CHAPTER 7. CONCLUSIONS AND SUGGESTED RESEARCH

This chapter answers the research questions listed in Chapter 1, summarizes other key findings and conclusions drawn from the analyses conducted in this project, and provides suggestions for further research.

RESEARCH QUESTIONS

• What are the most (and least) effective assignment methods?

  Most effective:

  • Cluster analysis (M10 and M13) is the most effective method in creating homogeneous adjustment factor groups. Some clustering algorithms such as the k-prototypes and hierarchical clustering allow the use of both continuous and categorical variables that help, to some extent, in creating clusters that share common characteristics. However, ultimately, analysts have to manually review and refine the produced clusters or employ other statistical or ML methods such as support vector machines (M15) and decision trees (M14) to assign counts to clusters. Support vector machines are slightly more effective than decision trees in assigning counts to clusters, but they have a more complex structure.
  • Further, grouping CCSs by functional class and rural/urban area type (M3) tends to produce slightly less accurate AADT estimates than cluster analysis, but the development of CCS groups and the assignment of counts to groups are easy and straightforward.
  • Annualizing counts using probe-based adjustment factors (M18) is a promising method that can potentially produce more accurate AADT estimates than traditional methods as long as the penetration rate of the probe data is sufficient and the correlations between probe-based and actual adjustment factors derived from CCSs are strong (>0.85).

  Least effective:

  • Among all methods examined in this study, not factoring counts (M1) is the least effective approach. Other methods that exhibit lower effectiveness include assigning counts to individual CCSs using decision trees (M16); developing clusters using inputs with limited or no temporal adjustment factors (M13); and using probe-based adjustment factors (M18) developed from raw probe data that have low penetration rates.
• What are the most (and least) important assignment attributes?

  Most important:

  • The most important attributes used in decision trees (M14) and support vector machines (M15) are the 8,760 hourly factors by day of year, followed by the 2,016 hourly factors developed by month and day of week. The most important attributes used to develop clusters are the 84 monthly day-of-week factors (M10), followed by the 12 monthly factors (M13).

  Least important:

• • Attributes with comparatively lower importance, particularly when they are used alone in a model, include census variables (e.g., population density, etc.) and traffic volumes (e.g., 12 monthly volumes and 365 daily volumes).
• Is it better to use functional classification or volume factor groups?
  • When CCSs from all functional classes were taken into consideration, all four methods—functional classification (M2), functional classification combined with area type (M3), five volume groups (M4), and 10 volume groups (M5)—resulted in similar AADT accuracy. FC_RU (M3) and FC (M2) performed slightly better than volume groups (M4 and M5) in lower functional classes, FC6 and FC7. This can be attributed to the fact that in M2 and M3, group adjustment factors are developed exclusively from CCSs belonging to FC6 and FC7. In contrast, volume groups may contain CCSs from all seven functional classes, making the groups more variable and less effective. Volume groups are suitable when CCSs are unavailable on lower functional classes, and therefore, functional classification (M2) cannot be applied.
  • Another relevant finding is that dividing five volume groups (M4) into 10 volume groups (M5) did not enhance AADT accuracy. Therefore, in most cases, five or fewer volume groups should be preferred over a high number of volume groups.
• What is the anticipated improvement in AADT accuracy if we use cluster analysis?
  • Cluster analysis tends to produce significantly more homogeneous clusters than the traditional grouping methods, though the improvement in AADT accuracy is not as pronounced. For instance, the overall within-group variability (WACV) of seven clusters (M10) is around 6.4 percent, while that of the seven FCs (M2) is much higher (11.8 percent). The MAPE of cluster analysis (M10) performed assuming that the cluster membership of counts is known was 6.7 percent, whereas the corresponding error from functional classification (M2) was higher, 9.3 percent. The MAPE obtained from using support vector machines (M15) to assign weekday counts to clusters was 7.8 percent.
• What is the expected increase/reduction in AADT accuracy if we combine two or more assignment methods?
  • It depends on the methods to be combined. Combining two or more methods may result in a small increase in AADT accuracy or slightly reduce the accuracy but mitigate individual method limitations. For example, across all states and years, factor groups developed by combining functional classification with rural/urban area type (M3) yielded slightly lower AADT estimation errors, around 0.5 percent, than developing groups only by FC (M2). In contrast, combining cluster analysis (M10) with decision trees (M14) tends to reduce the AADT accuracy of M10 by 1.0–2.0 percent, depending on the number of clusters created, but facilitates the assignment process, making it data driven. Another example is that when functional classification is combined with cluster analysis, the latter tends to become slightly less effective, but the produced clusters are better defined.
• Can clusters be easily defined from a practical/assignment standpoint?
  • Some clustering methods, such as the k-prototypes partitioning algorithm and hierarchical clustering, can handle both continuous and categorical variables, which,

• • to some extent, aid in improving the definition of clusters by incorporating some assignment characteristics. Clustering may be a good starting point for identifying similar patterns and determining factor groups. However, findings from this project and previous research studies have shown that there is no statistical or machine learning method nor a set of clustering variables that can automatically and easily produce clusters that are well-defined from a practical perspective without modifying them further. Ultimately, analysts have to manually review and refine the produced clusters. However, there may be cases where some of the clusters cannot be easily defined but have unique traffic patterns that need to be kept as separate factor groups. In this case, one option is to employ other statistical or machine learning methods such as decision trees (M14) to assign counts to clusters in a data-driven manner.
• Which assignment characteristics should be incorporated in cluster analysis?
  • The most important attributes used to develop clusters are the 84 monthly day-of-week factors (M10), followed by the 12 monthly factors (M13). Other variables that can potentially be used to better define the produced clusters include the roadway functional class, area type, land use characteristics (if available), and geographical coordinates of the CCSs. Census variables and distance or proximity variables can also be used but may not be as effective as the attributes stated above.
• Are there other data-driven methods to assign counts to existing clusters or individual CCSs?
  • Decision trees, support vector machines, discriminant analysis, random forests, gradient boosting, and artificial neural networks are examples of statistical and machine learning classification methods that can be used to assign counts to existing clusters or individual CCSs. This research project applied and validated decision trees and support vector machines and found that the latter are slightly more effective than decision trees but have a more complex structure. The validation also revealed that assigning counts to clusters using decision trees (M14) and support vector machines (M15) results in more accurate AADT estimates than assigning them to individual CCSs within the same functional class (M16 and M17).
• In the case of low-volume roads, what is the expected AADT accuracy of counts factored using adjustment factors from higher functional classes?
  • For FC6 and FC7, the average AADT estimation errors of weekday counts annualized using adjustment factors from FC5 (M6) or from 5R and 5U (M7) exceed 13 percent. In comparison, other traditional methods such as functional classification (M2) yield more accurate AADT estimates (MAPE = 11.8 percent) than M6 and M7. For rural local roads (7R), counts annualized using factors from 6R (M9) result in MAPEs of 11.3 percent, which are slightly higher than those (10.6 percent) obtained from M2.
• Is it worth factoring counts? If yes, which assignment method should be used and what is the anticipated level of improvement in AADT accuracy?
  • Yes, it is worth factoring counts. Among all methods examined in this project, not factoring counts (M1) yields the least accurate AADT estimates, with average errors ranging from 11.5 percent (for counts taken on Tuesday–Thursday) to 19.1 percent (for counts conducted on Saturdays and Sundays). Simple traditional methods such as functional classification (M2) and functional class combined with rural/urban code

• • (M3) are expected to improve the AADT accuracy of weekday counts by approximately 26 percent and 30 percent, respectively. More advanced methods such as cluster analysis combined with support vector machines (M15) are expected to improve the AADT accuracy of weekday and weekend counts by more than 37 percent and 50 percent, respectively. Further, annualizing counts using probe-based segment-specific adjustment factors (M18) that are strongly correlated with actual adjustment factors can considerably increase the AADT accuracy of unfactored counts (M1).
• Which assignment method should be used to factor counts on low-volume roads if there is no CCS on these roads?
  • Agencies should ideally operate CCSs on low-volume roads. If CCSs are not available on local roads (FC7), one alternative is to use group adjustment factors from 6R and 6U to annualize counts taken on 7R and 7U, respectively. A second option is to use volume factor groups. A promising option is to use segment-specific probe-based adjustment factors provided that the latter are strongly correlated (>0.85) with actual adjustment factors stemming from CCS data. The higher the correlations, the higher the anticipated accuracy of AADT estimates derived from counts factored using probe-based factors.
• Can temporal adjustment and axle-correction factors be developed from raw probe data?
  • Yes. Some probe data can be a viable data source for calculating temporal adjustment factors, but their effectiveness depends on several factors that are described below. However, as of the publication date of this report, it is more difficult to develop accurate axle-corrections factors due to the limited availability of probe data for specific vehicle classes. The analysis in this project revealed that the effectiveness of probe-based adjustment factors in annualizing counts depends on several interconnected factors, including (a) the penetration rate of the raw probe data, (b) the types of probe data (e.g., GPS devices, location-based service [LBS] devices, connected vehicles, etc.) and their representativeness of the entire vehicle population, (c) the correlations between probe-based factors and actual adjustment factors, and (d) the types of adjustment factors being used (e.g., monthly, daily, etc.). Though exceptions may occur, generally speaking, the higher the penetration rate of the probe data, the stronger the correlation between probe and actual factors, which, in turn, results in more accurate AADT estimates developed from counts annualized using probe-based factors.
  • For example, the results from the validation of probe-based factors (M18) showed that the accuracy of probe-based AADT estimates developed using probe data from only one of the three vendors (Vendor A) was comparable to that obtained from traditional methods, M2–M5. In the other two cases (Vendor B and Vendor C), the AADT estimation errors were higher than those obtained from the no-factoring method (M1). The probe data of Vendor A had an average penetration rate of 5.8 percent, and the Pearson correlation coefficient between the best performing set of probe-based adjustment factors (seven annual day-of-week factors) and actual adjustment factors was around 0.82. The corresponding penetration rates and correlations obtained for Vendors B and C were lower or significantly lower than those of Vendor A.

• • There may be cases where the penetration rates are high but the correlations and thereby the AADT accuracy are low. For instance, the results from M19 showed that the 12 monthly probe-based adjustment factors developed for medium- and heavy-duty trucks using probe data from Vendor C were highly ineffective (MAPE ≈ 42.7 percent), having weak correlations (≈0.32) with actual adjustment factors, despite the relatively high penetration rates (≈28 percent). The latter are likely a data artifact because Vendor C rounds raw probe trip counts to the nearest 1,000. This rounding leads to increased penetration rates but weak correlations and introduces significant AADT estimation errors.
• What is the accuracy of AADT estimates derived from SDCs that have been annualized using probe-based factors?
  • Table 36 shows the average penetration rates, correlations, and AADT estimation errors (MAPE) produced using probe data from different vendors and states (M18).

Table 36. Penetration Rates, Correlations, and AADT Accuracy of Probe-Based Adjustment Factors (M18).

• • These results hold true only for the three states and the probe datasets that were analyzed in this project. Different results may be obtained using data from different vendors or from the same vendor but different states.
• Can probe-based adjustment factors be used to factor counts taken on lower roadway functional classes?
  • Yes, as long as the probe-based factors are strongly correlated with actual adjustment factors. The correlation analysis performed using Vendor A’s seven annual day-of-week factors revealed that the correlations increased from higher to lower functional classes; however, a larger sample size is needed to draw more robust conclusions for Vendor A’s data. The AADT estimation errors slightly increased in the lower functional classes, but they remained less than 10 percent. This increase is expected, though, because the MAPE generally tends to increase as the volumes decrease. For this reason, more room for error is typically allowed in the lower functional classes. The main advantages of probe-based factors over other grouping and assignment methods include a) the ability to develop probe-based adjustment factors separately for every roadway segment of the entire network; b) the elimination of the need to create factor groups; and c) the elimination of the need to assign counts to groups.

• What is the relationship between the accuracy of probe-based adjustment factors and the penetration rate of raw probe data?
  • In general, higher penetration rates tend to more effectively capture temporal changes in traffic volumes, leading to stronger correlations between probe-based factors and actual adjustment factors. This in turn increases the accuracy of probe-based AADT estimates developed from annualized counts. However, many factors may affect the penetration rates, the correlations, and thus the AADT accuracy, such as the types of probe data being used and their representativeness of the entire population. If the probe data only capture a specific subset of the traffic or are biased toward certain conditions or a specific area, they may not be fully representative of the entire population over time, leading to discrepancies between estimated and actual AADT values.
  • Another challenge that can lead to poor correlations and AADT accuracy is vendors obtaining and adding new raw probe data from new sources to increase the penetration rate of their raw probe data and the accuracy of the data products that they develop. This ultimately may affect the temporal traffic patterns captured through raw probe data, particularly if the dates on which new data sources are added are unknown or are not taken into consideration when probe-based adjustment factors are developed. Something similar can happen when existing sources of probe data are discontinued or abandoned by a vendor.

OTHER KEY FINDINGS AND CONCLUSIONS

Other key findings and conclusions drawn from the analysis in this project are provided next for the existing, improved, and new methods examined in this project.

Existing Methods (M1–M9)

Nine traditional, widely used methods were applied and validated in this research. The first five methods (M1–M5) were applied to all FCs. The last four methods (M6–M9) involved using factors from higher FCs to annualize counts on FC6 and FC7. Some of the most important findings from the validation of all nine existing methods are:

• Not factoring weekday counts (M1) results in a significant overestimation of AADT, whereas unfactored weekend counts lead to AADT underestimation. This happens because the average weekend traffic on nonrecreational roads tends to be lower than the AADT. The opposite trend is typically observed for weekday counts.
• The remaining four traditional methods—FC (M2), FC_RU (M3), 5_VG (M4), and 10_VG (M5)—slightly overestimate AADT regardless of the days of the week on which counts are taken.
• There is a high variability in the results from one state to another. Accounting for the rural/urban area type in the factor grouping process tends to slightly improve AADT accuracy in many, but not all states. For example, FC_RU (M3) produced slightly more accurate AADTs, by 0.5 percent, than the other three methods, including FC. The statistical tests showed that despite the small differences (0.5 percent) in the AADT estimation errors produced from these methods, the MAPEs of M3 are statistically different from the errors of the other methods at the 95 percent confidence level. However, this finding does not hold true in all states and cannot be generalized.

• There is a certain point beyond which dividing a group into subgroups based on the same variable does not significantly improve the AADT accuracy (MAPE) even though the within-group variability (WACV) continues to decrease. For example, five volume groups (M4) are less homogeneous than 10 volume groups (M5); however, M4 results in slightly more accurate AADTs than M5. This happens because each volume group can contain highly variable monthly day-of-week patterns/factors, which depend on various spatiotemporal characteristics, not just the volume of each CCS. In general, a grouping variable (e.g., FC in M2 or volume in M4) can only capture some of the variability in the CCSs’ monthly day-of-week patterns.
• A comparison of the results across M2 and M5 reveals an interesting relationship among the three main performance metrics: AADT accuracy (MAPE), within-group variability (WACV), and group factor precision (WAP). In general, MAPE generally exhibits similar trends to WACV but opposite trends compared to precision, with a few exceptions. This means that when MAPE improves (i.e., decreases) from one method to another, WACV tends to improve (i.e., decrease) as well, but the precision tends to worsen (i.e., increase). For example, FC_RU (M3) is more accurate (i.e., lower MAPEs) and generates more homogeneous groups (i.e., lower WACV) than FC (M2). However, the latter produces more precise group factors than M3 because it contains more CCSs per group. Generally, the group factor precision (WAP) is affected to a greater extent by the more pronounced changes in the number of CCSs (i.e., sample size) within a group than from the smaller changes observed in the within-group variability. In other words, increasing the number of CCSs within a group tends to improve (i.e., lower) the group factor precision, even though the within-group variability tends to increase. The positive relationship between MAPE and WACV makes the interpretation of the WACV results more intuitive than those of the WAP.

Improved Methods (M10–M15)

Six methods (M10–M15) were applied to improve cluster analysis, which is recommended in the TMG, and many agencies use it to create adjustment factor groups. Key findings from the validation of the cluster-based methods are summarized next.

• Cluster analysis assuming that the cluster membership of counts is known:
  • When the cluster membership of sample counts was assumed to be known, cluster analysis (M10) outperformed all existing, improved, and new methods examined in this study, yielding more accurate AADTs, more homogeneous groups, and more precise group factors, regardless of the number of clusters created. However, in practice, the cluster membership of actual counts is unknown. Analysts have to refine the produced clusters to create more well-defined factor groups with identifiable characteristics to facilitate the assignment process.
  • The AADT accuracy (MAPE) improves at a high rate when the first few clusters are created. However, beyond a certain number of clusters, referred to in this report as the critical point, creating additional clusters provides negligible or no improvement in AADT accuracy. This critical point is typically between 3 to 7 clusters.
  • One of the factors that affects the critical number of clusters is the total number of CCSs within a state. Generally, a higher number of CCSs tends to lead to a higher critical number of clusters, and vice versa.

• • The within-group variability (WACV) also improves as the number of clusters increases. It exhibits significant improvement when the first few clusters are created, but unlike the AADT accuracy, it does not flatten after the critical point described above. Instead, it continues to decrease as the number of clusters increases, albeit at a lower reduction rate.
  • Unlike the MAPE and the WACV, which exhibit decreasing trends as more clusters are created, the group factor precision (WAP) follows an increasing trend up to a certain point, beyond which it tends to plateau. This happens because the precision is disproportionally affected by the number of CCSs per cluster, which is in the denominator of the precision formula, but proportionally influenced by the variability (i.e., coefficient of variation) of group adjustment factors, included in the numerator of the precision calculation. When more clusters are created, CCSs are moved from old clusters to new clusters. As a result, both the average number of CCSs per cluster and the coefficient of variation decrease. However, the reduction in the average number of CCSs per cluster is more significant and has a greater impact on the precision than the smaller reduction in the coefficient of variation.
• Clustering optimization criteria:
  • The pseudo-F statistic (M11) indicated that the optimal number of clusters in all 135 state-years examined in this research is three (3.0). However, as explained previously, the AADT accuracy may continue to improve when more than three clusters are created, particularly in states where the number of CCSs is high. Therefore, selecting the optimal number of clusters requires a manual review of the results combined with engineering judgment and cannot rely exclusively on the use of the pseudo-F statistic.
  • The elbow approach (M12) involved calculating the WCSS for a different number (k) of clusters, plotting WCSS against k, and then manually determining the optimal number of clusters. The results showed that the WCSS follows the same decreasing trend as the WACV. Both metrics capture the within-group variability. The advantage of the WACV over the WCSS is that the former is expressed as a percentage, similar to the MAPE, making the visualization and interpretation of the results easier. However, attention should be paid to the selection of the optimal number of clusters because both metrics continue to decrease after the critical point, beyond which the MAPE starts to plateau.
• Cluster analysis combined with assignment methods (DTs and SVMs):
  • Several DTs (M14) and SVMs (M15) were developed using six assignment inputs to assign sample counts to different clusters. Three types of clusters were developed using 2,016 HFs by month and day of week, 84 MDWFs, and 12 MFs, respectively. The results showed that the success rates (=correct assignments/total assignments) of all DTs and SVMs tend to decrease as the number of clusters increases. The success rates were significantly higher than the corresponding statistical probability of randomly assigning a count to the correct cluster. For example, in the case of three clusters, the probability of randomly assigning a count to the correct cluster is 33.3 percent, yet all DTs and SVMs had a much higher success rate. However, high success rates do not necessarily correspond to high AADT accuracy.

• • Among the six sets of assignment inputs, the 8,760 HFs developed by day of year yielded the highest success rates across all models. The second most effective set of DT inputs was the 2,016 HFs calculated by month and day of week.
  • The success rates tend to increase as the similarities between the clustering inputs and the assignment inputs increase. The highest success rates were obtained for the clusters created using the 2,016 HFs. This can be attributed to the fact that the 2,016 HFs capture some of the traffic variability within a day, similar to the assignment inputs. On the other hand, clusters developed using the 12 MFs do not capture any of the hourly or daily traffic variability, yielding slightly lower success rates.
  • Overall, the SVMs yielded higher success rates and increased AADT accuracy compared to the DTs. This improvement can likely be attributed to SVMs’ superior ability to more effectively model complex data relationships than the DTs, which have a simpler structure.
  • In M14 and M15, as more clusters are created, the AADT accuracy tends to remain stable or marginally decrease. This can be better understood by considering the fact that, as previously explained in M10, when the cluster membership of the counts is known, and therefore the success rate is 100 percent, the MAPEs tend to remain stable after a certain critical number of clusters. However, in M14 and M15, the success rates are not 100 percent and continue to decrease as the number of clusters increases. As a result, the MAPEs tend to remain stable or slightly increase as more clusters are created.
  • Among the three clustering inputs, the 84 MDWFs yielded slightly lower MAPEs than the 12 MFs. Though the 2,016 HFs had the highest success rates, they yielded, in most cases, slightly higher AADT estimation errors than the other two sets of clustering inputs. This is likely due to the fact that the 2,016 HFs have more noise in their seasonal patterns than the aggregated 84 MDWFs or 12 MFs.
  • The AADT accuracy of assignment methods such as DTs and SVMs is affected more by the assignment inputs than the clustering inputs provided that the clusters have been developed using temporal adjustment factors (e.g., 12 MFs or 84 MDWFs).

New Methods (M16–M19)

Two nontraditional assignment methods, DTs (M16) and SVMs (M17), were used to assign counts to individual CCSs within the same FC. Additionally, probe data from three different states and vendors—one state per vendor—were used to (a) calculate the penetration rate of each vendor’s raw probe data; (b) determine correlations between four sets of segment-specific probe-based factors against the corresponding actual adjustment factors; and (c) annualize sample counts extracted from CCSs using probe-based factors developed for all vehicle classes treated as a single group (M18), as well as for medium- and heavy-duty trucks (M19). Key findings from the validation of these methods are provided next.

• Assignment of counts to individual CCSs within the same FC using DTs and SVMs:
  • DT (M16) and SVMs (M17) were also developed to assign counts to individual CCSs within the same FC. The results revealed that assigning counts to clusters (M14 and M15) results in more accurate AADTs than assigning them to individual CCSs (M16 and M17). This is likely due to the fact that M16 and M17 base their assignments solely on similarities in hourly traffic patterns between the count location and the

• • CCS location on the specific day of the count. In other words, M16 and M17 do not consider any similarities in monthly or seasonal patterns between the two locations (count vs. CCS). Consequently, M16 and M17 assign a count to the CCS with the most similar hourly traffic patterns on the specific day of the count, yet there is no assurance that the monthly or seasonal patterns at the CCS location are similar to those at the count location. However, the factoring process involves using MDWFs—not HFs. Conversely, in M14 and M15, each cluster exhibits similar monthly and day-of-week traffic patterns. This characteristic proves to be more effective in the factoring process than not accounting for monthly and day-of-week patterns at all.
  • Most SVMs (M17) produced more accurate AADTs than did DTs (M16). This can likely be attributed to SVMs’ superior ability to model complex data relationships more effectively than DTs.
  • The best performing set of SVM inputs is the one that contains 8,760 HFs developed by day of year.
• Annualization of counts using probe-based adjustment factors:
  • The average penetration rates vary significantly (0.39–5.83 percent) from one vendor to another, primarily because each vendor obtains its probe data from different sources. The highest penetration rate was determined for Vendor A, which obtains connected vehicle data from several OEMs. The other two vendors use LBS data generated or collected from smartphones or other probe devices.
  • Vendor A’s probe-based factors had the highest correlations (0.795–0.844) with actual adjustment factors. Vendor B’s correlations decreased as the number of factors within each set increased. The highest correlation was 0.549 for the seven DOW factors, and the smallest was 0.311 for the 84 MDWFs. This trend suggests a sample size effect since the average penetration rate for Vendor B is low, at 0.39 percent. The probe data of Vendor C had a higher penetration rate (2.86 percent) than those of Vendor B, yet the correlation for the 12 MFs was lower. This likely happened because the probe counts of Vendor C are rounded up to the nearest thousandth.
  • The penetration rates tend to decrease from higher to lower FCs. On the other hand, the correlations tend to increase as one moves from higher to lower FCs. A larger sample size is needed, though, to draw more robust conclusions.
  • The AADT estimation errors slightly increase in the lower FCs, but this is expected because the MAPE tends to increase as the volumes decrease.
  • Within each FC, the correlations on urban roads are consistently higher than those on rural roads, even though the penetration rates on some rural FCs are higher than those on the corresponding urban FCs.
  • Likewise, within each FC, the AADT accuracy of annualized urban counts is higher than that of rural counts.
  • In M19, probe-based factors were separately developed for medium- and heavy-duty vehicles using Vendor C’s monthly probe data. The MAPEs for these two truck groups were significantly worse than those obtained for all vehicles (M18), likely because Vendor C rounds up raw probe trip counts to the nearest thousandth.

SUGGESTED RESEARCH

The following topics for further research emerged from the analysis conducted in this project:

• Develop tools and code that can be used to apply the most promising methods recommended in this project. The survey conducted in this project revealed that the majority of the respondents indicated that a guide, approval of new methods by FHWA, training materials, and coding in programming languages such as R and Python would be keys to adopting a different method in the future. There is a need to develop user-friendly tools and code that can be used by practitioners to apply the most promising methods recommended in this project.
• Compare the accuracy of different types of adjustment factors. The leave-one-out validation approach adopted in this project used 84 monthly day-of-week factors to annualize sample counts. There is a need to determine the accuracy of other types of adjustment factors including, but not limited to, daily, weekly, monthly, and annual day-of-week factors.
• Examine accuracy of probe-based temporal adjustment and axle-correction factors from other vendors and states. NCHRP 07-30 validated the accuracy of probe-based factors developed using probe data from three states and three vendors – one state per vendor. Considering the positive findings obtained for one of the three vendors (Vendor A), there is a need to validate factors from additional vendors, which may have higher penetration rates than those of Vendor A. In addition, the validation should expand to other states that may have diverse roadway, traffic, and geographic characteristics.
• Develop data-driven method to identify recreational patterns and assign counts to recreational factor groups. Currently, there is no generally accepted method for determining recreational patterns. Existing guidelines do not explicitly recommend any method to address this need. Simplified approaches have been proposed but have not been thoroughly validated. Research is needed to develop robust data-driven methods that can effectively identify different types of recreational patterns that exist in various parts of the country. This research should also determine how to identify which short-term count sites to assign to recreational groups. These methods will reduce the subjectivity and human bias in identifying recreational patterns manually.
• Examine more combinations of clustering inputs. Though this project validated over 20 single and combinations of clustering variables, there are many more combinations of variables that can potentially be used to develop clusters. Further research should examine more combinations of variables, particularly by combining qualitative variables (e.g., FC and RU) with quantitative variables (e.g., hourly factors). The research should also determine the impact of assigning different weights to each clustering variable and recommend variable weights or weight ranges for different combinations of variables.