Introduction: Choosing the Best "Guess"
In your previous studies, you learned how to create estimators using methods like Maximum Likelihood Estimation (MLE) or the Method of Moments (MoM). But once you have an estimator, how do you know if it is actually any good? If you have two different ways to estimate the same insurance claim frequency, which one should you trust?
In this chapter, we explore the four "Golden Rules" of estimators: Bias, Mean Square Error (MSE), Efficiency, and Consistency. These properties help actuaries decide which statistical tools are most reliable for making real-world predictions.
1. Bias: Are We Hitting the Target?
Imagine you are practicing archery. If your arrows are scattered all over the place but are centered exactly around the bullseye, you are "accurate" on average. In statistics, we call this being unbiased.
An estimator \(\hat{\theta}\) is a formula we use to guess an unknown population parameter \(\theta\). The Bias is the difference between the expected value of our estimator and the true value.
The Formula:
\(Bias(\hat{\theta}) = E(\hat{\theta}) - \theta\)
- If \(E(\hat{\theta}) = \theta\), the bias is zero. We call \(\hat{\theta}\) an unbiased estimator.
- If \(E(\hat{\theta}) > \theta\), the estimator overestimates the parameter on average (positive bias).
- If \(E(\hat{\theta}) < \theta\), the estimator underestimates the parameter on average (negative bias).
Quick Tip: Don't worry if an estimator is slightly biased. Sometimes a small bias is a price worth paying if it makes the estimator much more stable!
2. Mean Square Error (MSE): The Total Error
While bias tells us if we are "on target," it doesn't tell us how much our shots are "spread out." An estimator could be unbiased but have such a huge variance that any single guess is likely to be far from the truth.
The Mean Square Error (MSE) combines both Bias and Variance into one single measure of "total error." It is the most common way to compare two estimators.
The Formula:
\(MSE(\hat{\theta}) = E[(\hat{\theta} - \theta)^2]\)
Through some clever algebra (which you don't need to derive, but must remember), this simplifies to:
\(MSE(\hat{\theta}) = Var(\hat{\theta}) + [Bias(\hat{\theta})]^2\)
Why is MSE useful?
If you are comparing Estimator A and Estimator B, the one with the lower MSE is generally considered better because it is closer to the true value more often.
Key Takeaway: For an unbiased estimator, the MSE is simply equal to the Variance, because the bias term is zero.
3. Efficiency: Which Estimator is "Sharper"?
Efficiency is all about making the most of your data. If you have two unbiased estimators, you want the one that varies the least. This is the "Efficiency" of the estimator.
Relative Efficiency:
To compare two unbiased estimators, \(\hat{\theta}_1\) and \(\hat{\theta}_2\), we look at the ratio of their variances:
\(Relative\ Efficiency = \frac{Var(\hat{\theta}_2)}{Var(\hat{\theta}_1)}\)
- If this ratio is greater than 1, then \(\hat{\theta}_1\) has a smaller variance and is more efficient than \(\hat{\theta}_2\).
- In actuarial exams, you might be asked to find which estimator is "more efficient." Always look for the one with the smallest variance (assuming they are both unbiased).
Did you know? Maximum Likelihood Estimators (MLEs) are often favored by actuaries because, for large sample sizes, they tend to be the most efficient estimators possible!
4. Consistency: Does More Data Help?
Consistency is a "large sample" property. It asks: "As my sample size \(n\) gets larger and larger (approaching infinity), does my estimator eventually hit the true value exactly?"
An estimator \(\hat{\theta}\) is consistent if it converges to the true value \(\theta\) as \(n \to \infty\).
How to check for Consistency:
For the CS1 exam, an estimator is consistent if two things happen as \(n \to \infty\):
1. The Bias tends to zero: \(Bias(\hat{\theta}) \to 0\)
2. The Variance tends to zero: \(Var(\hat{\theta}) \to 0\)
Analogy: Imagine looking at a blurry photo. As you increase the resolution (the sample size \(n\)), the image becomes perfectly clear. If the image eventually shows the true object with no blur or distortion, your "camera" (the estimator) is consistent.
Summary Table: The Property Checklist
Use this quick review to keep the concepts straight:
| Property | What it measures | Goal |
|---|---|---|
| Bias | Distance between average guess and truth. | Bias = 0 (Unbiased) |
| MSE | Total error (Variance + Bias squared). | Minimize MSE |
| Efficiency | Comparison of variances. | Smaller variance is better |
| Consistency | Behavior as sample size \(n\) grows. | Error \(\to 0\) as \(n \to \infty\) |
Common Mistakes to Avoid
- Confusing Bias and MSE: Remember that an estimator can be unbiased but still have a huge MSE if the variance is high. Always check both.
- Forgetting to square the bias: In the MSE formula, \(MSE = Var + Bias^2\). A common exam error is forgetting to square the bias term!
- Consistency vs. Unbiasedness: An estimator can be biased for small samples but still be consistent if that bias disappears as \(n\) gets larger. (Example: The MLE for variance, \(\hat{\sigma}^2 = \frac{\sum(X_i - \bar{X})^2}{n}\), is biased, but it is consistent).
Next Steps: Now that you understand how to evaluate estimators, you can move on to learning about the Asymptotic distribution of MLEs, where we see how these properties behave in very large samples.