Multicollinearity and impact of individual features

Question

Assume the following scenario:

1. I have four features: $x_1$, $x_2$, $x_3$, and $x_4$
2. There are non-negligible multi-collinearity among the features.
3. I want to predict $y$ (response variable) with those 4 features.
4. I use simple multiple linear regression model: $y = a_1x_1 + a_2x_2 + a_3x_3 + a_4x_4$

Let's say that I want to understand the impact that different components of a chair have on the chair's retail price. For example:

$y\,\,\,$ = chair's retail price

$x_1$ = color of cushion used

$x_2$ = overall design of a chair

$x_3$ = strength of a chair

$x_4$ = softness of a chair

$x_1$ is completely independent, but other features are all somewhat impacted by the other features due to multicollinearity. For example, changing the color of cushion changes the design of a chair. Changing the design (structure) of a chair changes the strength of a chair.

I've heard that the analysis of regression coefficients are unreliable under severe multi-collinearity.

Assuming that the multiple regression model fits the chair price well, can I naively use each feature's regression coefficient to understand the impact of each feature on the response variable? If not, what technique should I use?

Ex 1: If I use a red cushion ($x_1$), I can increased the retail price by 3 dollars

Ex 2: If I use conference room style chair ($x_2$), I can increase the retail price by 12 dollars

Answer 1

When you face multicollinearity, your regression coefficients will likely be biased, because under multicollinearity, the regression can not tell apart the different effects: https://datascience.stackexchange.com/a/57118/71442.

When you use only one variable at a time, you will face the omitted variable bias, because there are no other confounders, to which the regression could attribute relevant effects. https://en.wikipedia.org/wiki/Omitted-variable_bias

To my best knowledge, there is no easy-to-go way to mitigate one or both of the described effects. You should carefully check the correlation between your $x$ and decide if multicollinearity is a problem. If this is the case, and you believe (for theoretical reasons), that all the highly correlated $x$ are important, you could try to find other representations of (some) $x$ to mitigate multicollinearity, e.g. dummy/indicator representations.

I guess your chair example is generic, so I don‘t speculate on this. Maybe you can give some more background regarding the actual problem.

The discussion above refers to causal modeling. If you are only interested in making predictions (not statistical inference), you could look into shrinking coefficients using lasso.