Hello people,
I am interested in one delicate stuff, whatever happened to me during the data analysis. My issue is the following:
I have one dependent variable - Trust In Ad. I am doing Linear Regression to measure the impact of Facebook Involvement on Trust in Ad and it goes smooth, with all the Sig. p=0.000 and stuff.
However, when I want to measure and compare impact levels of Product Quality Perception and Facebook involvement (thus, I am including to independents in analysis: Product quality perception and Facebook involvement), something goes wrong for Facebook Involvement variable. Sig. level, in coefficient table, is more than 0.05 (p=0.256). Therefore, it implies that this data is statistically improper to be used.
In short, when I do impact of Facebook Involvement solely everything seems normal and I guess Sig.=0.000 which is good, while including Quality perception variable with Facebook involvement messes with the latter's sig level and it becomes more that 0.05.
What can be the reason of that can you advise me? Am I doing something wrong?
I will Attach the SPSS output for both cases: 1. Facebook Involvement's impact on Trust and 2. Facebook Involvement and Product quality perception's impact on Trust. Both of them are done with Linear Regression.
i) I suspect that the DV is not interval scaled. It may still be ok to use the DV as "quasi-interval scaled", but it can also make the interpretation of sum-of-squares weird or misleading.
ii) the residual plots for the second (simpler) model show a group of "outliers" - almost 10% of the values. How does this group influence the result? How do the values of the other IV look like for these values?
iii) There is a considerably strong correlation between the two IVs. It can well be that a similar part of both IVs explain the same variation in the DV. Adding both IVs will make the model having to devcide how the variance reduction is to be attributed to the IVs. I don't known what type of sum-of-squares (SS) is calculated here; one possibility is that the second SS is the contribution of the second IV over and above the first. This means: if you would exchange the order of the predictors you will get the inverse order of the SS-contribution and p-values. Such effects can be attenuated by centering the IVs (if these are quantitative variables).
What you see is likely the result of the "problems" addressed i) and ii) and iii), in some combination.
Hello Levan,
I would say you are doing everything all right. Both predictors were sig. in correlation table, and when in regression when facebook was on its own.
But then you added another predictor, and facebook is not sig. anymore. That happens pretty often. It seems that there is some overlap between the predictors and the role of facebook is diminished when quality perception is added to the regression model. It looks as if quality is superior in this model and although facebook on its own is a sig predictor, both predictors explain a decent proportion of common variance in the DV, hence one of them is redundant.
In a nutshell: Nothing is wrong. Also if you added more predictors it is NOT impossible that quality would be redundant as well.
Best wishes
Hello Levan. Do a Google search on , and start reading!
Also, a more efficient way to get your two regression models of interest is to use two METHOD=ENTER sub-commands, with FB_Involvement entered on the first one and Quality entered on the second. I think you'll find the output easier to work with. With this approach, you can also add the keyword CHANGE to the /STATISTICS sub-command to get a change in R-sq and associated F-test comparing the full model to the reduced (step 1) model.
REGRESSION
/DESCRIPTIVES MEAN STDDEV CORR SIG N
/MISSING LISTWISE
/STATISTICS COEFF OUTS R ANOVA COLLIN TOL CHANGE
/CRITERIA=PIN(.05) POUT(.10)
/NOORIGIN
/DEPENDENT Trust_in_ad
/METHOD=ENTER FB_Involvement
/METHOD=ENTER Quality
/SCATTERPLOT=(*ZRESID ,*ZPRED)
/RESIDUALS DURBIN HISTOGRAM(ZRESID) NORMPROB(ZRESID)
/SAVE COOK ZRESID DFFIT.
HTH.
Hi,
First of all, you are not doing anything wrong. When you perform simple regression (one independent) variable, you see that Facebook Involvement is significant. When you add another variable (Product quality) regression accounts for effects of two variables at the same time. In presence of another variable (product quality), it's normal and possible for another variable to lose it's significance.
First clue is that Product quality has stronger influence on your depended variable. I suggest you perform partial correlation analysis, to see the correlation of one independent variable with the dependent, when controlling for the influence of another independent. Then you can compare them and see how strong their influence is (if you are skilled with R, you can progress to find the partial correlation coefficients).
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Regression