I am looking for the relation between weight evolution (continuous) and a binary factor adjusted for confounders. The regression gives an (adjusted) difference and its CI less than the planned a non-inferiortiy threshold. However I can't find how to perform a test for the non-inferiority. I guess the bétha, the CI, the sample size and the number of vairables used to adjuste the model should be important. I use Stata (... if this makes a difference).
If I understand your question correctly, what you want is the t-test for the constant term in your regression. Stata should provide it for you. The statistic has the same formula as for any other regression coefficient. See this link for the details, Regression test for the constant term - Bing
but Stata should provide the computed values. Best, David Booth
Thank you for your quick answer. I'm not sure to understand what you mean by constant term. So I should maybe better explain the problem. I want to show that the weight evolution for patient receiving a specific management, is non-inferior to the patients without this treatment.
In the linear regression the dependant variable is the weight evolution (continuous) and the dependant variables are the treatment (binary) and other variables supposed to confound the association. Thus, the model gives the Beta for the factor of interest (and for the confounders), and I want to show that the Beta and its upper confidence interval, is smaller than the non-inferiority threshold.
Hello Gregor John. If I follow, you could use Ben Jann's user-written program -coefplot- to plot the coefficient & CI for your dichotomous explanatory variable. And you could use the xline() option to plot a line indicating the non-inferiority margin. Here is an example that I think lines up with what you described--it uses variable Foreign (1=Yes, 0=No) as the dichotomous variable of interest. (If you do not have -coefplot- installed, type ssc install coefplot, or just ssc describe coefplot if you want to read about it before installing.)
clear *
sysuse auto
tabulate foreign
regress headroom foreign length
* Suppose NI-margin is at -0.5 inches
coefplot, keep(foreign) xlab(-.6(.1).6,grid) xline(-.5) ///
title(Coefficient for variable Foreign)
This does not give you the test you asked for. But I think you could "roll your own" test if you really need a test by using post-estimation info stored after the -regress- command. You can retrieve the coefficient for variable foreign as _b[foreign], and you can get the corresponding SE by pulling the variance of that coefficient from the e(V) matrix and taking its square root. The t-test you get in the output is computed as (b-0)/SE. I think the test you want would be (b-delta)/SE, where delta = the non-inferiority margin. Is that right?
If you've never done further calculations using post-estimation info, you might find this introduction to the topic helpful:
https://stats.idre.ucla.edu/stata/faq/how-can-i-access-information-stored-after-i-run-a-command-in-stata-returned-results/
HTH.
PS- Figure 2 here might also be helpful.
Article Through the looking glass: Understanding non-inferiority
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Sampling Design
- Sample Size