Comprehensive Study Notes on Confounding Variables and Control Variables

Confounding Variables:
- Definition: Any variable that distorts the relationship between a main effect (predictor variable) and a response variable.
- Importance: They can lead to misconceptions about the relationship between predictor(s) and response if not identified accurately.

Scenario: Use of the number of reviews of a product to predict total sales.
- Question: What kind of relationship exists?
- Answer: The relationship tends to be __.
- Noteworthy Trend: As the number of reviews increases, total sales tend to .
- Is it a causal relationship?
- Answer: __.
- Insight: Customers write reviews a product; more likely that __.
  - Key point: Customers who leave reviews are generally more __, which may not impact total sales.
  - Potential issues: ___ not being considered in this analysis may be driving sales instead.
Question: What variables may be causally related to both number of reviews and total sales?
- Answers:
- _.
- _ (likes, shares, etc.).
- Spending on __.
How to determine if the relationship between reviews and sales is influenced by other variables?
- Answer: Check for .

Definition: Confounding variable is a common cause of both the predictor (main effect) and the response variable, leading to a misinterpretation of their relationship.
Observed Relationship:
- The relationship observed between variables 𝑋 and 𝑌 exists because confounding variable 𝐶 is related to both 𝑋 and 𝑌 but was not included in the initial model.
- Notably: 𝐶 confounds the relationship if the effect of 𝑋 differs significantly when 𝐶 is included in the model.

Scenario: Including advertising spending in the model as a predictor.
- Question: What happened to the slope coefficients when including advertising spending with reviews?
- Answers:
  - Reviews: Coefficient shifted from to ___.
  - Advertising: __ predictor that the model, likely the relationship between reviews and sales.

Advertising as a Confounder:
- Reasoning:
- Sales tend to as reviews increase.
- Advertising spending correlates with .
- More advertising boosts , attracting customers and potentially leading to more reviews.
- With more people , a greater number write reviews after purchases.
- After including advertising, the relationship between reviews and sales appears .
- Mitigating Impact of a Confounding Variable: Include it in the model as a .

Definition: Control variable is a variable related to both the predictor(s) and response, included in regression to accurately assess the relationship among the variables of interest.
Difference from Confounding Variable:
- Confounding Variable: Distorts the relationship between 𝑋 and 𝑌.
- A confounding variable that remains unmapped influences the relationship as its impact can’t be computed.
- Control Variable: Adjusts for significant relationships and facilitates accurate assessments between 𝑋 and 𝑌.
- Control variable that can be measured becomes a control when included in the regression.

Reasons:
- To avoid misleading conclusions about relationships.
- Omitting a control variable may mislead to believe a significant correlation exists between predictors and responses, possibly indicating an indirect relationship.
- To accurately reflect the effect of the main predictor on the response.
- Omitting may result in incorrectly assigning significance to predictors, distorting predictions.
- To optimize the model and improve predictive power.
- Even if a control variable is not the primary interest, its inclusion enhances model robustness and reliability.

Scenario: Let 𝑋 be the primary variable, 𝑌 be the response variable, and 𝐶 be control variable.
Criteria for Inclusion:
1. Strong Correlation: When there is a strong correlation between 𝑋 and 𝑌 corrected by including 𝐶.
- This usually indicates that 𝐶 is confounding.
1. Effect Modification: When the effect of 𝑋 on 𝑌 is modified (usually increased) by including 𝐶 but remains significant.
2. Model Improvement: When incorporating 𝐶 refines model predictions and reduces error, even if the impact is minimal.

Scenario 1: Relationship between years of school (𝑋), income (𝑌, in thousands), and IQ (𝐶).
- Is IQ a Confounder? Answer: ____.
- Relationship Insight: Education as a predictor of income, IQ correlates highly with both.
Scenario 2: Analyzing study hours (𝑋), GPA (𝑌), and sleep (𝐶).
- Is Sleep a Confounder? Answer: ____.
- Insight on Correlations:
  - Study hours are a strong predictor of GPA.
  - Sleep Hours is somewhat correlated with GPA
  - Sleep and study hours are only weak correlation measured as only (𝑟 = 0.16 ).

Steps to Assess:
1. Fit the regression model of 𝑌 = $(\beta0 + \beta1 X)$ .
2. Obtain slope estimate $\beta_1$ .
3. Calculate a 95% confidence interval for the slope.
4. Fit the regression model $Y = (\beta0 + \beta1 X + \beta_2 C)$ .
5. Obtain slope estimate $\beta_1|C$ with 𝐶 included as predictor.
6. Compare $\beta1|C$ with the 95% confidence interval for $\beta1$ .
Results Interpretation:
- If contained in C.I.:
- 𝐶 is not confounder, and its model inclusion is not imperative but optional.
- If not contained in C.I.:
- 𝐶 is a confounder as $\beta1|C$ and $\beta1$ are meaningfully different and necessary for model accuracy.

Existing Regression Equation: $Y = \beta0 + \beta1 X$
- Original slope for reviews: eta1 = _
- Confidence interval C.I. = ___ = __.
Inference based on analysis:
- Advertising spending ____ the relationship, including __ impacts the __.
- Should number of reviews remain in the model? Answer: .

Existing Regression Equation: $Y = \beta0 + \beta1 X$
- Original slope for education: eta1 = _
- Confidence Interval C.I. = ___ = __.

Existing Regression Equation: $Y = \beta0 + \beta1 X$
- Original slope for study hours: eta1 = _
- Confidence Interval C.I. = ___ = __.

Venn Diagram Analysis: Controls: ; Confounders:
- Not Explicitly Discussed:
- Related to both 𝑋 and 𝑌, may influence the slope of the initial predictor.
- Not Control Variables:
- Difficult to quantify, tough to include in modeling processes.