Logistic Regression output not matching data

My research question was looking at what factors would best predict which patients are on twice daily insulin dosing.

The outcome of interest, twice daily dosing, was coded as 1 and once daily as 0. In a previous analysis other possible factors were excluded as not significant, and only BMI and Total Daily Insulin Dose were significant predictors. By a means comparison, patients on twice daily dosing received much higher doses therefore in a logistic regression, I would expect that the odds ratio for a change in 1 unit of insulin(very small change) would be also be small, BUT the odds ratio would be in the direction of Twice daily dosing rather than once daily, based on the data. My results show the opposite. Patients on larger daily doses of insulin have a odds ratio greater than one for patients on once a day dosing. BMI results are in the direction as expected, ie patients with higher BMI's typically require more insulin and would be on twice a day insulin.

My question is: should I believe the results or is there a problem? There are unbalance groups(28 twice daily versus 148 once daily) and their variances are different and large. There are some outliers based on diagnostics, but not more than I typically have seen on other analysis. BMI and Dose do coorelate(multicollinearity ??), but if I run the Logistic Reg without BMI, I get the same results.

Attached is a pdf output from SPSS. Any thoughts would be appreciated.