含时间数据的嵌套ANOVA分析疑问及正确方法咨询

Hey Ennio, let's break down what's going on with your nested ANOVA and fix those issues step by step.

First, let's clarify the confusion around degrees of freedom #

Your original model lm(y~season+season:month) is actually correctly specifying a nested design (since each month belongs to exactly one season), but your expectation of the degrees of freedom using the balanced design formula A(B-1) doesn't apply here—you have an unbalanced nested design!

For unbalanced nested ANOVA, the df for the nested factor (month within season) is calculated as the sum of (number of months per season - 1) for each season:

• Winter (3 months): 3-1=2
• Spring (3 months):3-1=2
• Summer (2 months):2-1=1
• Autumn (3 months):3-1=2
  Total: 2+2+1+2=7, which matches the df shown in your ANOVA table. The A(B-1) formula only works when every season has the same number of months (balanced design), which isn't your case here. So that part is actually correct!

Why you're getting NA values in the Tukey test #

The NA results are happening because you're asking Tukey's test to compare combinations that don't exist in your data—like Autumn:April (April is in Spring, not Autumn). These non-existent groups have no observations, so the test can't compute a comparison, hence the NA.

The correct analysis workflow #

Let's walk through the proper steps for your nested ANOVA:

1. Validate your nested structure first
   Double-check that each month is assigned to exactly one season (no overlaps or missing assignments) with a contingency table:

   table(your_data$season, your_data$month)
   

   Pro tip: If you accidentally excluded August from Summer, fix that grouping first if it's a typo!

2. Fit the nested model properly
   R has a dedicated syntax for nested designs that makes the relationship clearer. Use either of these equivalent formulas:

   # Option 1: Using the / operator (shortcut for season + season:month)
   modello_corretto <- lm(y ~ season/month, data = your_data)
   
   # Option 2: Explicitly specifying month nested in season
   modello_corretto <- lm(y ~ season + month%in%season, data = your_data)
   

   Both will give you the same ANOVA table as your original model, but they make the nested relationship explicit for anyone reading your code.

3. Run the ANOVA
   Use anova() as before to get the significance tests:

   anova(modello_corretto)
   

   As we confirmed earlier, the df values here are correct for your unbalanced design.

4. Targeted post-hoc tests (no more NAs!)
   Instead of running Tukey's test on all possible season:month combinations (most of which don't exist), use the emmeans package to test only meaningful comparisons:

   • Test differences between seasons:

     library(emmeans)
     # Get estimated marginal means for seasons
     season_ems <- emmeans(modello_corretto, ~ season)
     # Tukey test for season comparisons
     pairs(season_ems, adjust = "tukey")
     

   • Test differences between months within each season:

     # Get estimated marginal means for months, grouped by season
     month_in_season_ems <- emmeans(modello_corretto, ~ month | season)
     # Tukey test for month comparisons within each season
     pairs(month_in_season_ems, adjust = "tukey")
     

   This way, you only compare groups that actually exist in your data, so no more NA results.

5. Check ANOVA assumptions
   Don't forget to verify that your model meets the assumptions of normality and homogeneity of variances:

   # Residual plot (check for equal variance) and Q-Q plot (check normality)
   plot(modello_corretto, which = 1:2)
   

   If the assumptions are violated, you might need to transform your response variable y or use a non-parametric alternative.

内容的提问来源于stack exchange，提问作者Ennio