The sensitivity analysis was performed by changing the effect model (random-effect to
fixed-effect model) and comparing the results of different effect model.
Next, Columns (7) through (9) report results from the application of
fixed-effect model, Columns (10) through (12) provides a similar report from the estimation of a random-effect model, and, finally, the last three columns do the same job for a random-coefficient model, where coefficient randomness is assumed to apply only to the coefficient of our concern, the level of corruption (cp i).
We assessed potential heterogeneity of effect estimates within each group using I [sup]2 test, where I [sup]2 (%) >50% and P < 0.10 was considered significantly heterogeneous, then random effect model was applied to combine results from different trials, otherwise,
fixed-effect model was used.
When [I.sup.2] is less than 50% and p>0.10, the results were considered homogeneous and the
fixed-effect model was used; when [I.sup.2] is greater than 50% and less than 75%, results were considered heterogeneous and the random-effect model was used.
The
fixed-effect model eliminates all cross-county variations, including any unobserved ones.
The
fixed-effect model is preferred under such circumstance (Hsiao, 2003).
A
fixed-effect model was used first, the Q test and I2 statistic was performed to assess the heterogeneity, and P < 0.1 or I2 > 50% was considered as heterogeneity between studies.
Two approaches are developed to capture the unobserved heterogeneity:
fixed-effect model and random-effect model.
This result supports the preference to
fixed-effect model rather than pool model.
Data was analyzed with a
fixed-effect model. These results were input into a simulation model, the Evidence Based Medicine Integrator, in order to estimate their long-term implications in a real-world population from Kaiser Permanente (CA, USA).