Hi Shannon

Stepwise and backwards regression are generally not recommended. The reason is that it can exclude associations that for whatever reason aren't identified as statistically significant, and include associations that do not reflect reality and are only significant by chance. Article Step away from stepwise

My approach is generally to be theoretical and think about why I am including covariates. So first off, is your analysis exploratory or confirmatory? The former is just trying to test what "stands out" as a predictor of your outcome. It's crude, but can lead to further hypotheses down the road. You may want to think about multicollinearity, or a strong relationship between independent variables/predictors, which can skew results. You can exclude or combine variables with high VIF, based on what you think makes sense (e.g, education and earnings are associated and may therefore have high VIFs, so you may want to exclude one). See here for how to check: https://www.spsstests.com/2015/03/multicollinearity-test-example-using.html

Confirmatory analyses are about having a specific idea, e.g., I think x causes y, but I'm concerned that it's confounded by z. To give a basic example, you think smoking causes lung cancer, but you're worried a genetic confound may cause people to both smoke and independently cause cancer. So you adjust for the presence of that gene to see if the association between smoking and lung cancer attenuates to non-significance. If this is akin to what your want to do, you may want to think about what is an actual confound (i.e., affects the outcome and predictor like the smoking/gene example, or affects the outcome but is unequally distributed in the predictor; for instance, comparing school performance, an imbalance of females across schools could confound results). You want to avoid crude testing to identify confounders in this case, because some factors that are associated with the outcome and predictor should *not* be adjusted for, such as mediators (smoking [predictor] -> tar deposits [mediator] -> lung cancer [outcome]) and colliders (another complicated topic).

I think you have a lot to offer as a qualitative researcher in this regard. There's a false impression that stats people just run tests to see what is significant, but in reality, we need that theoretical or qualitative understanding to put together reliable statistical analyses.

Just some other points.

• The specific type of regression you conduct depends more on what your outcome is. If binary, you would use a logistic regression. If continuous, linear regression.
• You should consider how many predictors your data can support. The rule of thumb is 10 observations for every variable.
• I believe that if you can plug the categorical variable directly into a regression in SPSS, so you don't need to create dummy vars; SPSS will automatically generate the dummy vars. By default it will select the first value as the reference. You can manually override to make the last the reference, or use syntax to specify another, but it's been a while since I've used SPSS so it may have changed features. For more info: https://www.ibm.com/support/pages/can-i-specify-reference-category-other-first-or-last-spss-binary-logistic-regression

I hope that's helpful.

I agree with Tyler. Here's a couple more references that should be helpful to you. Stay away from stepwise regression and its variants. Best wishes, David Booth

@Shannon In your query, you have not mentioned the type of data you have for you dependent variable and also the number of dependent variables in your model. Type of suitable regression model is decided on the basis of the data you have for your dependent variable(s). If your dependent variable is also categorical, then you need to use multinomial logistic regression. If it's on Likert scale and you have use composite score to compute dependent variable (continuous) then go for linear regression. You can also use Smart-PLS as it works well with categorical data for both dependent and independent variables.

In regard to your query regarding stepwise, backward or forward regression, I would agree with @David and @Tyler to refrain from it. Wishing you best with your research.

You also haven't mentioned what your research question(s) is (are), and in particularly why you are not interested in the associations between pairs of variables (i.e., bivariate associations). Worth knowing your sample size, the number of categories in your IVs (so we get an idea of the model df), etc.

All good answers--especially Tyler's advice to think about the purpose of your analysis. I would only add that if you have a quantitative response variable, then your analysis is technically an Analysis of Covariance (ANCOVA) (with your quantitative predictor as a covariate). Computing the analysis in a regression is OK--the calculations are the same. But you may find the ANCOVA easier to interpret and easier to work with if you have interactions among your predictors to consider.

Thank you to all of you. Tyler J Lane , I was unaware that SPSS would create the dummy variables for me automatically. Very helpful! The research is exploratory in nature. Imran Anwar , I have one dependent variable that's continuous data. Daniel Wright , the research question is: "Which of the following factors best predict (dependent variable)?" My IVs have the following numbers of groups:

X1: 5 groups

X2: 4 groups

X3: 4 groups

X4: 4 groups

X5: 2 groups

Sample Size: 138

Any other info you can offer would be greatly appreciated.

Shannon

@Shannon Since you have your continuous dependent variable, you can use linear regression analysis to testify which independent variable is the best predictor of the outcome variable. Wishing you best with your work.

"Which of the following factors best predict (dependent variable)?"

First, multiple regression is wrong, though the forward method in some packages could get you the answer. The reason why multiple regression is wrong is that the individual betas are conditional on the other predictors, and as you have stated you are not interested in this. To answer your question it depends what you mean by "best," but let's say it is something like largest omega from an ANOVA. If you ran an ANOVA on each of these, you could just choose the largest, and in that sense it predicts it best. The metric predictor you could find a similar value. If you weren't just interest in this, but say two-way interactions, you could include these. However, you have about 600 cells, some of which have zero frequencies, and only n=138, so even two-way interactions are low on power.

If I was trying to do predictions like you are I would not be using standard regression as it is likely both to miss non-linearities and interactions and you really need multiple hold-out samples so as no to over-fit to the data you have.

In exploratory mode I would try using Tree regression; here is a worked example -

http://www2.stat.duke.edu/~rcs46/lectures_2017/08-trees/08-tree-regression.pdf

The cross validation is important as it gives some idea of how complex a model can be supported by the data you have. As you increase the number of terms in the model you usually get a better and better fit to the observed data but a worse and worse fit to the holdout sample - you are then modelling noise and what you have found will not reproduce, and can be very misleading. Software implementations of tree regression in R do this automatically.

https://cran.r-project.org/web/packages/tree/tree.pdf

It is also available in SPSS, but I have never used that version

https://www.ibm.com/uk-en/products/spss-decision-trees

Very helpful! Thanks to all of you!