Exact logistic regression and its performance to SAS system / 中华流行病学杂志

ABSTRACT

<p><b>OBJECTIVE</b>To explore the feasibility of exact logistic regression, used as a complemental method for the maximum liklihood estimation, and to analyse with data small sample, unbalanced structure and highly stratal nature under the situations of questionable results or inexistence of the maximum likelihood estimation.</p><p><b>METHODS</b>Data from 37 postoperative breast cancer cases were analyzed in 1997 by exact logistic regression under SAS system.</p><p><b>RESULTS</b>Data calculated by SAS software showed that Quasi-complete separation of data points was detected but the results of maximum likelihood estimation did not exist, SAS outputs conflicted the results of the last maximum likelihood iteration (likelihood Chi-square and score Chi-square have similar P, less than 0.05, but the Wald chi-square had a larger P, more than 0.05). Under conditional exact parameter estimation, it appeared that (1) the joint effect of conditional score statistics was 21.12 with P = 0.000 3; (2) for individual parameters, the effect conditional score statistics of histological classification (grades) was 5.80 with P = 0.020 8; axillary node metastasis (diversion) was 5.74 with P = 0.019 5; tumor size (size) was 0.79, with P = 0.647 6. The effects of tumor histological classification and axillary node metastasis were statistically significant on breast cancer tumour.</p><p><b>CONCLUSION</b>Exact logistic regression seemed to be a very useful method in analyzing data from small sample when the maximum likelihood estimation was either with no effect or did not exist.</p>