First of all, picking up only significant variables after univariate analysis and then including only significant variables in the multivariate analysis might have statistical consequences. Sometimes insignificant variables in the bi-variate analysis can become significant in a complex multivariate analysis! Model building calls for substantive literature support for a priori causality, if it exists. Mere significance cannot guarantee causality. However, in many published papers particularly in biological and health science, this practice has been observed. If this is an established practice for any reason then I would like to learn more about it, but statistically there is a problem in such practice, in my humble opinion.

In addition, it seems that the dichotomous conversion of polychotomous predictor variables might not have been necessary. The polychotomous classification variable for patients, when converted to dichotomous variable, some important information loss is expected. For a study in health context, such information loss in ASA scoring (even if very minor) might have important effect on the outcome variable, harming the predictive validity of the model. In addition, the nature of ASA score may qualify itself as an ordinal variable. Having said that, you may use logistic regression to find odd ratios, and Wald statistic that denote significance of variables in the model. About forward or backward variable selection, there is no one best approach to modeling, these methods follow inclusion or exclusion criteria based on p-value (SPSS has some default p-value for these criteria). However, there are evidences in logistic regression literature that backward selection is often less successful than forward selection because the full model fit in the first step is the model most likely to result in a complete or quasi-complete separation of response values. Thanks.

First of all, picking up only significant variables after univariate analysis and then including only significant variables in the multivariate analysis might have statistical consequences. Sometimes insignificant variables in the bi-variate analysis can become significant in a complex multivariate analysis! Model building calls for substantive literature support for a priori causality, if it exists. Mere significance cannot guarantee causality. However, in many published papers particularly in biological and health science, this practice has been observed. If this is an established practice for any reason then I would like to learn more about it, but statistically there is a problem in such practice, in my humble opinion.

In addition, it seems that the dichotomous conversion of polychotomous predictor variables might not have been necessary. The polychotomous classification variable for patients, when converted to dichotomous variable, some important information loss is expected. For a study in health context, such information loss in ASA scoring (even if very minor) might have important effect on the outcome variable, harming the predictive validity of the model. In addition, the nature of ASA score may qualify itself as an ordinal variable. Having said that, you may use logistic regression to find odd ratios, and Wald statistic that denote significance of variables in the model. About forward or backward variable selection, there is no one best approach to modeling, these methods follow inclusion or exclusion criteria based on p-value (SPSS has some default p-value for these criteria). However, there are evidences in logistic regression literature that backward selection is often less successful than forward selection because the full model fit in the first step is the model most likely to result in a complete or quasi-complete separation of response values. Thanks.

In addition to the Khandoker's nice explanation I would like to add a warning against using stepwise methods.

Since the interpretation of coefficients in a model depends on the other terms included, it seems unwise to let an automatic algorithm determine the questions we should ask about our data.The decision which variables to include into an analysis should be based on theory.

In their seminal book Judd and McClelland state the following:

Models identified by stepwise methods have an inflated risk of capitalizing on chance features of the data. They often fail when applied to new datasets. Furthermore, the F and chi-squared tests quoted next to each variable on the printout do not have the claimed distribution and the method yields confidence intervals for effects and predicted values that are falsely narrow.

Judd, C. M., & McClelland, G. H. (1989). Data analysis: A model comparison approach. San Diego, CA: Harcourt, Brace, Jovanovich.

The arguments against stepwise methods are summarized here:

http://www.stata.com/support/faqs/statistics/stepwise-regression-problems/

Regards, Karin

 Firstly, Khandoker Mahmudur Rahman, Karin Schermelleh-Engel and Alvin Vista thank you very much for your response. It was very helpful. however i still have some concerns which i would like to share.

Mr Khandoker as per your discussion i have concluded that i need to include all my variables, whether significant or insignificant, in my multivariable (MV) Logistic regression (LR) analysis to get more reliable results. i  I also understood and followed your suggestion to leave my polychotomous categorical variables as it is. However, i need you to explain to me the usefulness of Wald statistics and how to interpret my results from the column of wald for each variable (predictor). 

However, at my institution i consulted biostatistics department, who suggested me to do a univariate analysis on individual variable and out of those which became significant at p

I understood the inclusion criteria shouldn't be only statistic but the theory behind. It's the procedure I always consider to define which variable to include in the multi-variable model.

well, I had the same question but when implemented Enter and forward stepwise(wald) method it was clear that the 2nd method results in better classification rates.

Omaid Hayat Khan Karin Schermelleh-Engel Khandoker Mahmudur Rahman Thanks for this fruitful conversation, looks like it is very often met. I would actually to ask 2 Qs:

1- At the very end, if I had NO significant predictors in the (Enter mode) but significant in the (stepwise fwd conditional mode). What does it mean?

2- What to choose finally in the my conclusion?

Thanks a lot.

1. There may be multicollinearity resulting in small unique effects of each predictor variable if you include all predictors simultaneously.

2. Follow Angelina Lessa's advice.

Kind regards, Karin

Thanks a lot. Michael R. Frone Karin Schermelleh-Engel

As to me, there is no hard and fast rule to decide which methods to use. I recommend to decide based on the nature and characteristics of the variables to be included. Enter method has also some limitations especially when we have lots of predictor variables which may have correlation with each other. This time, while controlling the confounding effect of all other variables may end up with no association but in reality, there might have some association. In such cases, better to use stepwise either automatically (if we have no sufficient evidence to rank among the variables) or manually based on priori knowledge and existing scietific eveidence or theory. e.g. if you are in dilema to choose between variables which one to enter first, you can choose forward LR to choose the most significant first.