Residuals in transformed models: The differences between the transformed observed values and the transformed predicted values from the fitted model.
- Fit regression model to transformed data
- Calculate residuals: e_i = y'_i - ŷ'_i
- Plot residuals vs fitted values
- Check for patterns or systematic deviations
- Assess normality of residuals
- Identify outliers or influential points
Transformed y values: ln(y): 4.605, 5.521, 6.397, 7.244, 8.071
Regression: ln(ŷ) = 3.741 + 0.866x
Fitted values: 4.607, 5.473, 6.339, 7.205, 8.071
Residuals: 0.002, 0.048, 0.058, 0.039, 0.000
Check for randomness and constant variance
No obvious pattern observed
Mean of residuals ≈ 0 (as expected)
Standard deviation: ≈ 0.035
Residuals appear randomly scattered
Residuals appear roughly normally distributed
No extreme outliers identified
Residual analysis suggests the log transformation is appropriate
Model assumptions appear satisfied
The residual analysis indicates that the log transformation successfully linearizes the relationship. The residuals are randomly scattered around zero with constant variance, suggesting the model assumptions are satisfied.
• Randomness: Residuals should be randomly scattered
• Zero mean: Mean of residuals should be approximately zero
• Constant variance: No funneling or systematic change
• Random scatter around zero
• Constant variance across fitted values
• No systematic patterns detected
Square root transformation residuals: Differences between transformed observed values (√y) and predicted values from the fitted linear model.
Original y: 1, 4, 9, 16, 25
Transformed y: √y = 1, 2, 3, 4, 5
√ŷ = 0 + 1×x = x
This is a perfect fit since √y = x in the original data
Fitted values: 1, 2, 3, 4, 5
Residuals: 0, 0, 0, 0, 0
All residuals equal zero
Perfect fit achieved
Linearity: Perfectly satisfied
Constant variance: All residuals = 0
Normality: Trivially satisfied
The square root transformation completely linearizes the quadratic relationship
The original relationship is y = x²
The square root transformation achieves a perfect fit with all residuals equal to zero. This confirms that the original relationship is y = x², and the transformation completely linearizes it.
• Perfect fit: Zero residuals indicate perfect linearization
• Quadratic relationship: Square root transformation is ideal
• Validation: Residual analysis confirms transformation effectiveness
• All residuals are zero
• Perfect linear relationship achieved
• Transformation completely appropriate
Reciprocal transformation residuals: Differences between transformed observed values (1/y) and predicted values from the fitted model.
Original y: 100, 50, 33.3, 25, 20
Transformed y: 1/y = 0.01, 0.02, 0.03, 0.04, 0.05
1/ŷ = 0.01x
This approximates the relationship 1/y = 0.01x (from PV = 100)
Fitted values: 0.01, 0.02, 0.03, 0.04, 0.05
Residuals: 0, 0, -0.0003, 0, 0
(33.3 is slightly off from exact 33.333...)
Residuals are very close to zero
One small deviation due to rounding
Linearity: Well satisfied
Constant variance: Well satisfied
Normality: Well satisfied
The transformation confirms PV = constant (Boyle's Law)
Residual analysis validates the model
The reciprocal transformation successfully linearizes the inverse relationship. Residuals are near zero, confirming that the data follows Boyle's Law (PV = constant).
• Inverse relationships: Reciprocal transformation linearizes
• Physical laws: Residual analysis validates theoretical relationships
• Small deviations: May be due to measurement error
• Residuals very close to zero
• Confirms theoretical relationship
• Minor deviations due to rounding
Residual: The difference between the observed value and the predicted value from the fitted model.
Transformed residuals: Residuals calculated on the transformed scale after applying mathematical transformations.
Homoscedasticity: The assumption that residuals have constant variance across all fitted values.
Normality: The assumption that residuals are normally distributed.
Independence: The assumption that residuals are independent of each other.
- Calculate residuals: e_i = y'_i - ŷ'_i on transformed scale
- Residual vs fitted plot: Check for patterns and constant variance
- Normal probability plot: Assess normality assumption
- Residual vs predictor: Check for remaining non-linearity
- Time series plot: If data is ordered temporally
- Outlier detection: Identify unusual residuals
- Influence measures: Assess impact of individual points
- Conclusion: Determine if model assumptions are satisfied
• Random scatter: Good model fit
• Funnel shape: Heteroscedasticity (non-constant variance)
• Curved pattern: Remaining non-linearity
• Systematic trend: Missing variable or transformation
• Outliers: Individual points far from pattern
• Non-constant variance (heteroscedasticity)
• Non-normal residuals
• Remaining non-linearity
• Outliers or influential points
• Autocorrelation in time series data
Heteroscedasticity: A condition where the variance of residuals is not constant across all fitted values.
Funneling pattern in residuals vs fitted plot
Residual variance increases with fitted values
Violates the assumption of constant variance
Standard errors of estimates may be incorrect
Confidence intervals and hypothesis tests may be invalid
Original transformation may not be adequate
Need for additional transformation
Model specification error
Try square root transformation instead of log
Consider Box-Cox transformation for optimal power
Apply weighted least squares
Try √y transformation if log failed to stabilize variance
Re-analyze residuals after new transformation
Check if funneling pattern disappears
Confirm constant variance assumption
The funneling pattern indicates heteroscedasticity (non-constant variance). This violates regression assumptions. Try alternative transformations like square root or Box-Cox to better stabilize the variance.
• Pattern recognition: Funneling indicates increasing variance
• Assumption violation: Heteroscedasticity affects inference
• Remedial measures: Alternative transformations may help
Systematic pattern in residuals: Indicates that the model does not adequately capture the relationship between variables.
Curvilinear pattern in residuals vs fitted plot
Indicates remaining non-linearity after transformation
Drug elimination may have multiple phases
Initial rapid distribution followed by slower elimination
Add polynomial terms to the model
Consider bi-exponential model for two-phase elimination
Include interaction terms if appropriate
Try: ln(concentration) = β₀ + β₁×time + β₂×time²
Or use non-linear regression: concentration = Ae^(-k₁t) + Be^(-k₂t)
Fit proposed model
Perform residual analysis on new model
Check if systematic pattern is eliminated
Biological model should reflect actual elimination processes
Parameters should have clinical relevance
The curvilinear pattern in residuals suggests the simple first-order elimination model is inadequate. Consider a bi-exponential model reflecting distribution and elimination phases for better clinical accuracy.
• Pattern interpretation: Curvilinear residuals indicate model inadequacy
• Biological relevance: Model should reflect actual processes
• Clinical application: Parameters must be interpretable
Regression diagnostics: Procedures used to evaluate the appropriateness of a regression model and its assumptions.
Model adequacy: The extent to which a model appropriately captures the relationship in the data.
Outliers: Data points that deviate significantly from the overall pattern.
Influential points: Data points that have a disproportionate effect on the regression model.
Specification error: Occurs when the model does not correctly represent the relationship between variables.
- Calculate residuals: On the transformed scale used for modeling
- Residual vs fitted plot: Primary diagnostic for linearity and constant variance
- Normal probability plot: Assess normality of residuals
- Scale-location plot: Check for heteroscedasticity
- Residual vs predictor: Identify remaining non-linear relationships
- Time series plots: For temporal data to check independence
- Influence measures: Cook's distance, leverage values
- Conclusion synthesis: Integrate all diagnostic information
• Scale consistency: Analyze residuals on the same scale as the model
• Visual inspection: Primary tool for identifying patterns
• Multiple diagnostics: Use various plots for comprehensive assessment
• Context consideration: Account for sample size and practical significance
• Remedial action: Address identified problems systematically
• Transformations may affect the interpretation of residuals
• Back-transformed predictions may have different residual properties
• Consider robust methods when outliers are present
• Validation with new data is the ultimate test of model adequacy