# RE: st: Test for trend for SIR

 From "Harland Austin" To Subject RE: st: Test for trend for SIR Date Tue, 13 Nov 2007 18:55:16 -0500

```Roland,

Rather than change the ado file, you can change E (the expected number of
events) to an integer, - replace E= round(100*E) and then  - smrby D E,
by(timeperiod) trend -.

The trend test is still appropriate since it depends on the relative, not
the absolute, sizes of the timeperiod groupings.

Hope this helps.

I have individual data with multiple records per patient after I have split
the follow up time depending on year and age and have merged a variable for
the incidence rate per 100.000 personyears. With the command

xi:strate i.timeperiod,smr(inc) per (100000)

I get the following result

+-----------------------------------------------------------+
timeperiod    D      E      SMR   Lower    Upper
-----------------------------------------------------------
1                 7   0.43   16.102   7.677   33.777
2                 5   1.65    3.032   1.262    7.284
3               14   2.11    6.642   3.934   11.215
+-----------------------------------------------------------+

I want to test for the linear trend over the timeperiods.

If I try the smrby on this result I get:

. smrby D E, by(timeperiod) trend

Observed    Expected	-- Poisson	Exact --
timeperiod   D           E              O/E (%)	[95% Conf.	Interval]

1                   7      0.4300      1627.9***	655	3354
2                   5      1.6500       303.0	98	707
3                  14      2.1100       663.5***	363	1113
may not use noninteger frequency weights r(401);

Why do I get this error message?

According to your second choice I use the following model on this result.

glm D timeperiod, family(poisson) lnoffset(E) eform

Iteration 0:   log likelihood = -9.6767233
Iteration 1:   log likelihood = -9.5071336
Iteration 2:   log likelihood = -9.5068971
Iteration 3:   log likelihood = -9.5068971

Generalized linear models                          No. of obs      =	3
Optimization     : ML                              Residual df     =	1
Scale parameter =	1
Deviance         =  7.236772128                    (1/df) Deviance =
7.236772
Pearson          =  6.884656839                    (1/df) Pearson  =
6.884657

Variance function: V(u) = u                        [Poisson]
Link function    : g(u) = ln(u)                    [Log]

AIC             =	7.671265
Log likelihood   = -9.506897131                    BIC             =
6.13816

OIM
D         IRR   Std. Err.      z    P>z     [95% Conf.	Interval]

timeperiod    .7561818   .2108008    -1.00   0.316     .4378616	1.305917
E  (exposure)

ie there is no linear trend p=0.316

Is this a correct use of the glm model? Or can I use some other method on
the original dataset? If I collapse the dataset what happens with the
incidensvariable which should not be aggregated but stay the same.
Or do I have to collapse the dataset and merge the incidensvariable after
the collapse?