本研究測試及比較各橫斷面模型所檢測出來盈餘管理之顯著水準及檢定力,用其類似Brown and Warner (1980, 1985) 的模擬程序,且檢測公司有極端績效之樣本時產生的偏差所加入的控制變數。偵測盈餘管理的方式是加入人為已知的數額來測試檢定力。裁決性應計項目係數之估計採橫斷面季料資。主要測試橫斷面Jones模型與橫橫斷面Modified Jones模型,及其各自以資產報酬率或營運現金流量加入迴歸式中,與以資產報酬率或以營運現金流量作績效配對的模式,並且測試在各應計模型裡加入截距項之效果。 研究結果顯示,顯著水準的測試所有的模型的顯著水準都接近所設定的測試水準,亦即模型皆設定良好。Jones模型與其對應Modified Jones模型,其Modified Jones模型檢定力較Jones模型佳。收入操弄的檢定力方面顯示控制極端財務績效之變數是以CFO作績效配對最具檢定力。測試結果顯示績效配對法較變數加入迴歸式之方法較更具檢定力;控制的變數營運現金流量比資產報酬率更易偵測出盈餘管理。而資產報酬率變數加入模型其模型解釋力並未增加,對於偵測盈餘管理的檢定力也不佳。營運現金流量變數加入迴歸模型中當自變數與績效配對法比較,其模型解釋力上升,檢定力也較佳。有截距項的模型其模型的解釋能力較差,但偵測盈餘管理的檢定力較佳。本研究結果建議控制極端財務績效之變數採用營運現金流量比資產報酬率較易偵測出盈餘管理。
This study examines specification and power tests results of alternative remedies to detecting earnings management and control extreme performance problems in discretionary accrual estimation. We use a simulation procedure developed by Brown and Warner (1980, 1985) to assess the power of each model. The power of the models is based on the inclusion of an artificially induced amount of accruals manipulation and the assessment of the different models’ ability to detect it. We focus on the specification of cross-sectional models of expected accruals using quarterly as well as annual data. We examine the specification and power of tests based on the cross-sectional version of Jones and Modified Jones model ,we evaluated the effectiveness of including additional ROA or CFO regressor and performance matching on ROA or performance adjusted by CFO. We also include a constant in the estimation for competing typically does not include a constant in all of models. Our results show the all of the models appear well specified when applied to a random sample of firm-quarter. In terms of power results indicate, in revenue manipulation uses control variable CFO is better than variable ROA .Performance matching is more powerful than adding ROA or CFO as an additional regressor to the accrual regression models. Our study in expense manipulation all of modes are insignificant. Including a constant in the estimation models has less explanation for models, but has more power than does not include a constant in models.