New package Biological interaction/interaction contrast - ic. icp and ici on SSC

[Niels Henrik Bruun]

Hi
A new package is available on the ssc named ic.

Description:
=========
Tool for calculating 3 different measures of interaction contrast (biological interaction) on an additive scale:

• relative excess risk due to interaction [RERI]
• attributable proportion [AP]
• synergy index [S]

Corresponding 95% confidence intervals and two-tailed tests for no interaction (RERI=0, AP=0, S=1, respectively) are given as well.

Assume that we want to analyze the effects of 2 dichotome factors, A and B, on a dichotome outcome Y. Define Pij as the average risk (the incidence proportion) when A = i and B = j, i = 0,1 and j = 0,1.
And define the relative risk as RRij = Pij / P00.

The biological interaction measures is defined as

• RERI = RR11 - RR10 - RR01 - 1
• AP = RERI / R11
• S = (RR11 - 1) / (RR10 + RR01 - 2)

The estimation/approximation of the RRs are described in eg Hosmer & Lemeshow (1992)

Hosmer, David W., Lemeshow, Stanley, Confidence Interval Estimation of
Interaction, Epidemiology, Vol. 3, No. 5 (Sep., 1992), pp. 452-456

Enjoy !!

[Jesper Love]

Hi Niels Henrik,

Recognizing the importance of "biological interaction" in epidemiological analyses I was very glad to find this post. However, I am also a beginner in STATA so if possible, could you explain how to use this info/code in analyzing, e.g. RERI?

Best,
Jesper Löve

[Niels Henrik Bruun]

Hi Jesper
When you have Stata opened you just write the command

ssc install ic

to install the code on your Stata.
Then look at the help files for further guidances, ie the code

help ic

.
If you just want to see the help files without installation use the command

findit ic

.

If you have suggestion to the help files please do not hesitate to write.

Good luck

[Jesper Love]

Thank you very much for this. The help site was indeed extensive :-).

Best,
Jesper

[Niels Henrik Bruun]

Glad to hear

[mehrane mehramiz]

Is it possible to run multiplicative interaction and additive interaction by stata? what are the commands?

[Niels Henrik Bruun]

I'm a bit rusty on this.

This package does Biological interaction/interaction control. It uses odds, hazard ratios or similar as a base for estimating interaction on a additive scale beyond the underlying regression.
Multiplicative interactions can statistically be found from interaction terms in eg logit, cox or similar regressions.

BUT epidemiologically there are problems with the statistically found interactions, hence the need for the command -ic- where a contrast is used to quantify the additive interaction.

My suggestion is that that you read the references mentioned in my documentation.

Good luck

[Maria Gabriela M Pinho]

Dear Niels Bruun,

This tool it is really interesting, thank you for sharing!
I am trying to calculate RERI for multinomial logistic regression and I am wondering if I could use the ic command?
I have an outcome with 3 categories and two exposures with 3 categories which.
I have tried to run it with the ic comand, and the RERI was significant, but since the results from additive interactions were not really consistent, I thought using this command with multinomial logist regression might not be adequate.
I also would like to ask you with using multiple imputed would affect the estimates?
I would very much appreciate if you have a suggestion..

Best wishes,
Maria Gabriela

[Niels Henrik Bruun]

Hi Maria
The code and the theory isn't designed to work in a multinomial logistic regression.
I'm not shure what you mean with multiple imputed.
If you mean imputation, I guess it can go either way.
You'll have to look at RERI both with and without imputed data.

[Marianne Holm]

Dear Niels Bruun,

Thank you for this. This Tool seems to be just what I need as I am trying to study biological interaction between multiple risk factors for cancer.
I am trying to use the syntax in/after Cox regression but I experience some problems.

I stset the data
and then type "icp, show: stcox 'outcome' 'vars2' 'vars3' i.'vars4'..... etc. (I assume 'vars4' onwards are just the confounders I am adjusting for in my model but not part of studied interaction?)
the estimates are very strange and I get the notification: "'vars2'_NOT_'vars3' omitted because of collinearity"

-Do you have any suggestion to what I might be doing wrong?
-and is icp not the preferred choice for testing interaction in Cox models?

Thanks very much,
Marianne Holm

[Niels Henrik Bruun]

Hi Marianne
Sorry for not answering before - Have been on holiday

To answer properly it would have been nice with more information regarding the problem at hand.

First using -stcox- you already have set the outome with -stset-, so your command should be "icp, show: stcox 'vars2' 'vars3' adjustments", where you want to investigate the interaction between vars2' and 'vars3'.

Anyway, this seems to be a known regression problem, not an icp problem: That two or more regressors are collinear.
The variables behind 'vars2'_NOT_'vars3' in this dataset generates 'vars2'_NOT_'vars3' that is linear dependent of one or more of the other regressors.

See if the problem remains with "icp, show: stcox 'vars2' 'vars3' "?
If not the linearity is between 'vars2'_NOT_'vars3' and one or more of the adjustments.
Then try adding adjustments stepwise.
When the problem reoccur you have found one of the adjustments that is collinear.
Being systematic you'll probably be able to find all collinear adjustments.
After that you must decide which adjustments to keep in the regression.

If the problem remains with "icp, show: stcox 'vars2' 'vars3' " then interaction analysis with your dataset is probably not possible.

Whether icp is the preferred choice for testing interaction in Cox models, I can not say.
To my knowledge it is the only way in Stata if you do not want to do the calculations from the regression yourself.

Kind regards
nhb

[Nicholas Graham]

Hi there,
First time on the forum, please forgive any posting errors in advance. I am currently using stata 12 and having problems with the icp command.
Similar to the above i am wishing to do "icp, show: stcox 'vars2' 'vars3'" to assess the interaction between var2 and var3 after having stset the data in advance, however the output reads:

The regression is not of the proper layout
code=|stcox 'vars2' 'vars3'|

if i add a third variable, say age, the command progresses. However, the interaction being tested in the output is between var3 and age, which is not what i wish and did not appear to come up with anything sensible.
I had thought that this may be due to me using an older version of stata, causing the program to run a bit differently. Chancing my arm, I have tried putting in the outcome in a similar method after regress ie: "icp, show: stcox outcome 'vars2' 'vars3'", which does appear to select the correct variables to interact, however the stcox regression that runs prior to RERI is incorrect and the RERI results are nonsensical

ANy advice you could offer would be fantastic and many thanks to yourself for making coding such as this available.
Please let me know if you require any further information
Kind Regards
Nick Graham

[Niels Henrik Bruun]

Hi Nick and Marianne

Sorry! There is a bug in icp when using stcox regression.
The bug is found, but I have to test the code.
Then it has to be uploaded to ssc (Which usually goes quick).

I let you know when the update is ready
If it is urgent, you can have the files and install them yourself.

Thank you both for letting me know about the error.

Kind regards
nhb

[Nicholas Graham]

Many thanks for a speedy response!

[Tiago Pereira]

Hi, Niels.

Is there a way to adapt -ic- to work with -glm-?