Intention to Treat (ITT) Analysis

[Grace Gao]

Hi,
I would like to know how to do an ITT analysis for binary outcomes. Thank you.

[Leonardo Guizzetti]

This question has nothing to do with software, and that's perfectly ok. An ITT (or other such analyses) analysis are definitions used to identify the correct sample on which to do the analysis. The statistical analysis plan will define who is included (and excluded) from each analysis sample, such as intent-to-treat, safety, per protocol, etc. Let's say you have created an flag, taking the value of 1 for people in the sample, and 0 for those excluded, and further suppose that this flag is in a variable called -samp_itt-, then you could do the following.

This would keep only those people identified as included in the sample. Anything done after the -keep- command is applied to these people.

keep if samp_itt
... perform your analysis here ...

Alternatively, you can use if qualifiers for each analysis command. This work requires more typing, but keeps the full sample if you don't want to discard some members for some reason.

regress .... if samp_itt
logit ... if samp_itt

[Roman Mostazir]

ITT analysis is carried out based on randomisation regardless whether participants received the treatment or not. Follow the data example below:


For example, you have a control group (rand=0, n= 100) of which 70 experienced death (death=1) and an intervention/treatment group (rand=1) of which n=40 experienced death(death=1). However, you further inspected that while the control group received whatever was promised to them (treatment as usual/tau/placebo), in the Treatment group, uptake of intervention treatment was not optimal by the participants. Only 40 out of 60 who did not experience a death in the treatment group received the treatment and 20 out of 40 who died received the treatment. You can do analyses on several subgroups such as per-protocol/as-treated etc. based on treatment received but they are likely to produce biased estimates. ITT analysis will be the one which will do the analysis based on randomization (the logit model below) regardless the treatment receipt status.

    rand   death   numevent   treceived  
       0       0         30          30  
       0       1         70          70  
       1       0         60          40  
       1       1         40          20  


logit death rand [fw=numevent],

Iteration 0:   log likelihood = -137.62776  
Iteration 1:   log likelihood = -128.40497  
Iteration 2:   log likelihood =  -128.3876  
Iteration 3:   log likelihood =  -128.3876  

Logistic regression                                     Number of obs =    200
                                                        LR chi2(1)    =  18.48
                                                        Prob > chi2   = 0.0000
Log likelihood = -128.3876                              Pseudo R2     = 0.0671

------------------------------------------------------------------------------
       death | Coefficient  Std. err.      z    P>|z|     [95% conf. interval]
-------------+----------------------------------------------------------------
        rand |  -1.252763   .2988072    -4.19   0.000    -1.838414   -.6671117
       _cons |   .8472978   .2182179     3.88   0.000     .4195986    1.274997
------------------------------------------------------------------------------

[Grace Gao]

Thank you very much for your answers. I was asked to do an unadjusted (complete case, ITT) analysis for the primary outcome using stata. And the outcomes are binary. But I was only been told how to do an unadjusted analysis of continuous outcome. I used 'ttest' in stata to do the analysis for continuous outcome. But I do not know can I still use 'ttest' to analyse binary outcome. Or is there anything else used in stata to analyse binary outcome? I really appreciate your help.

[Clyde Schechter]

The simplest unadjusted analysis of a study with a dichotomous outcome compared across groups is:

tab outcome group, col chi2

For it to be an ITT analysis, all participants who were randomized must be included in this calculation, even if ultimately they did not receive the treatment they were randomized to.

[Grace Gao]

Thank you, Clyde. So you mean just use 'tab' code in stata to analyse the binary outcome?

[Clyde Schechter]

Yes, when the outcome is dichotomous (or discrete with any number of values) the comparison of groups is most simply carried out with the -tab- command.

[Monzur Morshed]

Dear Clyde
Thanks for your cooperation "tab outcome group, col chi2" by the above code I only get frequency and percentage. How do I get coefficient, p-value and margins of ITT. Thanks in Advance

[Joseph Coveney]

How do I get coefficient, p-value and margins of ITT.

Wouldn't it be something like

glm outcome i.group, family(binomial) link(<linkname>)

with <linkname> one of logit, log or identity, depending upon what your protocols' statistical methods section or your statistical analysis plan specified?