A causes C, then it is likely that B and C are correlated. If all I measure are B and C, my understanding of the world becomes very different. See Simpson's Paradox.
4) I have non-significant outcomes that are non-significant because the null hypothesis is true.
5) I have non-significant outcomes because of Type II error (false negative)
5a) Insufficient replication. This is generally insufficient replication relative to the variability in the process being measured.
5b) Inappropriate model
5c) Excessive error. I think of this as human error, sloppy measurements, wrong number written in data book, wrong number entered into computer, etc ....
So the bottom line is that a significant outcome may not be important at all, and a non-significant outcome could be very interesting if we can somehow identify true null-hypotheses. Pollutant A really has no effect on kelp species 3, but it does cause species 2 to decline and Species 1 to take its place. This is not the same as "I failed to detect a significant difference in species 3, but did find a significant difference in species 1 and 2." The easy way to run the test is to argue that if a p-value
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Statistical Modeling