摘要
背景 牙周病(Periodontal disease, PD)流行病學研究中，建議以社區牙周指數(Community Periodontal Index, CPI)作為測量值。然而，該指標應用於社區流行病學研究時，其CPI測量值的信度及效度是常被評論的，且不同區域間會因為檢查者及個人變異而影響。而且，在牙周病檢測資料中，其結果常因多階層資料特性(地區及個人)呈現相關性，CPI測量值於牙齒及部位在同一個人相關性更顯複雜。因此，本研究考慮利用統計模式針對上述的多階層資料之CPI測量值進行校正，並進一步評估測量誤差校正對於牙周病相關危險因子之危險對比值估計值偏差之影響。 目的 本研究利用全國性牙周病盛行調查前執行CPI校正之資料，評估檢查者其CPI在信度及效度之表現，利用混合效應模式分析多階層架構下重覆測量值之相關性資料，了解牙齒不同部位、不同人，及不同地理區域信度及效度之大小。進一步根據檢查者與黃金標準的測量誤差，進行檢查者CPI測量值校正以貝氏統計模式應用於牙周病危險因子之估計值校正。 材料與方法 本研究主要分為兩部份，第一部份為牙周病測量值校正，第二部份則是以第一階段研究結果針對大規模社區牙周病資料，在探討危險因子與牙周病之關係，其估計值(95%信賴區間)進行測量效度及信度校正。第一階段利用13位檢查者進行信度及效度評估，探討檢查者對牙周病測量的變異性。共計邀請31位受試者利用社區牙周指數測量之site，進行信度與效度測量，其site為包括不同地區(北、中、南、東)之個人、sextant和tooth階層。效度上，以牙周病學中資深牙醫師作為黃金標準(S)，對其他檢查者之測量值進行敏感度與特異度分析，也利用Kappa統計值和相關係數進行信度評估。另外，我們進一步應用一系列隨機效應混合模式評估考慮相關性資料下的效度與信度。第一階段之測量誤差校正後，應用於全國資料3,860位受測者，探討抽菸和牙周病相關估計值，以貝氏多階層統計模式校正後其偏差程度。 結果 本研究共計有28,008個重覆測量site資料進行分析。檢查者測量值之信度與效度隨區域而有不同。信度方面，加權kappa落在0.34至0.97間，整體值加權kappa是0.69，其95%信賴區間為(0.68, 0.70)，根據羅吉斯迴歸分析結果顯示整體信度並沒有顯著差異存在，但在各區域間仍具有信度差異。效度方面，以計畫主持人為黃金標準(S)，敏感度落在0.10至0.66間，特異度落在0.91至1.00，整體敏感度為0.44，特異度為0.96，以北區與東區檢查者之敏感度較高，此外各區皆有好的特異度。不考慮牙齒間相關性下，使用羅吉斯迴歸分析，檢查者間之估計值落於-1.82至1.04間，有4名檢查者未達統計顯著性。而以廣義線性混合效應模式下考慮資料相關性後，其估計值落於-1.85至1.00間，但各區之95%CI 皆變得較寬，有6名檢查者未達統計顯著性，主要集中於北區與東區。 利用貝氏多階層統計模式應用於抽菸習慣之危險因子對於牙周病之危險對比值影響，研究結果發現考量檢查者之測量誤差及受測者資料之相關後，顯示以傳統之統計模式估計抽菸習慣對於牙周病之危險對比值為1.60倍，但進行上述因素校正後，抽菸習慣對於牙周病危險對比值提高至3.11倍，顯示校正前及校正後差異確實存在。 結論 本研究結果顯示檢查者在site測量為單位之信度及效度較差，而以sextant為單位之測量信度與效度較好。牙周病測量誤差校正研究顯示，以多階層模式分析考慮資料相關性後，整體牙周測量指數在信度上並無顯著差異、但不同區域有顯著差異。效度則在區域間有極大差別，考量檢查者之測量誤差及受測者資料之相關後，顯示以傳統統計模式估計危險因子對於牙周病之危險對比值含有極大之影響。
並列摘要
Background: Community Periodontal Index (CPI) has been recommended in epidemiological studies on periodontal disease (PD). However, there are several thorny issues when this indicator is applied to a community-based epidemiological study. The validity and the reliability of its measurement across examiners are often criticized. The correlation between teeth and sextant in the same individual or same district characterized by multilevel data further complicates the accuracy of measurement of CPI and the analysis of the correlates accounting for the condition of PD. It is of great interest to do calibration of CPI measurement making allowance for such multi-level data. The measurement errors obtained from the calibration study were further applied to calibrating the odds ratio for the association between certain risk factor (such as smoking) andPD. Aim: The calibration study before a nationwide survey of CPI was conducted to assess the reliability and validity on the periodontal disease measurement taking the correlated property of CPI into account. Theses measurement errors were incorporated into Bayesian hierarchical model to correct the magritude of odds ratio between smoking and PD. Methods and Materials: Thirteen examiners were involved in intra- and inter-rater measurement of site-level on thirty-one subjects with CPI index at individual, sextant, and tooth level from different districts. The validity with sensitivity and specificity was conducted to compare the results from these rates against a gold standard senior specialist in periodontology. The reliability with Kappa and correlation coefficient was also assessed. We applied a series of random effect mixed model to assess the validity and reliability making allowance for correlated data. We develop a Bayesian hierarchical model with the consideration of validity and correlation arising from hierarchical data structure. Results: The total of 28,008 repeated measure site data were available for analysis. Without stratifying the districts, as far as the reliability is concerned, the kappa values for site, tooth and sextant level, respectively, were 0.71 (95% CI: 0.69, 0.74), 0.78(0.74, 0.80) and 0.78(0.72, 0.83). Accounting for subject effect, the estimate and 95% confidence interval of intra-examiner based on the gold standard using a generalized linear mixed-effect model were 0.01 (95% CI: -0.35, 0.37), -0.15 (95% CI:-0.54, 0.25) and -0.29(95% CI: -0.93, 0.35) at site, tooth and sextant level, respectively, indicating a lack of statistical significance. The sensitivity and the specificity based on the gold standard sample at site, tooth and sextant level, respectively, were as follows: 0.44 and 0.96; 0.57 and 0.90; 0.73 and 0.82. Using the generalized linear mixed-effect model, the estimate and 95% confidence interval adjusted subject effect of validity at site, tooth and sextant level, respectively, were -0.11 (95% CI: -0.51, 0.30), -0.16(95% CI:-0.58, 0.26) and -0.12(95% CI:-0.49, 0.25), without statistical significance. However, the reliability and validity of examiners measurement varied with districts. After accounting for correlation, the estimates of reliability for the examiners of Taipei at site level and the estimates of validity for the examiners of Tainan(I) at site and tooth level were statistical significance. After the calibration of measurement error and the consideration of correlation resulting from hierarchical data structure with Bayesian DAG model, the adjusted odds ratio was inflated from 1.60(1.44, 1.78) to 3.11(2.14, 4.38). Conclusion: The calibration study on PD shows that reliability and validity was in poor at site level but acceptable at sextant level. The reliability varied with districts after taking correlation into account. However, the validity of CPI has a large variation and requires for calibration, but the overall difference was not statistically significant. The results of calibration would provide the basis for calibrating the association between correlates and the condition of PD.