本文之主要目的在討論成就或能力測驗題庫發展時，題目IRT參數連結 之相關問題，本文以三個實際資料模擬研究，四個Monte Carlo研究及十個應用研 究所得結果為依據，首先簡要呈現17個研究的結果，再以綜合統整方式進行討 論。討論重點包含連結效益、線性假設、連結方法、考生樣本的特質和大小、共 同試題的特質和題數、評鑑標準、及測驗題庫的特質。實際資料顯示，在平行等 化的情境中，兩組題目參數估計值間的線性假設得到很強的支持，但在垂直等化 情境下，結果稍差(尤其是鑑別度參數)。連結方法一般以"特徵曲線轉化法"為最 佳，"平均數標準差法。其次，但"平均數標準差法"計算簡易，值得較不具備專 業及電腦資源的應用場合參考使用。在成就測驗兩樣本能力差異一個學期時，連 結對難度參數量尺的調整效益大於對鑑別度量尺的調整效益，三參數模式的連 結，樣本人故應以一千人以上為宜，至於考生能力特質的影響，應視其所在能力 點測驗訊息量而定，假如新樣本中所缺乏之考生座落在測驗訊息量微少的能力點 上，連結可能就不一定必要。一般而言，共同題數15題連結效益就相當穩定，如 果只單憑共同試題(而非整個測驗)對連結效益進行評鑑，其推論較不穩定有效。 整體來看，連結誤差比估計誤差小。關於"理想共同試題特質。仍不明朗，目前 看來以中等難度題目效果較佳。

　　　　　The purpose of this paper is to discuss the relevant issues of IRT item parameter linking when developing an achievement or ability item bank is the focus. The major issues included are the linking effect, linear assumption, linking methods, sample size of examinees, test length of anchor tests, evaluation criteria and characteristics of examinees, anchor tests and item pools. The results of 3 real data simulation studies, 4 Monte Carlo simulation studies and 10 real applications are included for discussion. First, the brief summaries of these studies are provided. Then, an integrating discussion follows. The real data results suggest that the linear assumption between two sets of parameter estimates hold pretty well for parallel settings but not as good for the vertical ones (especially for the discrimination parameter). Generally speaking, method of "Item Characteristic Curve" transformation performs best. The method of "Mean and Sigma" is also recommended for itscomputation simplicity. When the ability difference between twosamples exists (one semester difference of achievement tests),linking is more helpful for adjusting the difficulty parameter scale than for the discrimination parameter scale. Sample of 1,000examinee will be necessary for the linking of the 3-parameter IRTmodel. The effect of the characteristics of the examinees dependsupon the test information where the examinees are located. If theexcluded examinees are located on the area with little test information,linking may not be necessary. Anchor test of ten to fifteen itemsfunctions pretty well. Evaluating the linking results by only theanchor items (instead of the whole test) is less stable and sometimesnot valid. Linking error is rather small compared with estimationerror. The characteristics of ideal anchor test is not clear. Mediumdifficult items are suggested at this point.