9.2 Safety Performance Functions (SPFs) & Crash Modification Factors (CMFs)

9.2 Safety Performance Functions (SPFs) & Crash Modification Factors (CMFs)

PTOE Exam Focus: Safety Performance Functions (SPFs) and Crash Modification Factors (CMFs) constitute the mathematical core of modern traffic safety analysis. Master the Negative Binomial model structure, the calculation of crash variance using the overdispersion parameter $k$, segment versus intersection SPF functional forms, CMF combination rules, and standard error confidence intervals ($CMF \pm 1.96 \sigma$).

---

1. Statistical Foundations: Why Ordinary Linear Regression Fails

Historically, transportation engineers attempted to model crash frequency using Ordinary Least Squares (OLS) linear regression ($y = \beta_0 + \beta_1 x_1 + \dots$). OLS fails catastrophically in traffic safety due to three mathematical violations:

1. Continuous vs. Discrete Counts: OLS assumes a continuous dependent variable, predicting fractional crashes ($1.43\text{ crashes}$) or physically impossible negative crashes ($-2.1\text{ crashes}$) at low volumes.
2. Heteroskedasticity: OLS assumes constant error variance (homoskedasticity). In crash data, variance increases with traffic volume and segment length.
3. Non-Normal Error Distributions: Crash events are rare, random, non-negative discrete integers ($0, 1, 2, 3, \dots$) governed by Poisson arrival processes rather than symmetric bell-curve normal distributions.

The Poisson Model & The Overdispersion Problem

A discrete Poisson process defines the probability of observing $y_i$ crashes at site $i$ given an underlying mean crash rate $\mu_i$:

$P(Y = y_i) = \frac{\mu_i^{y_i} e^{-\mu_i}}{y_i!}$

A strict property of the Poisson distribution is equidispersion, where the conditional variance equals the conditional mean:

$E[Y_i] = \operatorname{Var}(Y_i) = \mu_i$

However, empirical roadway crash data almost universally exhibit overdispersion (variance strictly exceeds the mean: $\operatorname{Var}(Y_i) > E[Y_i]$). Overdispersion arises from unobserved site heterogeneity—unmeasured variations in driver behavior, localized weather, pavement micro-texture, sight obstructions, and access conflicts across physical sites.

---

2. The Negative Binomial (Poisson-Gamma) Generalized Linear Model

To account for overdispersion, the HSM utilizes the Negative Binomial (NB) regression model, formulated as a Poisson-gamma mixture. The crash rate for site $i$ is modeled as:

$\mu_i^* = \mu_i \exp(\epsilon_i)$

Where $\mu_i$ is the systematic mean predicted by geometric and operational covariates, and $\exp(\epsilon_i)$ is an unobserved gamma-distributed disturbance term with mean $1.0$ and variance $k$ (the overdispersion parameter).

The Quadratic Variance Equation:

Under the Negative Binomial model, the conditional variance of crash frequency is a quadratic function of the predicted mean:

$\operatorname{Var}(Y_i) = \mu_i + k \mu_i^2$

Where:

• $\mu_i$ = Predicted average crash frequency ($N_{\text{predicted}}$ or $N_{\text{spf}}$).
• $k$ = Overdispersion parameter of the SPF ($k \ge 0$).

  Variance Var(Y)
       ^
       |                                      * Negative Binomial: Var = μ + k·μ²
       |                                   * 
       |                                *   (Overdispersion Envelope)
       |                             * 
       |                          * 
       |                       * 
       |                    *                 - - Pure Poisson: Var = μ
       |                 *            - - - 
       |              *       - - - 
       |           *  - - - 
      0+-------*---------------------------------> Mean μ

Key Mathematical Properties of $k$:

• As $k \to 0$, the overdispersion term $k \mu_i^2 \to 0$, and the Negative Binomial model collapses into a pure Poisson model with $\operatorname{Var}(Y_i) = \mu_i$.
• A larger value of $k$ indicates greater unobserved site-to-site variability, meaning observed crash history deviates substantially from model predictions.
• In the Empirical Bayes method, the weight assigned to the model ($w$) depends directly on $k$.

---

3. Mathematical Formulations of Base SPFs

Safety Performance Functions in the HSM relate predicted crash frequency under nominal base conditions to traffic volume exposure using natural log-link generalized linear models.

A. Roadway Segment SPFs

For roadway segments, exposure is defined by segment length ($L$, in miles) and Annual Average Daily Traffic ($AADT$, in vehicles/day):

$N_{\text{spf, segment}} = \exp\left(a + b \ln(AADT)\right) \times L = e^a \times AADT^b \times L$

Where $a$ is the regression intercept and $b$ is the traffic volume elasticity exponent.

• Non-Linear Returns to Scale: If $b < 1.0$, crash frequency increases at a declining rate relative to volume growth (i.e., crash rate per vehicle-mile traveled decreases as volume increases due to congestion-induced speed dampening).
• Length Scaling: Segment SPFs scale linearly with length ($L$). In HSM models, the overdispersion parameter for segments is inversely proportional to segment length: $k_{\text{segment}} = \frac{1}{\alpha \times L}$ or $k = \frac{a}{L}$.

B. Intersection SPFs

For intersections, exposure is defined by the entering traffic volumes on both the major roadway ($AADT_{\text{maj}}$) and minor cross-street ($AADT_{\text{min}}$):

$N_{\text{spf, intersection}} = \exp\left(a + b_1 \ln(AADT_{\text{maj}}) + b_2 \ln(AADT_{\text{min}})\right) = e^a \times AADT_{\text{maj}}^{b_1} \times AADT_{\text{min}}^{b_2}$

Where $b_1$ and $b_2$ represent the volume elasticities for the major and minor street approaches, respectively. The intersection overdispersion parameter $k$ is a fixed scalar constant calibrated for each specific intersection control type (e.g., 3-leg stop-controlled, 4-leg stop-controlled, 4-leg signalized).

---

4. Crash Modification Factors (CMFs) & Crash Reduction Factors (CRFs)

A Crash Modification Factor (CMF) is a multiplicative factor representing the relative change in crash frequency expected after implementing a specific countermeasure or altering a geometric design element:

$CMF = \frac{N_{\text{treated}}}{N_{\text{untreated}}}$

A Crash Reduction Factor (CRF) expresses the expected safety change as a percentage reduction:

$CRF = (1 - CMF) \times 100\%$

$CMF = 1 - \frac{CRF}{100\%}$

Directionality Matrix:

• $CMF < 1.00$ ($CRF > 0%$): The treatment reduces crash frequency (e.g., $CMF = 0.72 \implies CRF = +28%$ crash reduction).
• $CMF = 1.00$ ($CRF = 0%$): The treatment has zero net effect on safety.
• $CMF > 1.00$ ($CRF < 0%$): The treatment increases crash frequency (e.g., $CMF = 1.15 \implies CRF = -15%$, representing a 15% crash increase, such as adding an unchannelized commercial driveway).

---

5. Multi-Countermeasure Combination Rules & Pitfalls

When multiple countermeasures are applied simultaneously to a single roadway segment or intersection, the HSM Part C baseline method multiplies individual CMFs:

$CMF_{\text{combined}} = \prod_{i=1}^m CMF_i = CMF_1 \times CMF_2 \times \dots \times CMF_m$

The Independence Pitfall & Benefit Overestimation:

Simple multiplication assumes that each countermeasure acts completely independently and that the mechanisms preventing crashes do not overlap. In professional practice, this assumption frequently overestimates safety benefits:

1. Overlapping Crash Types: If an agency installs shoulder rumble strips ($CMF = 0.85$ for run-off-road) and a High-Friction Surface Treatment (HFST, $CMF = 0.75$ for curve run-off-road), multiplying them ($0.85 \times 0.75 = 0.6375$) assumes HFST prevents 25% of crashes that already avoided running off the road due to rumble strips.
2. Target Crash Specificity: CMFs must be applied strictly to their target crash typologies (e.g., applying a left-turn lane CMF only to angle/left-turn collisions, not total collisions).
3. Diminishing Marginal Returns: Each subsequent countermeasure acts on an already-reduced pool of susceptible crashes.

---

6. CMF Statistical Precision & Confidence Intervals

Every published CMF in the HSM and the FHWA CMF Clearinghouse includes a standard error ($\sigma_{\text{CMF}}$ or $SE$) reflecting empirical estimation variance.

95% Confidence Interval Formula:

$CI_{95\%} = CMF \pm 1.96 \times \sigma_{\text{CMF}}$

Statistical Significance Criteria:

• Statistically Significant at $\alpha = 0.05$: The 95% confidence interval does not contain $1.00$ (e.g., $CMF = 0.78, \sigma = 0.05 \implies CI = [0.682, 0.878]$; since $1.00$ is excluded, the treatment significantly reduces crashes).
• Not Statistically Significant: The 95% confidence interval spans across $1.00$ (e.g., $CMF = 0.92, \sigma = 0.08 \implies CI = [0.763, 1.077]$; since $1.00$ is inside the interval, the observed safety improvement could be due to random chance).

Standard Error of Combined Independent CMFs:

$\sigma_{\text{comb}} = \sqrt{\sum_{i=1}^m \left( \sigma_i^2 \prod_{j \neq i} CMF_j^2 \right)}$

FHWA CMF Clearinghouse Star Quality Ratings:

The CMF Clearinghouse rates countermeasure studies from 1 to 5 stars based on sample size, study design (Empirical Bayes / Full Bayes vs simple before-after), selection bias control, and standard error magnitude.