How to analyze unequal samples in linear and multiple regression I have three variables (sample size is mentioned below), and I want to analyze data using regression models as recommended by Judd and Kenny (1981) to see if $cc$ mediates the relationship between $spdm$ and $cty$ (I'm using abbreviations; see below for variable names). I have a unequal sample size problem, and I don't know how to handle it. I have read a few books and went through many articles, but none have mentioned or analyzed samples of unequal sizes. 
I want to run 3 models, of which 2 are with unequal samples:
model 1: $spdm = \beta \cdot cty + e$
model 2: $cty = \beta_1 \cdot cc + \beta_2 \cdot spdm + e$
(where $e$ is error)   
My Questions:


*

*How can I analyze my data in regression? Is it possible to analyze samples of unequal size?

*I read somewhere on the web that you can draw a random sample from the big sample to make it equal to other samples. Can I do that?


Please provide the source of any book or article.
Sample sizes and variables:


*

*a = 100 ($spdm =$ supervisor's participation in decision making)

*b = 100 ($cc =$ creative climate)

*c = 500 ($cty =$ creativity)

 A: I have made some assumptions that were not explicitly stated in your question. Correct me if I'm wrong.
Let me rephrase your situation. If I understood correctly, you have three variables $x_1, x_2, x_3$. Your observations are one instance of $x1$, paired with one instance of $x_2$, which is paired with $N_y$ instances of $x_3$ (number is not certain based on your comment). Each observation can be written as:
$$X = (x_1,x_2,x_3^1,x_3^2,...,x_3^{N_y}) $$
Where there are actually $N_y$ observations of the final variable for each observation of the other two. 
From what I see, there is no sense in keeping the last observations as separate variables. They do not represent the same people, nor is there a specific order to the observations, nor are they necessarily the same size for each $x_1,x_2$, so it makes no sense to create a model from them separately.
One idea can be to take the average of these observations (suppose there are $N_y$ observations) to create a new variable, or:
$$X = (x_1,x_2,y) $$
Where $y=1/N_y\sum_{i=1}^{n_y} x_3^i$. If you construct a model, you will have 100 samples of each variable. Your three models now become:
$$ x_1 = \beta_0 + \beta_1 x_2 + \epsilon$$
$$ y = \beta_0 + \beta_1 x_1 + \beta_2 x_2+ \epsilon$$
$$ x_2 = \beta_0 + \beta_1 x_1 + \beta_2 y + \epsilon$$
