# Ordinary least squares - change response and explanatory variable

In a simple regression model using OLS, why is it not, at least in general, possible to move the $x_i$'s to the left-hand side of the model and the $y_i$'s to the right hand side when I want to switch the explanatory and response variable? Intuitively, it seems fairly clear that if ${\hat \beta ^{OLS}}\left( {Y,X} \right) = \mathop {\arg \min }\limits_{\beta \in \mathbb{R}} \sum\limits_{i = 1}^n {{{\left( {{y_i} - x_i^T\beta } \right)}^2}}$ and ${\hat \gamma ^{OLS}}\left( {X,Y} \right) = \mathop {\arg \min }\limits_{\gamma \in \mathbb{R}} \sum\limits_{i = 1}^n {{{\left( {{x_i} - y_i^T\gamma } \right)}^2}}$, the simple "switching of sides" would not work since I'm minimizing two different sums. But is there any more formal explanation why it cannot be done?

• How more formal than you have already wrote? – Alecos Papadopoulos Jun 14 '14 at 18:00

2. After the switch, if there are multiple $x$s on the left side, this would imply a possibly correlated errors (multivariate) model, or a mis-specified model if you start from least squares multiple regression, and do not move to a multivariate model.
3. If your data are derived from a study design that supports causal inference, then you may be violating causality by reversing causation (e.g. heart attacks causing people to have a smoking history). This can extend to cross-level fallacies depending on the kind of inference you make out of your results (e.g. if $x$s are compositional population-level variables and $y$ is an individual-level variable, you may be asserting that individual experiences are creating (non-compositional) population-level contexts).