績效預測模式普遍應用於鋪面設計、評估、維修及路網管理各個方面。當鋪面設計從純經驗法向力學經驗法發展時，估計交通荷重之單軸軸重當量(ESAL)概念在力學經驗法設計手冊中已不再採用。新的設計手冊成功與否有相當大的程度需要依賴準確的鋪面績效預測模式。因此本研究中利用美國長程鋪面績效(LTPP)資料庫(http://www.datapave.com)來探討與驗證現有鋪面糙度預測模式並改善之。 針對反應變數進行資料探索分析(EDA)時發現，由常態分配對於隨機誤差與連續變異情形，可能不適用傳統統計迴歸方法來構建預測模式。因此，本研究於後續分析時採用廣義線性模式(GLM)與廣義相加模式(GAM)對於反應變數的分佈情形均不給予任何假設，而是採用以概似估計法的方式測試分配適用度,其中以柏松分配之適用性較良好。並配合Box-Cox冪次轉換法、視覺圖的技術，以及李英豪教授建議的系統化之統計與工程分析方法應用於構建預測模式中。 經過多方嘗試，預測模式中將保留具有顯著影響與合理闡明物理現象之變數。本研究將構建完成之預測模式檢定其適合度及針對相關的參數進行敏感度分析，最後建構之模式其適用情形與過去模式相比，有的到良好的改善。對目前建立之模式於未來亦可作更進一步之改進，使其更為完善。 Performance predictive models have been used in various pavement design, evaluation, rehabilitation, and network management activities. As pavement design evolves from traditional empirically based methods toward mechanistic-empirical, the equivalent single axle load (ESAL) concept used for traffic loads estimation is no longer adopted in the recommended Mechanistic-Empirical Pavement Design Guide. The success of the new design guide considerably depends upon the accuracy of pavement performance predictions. Thus, this study will first investigate its goodness of fit and strive to develop improved performance prediction models for pavement roughness using the Long-Term Pavement Performance (LTPP) database (http://www.datapave.com or LTPP DataPave Online). Exploratory data analysis (EDA) of the response variables indicated that the normality assumption with random errors and constant variance using conventional regression techniques might not be appropriate for prediction modeling. Therefore, generalized linear model (GLM) and general additive model (GAM) along with Poisson distribution were adopted in the subsequent analysis. Box-Cox power transformation technique, visual graphical techniques, as well as the systematic statistical and engineering approach proposed by Lee were frequently adopted during the prediction modeling process. By keeping only those parameters with significant effects and reasonable physical interpretations in the model, various tentative performance prediction models were developed. The goodness of the model fit was further examined through the significant testing and various sensitivity analyses of pertinent explanatory parameters. The tentatively proposed predictive models appeared to reasonably agree with the pavement performance data, although their further enhancements are possible and recommended.