Multiple Regression in SPSS/PASW (The Simple Way)

[sankalpnavghare]

@bartonpoulson Big Fan You and Your SPSS, I ve even viewed you lynda.com SPSS videos, Man YOU ARE A GOD when it comes to SPSS.

Pls let me know if you do any training on Marketing Research.

Thanks for all the Training CAPTAIN!!

Sankalp N.

[MrMustav]

Thanks!

[mufumbihc]

hi bartob, i need help with running an analyis with many predictors to determine aassociation on a DV. the problem there are so many variables and not knowing which one's to chose. Do i run a correlation to determine the relationship of all variables and then chose which best ass with my DV? please help

[bphnextdoor]

this helped IMMENSELY!

[percdann]

Statistics is horrible.

[dawidnejman]

Thank you for this one :)

[MsRohan20]

Thanks a million....:)

[NelsonMandeIa]

God or mother nature, what ever you believe in :), bless you!

[Tomahawk1999]

Hi there, wish someone can help me. I have a 2 reflective constructs with 5 indicators each. i want to run a MR model. what am i supposed to do with the indicators? average the scores? please help

[bartonpoulson]

@Tomahawk1999 Generally, yes, you would average them as long as they're on the same scale and they're correlated with each other (do a cronbach's alpha to check them). Then you would want to see if the construct scores were correlated with each other. If not, you could just do two regular multiple regressions. If they are correlated, then you have a few choices. If the correlation is really high, see what the cronbach's alpha is for everything put together; you may just a have a single factor.

[cutakecare]

@bartonpoulson

Thank u for educating... can u plz also tell how to form a norm group in SPSS 17 of a personality test. I have data of 200 people and test has 35 variables. I have got the zscore and tsore and I do not know the next step... basically need to make a norm group and name it. Its really urgent for my project.

Plz help me on this.

[bartonpoulson]

... And if the correlation between scale scores is moderate (I leave that up to you to decide what is moderate!), then you might need to do a multivariate analysis like canonical correlation or SEM, but those are both pains that I try to avoid. Let me know what they look like.

Bart

[Tomahawk1999]

@bartonpoulson

Dear sir, ty so much. is it possible at all for you to upload a video related to this where u can take a simple model with two reflective constructs and a DV and walk all of us through? Most datasets are canned (one observation for IV and a corresponding one for DV) which makes it hard when real survey data comes in.

before i do the cronbach, should i be doing the factor analysis to decide which items to retain?

really appreciate your quick response and thanks for the help.

[hessenface]

You are a lifesaver. In all five flavours. I'm a big fan.

[mplm00]

This a GREAT tutorial, you won't believe how much these 9 minutes helped! Thanks!

[mysoukouss]

Thank you soo much for explaining the SPSS out put so well. You expalined everything in a very clear and simple way.

God Bless You

RICH

[jusmedic]

where could I get such a collection of world data ? I would like to do some analysis on my own...thanks

[wuzumaki]

I just finished my final exams because of this video. Thanks a lot!

[daisysaran]

thank you

[robbie1988]

Barton, you sir are a god.

I am sitting in a crowded IT lab with my headphones in listening to your excellent explanation of regression and I am learning a hell of a lot more than I was in my lectures. Thank you so much

[kanidaka]

thank you



[Tranqvilier]

Thank you so much! Your explanations are so clear and understandable.

This really helped me more than any other tutorial in spss.

[gbirbilis]

nice tutorial, thanks

[psychosoundsys]

THANK YOU... THIS IS REALLY HELPED ME :)

[Mrjumawan90]

Thank a Lot.. it is indeed a great help especially in may thesis..

[webofbelief]

Excellent! You should have a series on regression, and advanced regression for graduate students. I would personally invite you to a guest lecture.

[mackykamya]

Thank Bart your videos on using SPSS are extremely helpful for those of us who've never used SPSS and have to analysis data for assignments. I will use the tutorial on correlation and regression to predict my hypotheses.

KR

Macky

[czarnywroc1000]

does anyone knows how to check if multicollinearity is a potential problem with my model and how to make hypothesis, how to interpret in detail the numerical value of the coefficient of determination associated with my regression model and also how to generate suitable plots to assess assumptions? Thanks for your help

[wallflowerchica85]

Briliant, thank you very much for this!

[JKmcbic]

This is soooo helpful. I'm a doctoral student in Counselor Education and Supervision, and I am very grateful. Do you have more of these available?

[lpb2108]

I swear that I have learned more from watching three of your videos than I did in an entire semester of stats. Amazing.

[arniz00]

You're awesome!! AAA

[8trackninja]

ThankyouThankyouThankyouThanky­ouThankyouThankyouThankyouThan­kyou!!!!!!!!!!!!!!!!!!!!! This is so wonderful. I missed class and was lost until i found this!!! Thank you so much for making these!

[kennycafferkey]

Barton this is excellent you dont believe how helpful this is. would you consider doing a video of a mediation test using regression analysis :) Thanks again

[bartonpoulson]

@kennycafferkey Glad to hear it. As soon as I can get my act together, I'll try to do that (along with factor analysis and cluster analysis)!

[justin8910]

@bartonpoulson I can not reiterate how much this video helped me understand something I simply do not. My dissertation is due in 2moro morning and you have enabled me to blag something I never learned into my diss. I owe u a drink

[chatlil]

Hi , this is extremely helpful - cant thank you enough. I am a merket research professional in India and we use this a lot to understand what effects choice - the bit about coefficients was very very informative. Can you also tell us how to do factor /cluster analysis on SPSS

[bartonpoulson]

@chatlil Thanks for your message. I'm glad the videos are useful! Other people have asked for factor analysis, so I'll try to get to that next week (when I'm back at school). Cluster analysis is also a great idea, although that might take me just a little bit longer. I will, however, get to it ASAP.

[chatlil]

@bartonpoulson @bartonpoulson Thanks again . I also had one question ; at the end of this video you have pointed out that there is one variable where corr. changed when all the variables were regressed together against the dependent; my question is when I am constructing my regression equation now , do i need to rerun this regression after removing this variable (GDP) and keep doing this till all variables in the coefficients table show a sig <.05 .

[chatlil]

@bartonpoulson Till now as a practice, we used to check the correlations between independent variables in pairs at the beginning and if they were correlated we removed the variable that has lower correlation with the dependent variable before doing the multiple regression ; I wanted to know if this is the acceptable way to do it as i have been doing this for a while and am apprehensive that i was doing it the wrong way

[bartonpoulson]

@chatlil What you're doing sounds like a sort of stepwise-regression-by-hand procedure. If it's what you want, you can have SPSS add and remove things automatically by doing a stepwise procedure, but many people think that's a bad idea because such procedures really build on quirks in the data. But really, it all depends on what you're trying to accomplish; if you're just reporting the results of a study for work or school, I'd just group the significant and non-sig ones and leave them all in.

[bartonpoulson]

@chatlil On the other hand, if you're putting together a model to be used in other studies, you may consider keeping just the big ones (but you would really want to check out that reduced model's accuracy in another data set). Personally, I usually give both the regular correlations in a single column (showing how they're associated with the outcome variable) and then the results of one big multiple regression in the next column, making it easy to compare the two. Hope that helps!

Bart

[nathanielkolenberg]

Thanks so much for this helpful video! Just one quick question though. How do you do a multiple regression when there is a moderating variable involved? Let's say z might influenc the relationship between x and y. Can you just look at the correlation between z and x to determine the moderating effect when the regression shows that the variable z has significant explanatory power? Thanks in advance for your time. Kind regards.

[bartonpoulson]

@nathanielkolenberg Interestingly, that's the exact same question the last person asked! (As such, see my responses to katerinagn below.) The basic idea is to look at the correlation between x and y when controlling for y. If you're only concerned about those three variables, then the easiest way is to do regular zero-order correlations (the normal Pearson product-moment correlation coefficient) for all three and then get a partial correlation for x and y controlling (or partialling) for z.

[nathanielkolenberg]

@bartonpoulson Thanks so much for your quick reply. My apologies for not seeing the previous reactions earlier. I was so excited to find a good explanatory video of multiple regression that I just wrote the question right away. But thanks so much for the information. Kind regards.

[katerinagn]

yes my variables are all coded! i used a 7 Likert scale from 1=highly disagree to 7=highly agree! I will check the correlations and coefficient and hope everything goes as hoped! Thanks for sheding some light on it!

[katerinagn]

thanks for your fast reply! the study is about users'' perceptions on moderators' power use in online communities, and how moderators' power can have an impact on their intentions to share knowledge. I have 6 variables.INDEPENDENT: 1) commitment to community 2) reciprocity 3) enjoy helping others 4) reputation. DEPENDENT: intentions to share knowledge. MODERATING: perceived power use. I need to show that perceived power use will weaken the relationship between the independent variables&dependent

[bartonpoulson]

@katerinagn That's an interesting study! There are a few ways to go about this.

First, what are the correlations between perceived power use and the other variables? A negative correlation between power use and your DV provides some support for your hypothesis.

Second, you can just run a regular multiple regression where you DV is a function of the IVs and power use. Again, a negative coefficient for power use provides support....

[bartonpoulson]

@katerinagn Third, you should also do a multiple regression without power use so you can compare all of the IVs coefficients with the last model. You can do this in one step if you do a hierarchical regression. (I have another video called "Hierarchical regression in SPSS-PASW as a quasi-experimental method" that shows how to do this.)

[As a note for the future, I find it MUCH MUCH easier to keep track of things if variables are coded so "agree" has the high numbers and "disagree" is low.]

[katerinagn]

could you please tell me what you do when having a moderating variable? and i am also not sure if i have to use linear regression for my model. all of my items in the survey had scales (1-totally agree/ 7-totally disagree). and 1 last question: before i enter the data in SPSS, the means of the items concerning each variable are added to one value?? f.e if you have R1 equals 6, R2 equals 6 AND R3 equals 7 you add them to the value R equals 6.333333?

HELP!

[bartonpoulson]

@katerinagn Multiple regression works great with moderating variables. I'll need a little more information to help you but it should work out fine.

1. What is the study about?

2. What are the variables in the data set?

3. What is the moderating variable?

4. How do you think the moderator would affect the relationships?

Also, it's best to enter the data in raw format and do any averaging or such in SPSS. I'll wait to hear from you.

[Kittensbane01]

thank u very much

[oohjamaflip]

thanks so much for this - sent a request to your inbox! xx

[dogarizm]

thanx a loy! i wish my stats teacher was as good as you are ;-)

[dogarizm]

can anybody plz tell me that on what basis can we chose two variables from a large given data to be tested for regression? also how do we interpret the following results:

R value

R square

Sig

plz help i hav an spss final exam tomorrow ;-(

[bartonpoulson]

@dogarizm Which variables to choose really depends on your purpose. Basically, the dependent (or outcome) variables are what you trying to predict (like outcome, for example) and the independent (or predictor) variables are the ones that you use to predict outcome (like education, age, etc.). Regression is nice because it can pretty much do anything that any other procedure can, such as t-tests, ANOVA, chi-square, etc. (to be continued...)

[bartonpoulson]

@dogarizm (Continued) As far as interpreting r, r^2, and sig, goes, r is the correlation coefficient: 0 is no linear association, 1 is a perfect straight line, +/- indicate uphill or downhill. R^2 is good because it tells you how well you can predict the outcome, from 0% to 100% of the variance explained. Sig, or the statistical significance level (or p-value for probability value) tells you how likely the result is IF there really is no association (i.e., a false positive)...

[bartonpoulson]

@dogarizm (Part 3) A small sig. value (less than .05) is usually considered "statistically significant"—the result is unlikely to arise as a false positive, or, in other words, the difference between groups (or association between variables) is reliable. Also, as long as the sample size is consistent, bigger correlation (both r and R^2) have smaller p-values (or sig. levels).

Hope that helps! Watch the correlation and regression videos, and possibly the ones on comparing means. Bart

[madeleinemahin]

thank you for a quick response! It helped a lot! really cool that you answer so many comments. :)

[madeleinemahin]

Great video! :) I have a question about an assignment: they want to examine if children who are breast-fed after six months of age have higher IQ scores. Y = IQ score of children, and the x variables are: IQ score of mother, IQ score of father and if the child has been breast fed or not (yes=1 no=0) They ask why I think it is important to include IQ score for both mother and father, even though they are highly correlated , And b0 is 0, is that normal? Hope you can help, I would be very thankful!

[bartonpoulson]

@madeleinemahin Thanks, glad you like it! Your question is a good one; it's easy to measure the association between breastfeeding and IQ but it's hard to show that any differences are BECAUSE of breastfeeding. That is, mothers who breastfeed are probably different from those who don't in a lot of ways; this is the issue of confounded variables. Controlling the association by using the parents' IQs is a good idea because that will compensate for many of the confounds. And b0 NEAR 0 isn't unusual.

[pavorittiluvn]

Enjoyed looking @ the video. Question...what if you have 4 dummy variables...how would you include them into your analysis report

[bartonpoulson]

@pavorittiluvn That shouldn't be any problem. Just make sure that you have one less dummy variable than categories (e.g., you use only 3 dummy variables to represent 4 categories because the 4th category is implied by the others) so you don't have multicollinearity. All you would need to do is have a table with the names of all of the predictor variables (including your dummy variables) down the side and the coefficients (and their p-values) next to them.

[criticalclarity]

Studying for a Research Design and Interpretation course final - which includes the interpretation of a PASW multiple regression output - and this was IMMENSELY useful. Can't thank you enough!

[MrBluemirror]

Thank you!!

[justinfardin]

Its a good video to learn the explanation of outcome..

Guys, i am very new of analyzing data in spss. could you please help me how to make the average dependent variable befor analyzing the MR?

I need your kind attention, please..

Wait for your help...

[bartonpoulson]

@justinfardin Do you mean that you need to know the average of the dependent variable? If so, that one's easy: You just need the univariate descriptions, where you can get the mean, standard deviation, and so on. Go to "Analyze > Descriptive Statistics > Descriptives..." and select the variables you want in the left column so they go in the right column. You can then just click on "OK" because it will give you the mean by default, along with the N, Minimum, Maximum, and Standard Deviation.

[MrBluemirror]

What would you suggest if the constant value of Multple regression comes negative.? Is really a big problem...or any mistake done while running M regression..

[bartonpoulson]

@MrBluemirror No problem at all with a negative constant. What the constant represents is the value on the outcome if ALL of the predictor variables (that is, independent variables) have values of zero. In many cases, that can't happen in the data (for example, height can't ever be zero), so the constant just represents a mathematical starting point and not anything particularly meaningful.

[covercoverful]

omg thanks sooo much. this is great!

[DtotheBU]

dear bartonpaulson...i don't throw this word out alot..but you sir are a HERO. you saved my bacon so badly i cannot express it in words!!! great tutorial, it all makes sense now!

[duipraye]

your videos are amazing!! thank you so much! i cant believe it has taken me two years to finally fully understand this and all i need was a youtube tutorial!! i should just stop going to my classes and sit on youtube all day! id probably learn alot more! but seriously your videos have really helped me ALOT!! thank you :)

[bartonpoulson]

@duipraye Well, I wouldn't recommend stopping classes and watching YouTube all day (you'd probably get bed sores on your rear) but I AM very grateful for the lovely feedback. Hope things go well.

Bart

[souw1990]

Thank you very much, you were a great help! :)

One last question, do you know if there's a simple examination that I can carry out to decide whether multicollinearity might be a problem with this analysis?

Thanks

[bartonpoulson]

@souw1990 Probably the easiest is to just look at the correlations between your predictors—it they're big then you have a problem. A more technical approach is to click on "Statistics..." in the "Linear Regression" dialog box and check "Collinearity diagnostics." The second to last table in the output ("Coefficients") will have a last column called "VIF" for "Variance Inflation Factors." If that number is over 10, for example, then that variable is too redundant with the others.

[souw1990]

@bartonpoulson

What is considered too big regarding the correlation between the predictors?

At the moment I have 0.513 between two of my predictor variables which has been marked as significant at the 0.01 level. Does this suggest that multicollinearity will be a problem with my analsis?

Thanks

[bartonpoulson]

@souw1990 That's pretty big so, yes, there is multicollinearity but it doesn't kill things. What it does is make the regression coefficients less stable (i.e., with a new data set from a similar group, the coefficients could be very different). Not the end of the world, though. Mostly, it's something to keep in mind, especially when interpreting your results. That is, the coefficients that you get are ONLY valid in the context of all of the other predictors in that particular model. Have fun!

[berber85]

Can you help me with this?

I have 4 predictor variables and 3 dependent variables, however in each variable i have a number of questions. So say, satisfaction is my predictor variable and i have asked 3 questions regarding satisfaction. my dependent variable is impulse buying and i asked 7 questions. the question is how do i 'group' them in order to do the regression? do you get my question?

[bartonpoulson]

@berber85 Do you mean, for example, that "satisfaction" is a variable that is computed as the average of three other questions? If so, just create a new variable like this: "COMPUTE Satisfaction = MEAN(q1, q2, q3). EXECUTE." but put in the real variable names for q1, etc. Also, are you planning on doing 3 separate regressions, one for each outcome variable? Or are you asking about grouping the predictor variables (after the composites are made)?

[berber85]

@bartonpoulson yes that's what i mean. what you mean by putting in the real variables names for q1, etc? i am actually doing separate regressions but not for each "question" because other than satisfaction as a predictor variable, i have 3 more predictor variables, like hedonic shopping value is another predictor variable with 11 questions.

i actually read up on the "compute", but how to do it using the icons given in spss?

Thank you so much for the quick reply.

[souw1990]

Is it possible to use multiple regression with interval, ordinal and nominal data in the same analysis?

For example: if i wanted to predict a criterion variable using an ordinal, nominal and interval predictor variable all in the same test?

Thanks

[bartonpoulson]

At the risk of appearing statistically irresponsible, I'd say yes, you can, because the Pearson product-moment correlation coefficient for interval/ratio data (r), the Spearman rank-order correlation coefficient for ordinal data (r(s)), and the phi-coefficient for nominal data are actually all the same thing. That is, the Spearman and the phi coefficient are nothing more than computation simplifications of the Pearson that take advantage of the regularities in the data. Go for it.

[MilliVanilli2007]

Sometimes I believe God talks to me and answers my prayers. Finding this video is defenitely one of those moments. You saved me ...thanks!

[bartonpoulson]

My goodness, I'm blushing! I'm very, very glad to be of help. If you'd like to see anything else, just let me know.

Bart

[MilliVanilli2007]

@bartonpoulson Haha...well...you could help me by answering a question-- i have a small sample of only 32 people. I can't do multiple regression there, can I? I have done a simple linear regression now for every significant pair that came up in a Pearson's correlation.....is that right/ all I can do? thxx

[bartonpoulson]

Well, you do have a sample size issues but, really, it depends on how many variables you have. You probably wouldn't want to do it if you had more than 2 or 3 predictor variables. If you have more than that, I would probably just stay with the Pearson correlations (not even simple regressions), making sure to keep in mind that if the predictor variables are correlated with each other (which they usually are), then the associations in the multiple regression could be very, very different.

[yannis5]

you explain things better than my teacher

[frowningcat]

oh man, thank you soo soo much for posting this! i never thought in a million years that i'd find such an in-depth tutorial for this. you've helped me immensely!

[butterfly136923]

You have no idea how much this helps! I wish I would have discovered you earlier!

[diddle0111111]

very useful!

[ttminhttminh]

a very big THANK YOU

[senadoll]

Thank you. Thank you. THANK YOU.

[KevsLola]

Thank you so much for this video, I finally understand what I am doing. Everyone else I have asked confused me more, but you are explaining it so clearly and concise. Thanks!

[arnatwannasri]

Yeh, i got it.

I will examine whether treatment expectancy/credibility influence the treatment outcomes. This method might be a good one.

[rodkatxxx]

thank you so much, for the first time i actually understood what i am doing!

[kaare100]

You are awesome dude!

[jeetalshah]

Very clear explanation. Thank you so much!!

[MsSweetc1]

this is awesome. i read the book 3 times and still didn't get it--now i do

[gaicccp]

Thank you so much for this clear and useful explanation.

[arsalanrozy]

You are super for doing this! thank you.

[lb5962]

to do multiple regression do you need to have continuous variables?

[bartonpoulson]

No, that isn't a restriction; you can use standard multiple regression for a quantitative outcome (like income or IQ) or logistic regression for a categorical outcome (usually dichotomous, like yes/no).

In either case, you can use quantitative or categorical predictor variables. Categorical variables should be dichotomous indicator variables (0/1). You can use a variable with more than two categories if you turn it into several 0/1 variables. (I can show that in a video later, if you want.)

[lb5962]

That would be very helpful Barton.

I have a masters in pure maths, and I've been employed by a relative to do some statistics for a public health questionnaire regarding underage drinking. Because of my lack of statistical knowledge I am finding it very difficult to understand what I am expected to do. I have managed to pick up rather a lot in 2 weeks, but I'm still not confident. If you feel like you have time to give me more advice via messages I would be very grateful.

Thanks for your time.