Corrélation sérielle – symptôme d’une spécification fonctionnelle erronée
AI-extracted key points, takeaways & quotes
Serial correlation often indicates functional misspecification, as demonstrated when fitting a linear model to inherently nonlinear data like the relationship between age and wages. Correcting the model's functional form resolves the correlation and prevents biased estimates.
◆Main Points
Serial correlation frequently serves as a symptom of functional misspecification in modeling.
Fitting a straight line to inherently nonlinear data produces runs of positive and negative errors.
Adjacent data points in a misspecified model are likely to share the same sign of error.
Wages typically start low during youth, peak in middle age, and decline near retirement.
Naively estimating a linear function for age's effect on wages misrepresents the true relationship.
A linear fit creates alternating clusters of overpredictions and underpredictions across the data range.
Adding an age-squared term captures the diminishing returns to age.
The coefficient for the squared term is expected to be less than zero.
Specifying a more correct functional model eliminates the serial correlation problem.
Misspecification wrongly suggests constant positive returns to age.
Failing to model nonlinearity introduces significant bias into the coefficient estimates.
Recognizing serial correlation helps identify incorrect underlying functional forms.
✓Takeaways
Always investigate serial correlation as a potential sign of a misspecified model.
Real-world relationships, like age and earnings, are often inherently nonlinear.
Applying linear models to curved relationships creates systematic, clustered errors.
Including polynomial terms, such as a squared variable, can properly capture diminishing returns.
Correcting the functional form of a model is preferable to ignoring serial correlation.
Biased estimates result from forcing linear assumptions onto nonlinear processes.
“Quotes
"Serial correlation... is symptomatic of functional misspecification."
"I've tried to fit a straight line to something which is inherently nonlinear."
"We would have runs of basically positive errors then followed by runs of negative errors."
"Specifying a more sort of correct model that has enabled us to get rid of this serial correlation."
"It actually would have led us quite wrongly to conclude that there was some sort of constant returns to age."
"We would have had some sort of bias in our estimate of beta."
⚙Tools
Linear regression model
Polynomial regression model
Age-squared variable
Residual analysis
Functional form specification
Error clustering observation
✦Facts
Wages generally peak when individuals are in their 30s, 40s, or early 50s.
Young workers, such as those doing paper rounds, typically earn very little.
Wages tend to decrease as workers approach retirement age.
A positive error in a linear model makes an adjacent positive error highly likely.
The coefficient for an age-squared term must be less than zero to show diminishing returns.
Serial correlation can be entirely eliminated by choosing the correct functional form.
↗References
Wage rate modeling
Age-earnings profile
Linear function of age
Polynomial functional form (age and age-squared)
Diminishing returns to age
Beta coefficient estimation
→Recommendations
Check for serial correlation whenever you fit a regression model.
Visualize residuals to identify runs of positive and negative errors.
Avoid assuming linear relationships for lifecycle data like wages.
Incorporate squared terms to model diminishing or increasing marginal effects.
Re-specify your model's functional form before accepting biased coefficients.
Test whether adding nonlinear terms resolves detected serial correlation.
Summarize any YouTube video — free
Paste a link and get structured notes in seconds. No signup, no card.