Investigating the Relationship of As-Constructed Asphalt Pavement Air Voids to Pavement Performance (2021)

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.

19 C H A P T E R 2 Research Approach Objective and Scope The objective of this research was to determine the effect of in-place AV on the performance of asphalt pavements using data from the LTPP database and other appropriate sources. For clarity, the term in-place AV was defined as the as-constructed AV measured during construction. The research focused on four primary distress types related to asphalt pavement performance: rutting, fatigue cracking, transverse cracking, and ride. Three distress types used measures as described in the Distress Identification Manual (Miller, 2003) for the LTPP and ride was measured as international roughness index (IRI). Methodology The research team applied three analysis methods to the LTPP data and conducted validation efforts with data from other sources. The complexity of the influence of as-constructed AV on performance was examined with multiple methods to capture varying levels of available data and the impact of other variables, such as climate and traffic. Method 1 isolated the influence of as-constructed AV by dividing the projects into common subsets, Method 2 used a regression modeling technique to account for incomplete data, and Method 3 used a computer-based information processing technique commonly called artificial neural network (ANN). Overview of the LTPP Dataset A mining approach was used to extract data for this study. A query was created and the key parameter that eliminated the use of most of the LTPP database was the availability of as-constructed AV. The query included the following key boundaries: asphalt pavements, as-constructed AV, more than five years of LTPP monitoring, and more than three distress surveys. A total of 426 LTPP test sections met the criteria. Table 2-1 lists the number of sections from each LTPP group. As shown in this table, there is a reasonable split between new construction (226 sections) and rehabilitation projects (200 sections). Table 2-1. LTPP sections used in this study. LTPP Study Study Description No. of Sections GPS 1 Asphalt concrete on unbound granular base 7 GPS 2 Asphalt concrete on bound base 4 GPS 6B AC overlay with conventional asphalt cement on AC pavement, no milling 40 GPS 6C AC overlay with modified asphalt cement on AC pavement, no milling 16 GPS 6D Multiple AC Overlays with Conventional Asphalt Cement on AC Pavement, No Milling 6 GPS 6S AC Overlay on AC Pavement with Milling and/or Fabric Pretreatment 31

20 LTPP Study Study Description No. of Sections SPS 1 Strategic Study of Structural Factors for Flexible Pavements, New/Reconstructed AC pavements 149 SPS 5 AC Overlay of AC Pavement 90 SPS 8 Study of Environmental Effects in the Absence of Heavy Loads 33 SPS 9N Superpave Asphalt Binder Study (New AC Pavement Construction) 33 Geographic Distribution Figure 2-1 shows the range of the test sections from east to west and north to south. The distribution of sections favors the eastern half of the country, which reflects the distribution of all LTPP sections. Figure 2-1. Geographic distribution of selected LTPP sections (green=active, red=out of study). Climate Figure 2-2 shows that all four LTPP climate regions are represented. Both wet regions, specifically the wet-no freeze region, dominate the dataset.

21 Figure 2-2. Distribution of selected LTPP sections in each LTPP climate zone. Pavement Structure Figure 2-3 shows that the test sections cover a broad range of pavement structures based on the composite value of the computed SN. The majority of the sections are greater than SN=3.5 and less than SN=7.0, which is still a significant range. Figure 2-3. Distribution of selected LTPP sections by pavement structure. Traffic Figure 2-4 shows the cumulative ESALs over the first five years of service for each section. The range of values provides a ten-fold difference in traffic loading. A five-year cumulative value was used to mask any single year traffic anomalies similar to using a running average for quality control charts and agrees with the minimum years of measured performance applied for the LTPP data extraction.

22 Figure 2-4. Distribution of selected LTPP sections by traffic. Pavement Structure and Traffic Correlation Figure 2-5 shows the poor correlation between traffic and pavement structure indicating that both input categories were needed for the analysis. There appear to be pavements that are over-designed (high SN and low ESALs), and pavements that are under-designed (low SN and high traffic). This scatter in the dataset also reflects the criteria for some LTPP groups. For example, SPS 8 pavements are typically over designed so the impact of climate can be examined on common pavement structures but with the absence of traffic. Figure 2-5. Correlation of pavement structure and traffic for selected LTPP sections. As-Constructed Air Voids Figure 2-6 shows the range of as-constructed AV in the selected LTPP sections. The distribution is shown for both non-wheel path and wheel path measurements as a percent of the total for each set. There

23 are twice as many wheel path measurements as there are non-wheel path measurements. There is a slight shift (0.32 AV) between the means of the two datasets to lower wheel path AV. This reflects the delay between the time the pavement is open to traffic and the time the LTPP team arrived to collect as- constructed data on a portion of the projects. The difference in mean values between wheel path and non- wheel path data is negligible when considering the variation in test times and test methods reported in the LTPP database. Figure 2-6. Distribution of selected LTPP sections by as-constructed air voids. LTPP Monitoring Period The study’s dataset was restricted to LTPP sections with more than five years of LTPP monitoring to obtain sufficient data for the analysis. Figure 2-7 shows a good distribution of the years of evaluation. More than half of the sections have more than ten years of evaluation. Figure 2-7. Distribution of selected LTPP sections by duration of monitoring period.

24 Number of Performance Measures Figure 2-8 shows that more than two-thirds of the sections have more than eight increments of performance data collected. Figure 2-8. Distribution of selected LTPP sections by number of performance measures. Overview of Datasets for Validation Data for this study was also collected from the NCAT Test Track and Minnesota Department of Transportation pavement research test facility (MnROAD). Although the number of test sections was substantially smaller, the quality and depth of data for each section was very good. A total of 130 sections from the NCAT Test Track and 57 sections from MnROAD’s mainline, low volume road, and farm road were available. MnROAD’s construction, performance, and weather data were available online along with procedures used to collect the data for pavement distress, surface characteristics, and other non- destructive testing data. Traffic volume and loading data at MnROAD are recorded by weigh-in-motion (WIM) sensors in order to determine the equivalent single axle load (ESAL) values for test pavements. In order to damage NCAT’s pavement sections on the 1.7-mile test oval, a fleet of Class 8 heavy trucks applies 10 million ESALs during each three-year research cycle. Both test facilities use an automated pavement distress data collection vehicle to quantify roughness, macrotexture, rutting, and cracking in the same manner used by most state highway agencies for their pavement management systems. Additional information about dataset assembly is presented in the validation section of this report. Initial LTPP Data Extraction The research team determined that the following seven categories of LTPP data were necessary to accomplish the study. Each category was assembled into an individual spreadsheet and the LTPP section identification code (ID), consisting of the State code and section number, is included in all spreadsheets to tie the dataset together. This subdivision of the dataset into seven categories was used to perform the three analysis methods in this study. The seven categories are: • Pavement structure: A pavement structure is a combination of complex material layers. The LTPP pavement database includes a computed SN and attributes of each layer. This data category includes as-constructed AV. Values for each asphalt layer are needed to account for differences between

25 pavement surface performance (rutting and thermal cracking) and pavement full depth performance (fatigue cracking). • Climate: The LTPP climate data includes the standard four LTPP climate zones and a time history of annual values for temperature and precipitation. • Traffic: The LTPP traffic data includes a time history of average annual daily truck traffic (AADTT), annual ESALs, and cumulative ESALs. The research team used different traffic values for each analysis approach. • Rutting measurements: The maximum mean depth in inches from the left and right wheel paths was used to create a consistent method of measurement to create performance curves. • Fatigue cracking measurements: The LTPP program simply measures the amount and location of cracking, not the probable cause of the cracking. As such, the LTPP data is not classified as fatigue cracking, so the research team defined fatigue cracking as any measured cracking in both wheel paths expressed as a percentage of a 39-inch wheel path based on total lane width, which follows the FHWA 2016 Highway Performance Monitoring System (HPMS) Field Manual criteria (FHWA, 2016). LTPP distress data now distinguishes between types of longitudinal cracking. Longitudinal cracking in the wheel path, which is more logically associated with fatigue, is separated from cracking associated with longitudinal construction joints. Further, a possible cause of wheel path cracking on LTPP rehabilitation projects may be reflective cracking. • Thermal cracking measurements: The LTPP program simply measures the amount and location of cracking, not the probable cause of the cracking. As such, the LTPP data is not classified as thermal cracking, so the research team defined thermal cracking as measured length of transverse cracking per lane mile. Further, a probable cause of transverse cracking on LTPP rehabilitation projects is reflective cracking. • Ride measurements: This category of LTPP data is relatively straight forward and based on the mean value of IRI in both wheel paths. A data mining approach was used to extract data for this study. Queries of the LTPP database were created in order to build the analysis dataset required for this project. A critical key parameter was the availability of as-constructed AV from the LTPP database. The query included the following key boundaries: asphalt pavements, as-constructed AV, more than five years of LTPP monitoring, and more than three distress surveys. The initial queries identified 426 LTPP test sections that met the criteria to be considered for use in the Phase II analysis. Further review of available LTPP data was conducted, and the analysis dataset used on this project was refined. Additional LTPP Data for Analysis Methods 2 and 3 This section describes the initial processing of the LTPP data selected to develop an analysis dataset for Analysis Methods 2 and 3. The initial LTPP dataset for this study was comprised of 426 sections distributed across 44 States as displayed in Figure 2-9.

26 Figure 2-9. Location of LTPP sections initially selected for analysis. The main goal of the data mining and processing effort was to develop a complete and reliable dataset for the analysis of this study. The main steps carried out for developing the analysis datasets for Analysis Method 2 (regression analysis) and Analysis Method 3 (artificial neural network analysis) are listed as follows. For continuity, the research team elected to use the variable units extracted from the LTPP database, which are a mixture of metric and US customary (English) units. 1. Assemble climate data. This step consisted of identifying specific climate parameters for each LTPP section. 2. Combine layer attributes into section-level attributes for each construction event. This step consisted of generating the explanatory variables for analyses corresponding to each LTPP section (defined by the combination of STATE_CODE and SHRP_ID field) and construction event (CONSTRUCTION_NO) by processing and combining layer-level attributes data. For example, if a pavement had three AC layers, a representative AV value for the section was computed as the mean of the layers’ AV weighted by their layer thicknesses. 3. Remove or impute missing values in the dataset. This step consisted of analyzing the extent and type of missing values in the dataset to decide what observations to remove or impute (by inserting an approximate value) in order to obtain a complete section attributes dataset for analysis. 4. Define analysis periods for each test section. This step consisted of determining which construction events caused significant immediate changes in the surface condition or significant changes in the performance trend. These significant events defined periods of analysis with homogeneous performance characteristics. The resulting period of analysis for each section are referred to, in this report, as performance curves. 5. Assemble and process the condition data and time-dependent variables. This step consisted of generating the dependent variables for model development by processing the condition metric measurements data (pavement layer and materials) as well as gathering and processing time-dependent explanatory variables (climate and traffic) for each performance characteristic processed in the previous steps. 6. Merge the time-dependent performance, climate and traffic data, and section-level pavement structure attributes. The sixth and final step consisted of merging the time-dependent dataset and the complete

27 pavement section attributes dataset to obtain the final analysis dataset for each of the four performance characteristics in this study: ride, fatigue cracking, thermal cracking, and rutting. The following parts of this chapter describe each of the aforementioned data processing steps along with summary statistics of the intermediate datasets. Assemble Climate Data The first step identified and assembled climate data specific to each LTPP section in the study. The LTPP database has a climate module that contains general environmental information from weather stations located near LTPP test sections. In this study, the collected data included: • average annual air temperature, • average annual precipitation, and • freezing index, calculated by Equation 1 where 𝐹𝐼 is the freezing index (unit: degrees Celsius [°C] degree-days); 𝑇 is the average daily air temperature on the ith day when the average daily temperature is below freezing (unit: °C); and 𝑛 is the number of days in the specified period when average daily temperature is below freezing. 𝐹𝐼 = ∑ 0 − 𝑇 (1) Combine Layer Attributes for Each Construction Event The second data processing step had the objective of developing section-level attributes for each time period within construction events from the layer-level data. Pavement layer attributes were combined into three main groups: • asphalt (AC), comprising all asphalt layers in the pavement, • base (BSB), which includes all base and/or sub-base layers (unbound or treated) between the AC layers and the subgrade, and • subgrade (SG), representing the soil properties of the top portion of the subgrade (closer to the surface). LTPP Section General Attributes. The following pavement section attributes were gathered from LTPP database tables for each LTPP section and construction number: • Experiment type, ExpType: each test section and construction event period was assigned to an LTPP experiment type designated by combining the GPS_SPS field (which indicates whether the experiment is part of a general pavement study (GPS) or a specific pavement study (SPS)) and the EXPERIMENT_NO field (which has the experiment number) in the EXPERIMENT_SECTION table. • Pavement section type, SECTYPE: categorical variable equal to “New” for new pavements and “Rehab” for rehabilitated pavements. New pavements were defined as being part of experiments GPS_1, GPS_2, SPS_1, SPS_8, or SPS_9N. Rehabilitated pavements were defined as being part of experiments GPS_6A, GPS_6B, GPS_6C, GPS_6D, GPS_6S, SPS_1, SPS_5, or SPS_9O. • Opening to traffic date, OTDate: the date when the section was opened to traffic, obtained from the LTPP database TRAFFIC_OPEN_DATE field in the INV_AGE table. • Construction event assignment date, CEDate: the assignment date for the beginning of a construction event. When the test section is initially accepted to the LTPP program (i.e., CONSTRUCTION_NO = 1) then CEDate is also the acceptance date. Otherwise, the CEDate updates when the CONSTRUCTION_NO increases.

28 LTPP Section Material Attributes. Asphalt mixture consists of aggregate, asphalt binder, and AV. To characterize the mixture component, the research team collected the asphalt binder content, aggregate gradation, and as-constructed in-situ AV. As-Constructed Air Voids. The air void value for each section and construction event was obtained as a function of the AC layers’ AV. The air void value for each layer was obtained from data in the LTPP database TST_AIR_VOIDS_SECT table. The TST_AIR_VOIDS_SECT table contains summary statistics and other information from AV computed values by lane location (wheel path or non-wheel path) at different points in time for AC layers. The layer average AV value (AIR_VOIDS_AVG field) from this table was used in the computations. The representative as-constructed AV value for each section, layer, and construction event was computed as either the mean of the non-wheel path AV values for the construction event if there was at least one non- wheel path average AV value available; or, the mean of the wheel path AIR_VOIDS_AVG available for the construction event if there were no non-wheel path average AV values available. Once the as-constructed AV estimate for each layer and construction event was computed, the next step consisted of estimating a representative AV value for the section and construction event. Two types of section-level estimated as-constructed AV values were computed: • Surface air voids, AV_s, defined as the as-constructed AV value of the asphalt layer (at least 1- inch thick) closest to the surface, and • Mean air voids, AV_m, computed as the average of the asphalt (at least 1-inch thick) layers’ as- constructed AV values weighted by the layers’ thicknesses, expressed as %Gmm_m. %𝐺𝑚𝑚 = ∑ % ∑ (2) Where: %Gmmm = mean as-constructed AV for the section; thl = thickness of AC layer l, with l being the total number of AC layers in the pavement; %Gmml = as-constructed AV of layer l; and %Gmmm was used as an explanatory variable for the analysis of fatigue cracking; while %Gmms was used for the analyses of IRI, thermal cracking, and ride. Layer Thickness. The following thickness variables were computed for each section and construction event using the LTPP database REPR_THICKNESS field from the TST_L05B table: • AC total thickness, ThAC: computed as the sum of all AC layer thicknesses. AC layers were identified as having LAYER_TYPE field value of “AC”. • BSB total thickness, ThBSB: computed as the sum of all base and sub-base layer thicknesses (base and sub-base layers were identified as having LAYER_TYPE field values of “GB”, “TB”, “GS”, or “TS”. • AC Overlay total thickness, OvThAC: computed as the difference in ThAC before and after a construction event. If the difference in ThAC was negative (milling), then OvThAC was set to zero. Asphalt Mixture Binder Content. The binder content value for each AC layer and construction event was estimated as the average of the LTPP database ASPHALT_CONTENT_MEAN measurements in the TST_AC04 table. Two representative binder content values for section and construction event were computed from the layer estimates: • Surface binder content, BCs, defined as the binder content value of the asphalt layer (at least 1-inch thick) closest to the surface used as a variable for the analysis of rutting, thermal cracking, and ride, and

29 • Average binder content, BCm, computed as the mean of all asphalt layers’ binder content values (𝐵𝐶 ) weighted by the layers’ thickness used as a variable for the analysis of fatigue cracking. 𝐵𝐶 = ∑∑ (3) Where: BCm = mean binder content for the pavement section; thl = thickness of AC layer l, with l being the total number of AC layers in the pavement; and BCl = binder content of layer, l. Asphalt Mixture Aggregate Gradation. The AC gradation variables for each pavement section and construction event were estimated from the surface AC layer gradation data. The surface AC layer was defined as the AC layer (at least 1-inch thick) closest to the surface. Two types of variables were used to characterize the aggregate gradation for the surface AC layer: • percent passing sieves No.4 Gr#4 and No.200 Gr#200, obtained from INV_GRADATION table, and • parameters Grk,l and Grλ,l, obtained from fitting a Weibull distribution (equation below) to the aggregate gradation data. (Masad et al, 2009) 𝐹(𝑥; 𝑘, 𝜆) = 1 − 𝑒 ( / ) (4) Where: Fx;k,λ = cumulative distribution function of Weibull distribution; x = aggregate size, mm; and k,λ = Weibull parameters, two model coefficients known as shape and scale parameters, respectively. Asphalt Mixture Dynamic Modulus. The determination of material stiffness considered the viscoelasticity of asphalt mixture and the stress-dependency of unbound material. In the LTPP database, the dynamic moduli of an asphalt mixture represents its stiffness at various temperatures and loading frequencies. Kim et al. (2011) utilized the ANN approach to develop five hierarchical models to estimate the dynamic moduli of asphalt mixtures. Model 1 is based on LTPP laboratory indirect resilient modulus tests performed at three temperatures, which has the highest prediction accuracy; Model 2 is a binder viscosity based model; Models 3 and 4 are based on dynamic shear modulus of asphalt binder with different aging inputs; and Model 5 is only based on asphalt binder grade, which has the lowest prediction accuracy. The LTPP database uses these models to estimate the dynamic moduli of an asphalt mixture when the laboratory testing results are missing. Note that the LTPP database sometimes uses two or more prediction models to estimate the dynamic moduli of the same material. In that case, the dynamic moduli with higher or highest reliability were extracted for this study. To better associate material conditions for each performance category, the following three dynamic modulus values for asphalt mixture were determined: • dynamic modulus at 14°F and 0.5Hz applied to the transverse cracking model; • dynamic modulus at 70°F and 10Hz applied to the fatigue cracking and ride models; and • dynamic modulus at 130°F and 10Hz applied to the rutting model. The representative dynamic modulus value for the section used for the analysis of ride (E*IRI), thermal cracking (E*ThC), and rutting (E*Rut) performance was defined as the E* values of the surface AC layer, defined as the AC layer (at least 1-inch thick) closest to the surface. The representative dynamic modulus value for the section used for the analysis of fatigue cracking (E*FtC) was computed as a function of all AC layers (at least 1-inch thick) dynamic modulus using the following compound equation.

30 𝐸∗ = ∑∑ / ,∗ (5) Where: E*FtC = compound AC dynamic modulus for analysis of fatigue cracking; thl = thickness of AC layer 𝑙, with 𝐿 being the total number of AC layers in the pavement; and E*FtC,l = AC dynamic modulus at 70F and 10Hz of AC layer l. Base/subbase and Subgrade Resilient Moduli. The resilient moduli of unbound materials including unbound aggregates and untreated subgrade soil are dependent on the stress level, which are determined in compliance with AASHTO T 307. The resilient modulus of each BSB layer (MrBSB,l) or SG (MrSG) was computed from the layer-level information as follows. For unbound materials (either BSB layers or SG) having triaxial test data available, the Mr value was estimated using the 2002 Pavement Design Guide constitutive equation (shown below). The 𝑘 , 𝑘 and 𝑘 model parameters were estimated using the triaxial test results from the LTPP database TST_UG07_SS07_WKSHT_SUM table. Witczak (2003) suggested reporting the resilient modulus of subgrade soil at 14 kPa (2 psi) confining pressure and 41 kPa (6 psi) deviatoric stress, and the resilient modulus of unbound aggregates at 35 kPa (5 psi) confining pressure and 103 kPa (15 psi) deviatoric stress. These resilient modulus values were used to represent the stiffness of unbound materials. 𝑀𝑟 = 𝑘 𝑃 oct + 1 (6) Where: Mr = laboratory resilient modulus; σ1 = major principal stress = NOM_MAX_AXIAL_STRESS + CON_PRESSURE; σ2 = intermediate principal stress = CON_PRESSURE; σ3 = minor principal stress = σ2; θ = bulk stress = σ1 + σ2 + σ3; σd = deviatory stress = σ1 - σ3; 𝜁oct = octahedral shear stress = √ 𝜎 ; Pa = atmospheric pressure = 101 kPa (14.7 psi); and k1,k2,k3,k6 = regression parameters. For BSB layers with unbound materials and no triaxial test data available the following default values were used. • For unbound BSB, MrBSB = 20 ksi was used. • For treated aggregate base, Mr = 100 ksi was used. • For treated soil subbase, MrSG = 15 ksi was used. • All SG materials in the experiment had triaxial test data available. Considering that LTPP pavement sections have various types of underlying layers, the concept of composite moduli was used to reflect the overall stiffness of these materials. Based on the layered elastic theory, the following equation was used to calculate the composite moduli of the materials between asphalt and untreated subgrade. The BSB resilient modulus (𝑀𝑟 ) for each section and construction event was computed from the layer-level BSB resilient moduli (𝑀𝑟 , ) using the following compound equation: 𝑀𝑟 = ∑∑ / , (7)

31 Where: MrBSB = compound BSB resilient modulus of section; thl = thickness of base or sub-base layer l, with L being the total number of base or sub-base layers in the pavement section; and MrBSB,l = BSB resilient modulus of layer l. Remove or Impute Data The resulting section attribute dataset contained missing values for some of the variables. The extent and type of missing data for each variable were analyzed to determine whether the LTPP section construction event would be removed from the dataset or if imputation of the missing data would be used in order to keep that section in the analysis. This segment describes how the missing data was treated including the methodology and results from the imputation of missing data. Treatment of Missing Data. The first variables analyzed were those related to as-constructed AV. A total of 106 construction events had missing AV_s or AV_m values, which represented 12.3% of the 861 construction events in the section attribute dataset. Since as-constructed AV is the main factor under analysis in this study, all construction events with missing AV data were removed from the dataset. The resulting dataset had 393 test sections that included a total of 755 construction events. The percentage of missing data for each variable in the section attribute dataset after removing those fields with missing AV data is presented in Figure 2-10. Four groups of variables had missing values: gradation, AC dynamic modulus, BSB resilient modulus, and binder content. The highest percentage of missing data among these variables was 3.6%. Based on the low percentage of missing data observed for each variable, the research team decided to use the impute data. Figure 2-10. Percent of missing data for each variable. Imputation of Missing Data. The imputation of missing data was carried out by developing predictive models for each variable with missing data as a function of all other variables. This process was

32 implemented using R programming language (R) “mice” (Mice: Multivariate Imputation by Chained Equations) package (R Core Team, 2019 and Van Buuren and Groothuis-Oudshoorn, 2011). The first step consisted of determining what type of model had the best prediction accuracy by estimating different model types offered by the package and comparing their MAPE (mean absolute percentage error). The best performing model type (lowest MAPE) was random forest. The number of multiple imputations in the mice function was set to 10 and the number of iterations was set to 20. (Refer to Van Buuren (2018) for more information about imputation method parameters.) The missing values were replaced by the mean of all imputed values for the different iterations using the R “sjmisc” package (Lüdecke, 2018). Figure 2-11 shows the distributions of the mean of the imputed values for each observation and of the observed values for the variables with missing data. The almost exact overlap between these distributions shows that the resulting variables’ distribution after merging the imputed values are consistent with the distribution of the observed portion of the variables.

33 Figure 2-11. Distribution of original values and with imputed values for selected variables. Define Analysis Periods During the time a test section is monitored under the LTPP program (between the LTPP Acceptance Date and the Out-of-study Date), every construction event is registered as a change in CONSTRUCTION_NO. Depending on the type of construction, these changes may result in an immediate change in one or more of the performance characteristics, a change in the pavement performance trend (change in deterioration rate Variable D ist rib ut io n (% ) With imputed values Original values Legend

34 over time), a change in both, or in no significant changes to the performance of the section. Consequently, an important data processing step consisted of identifying the time periods (comprising either one or multiple consecutive CONSTRUCTION_NOs) under which the pavement section followed a continual performance trend without sudden improvements in condition. Segmenting of Analysis Periods. Figure 2-12 shows, as an example, the timeline with construction events and other milestone dates corresponding to section 20-0110 in Kansas. In this example, the construction and opening to traffic dates (date at which the analysis period for new pavements started) coincided. The initial CONSTRUCTION_NO value was assigned when material testing and climate data collection started on the LTPP acceptance date (which preceded the test section construction date). The first change in CONSTRUCTION_NO occurred approximately five years later, in 1998, when the section received crack sealing and the second change in CONSTRUCTION_NO occurred in 2001 when the section was milled and overlaid. From the analysis of the effects that each of these construction events had, it was determined that the first event did not lead to significant changes of condition or performance while the second one did. Therefore, two analysis periods were defined for the section: one spanning from the section’s opening to traffic date and the crack sealing date (including both CONSTRUCTION_NOs 1 and 2), and the second spanning from the mill-and-overlay treatment date to the out-of-study date (including only CONSTRUCTION_NO 3). Figure 2-12. Timeline of construction events and milestone dates for Kansas LTPP section 20-0110. In order to determine if a construction event resulted in a significant change in immediate condition and/or performance trend, the following data checks were made. • If the construction work involved milling and/or application of overlay, the event was flagged as a candidate for causing significant change in at least one condition metric and/or performance trend. The CN_CHANGE_REASON codes flagged in the dataset were 12, 19, 43, 51, 55, and 56. • If the AC total thickness increased by 1 inch or more (overlay) or decreased by 1 inch or more (milling), or if the experiment type changed after a construction event, the event was flagged as candidate for change in analysis period. • If the visual inspection of a performance condition plot over time for each performance type, as part of Analysis Method 1, identified a sudden improvement in condition or performance trend, the event was flagged as candidate for change in analysis period. The segmenting of each of the 393 test sections into analysis periods with continual performance trends resulted in a total of 422 periods of analysis (or individual performance curves) for model development comprised of 221 new pavements and 201 rehabilitated pavements. The test section pavement attributes were held constant between the starting and ending dates of each analysis period. Summary statistics of the test section attributes for the 422 analysis periods are presented in Table 2-2. Note that the reported

35 minimum and maximum values for some attributes represent extreme values that would not be considered normal and have very little influence on the analysis as 1 of 400 values. Obvious examples are 221 inches of base/subbase and 4400 ksi for asphalt mixture E* at 130°F. Table 2-2. Summary statistics of complete section attributes dataset for all analysis periods. Variable Units Mean Standard Deviation Minimum p05% p50% p95% Maximum AV_s %Gmm 6.1 3.0 0.1 1.9 5.8 11.2 21.8 AV_m %Gmm 6.0 2.4 1.4 2.6 5.8 10.4 15.2 Th_AC in 7.2 2.8 1.4 3.9 6.9 12.8 21.9 Th_BSB in 18.4 16.5 0.0 4.2 14.9 42.8 221.5 OvTh_ACa in 3.6 1.7 0.9 1.5 3.3 6.8 9.2 E*_IRI ksi 1,082.4 456.3 406.0 575.0 975.1 1,781.2 4,398.9 E*_FtC ksi 1,074.2 411.7 225.0 575.0 1,046.1 1,688.6 4,398.9 E*_TrC ksi 3,167.8 644.8 227.6 2,369.5 3,196.2 4,058.1 5,599.0 E*_Rut ksi 99.2 216.2 16.9 43.5 73.7 156.6 4,398.9 Mr_BSB ksi 32.3 26.8 9.1 14.7 21.1 100.0 100.0 Mr_SG ksi 9.1 3.0 3.3 4.9 8.8 15.1 17.9 BC_s % 5.0 0.7 3.6 4.0 4.9 6.4 9.0 BC_m % 4.8 0.7 3.5 3.8 4.8 6.0 9.0 Gr_4 % 55.9 10.1 22.0 39.0 55.0 72.9 90.0 Gr_200 % 6.1 1.6 0.9 3.7 6.4 8.9 13.0 Gr_k - 5.3 2.2 2.2 3.0 5.1 8.1 15.3 Gr_lambda - 0.8 0.2 0.6 0.7 0.8 1.1 1.7 aThe statistics shown correspond to the positive AC Overlay total thickness values. Process Condition and Time-dependent Variables The next processing step consisted of gathering and processing the condition measurements to be used as dependent variables in the performance models. These measurements throughout time, as well as the time-dependent explanatory variables, were considered for model development for each year with measured condition data. Performance Measurements. Four types of pavement performance characteristics were used for analysis: rutting, fatigue cracking, thermal cracking, and ride. The following bullets describe the different metrics used for model development by performance characteristic type: • Rutting was captured as the maximum measured value between the right and left wheel path mean value using a 1.8-m straight edge rut depth following LTPP rutting definitions. The rutting measurement at each point in time (Rutt) was obtained from the LTPP database MAX_MEAN_DEPTH_1_8 field in the MON_T_PROF_INDEX_SECTION table. • Fatigue cracking was captured as percent cracking in the wheel path defined in the 2016 FHWA HPMS Field Manual (FHWA, 2016) for asphalt pavements (see equation below). The cracking percent measurements were obtained from the LTPP database HPMS16_CRACKING_PERCENT_AC field in the MON_DIS_AC_REV table. The amount of fatigue cracking at each point in time was referred to as wheel path cracking (WpCt) for new pavements and as reflected fatigue cracking (RfCt) for rehabilitated pavements.

36 𝐶𝑃 = ( ) (8) Where: CP2016HPMS = cracking percent, in percent; wpwidth = wheel path width, 39 inches (1.0 m) as defined in the 2016 HPMS Field Manual; Lcrklwp = length of left wheel path with fatigue cracking; Lcrkrwp = length of right wheel path with fatigue cracking; W = lane width; and L = lane length. • Thermal cracking was captured as the total length of transverse cracks greater than 6 ft (1.83 m) per mile following the LTPP transverse cracking definition. This metric was computed as the total length of transverse cracks obtained from the LTPP database TRANS_CRACK_L_GT183 field in the MON_DIS_AC_REV table divided by the section length, then multiplied by 5280 to achieve feet per mile units. The amount of thermal cracking at each point in time was referred to as transverse cracking (TrCt) for new pavements and as reflected thermal cracking (RtCt) for rehabilitated pavements. • Ride quality of the pavement was captured as IRI measurements. The IRI value at each point in time, IRIt, was computed as the average of the left and right wheel path Mean Roughness Index (MRI) values for each data collection date (VISIT_DATE) reported in the LTPP database MON_HSS_PROFILE_SECTION table. The last step in the processing of the performance characteristic data was to compute annual estimates for each distress as the average of all measurements available for the different years with performance data. For example, the annual rutting value for a given section s and year yr, Rutyr,s, was computed as Rutyr,s = ∑ Rutt / N; N being the number of rut measurements available for year yr. Thus, a unique value per year was available for each surface performance characteristic in the resulting dataset. Age Variables. The age of the pavement for each year with surface performance data was computed as the number of years elapsed between the initial date of the analysis period and the year of the performance condition measurement. For performance curves corresponding to new pavements, the initial date of the analysis period was defined as the date when the LTPP section was opened to traffic (i.e., OTDate of performance curve). For performance curves corresponding to rehabilitated pavements, the initial date of the analysis period was defined as the date when the construction event (CONSTRUCTION_NO) of the LTPP section was assigned to the LTPP program (i.e., minimum CEDate of performance curve). While the initial date of the analysis period was independent of the pavement performance measurements under analysis, not all performance measurements were collected on the same dates nor did all performance measurement types have the same number of measurements over time. Consequently, each condition data table may present a different range of ages computed for each test section performance characteristic. Annual Climate Variables. Three variables were considered to capture the effect of local climate on performance: temperature, freezing index, and precipitation. The LTPP tables used to gather these variables were populated with virtual weather station data. Annual estimates of these climate variables are described as follows: • Temperature, Tempyr, was captured by the average of daily mean air temperatures for the year, obtained from the LTPP database MEAN_ANN_TEMP_AVG field in the CLM_VWS_TEMP_ANNUAL table. • Freezing Index, FIyr, was captured by the calculated freezing index for the year following LTPP definitions obtained from the LTPP database FREEZE_INDEX_YR field in the CLM_VWS_TEMP_ANNUAL table.

37 • Precipitation, Precyr, was captured by the total precipitation for the year obtained from the LTPP database TOTAL_ANN_PRECIP field in the CLM_VWS_PRECIP_ANNUAL table. Annual Traffic Variables. The effect of traffic on pavement performance was defined by ESALs. The two types of ESAL data available in the LTPP database tables at the time data processing was initiated were the computed ESAL (estimated from traffic counts and limited measured ESAL data) and the measured ESAL data (from WIM or truck scales). The quality and main characteristics of the two ESAL data types available for the sections in the study were assessed by the research team prior to their processing. The main findings from this assessment were that neither dataset was consistently reported through the performance analysis periods for any one LTPP section, and that the computed and measured ESAL data for the same section and year presented significant differences in magnitude and/or trend for many of the study sections and periods of analysis. Given the issues listed above, the research team explored alternative approaches to obtain more complete and reliable ESAL data. The selected approach was the use of a newly estimated ESAL dataset developed as part of a LTPP analysis contract (Selezneva and Hallenbeck, 2018) that merged the computed and measured datasets. This new ESAL data type was referred to in this study as trend ESAL data. The review of the new traffic data showed completeness and reasonable ESAL estimates and trends for all test sites. Consequently, the trend ESAL data was deemed acceptable for the purposes of this study. The annual traffic variables extracted and generated from the trend ESAL dataset for each year with condition data were: • Annual ESAL ESALyr: estimated ESALs for the section at year yr–from the traffic dataset developed by Selezneva and Hallenbeck (2018). • Cumulative ESALs CumESALyr: cumulative ESALs from the initial year of the analysis period to year 𝑦𝑟, computed as 𝐶𝑢𝑚𝐸𝑆𝐴𝐿 = ∑ 𝐸𝑆𝐴𝐿 For rehabilitation projects, the research team also proposed to use a cumulative traffic value, PreTRF, to represent the age and condition of the pavement prior to the rehabilitation. Unfortunately, the LTPP database only has this data on a limited number of projects, so the research team did not pursue this parameter for the study. Final Analysis Data Tables The final step in developing the analysis datasets for Analysis Methods 2 and 3 consisted of merging the time-dependent datasets (one for each of the four performance characteristic types) and the complete section attributes dataset for each performance curve. The main characteristics of the resulting data tables are presented in Table 2-3. A more detailed description of each of these analysis data tables can be found in in the following sections. Table 2-3. Main characteristics of the data tables developed for Analysis Methods 2 and 3. Analysis Dataset Count Type New Pavements Rehab. Pavements Total Rutting Observations 1,827 1,382 3,209 Test Sections 219 156 375 Perf. Curves 219 172 391 Fatigue Cracking Observations 1,534 1,142 2,676 Test Sections 213 165 378 Perf. Curves 213 167 380 Thermal Cracking Observations 1,534 1,120 2,654Test Sections 215 164 379