@Estee wrote:
Hi,
I am unsure if I am using the best method to analyze my dataset and would like some help/feedback on this.
I am looking at 8 states from 2010 to 2014 at a monthly level. 5 out of the 8 states eventually receive the treatment, which is the implementation of a mandatory state policy. The time at which the policy was implemented varies from state to state. Also, in some of these 5 states, there is a prior non-mandatory phase, where the policy is suggested but not required. So there are a total of 3 phases total: no policy, non-mandatory phase, mandatory phase.
The outcome is child deaths per state-year-month (these deaths are related to the policy being considered).
The research question is as follows:
1) Is this policy effective in reducing child deaths?
Answers to "What type of indication would indicate effectiveness?" might get you started. Then you would at least have a starting point of likely tests.
Are you more concerned with "within state" or "all states combined"?
Are concerned with just no policy/non-mandatory/mandatory or how long after it became mandatory?
Likely you need to be considering standardized rate of some sort, possibly within age groups and actual raw counts could be misleading. If the number of deaths increases but is less relative to change in population then even though the numbers are larger the rate would be declining. Or have two "counts" live vs death and event to tell which.
A simplistic approach since the result is basically live/death might lead to logistic regression similar to:
data example;
input stateid $ when death count ;
datalines;
A 1 0 100000
A 1 1 445
A 2 0 75000
A 2 1 123
A 3 0 230000
A 3 1 333
B 1 0 400000
B 1 1 812
B 2 0 100000
B 2 1 222
B 3 0 600000
B 3 1 444
; .
run;
proc logistic data=example;
class when (ref='1');
model death(event='1') = when;
weight count;
freq count; /* realized later that freq is more appropriate than weight*/
run;
Where the when coding is 1 the earlier, 2 is non-mandatory and 3 mandatory.
Odds ratios of when 3 vs 1 with values less than one might be an indicator of reduction.
I didn't include the "year" of data but that should be included in the data at least if not necessarily the model.
Some tricky parts are addressing the counts by year as the event likely didn't change on the calendar year boundary.
Children's age or age group likely should be included in the data and the model if there is reason to believe age groups may be affected differently.
And if the most at risk are the very young you may need to consider birth status such as premature or low birth weight (or other health scores)
