Running regression on the entire dataset vs running on smaller dataset then take average of coefficient? Suppose we have a big data set where even if I break it into 10 smaller pieces the number of data points in each piece still far outnumbers the number of variables. Now if I run regression using two methods:

*

*run OLS on the entire dataset

*break data set into 10 random subsets, run a separate OLS on each. Then take the average of the coefficients.

I would like to know

*

*Would the coefficients from the two approach would be the same?

*Would the standard errors of the coefficients be the same?

*Overall is there any advantage of one over the other? (outside of computation considerations)

Thank you all in advance!
 A: Imagine you have the following dataset of measured $(x,y)$ pairs (each row is a measured $(x,y)$ pair):
$$
\begin{pmatrix}
0 & 0 \\
0 & 0 \\
0 & 1 \\
1 & 0 \\
1 & 1 \\
1 & 1 \\
\end{pmatrix}
$$
I.e., you have three $y$-values for $x=0$, two of them zero and one equal to one, and at $x=1$ you have one $y$-value equal to zero and two equal to one.
Next, presume that your random partitioning creates the following two partitions:
$$
P_1 = \begin{pmatrix}
0 & 0 \\
0 & 0 \\
1 & 0 
\end{pmatrix}, \quad
P_2 = \begin{pmatrix}
0 & 1 \\
1 & 1 \\
1 & 1 
\end{pmatrix}.
$$
Let's compare the regression results.
The important point is that OLS takes at each $x$ the average. So, while OLS for the complete dataset gives you the approximations
$$
\begin{align}
\hat y(x=0) &= 1/3\\
\hat y(x=1) &= 2/3
\end{align}
$$
the average of the approximations over the partitioned data gives:
$$
\begin{align}
\hat y^p(x=0) &= \frac{avg\{y \,|\, (x,y)\in P_1, x=0\} + avg\{y\,|\,(x,y)\in P_2, x=0\}}{2}\\
              &= \frac{0+1}{2}\\
              &= \frac{1}{2}\\
\hat y^p(x=1) &= \frac{avg\{y \,|\, (x,y)\in P_1, x=1\} + avg\{y\,|\,(x,y)\in P_2, x=1\}}{2}\\
              &= \frac{0+1}{2}\\
              &= \frac{1}{2}\\
\end{align}
$$
So we see that averaging the OLS results from partitions gives wrong results, and the method without partitioning is to be preferred.
Note, that the wrong results from the method using partitioning can clearly have arbitrarily large deviations from the correct estimations which obviates the question w.r.t. standard errors of the coefficients.
The deeper reason for this phenomenon is that the averaging of sub-averages screws up the proper weighting of your data. But you cannot fix this by using weighted averages because the weights are usually different for different $x$.
