Improved Guidelines for Recalibration of Predictive Models over Time Based on Model Uncertainty
Bibliographic record
Abstract
The Highway Safety Manual (HSM) summarizes the safety performance functions (SPFs) of various facility types. The primary use of SPFs is to estimate the safety performance (i.e., the number of crashes by severity level) of different facilities based on geometric and traffic variables. The SPFs were developed using the negative binomial (NB) regression model based on crash data obtained from a selected number of states and cities in the United States and Canada. Applied directly to the local jurisdictions, SPFs may yield biased or incorrect results. Therefore, calibration of the SPFs or predictive models is an important step before applying them to local jurisdictions. Moreover, it is also necessary to recalibrate SPFs over time to account for variations in factors that cannot be accounted for directly in SPFs, such as changes in driver behavior, crash-reporting thresholds, etc. The calibration factor (for a specific facility type) is defined as the ratio of the observed number of crashes to the predicted number of crashes. The HSM recommends that SPFs be recalibrated every 2 to 3 years. However, these guidelines are not based on sound research or reliable criteria. The lack of appropriate guidelines can lead to two types of errors: recalibrating of the models when it is not needed, and not recalibrating them when such a need arises.\n \nThe aim of this thesis is to develop guidelines regarding when or how often SPFs should be recalibrated. To this end, two methodologies were created related to the variance or uncertainty associated with the SPFs, and the guidelines were developed using statistical principles. These guidelines are that SPFs should be recalibrated when (i) the total number of crashes that occur in a network of similar types of facilities falls beyond the prediction intervals of the predicted or estimated total number of crashes in that same network; or (ii) the calibration factor developed in a specific year is statistically significantly different than 1 (based on coefficient of variation (CV) of the SPF and the Calibration Factor C). \n\nBoth approaches were tested on several intersections and segment datasets from Michigan and Toronto. The results show that both approaches are feasible and could provide safety analysts with better and more reliable guidelines regarding when SPFs should be recalibrated. However, the methodologies developed in this thesis cannot be applied to the SPFs developed in the HSM since the information needed to evaluate the variance of SPFs is not available in this manual. The results of both the methodologies were compared to the results of a methodology recently proposed in the literature that can be applied to HSM SPFs and uses a fixed threshold value of C-factor error estimate (say 10%). This study indicated that the 10% error is a reasonable value to use for re-calibrating models.\n\nThe shortcomings of these methodologies include the need to develop a new SPF (which is time-consuming and work-intensive process) and to collect extensive data every year. When data is available every year, the practitioner might as well estimate a new calibration factor every year instead of needing to know the frequency of recalibration or use an approximate method (Cproxy). Future research in this area should focus on identifying the minimum data requirements for both methodologies proposed in this thesis.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.061 | 0.299 |
| Meta-epidemiology (narrow) | 0.004 | 0.003 |
| Meta-epidemiology (broad) | 0.003 | 0.003 |
| Bibliometrics | 0.007 | 0.003 |
| Science and technology studies | 0.002 | 0.002 |
| Scholarly communication | 0.006 | 0.008 |
| Open science | 0.009 | 0.006 |
| Research integrity | 0.003 | 0.011 |
| Insufficient payload (model declined to judge) | 0.005 | 0.003 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".