This is a study conducted in 16 vineyards over 3 years using 3 wine grape cultivars. For simplicity, let's just say I want to know how the boron concentration in leaf blades collected at bloom is related to the sugar concentration of the grape juice at harvest. (There are more comparisons I want to make, but the design is the same for all of them, so let's just look at the one, chosen because the only significant effect in the model - alpha = 0.05 - is the effect I care about, so it's simple.) After consulting with a statistician, we came up with the following code:

proc mixed data=work.grapes;

class variety vineyard year;

model brix = BBB variety year variety*year BBB*variety BBB*year BBB*variety*year / s ddfm=satterth;

random vineyard vineyard*variety vineyard*year vineyard*variety*year;

run;

"brix" is juice sugar concentration. "BBB" is boron in the blade at bloom.

I really don't care much about the effects of variety or year, much less the random effects. I only care about how brix and BBB are related.

To assess the one relationship I really care about in the model, the statistician had me add the "s" option to generate parameter estimates for each variable. He told me to use the parameter estimates to find the direction of the relationship between BBB and brix, but to use the P-value from the Type III sum-of-squares output to determine the significance of the relationship.

Trying to apply this method across all my comparisons has led me to conclude that it doesn't work. I can have a very low P-value in the Type III sum-of-squares output while the P-value of the parameter estimate is over 0.9. That tells me that there's a strong relationship between the dependent variable and the predictor of interest, but I can't be sure in which direction.

That won't do. I can't just say that juice sugar is related to blade boron at bloom without being able to say whether it's positively or negatively related.

So how do I reliably determine the direction of the relationship in this scenario?