Obesity is a confounder and that is why you see the marginal > stratum specific effects. The noncollapsibility effect is always away from the null for prognostic covariates and it is present but likely masked by the confounding. The unequal effects per stratum is likely an artifact of the sample. see this preprint

Obesity is a confounder and that is why you see the marginal > stratum specific effects. The noncollapsibility effect is always away from the null for prognostic covariates and it could be present too but perhaps masked by the confounding. The unequal effects per stratum is likely an artifact of the sample but could also be an interaction as that can coexist with confounding.