9.3 Empirical Bayes (EB) Method & Regression-to-the-Mean (RTM) Correction

9.3 Empirical Bayes (EB) Method & Regression-to-the-Mean (RTM) Correction

PTOE Exam Focus: The Empirical Bayes (EB) method is the cornerstone of advanced traffic safety analytics tested on the PTOE examination. You must be able to calculate the statistical weight $w$, determine expected average crash frequency ($N_{\text{expected}}$), compute Excess Expected Crash Frequency ($EECF$ / Potential for Safety Improvement), and execute EB Before-After countermeasure evaluations to eliminate Regression-to-the-Mean (RTM) bias.

---

1. The Regression-to-the-Mean (RTM) Phenomenon

Traffic collisions are rare, discrete, stochastic events that fluctuate randomly over time at any given location. When transportation agencies screen a highway network using raw historical crash counts, sites identified for funding are almost invariably those experiencing extreme upward statistical spikes in recent years.

In subsequent years, even if zero engineering improvements are made, crash counts at these high-crash anomaly sites will naturally decline toward their true long-term statistical mean. This statistical tendency is known as Regression-to-the-Mean (RTM).

 Observed Crashes (N)
     ^
     |         * Random Spike (Site Selected for Treatment)
     |        / \
     |       /   \    <- Apparent Drop Misattributed to Countermeasure
     |      /     \      in Naive Before-After Studies
     |     /       *---------------------------------- True Long-Term Mean
     |    /         \
     |   *           *   * 
    0+-------------------------------------------------> Time (Years)
       Before Period   | After Period

The Failure of Naive Before-After Studies:

A simple naive before-after evaluation calculates treatment effect as $\Delta N = N_{\text{observed, before}} - N_{\text{observed, after}}$. If a site experienced 20 crashes in the before year and 8 crashes after installing a flashing beacon, naive analysis claims a $60%$ reduction ($12\text{ crashes prevented}$). However, if 8 of those 12 crashes were eliminated simply due to natural RTM regression back to a long-term mean of 10 crashes/year, the true engineering benefit was only 2 crashes prevented ($20%$ reduction), while $80%$ of the apparent drop was a statistical illusion.

---

2. The Empirical Bayes (EB) Mathematical Formulation

The Empirical Bayes method corrects for RTM bias by synthesizing two separate, independent streams of statistical information:

1. The Regional Prior Distribution ($N_{\text{predicted}}$): The predicted average crash frequency derived from a calibrated Safety Performance Function representing hundreds of similar peer sites with identical geometric and operational features.
2. The Site-Specific Likelihood ($N_{\text{observed}}$): The actual recorded crash history at the specific site over $n$ observation years.

The EB Expected Crash Equation:

$N_{\text{expected}} = w \times N_{\text{predicted}} + (1 - w) \times N_{\text{observed}}$

Where:

• $N_{\text{expected}}$ = Expected average crash frequency for the site (crashes/year).
• $N_{\text{predicted}}$ = Predicted average crash frequency from the calibrated SPF (crashes/year).
• $N_{\text{observed}}$ = Observed average historical crash frequency at the site over the study period (crashes/year).
• $w$ = Empirical Bayes statistical weighting factor ($0 \le w \le 1$).

The EB Weighting Factor ($w$):

$w = \frac{1}{1 + k \times \sum_{i=1}^n N_{\text{predicted}, i}}$

Where:

• $k$ = Overdispersion parameter of the Safety Performance Function.
• $\sum_{i=1}^n N_{\text{predicted}, i}$ = Cumulative predicted crashes across all $n$ years of the historical observation period ($n \times N_{\text{predicted}}$ if annual volume is constant).

+---------------------------------------------------------------------------------------+
|                       EMPIRICAL BAYES WEIGHT BEHAVIOR DYNAMICS                        |
+-----------------------+-----------------------+---------------------------------------+
| Condition             | Weight Shift          | Mathematical & Engineering Meaning    |
+-----------------------+-----------------------+---------------------------------------+
| Overdispersion k -> 0 | w -> 1.0              | Low site heterogeneity; trust SPF     |
| Overdispersion k >> 0 | w -> 0.0              | High site heterogeneity; trust data   |
| Short history (n=1 yr)| w is larger           | Limited data; rely more on SPF        |
| Long history (n=10 yr)| (1 - w) -> 1.0        | Large sample; site history dominates  |
+-----------------------+-----------------------+---------------------------------------+

Variance and Standard Error of Expected Crashes:

$\operatorname{Var}(N_{\text{expected}}) = (1 - w) \times N_{\text{expected}}$ $\sigma(N_{\text{expected}}) = \sqrt{(1 - w) \times N_{\text{expected}}}$

Notice that the variance of the EB estimate is strictly smaller than the variance of either the SPF model prediction alone or the raw observed count alone, proving that EB maximizes statistical precision.

---

3. Network Screening: Excess Expected Crash Frequency ($EECF$ / PSI)

In Highway Safety Manual Part B network screening, ranking sites by raw crash counts or crash rates is statistically flawed due to RTM bias and traffic volume distortion. The premier, state-of-the-art metric for identifying high-priority safety locations is Excess Expected Crash Frequency ($EECF$), also termed the Potential for Safety Improvement (PSI):

$EECF = PSI = N_{\text{expected}} - N_{\text{predicted}}$

Engineering Interpretation of $EECF$:

• $EECF > 0$ (Positive Excess): The site experiences more crashes than expected for a typical, well-performing peer facility with identical traffic volume and geometric alignment. This excess reflects correctable engineering deficiencies (e.g., poor sight lines, lack of turn lanes, slippery pavement), making it a prime candidate for cost-effective safety treatments.
• $EECF \le 0$ (Zero or Negative Excess): The site performs at or better than the regional baseline expectation. Any crashes occurring at the site are primarily baseline ambient risk rather than localized geometric defects; capital improvement funds will yield low returns.

---

4. Empirical Bayes Before-After Safety Effectiveness Evaluation

When evaluating the actual safety performance of an installed engineering treatment (e.g., modern roundabout conversion), the EB Before-After method calculates the counterfactual: How many crashes would have occurred in the after period if the treatment had NOT been installed?

Step-by-Step EB Before-After Algorithm:

1. Step 1 — Calculate Expected Before Crashes ($N_{\text{expected}, B}$): Apply the standard EB formulation to the before-period data: $w_B = \frac{1}{1 + k \times \sum N_{\text{predicted}, B}}$ $N_{\text{expected}, B} = w_B \sum N_{\text{predicted}, B} + (1 - w_B) N_{\text{observed}, B}$

2. Step 2 — Calculate Exposure Adjustment Factor ($C_{\text{adj}}$): Adjust for changes in traffic volume, study period duration, or regional trends between before ($B$) and after ($A$) periods: $C_{\text{adj}} = \frac{\sum N_{\text{predicted}, A}}{\sum N_{\text{predicted}, B}}$

3. Step 3 — Compute Expected Crashes in After Period Without Treatment ($N_{\text{expected}, A, \text{without}}$): $N_{\text{expected}, A, \text{without}} = N_{\text{expected}, B} \times C_{\text{adj}}$

4. Step 4 — Calculate Treatment Index of Effectiveness ($\theta$): $\theta = \frac{N_{\text{observed}, A}}{N_{\text{expected}, A, \text{without}}}$

5. Step 5 — Calculate Overall Percentage Crash Reduction: $\text{Percent Crash Reduction} = (1 - \theta) \times 100\%$

6. Step 6 — Compute Standard Error of Index of Effectiveness ($\sigma_\theta$): $\sigma_\theta = \frac{\theta \sqrt{\frac{1}{N_{\text{observed}, A}} + \frac{\operatorname{Var}(N_{\text{expected}, A, \text{without}})}{(N_{\text{expected}, A, \text{without}})^2}}}{1 + \frac{\operatorname{Var}(N_{\text{expected}, A, \text{without}})}{(N_{\text{expected}, A, \text{without}})^2}}$

If the interval $\theta \pm 1.96 \sigma_\theta$ does not overlap $1.00$, the treatment effectiveness is statistically significant at $\alpha = 0.05$.

---

5. Worked PTOE Calculation Example

Problem Statement:

A high-crash suburban intersection experienced a total of $N_{\text{observed}} = 45\text{ crashes}$ over a 3-year historical period ($15.0\text{ crashes/year}$). A calibrated regional SPF for this intersection type predicts $N_{\text{predicted}} = 6.0\text{ crashes/year}$ ($18.0\text{ crashes}$ total over 3 years) with an overdispersion parameter $k = 0.20$.

1. Calculate the Empirical Bayes statistical weight $w$.
2. Determine the 3-year total and annual expected crash frequency ($N_{\text{expected}}$).
3. Calculate the annual Excess Expected Crash Frequency ($EECF$ / PSI).
4. Compute the standard error of the expected annual crash frequency.

Solution:

1. EB Weight ($w$): $\sum N_{\text{predicted}} = 3 \times 6.0 = 18.0\text{ crashes}$ $w = \frac{1}{1 + k \times \sum N_{\text{predicted}}} = \frac{1}{1 + 0.20 \times 18.0} = \frac{1}{1 + 3.60} = \frac{1}{4.60} = 0.2174$

2. Expected 3-Year Crash Frequency ($N_{\text{expected, 3-yr}}$): $N_{\text{expected, 3-yr}} = w \left(\sum N_{\text{predicted}}\right) + (1 - w) N_{\text{observed}}$ $N_{\text{expected, 3-yr}} = 0.2174(18.0) + (1 - 0.2174)(45) = 3.913 + 0.7826(45) = 3.913 + 35.217 = 39.13\text{ crashes}$ $\text{Annual } N_{\text{expected}} = \frac{39.13}{3} = 13.04\text{ crashes/year}$

3. Annual Excess Expected Crash Frequency ($EECF$ / PSI): $EECF = N_{\text{expected}} - N_{\text{predicted}} = 13.04 - 6.00 = +7.04\text{ excess crashes/year}$

4. Standard Error of Expected Annual Crashes: $\operatorname{Var}(N_{\text{expected, annual}}) = \frac{(1 - w) \times N_{\text{expected, annual}}}{n} = \frac{(1 - 0.2174) \times 13.04}{3} = \frac{0.7826 \times 13.04}{3} = \frac{10.205}{3} = 3.402$ $\sigma(N_{\text{expected, annual}}) = \sqrt{3.402} = 1.84\text{ crashes/year}$

Key Insight: Notice how the EB method corrected for RTM bias, adjusting the raw observed rate of $15.0\text{ crashes/year}$ downward to an expected long-term mean of $13.04\text{ crashes/year}$, while demonstrating a significant safety deficiency of $+7.04\text{ excess crashes/year}$ above peer baseline.