cii means poisson: #exposure must be integer. Why?

[Andrea Discacciati]

Before the Stata update released on the 29 October 2015, the 95% exact confidence interval for a Standardized Mortality Ratio where, for example, 15 deaths were observed and 34.5 deaths were expected, could be obtained as follows (under the common assumption that the number of observed deaths follows a Poisson distribution):

version 14
cii 34.5 15, poisson

Following the 29oct205 update, this is no longer possible, because the #exposure argument (that is, the first argument) must be a positive integer. Therefore, the following code:

version 14.1
cii means 34.5 15, poisson

will return the following error message: first argument, #exposure, must be a positive integer.

Could anyone explain to me the reasons behind this choice?

[Carlo Lazzaro]

Andrea:
your first code still works for me in Stata 14.1 (Current update level: 01 Feb 2016):

. cii 34.5 15, poisson

                                                         -- Poisson  Exact --
    Variable |   Exposure        Mean    Std. Err.       [95% Conf. Interval]
-------------+---------------------------------------------------------------
             |       34.5    .4347826    .1122604        .2433445    .7171078

whereas your second code does not:

. cii means 34.5 15, poisson
means not found
r(111);

[Clyde Schechter]

Odd, I get the same response as Andrea, and I, too, am updated to 1 Feb 2016:

. about

Stata/MP 14.1 for Windows (64-bit x86-64)
Revision 01 Feb 2016
Copyright 1985-2015 StataCorp LP

Total physical memory:     8269900 KB
Available physical memory: 3264048 KB

Single-user 2-core Stata perpetual license:
       Serial number:  REDACTED
         Licensed to:  Clyde Schechter
                       Albert Einstein College of Medicine

. cii 34.5 15, poisson
you must specify one of means, proportions, or variances following cii
r(198);

. cii means 34.5 15, poisson
first argument, #exposure, must be a positive integer
r(7);

Is this a Windows vs Mac or Unix thing perhaps?

And I agree with Andrea's objection: I can't think of any reason in principle why the exposure for a Poisson rate calculation should be constrained to integer values.

[Carlo Lazzaro]

Dear Andrea and Clyde:
very odd indeed!

. about

Stata/SE 14.1 for Windows (64-bit x86-64)
Revision 01 Feb 2016
Copyright 1985-2015 StataCorp LP

Total physical memory:     8298808 KB
Available physical memory: 5316556 KB

As recommended by an old Statalist post (http://www.stata.com/statalist/archi.../msg00398.html) I omit licence details.

[William Lisowski]

Here's an explanation - not for the why the change was made, but for Carlo's and Clyde's apparent differing experience. I hypothesize that Carlo included the version 14 command in his initial run, which causes the second run to be diagnosed with a syntax error, since the mean clause was introduced in version 14.1.

. about

Stata/SE 14.1 for Mac (64-bit Intel)
Revision 01 Feb 2016
Copyright 1985-2015 StataCorp LP

. cii mean 34.5 15, poisson
first argument, #exposure, must be a positive integer
r(7);

. version 14

. cii 34.5 15, poisson

                                                         -- Poisson  Exact --
    Variable |   Exposure        Mean    Std. Err.       [95% Conf. Interval]
-------------+---------------------------------------------------------------
             |       34.5    .4347826    .1122604        .2433445    .7171078

. 
. cii mean 34.5 15, poisson
mean not found
r(111);

[Carlo Lazzaro]

William spotted the issue about my seeming success in running Andrea's first code. Thanks.

[Clyde Schechter]

Yes, that mystery is solved.

But Andrea's question about why the exposure is restricted to integer values remains unanswered. I can't think of any good reason for this; yet it was explicitly introduced as a change to the way the command works, so I imagine Stata Corp. has a reason for this. Generally if Stata prevents you from doing something, it's because there is something problematic about doing it. But in this case, the rationale escapes me. I hope someone from Stata Corp. will chime in on this.

[William Lisowski]

And now I've found the answer. Clyde's assurances in posts 3 and 7 gave me the confidence to look further. After examining cii.ado, I saw (a) where cii.ado checks that the first argument is an integer, and (b) that cii.ado seems to have a separate code base from ci.ado. So I duplicated our example from the "immediate" command cii by created a one-observation dataset and applying the ci command.

. about

Stata/SE 14.1 for Mac (64-bit Intel)
Revision 01 Feb 2016
Copyright 1985-2015 StataCorp LP

. version
version 14.1

. set obs 1
number of observations (_N) was 0, now 1

. generate n = 15

. generate e = 34.5

. ci means n, poisson exposure(e)

                                                         -- Poisson  Exact --
    Variable |   Exposure        Mean    Std. Err.       [95% Conf. Interval]
-------------+---------------------------------------------------------------
           n |       34.5    .4347826    .1122604        .2433445    .7171078

I think this inconsistency needs to be brought to the attention of Stata Technical Support. I'll leave that to someone who actually uses the command; I don't.

[William Lisowski]

The following reproducible example demonstrates the problem with the version of cii (but not ci) introduced in Stata 14.1.

about
clear
set obs 1
generate n = 15
generate e = 34.5

version 14
ci n, poisson exposure(e)
cii 34.5 15, poisson

version 14.1
ci means n, poisson exposure(e)
cii means 34.5 15, poisson

. about

Stata/SE 14.1 for Mac (64-bit Intel)
Revision 01 Feb 2016
Copyright 1985-2015 StataCorp LP

. clear

. set obs 1
number of observations (_N) was 0, now 1

. generate n = 15

. generate e = 34.5

.
. version 14

. ci n, poisson exposure(e)

                                                         -- Poisson  Exact --
    Variable |   Exposure        Mean    Std. Err.       [95% Conf. Interval]
-------------+---------------------------------------------------------------
           n |       34.5    .4347826    .1122604        .2433445    .7171078

. cii 34.5 15, poisson

                                                         -- Poisson  Exact --
    Variable |   Exposure        Mean    Std. Err.       [95% Conf. Interval]
-------------+---------------------------------------------------------------
             |       34.5    .4347826    .1122604        .2433445    .7171078

.
. version 14.1

. ci means n, poisson exposure(e)

                                                         -- Poisson  Exact --
    Variable |   Exposure        Mean    Std. Err.       [95% Conf. Interval]
-------------+---------------------------------------------------------------
           n |       34.5    .4347826    .1122604        .2433445    .7171078

. cii means 34.5 15, poisson
first argument, #exposure, must be a positive integer
r(7);

.

[Houssein Assaad (StataCorp)]

Andrea Discacciati discovered a bug in cii means, poisson. We were able to fix it for the Stata update that came out today (16feb2016). Please type

. update all

to update your Stata. Thank you!

[Andrea Discacciati]

Carlo, Clyde, William: thank you for looking into this.

Houssein: thank you very much for the extremely quick fix!