5

I’ve got two main groups, which I am comparing, a) rats and b) rabbits. Each group consists of two subgroups – breed 1 and breed 2.

Using linear regression, I would need to find out, does a model taking into account the subgroups account for more variance than a model with only the main groups.

So I have created two models: \begin{align} Y &= a_1+b1_1(\text{rab})+e_1 \tag{1} \\ Y &= a_2+b1_2(\text{rat2})+b2_2(\text{rab1})+b3_2(\text{rab2})+e_2\ \tag{2} \end{align} Where all variables are dummy variables:

  • $\text{rab}\ \ = 1$ for a rabbit and zero otherwise (e.g., rats)
  • $\text{rab1} = 1$ for rabbit breed 1
  • $\text{rab2} = 1$ for rabbit breed 2
  • $\text{rat2}\, = 1$ for rat breed 2

In order to compare the models statistically, I would need to know, whether my models can be considered nested or not.

If the models are considered nested, I would also need advice in how to compare them statistically - preferably with SPSS. With non-nested models I can handle the rest.

amoeba
  • 93,463
  • 28
  • 275
  • 317
Ina
  • 51
  • 1
  • 4
  • 2
    It is safe to make the presumption that breeds between species have nothing to do with each other? That is there is no relationship between rat breed 1 and rabbit breed 1? – Andy W Mar 20 '14 at 12:47
  • How many breeds of rabbit are there? – Penguin_Knight Mar 20 '14 at 12:50
  • I assume it's safe to consider there's no relationship. However, I can't come up with any examples, what could be such a relationship. – Ina Mar 20 '14 at 12:52
  • Two breeds of rabbit, two breeds of rat. – Ina Mar 20 '14 at 12:53
  • If you have only 2 breeds of rabbits, you can't have 2 dummy codes for the 2 breeds, the design matrix will be singular. 1 dummy code is enough for a categorical variable w/ 2 levels. – gung - Reinstate Monica Mar 20 '14 at 18:49
  • 2
    @gung, if you look at the models closely there are 3 variables for 4 groups (2 species by 2 breeds), the 2 rabbit variables are a different coding of an interaction term. The design matrix will not be singular (as long as all 4 cells have observations). – Greg Snow Mar 20 '14 at 18:51
  • Thanks for the pointer, @GregSnow, I think you're right. It's an unusual way to use the dummy coding. I guess it is what's sometimes called the 'flat' (one-way) approach to factorial ANOVA. I didn't recognize it at first. – gung - Reinstate Monica Mar 20 '14 at 18:56
  • @gung, I actually made the same mistake (the similarity of `rab` and `rat2` visually made the mistake easier), just managed to see the difference before posting. – Greg Snow Mar 20 '14 at 19:03

2 Answers2

5

While your data and model don't fit the formal definition of nested, they are in essence nested because you can redefine the variables in such a way that they do fit the definition. For example change to 2 variables:

rab is 0 for rats and 1 for rabbits

breed is 0 for breed 1 (either species) and 1 for breed 2 (either species)

then fit the model:

$ y = a + b1(rab) + b2(breed) + b3(rab \times breed) + e$

Now you can see that your first model is clearly nested in this model and this model will give the exact same predictions and therefore the exact same sums of squares, likelihood, etc. so any test using these values will give the same results for your model and my model, so your 2 models can be treated as nested.

Greg Snow
  • 46,563
  • 2
  • 90
  • 159
  • Nice! And moreover, you can do an F test to check if b2 and b3 are equal to zero. I'm going to delete my solution. – veryshuai Mar 20 '14 at 19:11
  • Great, this makes sense. Thank you both for your effort and help. – Ina Mar 20 '14 at 22:15
  • When slightly modifying the setting and comparing rats (only one group) to rabbits (still two different breeds), is the following model correct? y=a+b1(rab)+b2(rab×breed)+e – Ina Apr 01 '14 at 11:08
0

They are in fact nested models. Model 1 can be considered a special case of Model 2 by imposing the following parameter restrictions:

$$ b1_2=0 $$ $$b2_2=b3_2$$

In effect, this will absorb $rat2$ into the intercept (which already represents $rat1$), and $rab1+rab2$ will be combined into a single vector equivalent to the original $rab$.

Joby D
  • 33
  • 4