255 Abstracts a variety of interpretations. The objectives of this paper are to critically evaluate the alternatives and present a set of statistics with known psychometric properties and unambiguous inter- pretation. METHODS: Data from several cancer registries and retrospective studies were mined to identify and categorize the various naturally-occurring scenarios impacting ARDI. RESULTS: Three statistics were derived from these samples, which discriminate among three key ARDI aspects, labeled “planned ARDI”, “delivered ARDI” and “% Optimal Dose”. They measure, respectively, the physician’s prescribed dose inten- sity, the actual delivered dose intensity, and the total delivered dose independent of time, all relative to the associated standard. Several visualization and analysis techniques are also presented that employ these measures to determine the relative contribu- tion of the various fundamental causes of suboptimal dose administration. These causes include cycle delay, dose reduction, treatment attenuation and planned deviation. CONCLUSIONS: The methods presented provide those engaged in naturalistic research and clinical performance improvement with a validated set of statistics and a concise, unambiguous terminology to measure and interpret the complex treatments involved in the study of chemotherapy effectiveness. PCN30 USING THE DIFFERENCE IN DIFFERENCE METHOD TO UNDERSTAND OUTCOMES IN PROSTATE CANCER Lee WC 1 , Pashos CL 2 , Brandman J 3 ,Wang Q 1 , Botteman MF 1 1 Abt Associates Inc, Bethesda, MD, USA; 2 Abt Associates Inc, Cambridge, MA, USA; 3 Novartis Pharmaceutical, Florham Park, NJ, USA OBJECTIVES: The “difference in difference” method (DD) is commonly used in health policy-oriented research. However, it is seldom used to design and analyze cohort outcomes studies. We applied the DD method to assess an independent association between androgen deprivation therapy (ADT) and bone compli- cations among non-metastatic prostate cancer patients receiving ADT. METHODS: Using medical claims data from a 5% national random sample of Medicare beneficiaries, prostate cancer patients who initiated ADT in 1992–94 without bone metastasis at baseline were identified (the “ADT” group, N = 3887). Prostate cancer patients without ADT matched on a 1 : 2 ratio on the basis of age, race and Charlson comorbidity index constituted the “comparison” group (N = 7774), a group similar to the ADT group but unaffected by ADT. We analyzed seven subsequent years of inpatient, outpatient, and physician claims data to identify rates of bone complications (e.g., fractures, osteoporosis/osteopenia) conditional on patient survival. RESULTS: Fracture incidence rates for the initial baseline two years and 7 years respectively (conditional on survival) were 11.3% and 83.3% for the ADT group versus 10.4% and 56.3% for the comparison group. As the temporal effect from the com- parison group may reflect change that would have occurred in the absence of ADT over time due to aging and disease pro- gression, we subtracted the change (56.3% – 10.4%) for the comparison group from the corresponding change (83.3% – 11.3%) for the treatment group, in an effort to account for the unmeasured time effects. Thus, the difference in difference (DD) estimate, 26-percentage point change (72% – 45.9%), reflects the association of ADT with fracture. CONCLUSION: This esti- mate will be valid if the time varying factors (e.g., disease pro- gression) are consistent or equivalent in treatment and comparison groups. Future research using clinically detailed data should assess whether such time-varying factors are different between those undergoing ADT and those not. PCN31 WEIGHT OR NOT TO WEIGHT? Baser O 1 , Given C 2 1 The MEDSTAT Group, Ann Arbor, MI, USA; 2 Michigan State University, East Lansing, MI, USA OBJECTIVES: Methods from the traditional survival analysis are not directly applicable to estimate medical costs since patients accumulate costs with different rate functions over time, leading to negatively biased estimates. A number of authors have incorporated inverse probability weightiness (IPW) technique to correct for this bias. None of these authors, however, compare their result with the method, which supposedly yields bias esti- mates, i.e. OLS over uncensored observations. In this paper, we test the differences between the coefficient estimates of OLS over uncensored observations and that of proposed model to deter- mine whether using weight yields statistically different results. Moreover, we compare the estimation power of the proposed alternative models. METHODS: A Hausman kind of test is pro- posed to compare the weighted estimator and unweighted esti- mators. Predictive Power tests are used to choose between alternative models. RESULTS: Our data set consists of an incep- tion cohort of 773 patients with incident cases of prostate, colon, lung and breast cancer from 24 Michigan community hospitals and their affiliated oncology units between the years 1994–1997. Hausman test indicated the results are statistically different. Pre- dictive Power tests yield that Lin [2003] model is better than Lin[2000], Carrides et al. [2000] and Bang and Tsiatis [2000]. CONCLUSION: Two conclusions are as follows: 1. If the error terms are homoskedastic and we fail to reject Hausman test use unweighted simple OLS over complete observations. 2. Other- wise, weighted estimators yield consistent results and predictive power tests can be used to choose among them. PCN32 CENSORED MEDICAL COST ESTIMATION Baser O 1 , Gardiner J 2 , Bradley C 2 , Given C 2 1 The MEDSTAT Group, Ann Arbor, MI, USA; 2 Michigan State University, East Lansing, MI, USA OBJECTIVES: To propose a method to estimate total medical cost from censored data. METHODS: In this paper, the inverse probability of weighted (IPW) least squares method is used to assess the effect of covariates (e.g. patient and clinical charac- teristics) on medical cost with censored data. We outlined IPW least squares as applied to censored medical cost data, including the statistical properties of the estimation, then introduced Hausman type of test to compare the estimators calculated by using IPW least squares and OLS over uncensored data and applied our method to the estimation of cancer costs RESULTS: Medicare claim files are examined to apply our method. Each patient is followed two years after diagnosis of cancer (breast, colon, prostate or lung). For patients who have less than two years of cost and still alive at the end of the study were consid- ered censored. The reference group fro treatment modalities is surgery plus adjuvant therapies, the reference group for site of cancer is lung. Variables that reach statistical significance (p < 0.05) include physical function, type of cancer (except colon), surgery and radiation, radiation only, and chemotherapy and radiation. Ten additional points in patient’s prior physical func- tion score decreases total medical cost by 0.7 percent. Prostate cancer patients and breast cancer patients cost 1.36 and 2.46 times lower than lung cancer patients respectively, these esti- mates are 1.16 times and 2.40 times according to IPW least square estimation. CONCLUSIONS: Hausman test suggests that IPW estimates are significantly different (p < 0.05) and suggested