|
Guest
|
Dear all,
Many thanks in advance for any suggestions you can provided in the
following 3 questions.
1. Should I treat the following outcomes as continuous variables using
linear regression modelling, or treat it as count variables using
Poisson model, please?
outcome1 = sick-days/PersonYear
outcome2 = Amount-of-dollars/PersonYear
2. How to handle the sample size calculation for a Poisson distributed
outcome, please? such as: the CONTROL group rate is 10 events/
personyear, if I think TREATMENT group with 8 events/personyear
[(10- /10=20% decrease] will be good enough to concluded that the
TREATMENT is significant better than CONTROL, how many 'personyear' in
each group at 0.05 confidential level with 80% power should I have,
please? any suggestion (software, paper) is very appreciated!
3. To compare whether or not the rate_2005 (10/PersonYear) vs.
rate_2006 (20/PersonYear) vs. rate_2007 (40/PersonYear) in whoe BC
province, which options do you perfer, please?
(1) No P value avaialble because no "sampling", the rate_2005, 2006,
2007 were calculated based on the all events and all people in whole
province. We can simply concluded:
- compared to 2006, rate_2007 increased 50% [(40%-20%)/40% = 50%]
-compared to 2005, the rate-2007 increased 75% [(40%-10%)/40% = 75%].
(2) use Poisson model to get a P value. GEE in the Poisson model to
handle clustering within subject (one person may have injury in 2005,
and then in 2006, and then in 2007-multipel innury) .
Sincerely
Helen |
|
|
