Graphing Poisson regression for segmented regression

[syoung15]

I am doing a segmented poisson regression to look at counts pre- and post- an intervention.

There is therefore a poisson regression line pre-intervention and post-intervention. The delineation is by week - 141 weeks total, the break point is week 105.

I can create a variable ("preintervention") and code it as 1 if week < 105 and 0 if week >=105.

I want to graph the poisson regression lines pre- and post- intervention.

The overall regresson would be:

proc genmod data=have;

model count=week / dist=poisson link=log;

run;

My questions is - how do I transform this into a visual graph (presumably using proc sgplot) that will divide into 2 separate poisson regression lines - 1 pre intervention and 1 post.

Thank you for any assistance!

[Rick_SAS · Accepted solution]

1. Add a CLASS statement and list the PreIntervention variable.

2. Include the PreIntervention variable in the model

3. Use the OUTPUT statement to get the predicted values in a SAS data set.

4. Graph the predicted value vs weeks

/* simulate data as an example */
data have;
call streaminit(1);
preintervention = 1;
do week = 1 to 105;
   count = rand("Poisson", 0.8);
   output;
end;
preintervention = 0;
do week = 106 to 150;
   count = rand("Poisson", 2.4);
   output;
end;
run;
/* end simulation */

/* read whatever data you have and graph the results */
proc genmod data=have plots=none;
class preintervention;
model count=week preintervention / dist=poisson link=log;
output out=PoiOut pre=Pred;
run;

proc sgplot data=PoiOut;
scatter x=week y=count / group=preintervention transparency=0.5;
series x=week y=Pred / group=preintervention;
run;

[syoung15]

Thank you for this. However, the graph does not look right.

Could you please clarify what the statements:

count=rand ("Poisson", 0.8);

and

count=rand ("Poisson, 2.4);

mean? It seems to be a random number generation, but I'm not sure why you would do that? And why are 0.8 and 2.4 chosen, in particular.

I want the graph to show the true count (which are outcomes per week and are in the range of 100-750, varying per week) and overlay the Poisson regression line, divided by pre and post intervention. When I use this code, "count" becomes a number between 1-3 or so...

Thank you for all of your assistance!

[Rick_SAS]

You did not provide any data, so I simulated some to illustrate the process. Ignore the DATA step and just focus on the PROC GENMOD and PROC SGPLOT steps.