![[ConfoundingAndSelectionBias.png#invert|center]]
> [!tldr] Confounding Bias
> **Confounding bias** occurs when there is a **confounding variable** with causal effect on both the treatment and the response. In most cases, it is **pretreatment**: in terms of a [[Causal Graphs|causal graph]],
> ```mermaid
> flowchart LR
> X[confounder X] --> T[treatment T]
> X --> Y[response Y]
> T --> Y
> ```
> See [[Causal Graphs#Indirect Influences]] for an example.
>
> Another possibility is a **intermediate/post-treatment confounder**, which only occurs when the treatment is prolonged (e.g. education; otherwise the confounder cannot be later than its effect, i.e. the treatment):
> ```mermaid
> flowchart LR
> T[earlier treatment T] --> X[confounder X] --> TT[later treatment T]
> X --> Y[response Y]
> T --> Y
> TT --> Y
> ```
> For example $X$ is education, $T$ is good study habits (which is fostered by previous education and contributes to attaining higher education in the future), and $Y$ is salary.
>
> However, if $X$ can be controlled, (at least in this causal model) we can inference causality in $Y_{0},Y_{1} \perp T ~|~ X.$
> [!tldr] Selection Bias
> Selection bias occurs when an inappropriate variable is controlled for when studying causality, e.g. when trying to prevent [[Confounding and Selection Bias in Causal Inference]].
>
> In a [[Causal Graphs|causal graph]], this can occur like:
> ```mermaid
> flowchart LR
> T[treatment T] --> Y[response Y]
> T --> X[controlled variable X]
> Y --> X
> ```
>
> Alternatively, in a case of **mediators**, the causal graph looks like
> ```mermaid
> flowchart LR
> T[treatment T] --> Y[response Y]
> T --> X[mediator X]
> X --> Y
> ```
> In both cases, marginally we have $Y_{0},Y_{1} \perp T$, but by controlling $X$, this may not hold, and causal inference breaks down.
For example, consider the [[Simpson's Paradox#IRL Example UCB Admission Rate by Sex]] example:
- Let $T$ indicate sex ($1=$ woman), and assume for example's sake being a woman causes the student to apply for harder programs, whose difficulty is coded by $X$, so $T \to X$.
- Let the response $Y$ be a binary $0/1$ encoding of admission/rejection, so obviously $X \to Y$.
- Now $X$ is a mediator, so if we controlled for $X$ (e.g. looking at the data at the per-department level), we would conclude that being a woman "caused" higher admission rates; if we do not control for $X$ (i.e. looking at the data pooled across UCB), we would conclude the contrary.