Consideration of Preservation in Pavement Design and Analysis Procedures (2015)

Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

21 C H A P T E R 5 This approach considers pavement preservation by calibrat- ing the MEPDG local models. Calibration is a systematic pro- cess for eliminating any bias and minimizing the residual errors between observed or measured results from the real world and predicted results from the model (AASHTO 2010). The approach assumes that the MEPDG distress prediction models do not account for the effects of pavement preservation and that these effects can be considered by modifying the calibra- tion coefficients. The modified calibration process lends itself to models that directly calculate the magnitude of distress from pavement response (e.g., rutting) and those that calculate the incremental damage index from pavement response and then use a transfer function to convert damage to a distress type (e.g., fatigue cracking). Preservation-based calibration requires a sufficient amount of performance data for pavements subjected to a specific preservation treatment or strategy (preferably on a variety of sites subjected to different levels of climate, traffic, etc.). These data are used to recalibrate the performance prediction mod- els (e.g., roughness, rutting, cracking, and faulting) to account for the effect of the treatment or strategy using the procedures described in the AASHTO Local Calibration Guide (AASHTO 2010). The performance data derived from either in-service pavement sections or test sections specifically constructed and monitored are used in this calibration. The calibration procedure considers the coefficients and exponents of the MEPDG flexible and rigid pavement transfer functions or distress/smoothness models and adjusts one or more of these coefficients to result in better agreement between predicted and observed distress/smoothness (Kim et al. 2011). Although preservation treatments may affect other surface condition parameters (e.g., raveling, bleeding, segregation, distortions) and performance indicators (e.g., friction, noise), this approach only addresses the effects of treatments on the performance prediction models included in the MEPDG. The preservation-based local calibration effort requires developing input values for the selected pavement/test sec- tions and performing multiple runs of the AASHTOWare Pavement ME Design software. The process will then establish a unique set of calibration parameters (k, b, and C) for use in the MEPDG models to better reflect the performance of spe- cific preservation-treated pavements. Figure 5 illustrates the calibration effect using smoothness as an example. The IRI values predicted by the MEPDG (default) model are mostly greater than the measured IRI values (overprediction), and the amount of overprediction increases as IRI increases. Also, there is a wide amount of scatter (high variability/error) in the linear trend line fitted through the predicted versus mea- sured data points. Calibrating the model using the data for the preservation-treated sections will account for the effect of preservation more appropriately. Tables 10 and 11 list the calibration parameters of the MEPDG flexible and rigid pavement transfer functions or distress/smoothness models and their default values as given in the AASHTOWare Pavement ME Design software program. These parameters, typically considered in the local calibration process, will be used in the preservation-based model calibra- tion procedure. The preservation-based model calibration can be performed using one of four approaches detailed in the Local Calibration Guide (AASHTO 2010): • Full Sample: All sections (i.e., n data sets) are used in the calibration process; no sections remain for validation. • Traditional Split Sample: A portion of the total number of sections (usually more than half) is used to calibrate the models; the remainder is used to validate model accuracy. • Jackknife testing: A rolling set of calibrations and validations are performed using n-1 data sets. • Split-Sample Jackknife Testing: A combination of split-sample testing and jackknife testing is performed that uses an n/2 jackknifing scheme. Process Description The process for calibrating the MEPDG models to account for preservation effects, as summarized in the following, is similar to the process for calibrating MEPDG models to local Calibrating MEPDG Models to Account for Preservation

22 conditions described in the AASHTO Local Calibration Guide (AASHTO 2010). 1. Select Hierarchical Input Level for Each Input Parameter: The level for each input parameter is selected consider- ing field and laboratory testing capabilities, material/ construction specifications, and traffic data collection procedures/equipment. Different input levels are likely to be selected for different input parameters. 2. Develop Experimental Plan or Sampling Template: A detailed, statistically sound experimental matrix is devel- oped to represent the different conditions, materials, and Figure 5. Illustration of effect of model calibration on accuracy of performance prediction. Distress Eliminate Bias Reduce Standard Error Total Rutting Unbound Materials1 and HMA Layers kr1 = −3.35412 βr1 = 1 βs1 = 1 kr2 = 1.5606 kr3 = 0.4791 βr2 = 1 βr3 = 1 Load-Related Cracking Bottom-Up Alligator Cracking kf1 = 0.007566 C2 = 1 kf2 = 3.9492 kf3 = 1.281 C1 = 1 Top-Down Longitudinal Cracking kf1 = 0.007566 C2 = 3.5 kf2 = 3.9492 kf3 = 1.281 C1 = 7 Semi-Rigid Pavements (CTB layer) βc1 = 1 C2 = 1 C1 = 1 C2 = 1 C4 = 1,000 Non-Load-Related Cracking Transverse Thermal Cracking βt3 = 1 kt3 = 1.5 βt3 = 1 kt3 = 1.5 Smoothness/IRI C4 = 0.015 (new/reconstructed HMA) C4 = 0.00825 (HMA overlay) C1 = 40 (new/reconstructed HMA) C1 = 40.8 (HMA overlay) C2 = 0.4 (new/reconstructed HMA) C2 = 0.575 (HMA overlay) C3 = 0.008 (new/reconstructed HMA) C3 = 0.0014 (HMA overlay) Notes: Unless otherwise noted, the calibration coefficients pertain to both new/reconstructed HMA pavements and HMA overlays. CTB = cement-treated base. 1 Includes unbound materials for base, subbase, and subgrade layers. Table 10. Calibration parameters for flexible pavement transfer function (AASHTO 2010).

23 practices. The experimental matrix would ideally include key factors, such as design type (i.e., new/reconstructed, rehabilitation), pavement type/design (e.g., conventional HMA pavement, HMA overlay on existing PCC pave- ment), preservation strategy (e.g., preservation with one-time application of a specific treatment type, pres- ervation with multiple treatment applications, no pres- ervation), traffic level or facility type, and climate. The availability of sufficient in-service or experimental test sections (both treated with preservation and not treated) is required. An example of such an experimental matrix is shown in Table 12. 3. Estimate Sample Size for Specific Distress Prediction Models: The sample size or number of pavement sections needed to verify/calibrate the coefficients needs to be determined. Both the bias and precision of the prediction models are considered, and a level of significance (typically 90%) must be selected to determine the required sample size. Gener- ally, some sections are used to calibrate all models, and rep- licate sections are used to provide an estimate of the pure Distress Eliminate Bias Reduce Standard Error JPC Transverse Joint Faulting C1 = 1.0184 C1 = 1.0184 JPC Slab Cracking C1 = 2 C4 = 1 C2 = 1.22 C5 = −1.98 CRC Punchouts Fatigue C1 = 2 C2 = 1.22 Punchouts C3 = 216.842 C4 = 33.1579 C5 = −0.58947 Crack Widths C6 = 1 C6 = 1 Smoothness/IRI JPC J4 = 25.24 J1 = 0.8203 CRC — C1 = 3.15 C2 = 28.35 Note: Unless otherwise noted, the calibration coefficients pertain to both new/reconstructed JPC/CRC pavements and JPC/CRC overlays. Table 11. Calibration parameters for rigid pavement transfer function (AASHTO 2010). Pavement Type/Design Preservation Treatment/Strategy Interstate and Major Arterial Routes Minor Arterial Routes Climate 1 Climate 2 Climate 1 Climate 2 New/Reconstructed Conventional HMA (0) Untreated control (1a) Treatment A @ Year 3 (1b) Treatment A @ Year 4 (2) Treatment B @ Years 3 and 6 (3) Treatment B @ Year 3 and Treatment C @ Year 6 New/Reconstructed Deep-Strength HMA (0) Untreated control (1a) Treatment A @ Year 4 (1b) Treatment A @ Year 5 (2) Treatment B @ Years 4 and 8 (3) Treatment B @ Year 4 and Treatment C @ Year 8 HMA Overlay on Existing Flexible Pavement (0) Untreated control (1) Treatment A @ Year 4 (2) Treatment B @ Year 4 (3) Treatment C @ Year 4 HMA Overlay on Existing Rigid Pavement (0) Untreated control (1) Treatment A @ Year 3 (2) Treatment C @ Year 3 (3) Treatment D @ Year 3 Table 12. Example experimental/sampling matrix for preservation-based local calibration.

24 error. The suggested minimum numbers of sections for analysis of each distress type over the entire experimental/ sampling matrix are as follows (AASHTO 2010): – Distortion (rutting, joint faulting): 20 sections. – Load-related cracking (bottom-up alligator and top- down longitudinal cracking, transverse slab cracking): 20 sections. – Non-load–related cracking (transverse thermal crack- ing): 26 sections. – Reflection cracking: 26 sections. A more refined estimate of the sample size requirements can be obtained using the following equations (AASHTO 2010): n Z et = ×    α σ2 2 Eq. 1 e Z St e= ×α 2 Eq. 2 where: n = Minimum number of sections required for a given distress/IRI prediction model calibration/ validation. Za/2 = 1.601 for a 90% confidence interval. s = Performance indicator threshold/design crite- ria (to be selected by the agency; typical values include 0.4 in. for rutting, 20% for fatigue crack- ing, 1,500 ft/mi for transverse thermal cracking, 10% for slab cracking, 0.1 in. for joint faulting, and 130 in./mi for roughness). et = Tolerable bias at 90% reliability. Se = Standard error of estimate (reasonable values include 0.1 in. for rutting, 7% for alligator crack- ing, 600 ft/mi for longitudinal cracking, 250 ft/mi for transverse thermal cracking, 7% for slab crack- ing, 0.05 in. for joint faulting, and 18 in./mi for roughness). The same test sections could be used for calibrating multiple models to keep the number of sections to a minimum. Also, because IRI is a function of the other distresses, calibrating the IRI model using the same sec- tions used for calibrating the model requiring the largest sample size would be desirable. The experimental matrix can be developed if an ade- quate number of sections with the required types and ranges of performance data are available. Otherwise, other options must be considered, such as combining LTPP or other test sections with the available sections, limiting the analysis only to those factors represented by the available sections, or expanding the acceptable range for some input parameters. In the situations where recently constructed sections are included and no or limited performance data are available, calibrations can be performed at a future time when the required data have become available. 4. Select Roadway Segments: In-service pavement or test sec- tions (e.g., LTPP sections) appropriate to fill the cells in the experimental matrix are identified. Although some consideration should have been given to performance data, sections that have at least three time-series distress/ smoothness data points (from condition surveys) covering a 10-year period are generally required (AASHTO 2010). However, for preservation-treated sections, at least four time-series points (two points prior to the preservation treatment and two points after) and at least a 5-year period following the preservation treatment are desired. 5. Extract and Evaluate Distress and Project Data: The data needed to conduct MEPDG design runs for the cells of the experimental/sampling matrix (herein referred to as analysis cells) are collected and examined. It is necessary to ensure that the collected distress/smoothness data (likely obtained from the agency PMS database) are consistent with the formats used by the MEPDG. Discrepancies in the data formats may be addressed by developing and applying conversion equations or algorithms. Another important consideration is ensuring that the pavement sections cover a range of data for a particular distress and smoothness. It is generally recommended that the aver- age maximum distress/roughness level for the sections exceed 50% of the design criteria (AASHTO 2010). For example, for a rutting design threshold of 0.75 in., the average maximum rut depth for the sections should be at least 0.375 in. Gaps in data should be identified and addressed. 6. Conduct Field and Forensic Investigations: The data needed to fill the identified gaps are obtained. This may be done by conducting field or laboratory investigations (pavement surveys and/or forensic testing of materials and pavement structure), reviewing construction practices and specifica- tions, or by other means. 7. Assess Bias: Distress/smoothness for each analysis cell in the experimental matrix is predicted from MEPDG design runs using the MEPDG default calibration factors. (Details are provided in Appendix C.) The predicted val- ues (at a 50% reliability level) for a set of cells represent- ing a particular treatment type/strategy are then plotted and compared to the measured values, and the bias and standard error of the estimate for each particular distress/ smoothness model are determined. Figure 6 illustrates examples of predicted versus mea- sured rut depth for asphalt pavements with different mixes. The need for calibrating a specific model is determined from null hypothesis statistical testing of a paired t-test

25 that determines if there is a significant difference between sets of measured and predicted distress/smoothness and from an analysis of the intercept and slope estimates in the measured versus predicted linear regression model. In the example shown in Figure 6, the trend lines of the three data sets are statistically analyzed to determine if they are significantly biased in relation to the line of equality, which represents perfect prediction accuracy; calibration of the prediction model is required only if the trend line is found to be statistically different. Model prediction capability is assessed by perform- ing a linear regression of the measured (y) and predicted (x) values (model form yi = bo + m(xi), where bo is the y-intercept and m is the slope) and computing the coef- ficient of determination (R2). In general, models with R2 values above 65% are considered to have good predic- tion capabilities, and those with values below 50% are considered to have poor prediction capabilities. A poor correlation indicates the need for calibration. Model accuracy is estimated by means of the standard error of the estimate (Se), which is computed as the square root of the average squared error of prediction. The reason- ableness of Se can be compared with the Se values obtained from the national/global model calibration (Titus Glover and Mallela 2009); these values are shown in Table 13. Model bias (er) is determined through the following series of hypothesis testing (AASHTO 2010): • Hypothesis 1: There is no bias or systematic difference between the measured and predicted values of distress/ smoothness. A paired t-test is performed to test the fol- lowing null (H0) and alternative (HA) hypotheses: – H0: S(ymeasured - xpredicted) = 0, where ymeasured equals the measured value, and xpredicted equals the predicted value from the model. – HA: S(ymeasured - xpredicted) ≠ 0. • Hypothesis 2: The linear regression model developed using measured and predicted distress/smoothness has Figure 6. Example plot of predicted versus actual distress (AASHTO 2010). Pavement Type Performance Model Model Statistics Coefficient of Determination, R Standard Error of Estimate, S Number of Data Points, N New HMA Alligator cracking 0.275 5.01% 405 Transverse thermal cracking Level 1*: 0.344 Level 2*: 0.218 Level 3*: 0.057 — — Rutting 0.58 0.107 in. 334 IRI 0.56 18.9 in./mi 1,926 New JPC Pavement Transverse slab cracking 0.85 4.52% 1,505 Transverse joint faulting 0.58 0.033 in. 1,239 IRI 0.60 17.1 in./mi 163 Note: * Level of inputs used for calibration. Table 13. Statistics for new asphalt concrete (AC) and JPC pavements performance prediction models (Titus Glover and Mallela 2009).

26 an intercept of zero. Statistics from the linear regres- sion analysis are examined to test the following null and alternative hypotheses: – H0: bo = 0. – HA: bo ≠ 0. • Hypothesis 3: The linear regression model developed using measured and predicted distress/smoothness has a slope (m) of 1.0. Statistics from the linear regression analysis are examined to test the following null and alternative hypotheses: – H0: m = 1.0. – H0: m ≠ 1.0. If any of these null hypotheses are rejected, then the specific distress/smoothness prediction model should be recalibrated. If the null hypotheses are accepted (indicat- ing no bias), the standard error of the estimate for the data set should be compared to the global calibration data set. Figure 7 and Table 14 provide an example for a rutting model using hypothetical data for several full-depth HMA pavement/test sections, with and without preservation. Figure 7 compares the predicted (using the national cali- bration coefficients in the Pavement ME Design software) and measured values of total rutting for three sets of sec- tions (untreated sections, sections treated with preserva- tion type A, and sections treated with preservation type B). The figure shows that the overall (all sections combined) rutting model prediction capability is poor (R2 = 0.29) but that the overall standard error of the estimate (Se) for the model is lower than the national calibration coefficients (0.057 in. versus 0.107 in.). The results of hypothesis test- ing for overall model bias presented in the table show that each null hypothesis was rejected at the 10% significance level such that model recalibration is required to account for the effects of preservation or other factors. Table 14 summarizes the results of similar testing per- formed for each individual set of sections (untreated, preservation A-treated, and preservation B-treated). Although some improvement was observed in the model prediction capability and Se, each of these models was also shown to be locally biased (at least one of the three null hypotheses rejected) and requires recalibration. 8. Eliminate Bias of Distress and IRI Prediction Models: The cause of the bias, if it exists, is first determined through careful evaluation of the bias statistics. The bias that may exist for a given distress/smoothness model (er, Se, residual errors [ymeasured - xpredicted]) is then reduced or eliminated by running the Pavement ME Design software using adjusted calibration factors. The AASHTO Local Calibration Guide (AASHTO 2010) identifies the coefficients of the MEPDG models that should be targeted for bias adjustment. The bias in the prediction mode is described in one of three scenarios (AASHTO 2010): (1) high precision and high bias, (2) low precision and low bias, or (3) low precision and high bias. Scenario 1 requires less effort to reduce the bias than Scenarios 2 and 3. Bias testing that 0.000 0.050 0.100 0.150 0.200 0.250 0.300 0.350 0.400 0.000 0.050 0.100 0.150 0.200 0.250 0.300 0.350 0.400 M ea su re d To ta l R utti ng , i n. Predicted Total Rutting, in. Untreated Pres-Treated A Pres-Treated B H0: bo = 0 and H0: m = 1.0 (Linear Regression) R^2 = 0.29 Se = 0.057 Sy = 0.068 Se/Sy = 0.838 p-value (int) = 0.003 p-value (slope) = 0.000 Reject H0 (p-value <0.1) H0: ∑(ymeas – xpred) = 0 (T-test @10% significance level) n = 197 Avg Pred Total Rut = 0.166 in. Avg. Meas Total Rut = 0.098 in. Bias = -0.068 in. p-value = 0.000 Reject H0 (p-value <0.1) y = 0.4322x + 0.0265 R^2 = 0.2977 Figure 7. Hypothetical illustration of predicted versus measured total rut depth for full-depth HMA pavements.

27 Null Hypothesis Parameter Untreated Sections Preservation A- Treated Sections Preservation B- Treated Sections Number 81 55 61 Avg. predicted rutting, in. 0.145 0.175 0.186 Avg. measured rutting, in. 0.118 0.092 0.077 Bias (er), in. -0.027 -0.082 -0.109 R2 0.37 0.50 0.40 Se 0.061 0.045 0.039 H0: ∑(ymeasured – xpredicted) = 0 T-test p-value 0.001 0.000 0.000 Accept/reject H0 Reject Reject Reject H0: bo = 0 Regression p-value (intercept) 0.002 0.974 0.934 Accept/reject H0 Reject Accept Accept H0: m = 1.0 Regression p-value (slope) 0.000 0.000 0.000 Accept/reject H0 Reject Reject Reject Note: Hypothesis testing performed at 10% significance level. Table 14. Bias statistics for a hypothetical rutting model. focuses on traffic, climate, pre-treatment pavement con- dition, and treatment material/mix characteristics should provide a basis for adjusting the calibration coefficients. Tables 15 and 16 list the model coefficients that can be adjusted to reduce bias. Figures 8 and 9 show the Pave- ment ME Design program menu screens where the model calibration adjustments can be made for new flexible and new rigid pavements, respectively; similar menu screens are available in the program for HMA overlays and PCC rehabilitation treatments. Different approaches have been used to adjust the model coefficients and improve prediction accuracy and reduce prediction bias. One frequently used approach involves performing numerous Pavement ME Design runs using a large factorial of values for key coefficients (e.g., br2 and br3 for rutting, bf 2 and bf 3 for fatigue cracking) and Null Hypothesis Parameter Untreated Sections Preservation A- Treated Sections Preservation B- Treated Sections Number 81 55 61 Avg. predicted rutting, in. 0.131 0.112 0.080 Avg. measured rutting, in. 0.118 0.092 0.077 Bias (er), in. -0.013 -0.019 -0.003 R2 0.79 0.93 0.89 Se 0.034 0.017 0.017 H0: ∑(ymeasured – xpredicted) = 0 T-test p-value 0.002 0.000 0.177 Accept/reject H0 Reject Reject Accept H0: bo = 0 Regression p-value (intercept) 0.501 0.210 0.531 Accept/reject H0 Accept Accept Accept H0: m = 1.0 Regression p-value (slope) 0.000 0.000 0.000 Accept/reject H0 Reject Reject Reject Note: Hypothesis testing performed at 10% significance level. Table 15. Summary of rutting model bias statistics for untreated and preservation-treated sections following bias elimination/reduction.

28 then using Microsoft Excel Solver to determine the opti- mal values for all coefficients that give the smallest sum of squared error (SSE) between the predicted and measured distress/smoothness. Another approach involves optimiz- ing all model coefficients simultaneously using the genetic algorithm (GA) optimization technique within MATLAB (Kim et al. 2011). In the hypothetical example presented earlier, the untreated pavement group exhibits low precision and low bias, and the two preservation-treated groups exhibit high precision and high bias. After a detailed evaluation of the effects of different factors on bias, the rutting cal- ibration coefficients (bs1, br1, br2, br3) were modified to reduce the difference between measured and predicted rutting values; the resulting predicted versus measured plots are shown in Figure 10, and the corresponding bias statistics are listed in Table 15. Hypothesis testing still indicates an unacceptable level of bias for each group, but the prediction capability and accuracy of each has been greatly increased, and the bias has been greatly decreased. 9. Assess the Standard Error of the Estimate: The standard error of the estimate for each recalibrated model and each analysis cell is compared with reasonable values of the standard error of the estimate provided in the MEPDG Pavement Type/Design Preservation Treatment/Strategy Interstate and Other Freeway Routes (NFC-1 and NFC-2) Other Principal Arterial and Minor Arterial Routes (NFC-3 and NFC-4) Climate Zone 1 (Severe) Climate Zone 2 (Moderate) Climate Zone 1 (Severe) Climate Zone 2 (Moderate) New/Reconstructed Flexible Pavement or HMA-Overlaid Flexible Pavement (0) Untreated control 2 2 2 2 (1) Double microsurfacing 2 2 2 2 (2) Thin HMA overlay (1.5–2.0 in.) 2 2 2 2 Table 16. Experimental/sampling matrix for Michigan preservation-based local calibration. Figure 8. Distress model calibration settings—new flexible pavements.

29 Figure 9. Distress model calibration settings—new rigid pavements. Figure 10. Comparison of predicted and measured total rut depth for full-depth HMA pavements following bias elimination/reduction. Note: Pres-treated = preservation-treated.

30 Manual of Practice (AASHTO 2008); these values are listed in the following. • HMA-Surfaced Pavements – Bottom-Up Alligator Cracking: 7% of total lane area. – Top-Down Longitudinal Cracking (confined to wheel paths): 600 ft/mi. – Reflective Cracking (confined to wheel paths, and combined with alligator and longitudinal cracking in wheel paths): 600 ft/mi. – Rut Depth: 0.10 in. – Transverse Thermal Cracking: 250 ft/mi. • PCC-Surfaced Pavements – Transverse Joint Faulting in JPC (mean): 0.05 in. – Transverse Slab Cracking in JPC (bottom-up and top-down): 7% cracked slabs. – Punchouts in CRC: 4 punchouts/mi. Null hypothesis statistical testing for the experimental/ sampling matrix will result in one of three possible out- comes. These outcomes and recommended courses of action are: • Errors are not significantly different: The calibrated factors can be used (no attempts to reduce standard error are required). • Errors are significantly different, but the errors of the calibrated factors are smaller than those of the MEPDG-calibrated factors: the locally calibrated fac- tors can be used (no attempts to reduce standard error are required). • Errors are significantly different, but the errors of the calibrated factors are greater than those of the MEPDG- calibrated factors: the model should be recalibrated to lower the standard error (unless a higher standard error is considered acceptable). 10. Reduce Standard Error of the Estimate: A high standard error can be reduced by (a) computing the standard error within each cell of the experimental/sampling matrix and determining if the local standard error term is depen- dent on any of the matrix factors (such as preservation strategy), and (b) adjusting the calibration values of the distress transfer functions to reduce the standard error of the recalibration data set considering the coefficients of the MEPDG models identified in the AASHTO Local Calibration Guide. The values for the coefficients of the model are then improved by evaluating the goodness of fit using either an analytical approach (for models that suggest a linear relationship) or a numerical optimiza- tion approach (for models that suggest a nonlinear rela- tionship). If the standard error cannot be significantly reduced due to large measurement error, then proceed with Item 11. 11. Interpretation of Results, Deciding on Adequacy of Calibra- tion Parameters: The standard error of the estimate for each distress/smoothness prediction model is evaluated to determine the effect on the resulting designs at dif- ferent reliability levels. This is done by determining the expected design lives (for different reliability levels) for typical site features and pavement structures or rehabili- tation strategies; results are checked for reasonableness. Attempts to reduce the standard error of the estimate for specific models should take into consideration adjusting the calibration factors or possibly modifying the failure criteria or trigger values for these models. Feasibility Assessment Model calibration to account for preservation resembles the concept of calibrating the MEPDG performance models to account for local conditions. The design analysis would use the Pavement ME Design software program and calibra- tion factors for the various performance prediction models (MEPDG models only) to reflect the effects of preservation. Implementing this approach requires a significant level of effort to identify pavement test sections that cover a range of pavement types, preservation treatments/strategies, and traf- fic and climatic conditions and to gather relevant performance and other data. The calibration process requires statistical analyses of prediction model bias and error and identifying new calibration factors through iterative runs of the Pavement ME Design software or other means. The SHA interviews suggested that several agencies have the components needed for implementing this approach. The vast majority deal with new/reconstructed HMA and JPC, as well as HMA overlays of existing flexible and rigid pave- ments, and use three or more preservation treatment types for flexible pavements and at least two treatment types for rigid pavements. Some of the LTPP or PMS sections in these agencies could serve as calibration sections for local condi- tions but not for a variety of climate and traffic conditions. This approach requires no modifications to the Pavement ME Design software and entails no added complexity in the use of the program. It simply uses the preservation-based calibration coefficients in the design analysis computations. Because of the requirement for extensive data covering the long-term performance of a variety of preservation treatments subjected to different levels of traffic and cli- mate, this approach is also likely to be implemented as part of a national research effort or a multi-agency cooperative research program. An example illustrating the process for calibrating MEPDG models for preservation is presented in the following.

31 Example of Implementation Process The Michigan Department of Transportation (MDOT) maintains a database covering many years of preservation data for hundreds of pavement sections located through- out the state on roads with different functional classes. For the most part, the underlying pavements were constructed between 1985 and 2002 as part of major rehabilitation, resur- facing, or reconstruction projects. The preservation treat- ments were applied between 1992 and 2008. (In some cases, two or three treatments have been applied during that period.) Some data from this database, together with other data derived from agency specifications, manuals, and reports or otherwise estimated/assumed, are used in a hypothetical example to illus- trate the calibration of MEPDG models to account for the effect of preservation. Untreated sections used in this exam- ple were constructed between 1985 and 2005. The example follows the process described in this chapter and incorporates certain assumptions. Step 1: Select Hierarchical Input Level for Each Input Parameter Because the majority of the analyzed sections were more than 10 years old, and no detailed mix design or materi- als testing data were available for these projects, Level 3 materials inputs were used for the Pavement ME Design runs. Also, because detailed information regarding the traffic used in designing these pavements was not available, the available basic traffic data (e.g., average daily traffic [ADT], percent trucks) were used in combination with the national/default values (i.e., a combination of Levels 1 and 3) for the other traffic parameters. Climate data were clas- sified as Level 1 as they were available from the nearest of 19 weather stations. Step 2: Develop Local Experimental Plan or Sampling Template Performance analysis was only feasible for preservation treatments placed on HMA-surfaced pavements (i.e., new/ reconstructed flexible pavements, HMA-overlaid flexible pavements, and HMA-overlaid rigid pavements) as only a few preservation treatments were placed on PCC-surfaced pave- ments to provide sufficient performance data. This example considers two treatment types for HMA-surfaced pavements: double microsurfacing and thin HMA overlay. The preserva- tion treatments were applied to pavements that were neither severely distressed nor severely distorted in terms of cross- section; over 200 sections/projects of each were available for consideration. In developing the experimental/sampling matrix, the fol- lowing types of traffic, climate, and pavement were considered: • Climatic Zone – Moderate: Hot summers and cold winters (southern and central parts of the Lower Peninsula). – Severe: Warm, but shorter summers and longer, cold to very cold winters (northern part of Lower Peninsula and entire Upper Peninsula). • Traffic – Moderate to High: Interstates and other freeways (National Functional Classification [NFC] Categories 1 and 2). – Low: Other principal arterials and minor arterials (NFC Categories 3 and 4). • Pavement Type – New/Reconstructed Flexible Pavements: HMA on aggre- gate base and subbase. – HMA-Overlaid Flexible Pavements: Structural HMA overlays of existing flexible pavements. Detailed pavement cross-section data were not readily avail- able for many of the sections; only information on the basic pavement type (i.e., flexible, composite, or rigid) was available. New/reconstructed flexible pavements and HMA-overlaid flexible pavements were combined into one category. Considering the recommended minimum numbers of sec- tions of 20 for rutting and the availability of sections with adequate performance data, a goal of at least two pavement sections for each cell (i.e., combination of traffic, climate, and preservation treatment/strategy) was established for a total of 24 sections, as shown in Table 16. Step 3: Estimate Sample Size for Specific Distress Prediction Models According to Equation 1, the number of sections required for analysis for a 90% level of significance (Za/2 = 1.601), a 0.5-in. rut depth threshold (s, the assumed threshold value for this example), and a 0.1-in. rut depth standard error of the estimate (Se) is 25. This number is very close to the goal of 24 presented in Table 16; therefore, attempts were made to locate these sections. Step 4: Select Roadway Segments Table 17 shows the pavement sections identified for cali- brating the rutting model. All of these sections had a major structural improvement performed between 1986 and 1999, consisting of either a conventional overlay (structural HMA overlay on existing flexible pavement) or a crush/shape-and- overlay (pulverization, mixing, and replacing of existing HMA layers followed by structural HMA overlay). Other improve- ment types included mill-and-HMA overlay, rubblize-and-

32 HMA overlay, and reconstruction with conventional HMA pavement. A preservation treatment was later placed on the improved pavement sometime between 1999 and 2007. Rutting data were available for several years before and after preservation treatment application and were considered sufficient for the calibration. Step 5: Extract and Evaluate Distress and Project Data Rutting and other pavement performance data for the selected sections were obtained and reviewed. The data for total rutting in the pavement structure were obtained from automated surveys performed biennially on the state’s trunk- line roads in accordance with the Distress Identification Man- ual for the Long-Term Pavement Performance Program (Miller and Bellinger 2003). Transverse profiles were measured con- tinuously over the length of testing, and average rut depths for the left-wheel path, right-wheel path, and both-wheel paths were computed for 0.1-mi-long segments. The total rutting for most of the untreated sections exceeded 0.25 in. (50% of the 0.5-in. threshold criterion). For about half of the preservation-treated sections (i.e., double microsurfacing and thin HMA overlay sections), total rutting was about 0.25-in.; the remaining sections had total rutting of at least 0.15 in. Available traffic (average annual daily traffic [AADT], per- cent commercial trucks), pavement cross-section, and sub- grade (soil type) information for the various sections was compiled and reviewed. Some materials data (e.g., asphalt binder grade) were available, but other materials inputs (e.g., HMA mix volumetrics and dynamic modulus, aggregate base, subbase, and subgrade soil resilient moduli) were esti- mated from data in related reports (Buch et al. 2008, Baladi et al. 2009, Von Quintus and Perera 2011) or the LTPP data- base (DataPave). The national/default values contained in the AASHTOWare Pavement ME Design software were used for the remaining materials and traffic input data. Step 6: Conduct Field and Forensic Investigations No supplemental testing was required or performed. Step 7: Assess Local Bias (Verification of Global Calibration Values to Preservation) The performance of the treated pavement structure was computed for each section using the Pavement ME Design software. The computed values of total rutting (at 50% reli- ability) and the measured values for similar sections (i.e., untreated, double microsurfacing, and thin HMA overlay sec- tions) were then plotted for comparison. Figure 11 shows the plots for the untreated, double microsurfacing, and thin HMA overlay sections. The figure also shows a linear trend line fit- ted through all of the predicted versus measured data points (data for all three sets of sections) and lists the relevant sta- tistics for a combined/overall rutting model. These statistics indicate very poor model prediction (R2 = 0.03) and that each null hypothesis regarding model bias was rejected at the 10% significance level. Thus, model recalibration was necessary. Table 18 summarizes the results of similar testing for indi- vidual sets of sections in which the predicted rutting was considerably greater than actual rutting (>0.5 in. versus <0.25 in.), and model prediction capabilities were very poor (R2 ≤ 0.11) such that recalibration was required. Pavement Type/Design Preservation Treatment/ Strategy Interstate and Other Freeway Routes (NFC-1 and NFC-2) Other Principal Arterial and Minor Arterial Routes (NFC-3 and NFC-4) Climate Zone 1 (Severe) Climate Zone 2 (Moderate) Climate Zone 1 (Severe) Climate Zone 2 (Moderate) New/Reconstructed Flexible Pavement or HMA-Overlaid Flexible Pavement (0) Untreated control U-2: U.S. 41 Baraga Co. U-5: U.S. 10 Mason Co. U-9: U.S. 131 Mecosta Co. U-10: M-46 Montcalm Co. U-4: M-69 Dickenson Co. U-8: M-66 Missaukee Co. U-11: M-90 Lapeer Co. U-12: M-50 Lenawee Co. (1) Double microsurfacing DM-1: I-75 Crawford Co. DM-2: I-75 Crawford Co. DM-5: I-196 Van Buren Co. DM-6: U.S. 12 St. Joseph Co. DM-3: M-183 Delta Co. DM-4: M-55 Ogemaw Co. DM-7: M-50 Monroe Co. DM-8: M-40 Van Buren Co. (2) Thin HMA overlay (1.5–2.0 in.) TO-1: M-72 Oscoda Co. TO-2: U.S. 41 Houghton Co. TO-5: M-46 Montcalm Co. TO-6: U.S. 131 Mecosta Co. TO-3: U.S. 41 Keweenaw Co. TO-4: M-113 Grand Traverse Co. TO-7: M-57 Kent Co. TO-8: M-52 Ingham Co. Notes: Climate Zone 1 is represented by the Upper Peninsula and the northern half of the Lower Peninsula and consists of MDOT Regions 1 and 2. Climate Zone 2 is represented by the southern half of the Lower Peninsula and consists of MDOT Regions 3 through 7. DM# = double microsurfacing section ID; TO# = thin HMA overlay section ID; U# = untreated section ID. • • • • • • • • • • • • • • • • • • • • • • • • Table 17. Experimental/sampling matrix.

33 Step 8: Eliminate Local Bias of Distress and IRI Prediction Models Because the data presented in Figure 12 and Table 19 indicate high bias and low precision for each set of pavement sections, the data were reviewed to determine if certain factors (e.g., traffic, climate, pavement cross-section, or improvement year) caused these levels of bias and error. No specific factors were identified, but some data inconsistencies, possibly because of the use of different materials (e.g., rubblize-and-HMA overlay, reconstruction with HMA), were observed. These and other data that were found to be in error due to misalignment in the section limits were removed from the analysis. To conduct the calibration, an optimization routine was developed in a Microsoft Excel spreadsheet. The routine included the MEPDG HMA rutting model and the various inputs required to calculate HMA rutting. For expediency, it was assumed that HMA rutting is 25% of the total rutting Figure 11. Predicted versus measured rutting. 0.00 0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 0.90 1.00 0.00 0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 0.90 1.00 M ea su re d To ta l R utti ng , i n. Predicted Total Rutting, in. Untreated Double Microsurfacing-Treated Thin HMA Overlay-Treated H0: bo = 0 and H0: m = 1.0 (Linear Regression) R^2 = 0.03 Se = 0.111 Sy = 0.113 Se/Sy = 0.982 p-value (int) = 0.000 p-value (slope) = 0.083 Reject H0 (p-value <0.1) H0: ∑(ymeas – xpred) = 0 (T-test @10% significance level) n = 97 Avg Pred Total Rut = 0.582 in. Avg. Meas Total Rut = 0.177 in. Bias = -0.405 in. p-value = 0.000 Reject H0 (p-value <0.1) y = -0.1326x + 0.2538 R^2 = 0.0312 Table 18. Summary of rutting model bias statistics. Null Hypothesis Parameter Untreated Sections Double Microsurface- Treated Sections Thin HMA Overlay–Treated Sections Number 47 26 24 Avg. predicted rutting, in. 0.546 0.566 0.668 Avg. measured rutting, in. 0.221 0.148 0.112 Bias (er), in. -0.325 -0.418 -0.548 R2 0.00 0.11 0.01 Se 0.133 0.076 0.057 H0: ∑(ymeasured – xpredicted) = 0 T-test p-value 0.000 0.000 0.000 Accept/reject H0 Reject Reject Reject H0: bo = 0 Regression p-value (intercept) 0.003 0.000 0.273 Accept/reject H0 Reject Reject Accept H0: m = 1.0 Regression p-value (slope) 0.868 0.100 0.581 Accept/reject H0 Accept Accept Accept Note: Hypothesis testing performed at 10% significance level.

34 (i.e., predicted total rutting was computed as four times the predicted HMA rutting). The Microsoft Excel Solver function was used to determine the optimal values of br1, br2, and br3 that give the smallest SSE between the predicted and measured values of total rutting. The resulting plots of predicted versus measured total rutting are shown in Figure 12, and the cor- responding bias statistics are provided in Table 19. Although hypothesis testing indicated an unacceptable level of bias for each group, the prediction capability and accuracy of each have been greatly increased, and the bias has been greatly reduced. Step 9: Assess the Standard Error of the Estimate As Table 19 indicates, the Se value for the double- microsurfacing and thin HMA overlay sections was lower Null Hypothesis Parameter Untreated Sections Double Microsurfacing- Treated Sections Thin HMA Overlay–Treated Sections Number 441,3 252,3 243 Avg. predicted rutting, in. 0.182 0.105 0.107 Avg. measured rutting, in. 0.184 0.107 0.108 Bias (er), in. 0.002 0.002 0.001 R2 0.15 0.14 0.09 Se 0.096 0.042 0.045 H0: ∑(ymeasured – xpredicted) = 0 T-test p-value 0.867 0.770 0.937 Accept/reject H0 Accept Accept Accept H0: bo = 0 Regression p-value (intercept) 0.549 0.436 0.758 Accept/reject H0 Accept Accept Accept H0: m = 1.0 Regression p-value (slope) 0.009 0.071 0.158 Accept/reject H0 Reject Reject Accept Notes: Hypothesis testing performed at 10% significance level. 1 Sample size from Step 7 reduced by three due to removal of data outliers. 2 Sample size from Step 7 reduced by one due to removal of data outlier. 3 Adjustments made to a few of the measured rutting values in Step 7 to correct for misalignment in section limits. Table 19. Modified summary of rutting model bias statistics. 0.00 0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 0.90 1.00 0.00 0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 0.90 1.00 M ea su re d To ta l R utti ng , i n. Predicted Total Rutting, in. Untreated Double Microsurfacing-Treated Thin HMA Overlay-Treated H0: bo = 0 and H0: m = 1.0 (Linear Regression) R^2 = 0.31 Se = 0.072 Sy = 0.087 Se/Sy = 0.828 p-value (int) = 0.528 p-value (slope) = 0.000 Reject H0 (p-value <0.1) H0: ∑(ymeas – xpred) = 0 (T-test @10% significance level) n = 93 Avg Pred Total Rut = 0.142 in. Avg. Meas Total Rut = 0.144 in. Bias = 0.002 in. p-value = 0.791 Accept H0 (p-value >0.1) y = 0.9173x + 0.0137 R^2 = 0.3099 Figure 12. Modified predicted versus measured rutting.

35 than the reasonable value reported for MEPDG rutting model (0.076 in. and 0.057 in., versus 0.10 in.), and Se for the untreated sections was higher (0.133 in. versus 0.10 in.). As Table 19 shows, calibration of the rutting model for each set of sections resulted in lower Se values than the reasonable value reported in the MEPDG rutting model. The errors of the calibrated coefficients appear to be statistically significantly lower than those of the nationally calibrated coefficients. Step 10: Reduce Standard Error of the Estimate Because the Se values were lower than the reasonable values reported in the MEPDG rutting model, no further reductions were necessary. Step 11: Interpretation of Results, Deciding on Adequacy of Calibration Parameters Table 19 suggests some issues with the calibrated coef- ficients for the double-microsurfacing–treated sections, including the evident statistical bias with respect to the inter- cept value for the predicted versus measured relationship (i.e., the intercept is not zero) and the poor predictive capa- bility of the calibrated models (R2 < 50%). Also, an acceptable model could not be developed for the microsurfacing-treated and the thin HMA overlay–treated sections because of the limited range of measured data. (Only about half of the sec- tions exhibited total rutting values at or near the 0.25-in. criterion.)