That doesn't really matter. You only have to take care (in any case, btw) to interpret the coefficients correctly. In your case (responser = 0, nonresponder = 1), a positive coefficient (what is a log odds ratio) indicates an increase of the odds being a nonresponder. Exchanging the coding is the same as changing the sign of the coefficient.

Your coding should be such that it matches your question,

and as regards the software it will have a particular way of using a reference. SO it is back to you on your coding and the software.

Two things will help: first look at an example Eg mortality rate = died/exposed.

In your case you have intervened and are looking at a response. Did you also have a control group? What is your data is it really a 0/1 responder/non responder as you mentioned response rate ?

You look at the results as the log-odds of a "success" compared with a 'failure'

You also need to look at the meaning of the odds ratio all which will help you understand if this is right method. I suggest you look at a text book and a paper in your field to see whether the method is relevant.

Also not the reference notation varies with the software.

That doesn't really matter. You only have to take care (in any case, btw) to interpret the coefficients correctly. In your case (responser = 0, nonresponder = 1), a positive coefficient (what is a log odds ratio) indicates an increase of the odds being a nonresponder. Exchanging the coding is the same as changing the sign of the coefficient.

The outcome of the analysis is a model that predicts the probability of a particular outcome (in your case responder or non-responder), where Prob(Responder) = 1 - Prob(Nonresponder). If you ask the program save the probabilities in the two scenarios (where the response is coded as 1;0 or 0;1) the predicted probabilities will be identical. If your aim is predict the probability that an experimental unit will be a responder, the model will be easier to interpret if you code the outcome as 0=nonresponder and 1=responder.

Dear M. A.,

I largely agree with Jochen Wilhelm and Ikram Sghaier. The coding has nothing to contribute in your results here. Either way will provide you same results, but just in an opposite directions of the effect size.

Best...

Razak M Gyasi Jochen Wilhelm Miranda Mortlock Ikram Sghaier

In my view its up to you how would like to code. In my case I generally avoid '0' while coding. Generally 1, 2, 3, ...

Hom Nath Chalise,

Coding with 0 and 1 has the benefits that

• the resulting variable is a "count variable", counting the occurences of the values coded with 1,
• the intercept in the model is the estimate for the reference group (coded with 0) (here this estimate is a log odds score), and
• the simple contrast coefficient is a log odds ratio, estimating the effect of "being in group '1'" vs. "being in group '0'") (here: a log odds ratio)

Using a different coding scheme is not wrong, but it seems less elegant in most applications and makes interpretations more difficult.

If you have more than two groups, coding the groups in a single variable with 1, 2, 3, ... is not recommended and can lead to serious misinterpretations. Most software would interpret the variable as numeric and fit a simple regression model, assuming that group #1 refers to x=1, group #2 to x=2 etc. The model fits one intercept and one slope, and neither coefficient is meaningful.

I agree with the comments above, in particular that your coding should align with your question of interest. If you wish to understand what drives a response to the intervention, then coding responders as '1' makes the most intuitive sense to me. I believe coding decisions should be intentional, so I would avoid letting the software assign default values.

However, from a coding perspective, you don't necessarily need to recode the 1,0 designation to get the flip result. If you wish to model from the perspective of what drives a response, but responders are coded as '0', then you can simply adjust your MODEL line syntax as follows:

MODEL DV (EVENT='0') = IV;

The key in any circumstance, is keeping your underlying coding decisions in mind as you go to interpret the results. The significance of your results will not change based on this decision, only the direction of your interpretation.

Jochen and others are right - so long as you know/remember what groups 0 and 1 represent it won't influence the outcome of the analysis

Con la evolución de los sistemas, programas como R, no requieren codificar, la presencia o ausencia de un atributo se maneja de misma forma que si estuviera codificado con dos dígitos cualesquiera. Vale la pena insistir que si se manejan censos los resultados, tipo proporciones, son probabilidades, mientras si parte de muestras los indicadores corresponden a frecuencias relativas que a la vez son estimadores de la medida de probabilidad.

I too think it is important to think about the coding and I often choose a coding for the outcome and the categorical predictors that gives the lowest relative odds of 1. This is meant to help the reader as many people have difficulty interpreting (non-linear) odds in that a relative odds of 0.33 is equivalent to an odds of 3.

I use the MLwiN software which we have designed so that you can change the base category of the coding after estimation of a given model ( it does this by using the point estimates and standard errors as a basis for simulation) which can be much faster than re-estimating a complex model.

see customised predictions in this

Article Manual supplement for MLwiN Version 2.14

Some of the above answers may not be referring to binary logistic regression.

In binary logistic we have 0 and 1 coding of the outcome variable. Typically this could be for dead/alive; healthy/diseased; etc so widely used in medical studies.

In medicine it may be the outcome from some treatments and the 0/1 would indicate the disease or no disease for particular treatments..

To really engage with this method I would suggest read a good statistics methods book explaining the technique.

The outputs are typically reported as odds ratios, and so you need to understand the binomial data . The theory uses the Generalized Linear models using a logit link. (to fully plan and analyze) it is best to have a sound grounding in the theory. This space is to small to expand.