rate vs. IRR using multivariable poisson regression

[docfermi]

Using the national dataset, my study aim is to obtain cdi infection rate and mortality over 12-year period.

While I am doing so, I saw significant differences in rates among crude, adjusted for age/sex, and adjusted for age/sex/comorbidities.  Then I went ahead and obtained mortality for entire sample regardless of primary or secondary diagnoses.   Below are the results.   You can see almost 10x differences between 2nd (adjusted for age and sex) and 3rd  (adjusted for age/sex/race/comorbidities) results.

	crude	adjusted for age and sex	adj for demographics and comorbidities
2003	0.02202	0.01558	0.1463
2004	0.02096	0.01498	0.1376
2005	0.02043	0.01444	0.1294
2006	0.02018	0.01415	0.1234
2007	0.01916	0.01361	0.1149
2008	0.02014	0.01379	0.1174
2009	0.01893	0.01307	0.1084
2010	0.01862	0.01286	0.1034
2011	0.01872	0.01253	0.0968
2012	0.01845	0.01259	0.09664
2013	0.0189	0.01286	0.09651
2014	0.01901	0.01301	0.09573

My sas codes are below.

Questions:

1.I don’t know why numbers are significantly different.  Maybe Poisson is inappropriate for obtaining adjusted rates, but appropriate for IRR in multivariable model…?  

2.What would be the best method(s) to perform trends analyses for rates and IRR…?

Best,

Sun

Crude mortality

proc genmod data=NIS.cdi;

class year female agegroup / param=glm;

model died (event='1')=year  / type3 dist=poisson link=log offset=log_discharge;

weight trendwt;

store plmsourcenis;

run;

proc plm source=plmsourceni;

lsmeans year / ilink cl;run;

Mortality after adjusting for sex and age

proc genmod data=NIS.cdi;

class year female agegroup / param=glm;

model died (event='1')=year female agegroup / type3 dist=poisson link=log offset=log_discharge;

weight trendwt;

store plmsourcenis_sexage;

run;

proc plm source=plmsourcenis_sexage;

lsmeans year / ilink cl;run;

Mortality after adjusting for sex, age, and comorbidities

proc genmod data=NIS.cdi;

class year female agegroup race_three

CM_AIDS CM_alcohol CM_anemdef CM_arth CM_bldloss CM_CHF CM_chrnlung CM_coag

CM_depress CM_DM CM_dmcx CM_drug CM_HTN_C CM_hypothy CM_liver CM_lymph CM_lytes

CM_mets CM_neuro CM_obese CM_para CM_perivasc CM_psych CM_pulmcirc CM_renlfail

CM_tumor CM_ulcer CM_valve CM_wghtloss / param=glm;

model died (event='1')=year female agegroup race_three

CM_AIDS CM_alcohol CM_anemdef CM_arth CM_bldloss CM_CHF CM_chrnlung CM_coag

CM_depress CM_DM CM_dmcx CM_drug CM_HTN_C CM_hypothy CM_liver CM_lymph CM_lytes

CM_mets CM_neuro CM_obese CM_para CM_perivasc CM_psych CM_pulmcirc CM_renlfail

CM_tumor CM_ulcer CM_valve CM_wghtloss / type3 dist=poisson link=log offset=log_discharge;

weight trendwt;

store plmsourcenis_com;

run;

proc plm source=plmsourcenis_com;

lsmeans year / ilink cl;run;

[Rick_SAS]

Without the data, it is difficult to know how the response might be affected by the 'female' and 'agegroup' variables. However, in general, what you describe can when you add or exclude a categorical variable to a model. It is known as Simpson's Paradox and it means that the within-group relationships between variables are different from the between-group relationships. There are some pictures in the Wikipedia article that show why it occurs.  You might try graphing your data in a similar fashion to see if it reveals whether Simpson's paradox is responsible for what you are seeing.