very huge IRR

[Fredrick orwa]

E]
* Example generated by -dataex-. To install: ssc install dataex
clear
input int yoa byte(ipd nvt vt) int bcd float midyrpop str3 spn_serotype byte n str12 type byte _merge float(time vaxera vaxera2017 postslope postslope2017 vacc bcdcn)
1999 24 8 16 3490 33620 "13" 0 "non-vaccine" 3 1 0 0 0 0 0 3466
2000 47 12 35 3482 34505 "13" 1 "non-vaccine" 3 2 0 0 0 0 0 3435
2001 46 14 32 3394 35411 "13" 1 "non-vaccine" 3 3 0 0 0 0 0 3348
2002 28 4 24 2286 36340 "13" 0 "non-vaccine" 3 4 0 0 0 0 0 2258
2003 47 9 38 2634 39747 "13" 0 "non-vaccine" 3 5 0 0 0 0 0 2587
2004 36 11 25 2524 41985 "13" 0 "non-vaccine" 3 6 0 0 0 0 0 2488
2005 31 12 19 2109 42738 "13" 0 "non-vaccine" 3 7 0 0 0 0 0 2078
2006 23 9 14 2269 43916 "13" 0 "non-vaccine" 3 8 0 0 0 0 0 2246
2007 34 8 26 1832 44536 "13" 0 "non-vaccine" 3 9 0 0 0 0 0 1798
2008 19 3 16 1687 44820 "13" 0 "non-vaccine" 3 10 0 0 0 0 0 1668
2009 20 5 15 1879 46343 "13" 0 "non-vaccine" 3 11 0 0 0 0 0 1859
2010 46 7 39 1648 47714 "13" 0 "non-vaccine" 3 12 0 0 0 0 0 1602
2011 16 5 11 1438 46961.64 "13" 0 "non-vaccine" 3 13 . 0 0 0 . 1422
2012 7 4 3 1017 40730.05 "13" 0 "non-vaccine" 3 14 1 0 1 0 1 1010
2013 4 4 0 805 43214.2 "13" 0 "non-vaccine" 3 15 1 0 2 0 1 801
2014 10 9 1 1274 47807 "13" 0 "non-vaccine" 3 16 1 0 3 0 1 1264
2015 9 7 2 1228 47921 "13" 0 "non-vaccine" 3 17 1 0 4 0 1 1219
2016 4 3 1 1029 42048.97 "13" 0 "non-vaccine" 3 18 1 0 5 0 1 1025
2017 2 2 0 365 18128.926 "13" 0 "non-vaccine" 3 19 1 1 6 1 2 363
2018 6 2 4 1005 46210 "13" 0 "non-vaccine" 3 20 1 1 7 2 2 999
2019 6 5 1 1122 45972 "13" 2 "non-vaccine" 3 21 1 1 8 3 2 1116
end
label values _merge _merge
label def _merge 3 "matched (3)", modify
[/CODE]
------------------ copy up to and including the previous line ------------------

Listed 21 out of 21 observations

Dear team,
Can someone tell me why I get ver large IRR when I adjust for time in my model? below is my code

**unadjusted
nbreg n i.vacc, irr exp(midyrpop)
**final model(adjusted bcd control)
nbreg n i.vacc,irr exp(bcdcn) diff

**segmented
nbreg n i.vacc time, irr exp(midyrpop)


I have tried even zero inflated negative binomial but the irr is even worse
zinb n i.vacc time, irr exp(midyrpop) inflate(nvt) diff

What can I do to solve this.we discussed with my principal statistician to adjust for time in all serotype specific analysis.Am only having problem with this serotype

[Clyde Schechter]

Your n's are overwhelmingly zero, and the vacc variable is also overwhelmingly zero. The rest of the data is very sparse. I find that even in the unadjusted model, the IRR for vacc = 1 is almost zero, and the IRR for vacc = 2 is a large number (about 4.5, which seems possible, but is a much stronger effect than most things in real life produce). Your "final" model doesn't converge. And when you add the time trend variable, you get an IRR for vacc = 2 over 1,000, which is pretty ridiculous, I agree. Moreover, its CI goes from about 0.04 to about 30 million.

I think that the problem is just that you don't have enough information about vacc = 2 in your data to get a reasonable analysis. There's only a total of 3 cases with vacc = 2, split 2:1 between n = 0 and n = 2. When you then try to tease out a time effect, it gets even worse because the three cases occur in times 19, 20, and 21. So that is attempting to split almost no data into even less. There is nothing you can do to truly salvage this other than getting more data. If that isn't possible, I would omit vacc = 2 from the analysis altogether, or combine vacc = 1 with vacc = 2 or combine vacc = 0 with vacc = 2 if either of those makes sense to do from a clinical perspective.