Take the 2-minute tour ×
Cross Validated is a question and answer site for people interested in statistics, machine learning, data analysis, data mining, and data visualization. It's 100% free, no registration required.

I have a model that attempts to predict a nation's quality of life index by it's moral indifference to contraception and moral rejection of gambling. Initially the model contained several predictors, but I eliminated most using backwards elimination via AIC. Here is a summary of the model (generated using R):

> summary(fit1)

Call:
lm(formula = Quality.of.life.index ~ Morally.unacceptable.ga + 
    Not.a.moral.issue.co, data = qli_and_moral_ind)

Residuals:
    Min      1Q  Median      3Q     Max 
-89.670 -25.443  -4.732  36.129  64.441 

Coefficients:
                        Estimate Std. Error t value Pr(>|t|)    
(Intercept)             143.1410    32.7499   4.371  0.00019 ***
Morally.unacceptable.ga  -1.7690     0.3603  -4.910 4.71e-05 ***
Not.a.moral.issue.co      1.4471     0.7925   1.826  0.07981 .  
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Residual standard error: 40.39 on 25 degrees of freedom
Multiple R-squared:  0.6079,    Adjusted R-squared:  0.5765 
F-statistic: 19.38 on 2 and 25 DF,  p-value: 8.266e-06

There are two plots of the model that I can't interpret:

Norm QQ Plotresiduals versus fitted plot

According to the web, the residuals plot above may indicate predictable error, ie that I'm missing some variable in my model. Is that assessment correct? If so, what should I consider adding to the model? It kind of looks like the $y = x^3 - x$ graph - maybe add a cubed term?

share|improve this question
    
The residual plots look fine for such small sample size. Note that you chose the variables based on their relationship with the response variable. This is a quite easy way to destroy validity of the model. (See e.g. the book 'Regression modelling strategies' by Frank Harrell) –  Michael Mayer May 11 at 21:12
    
Thanks for the info. I'm not sure what you meant by your comment - is there an alternative to what I did? I chose predictors that contribute the most important information to correctly predicting values - is that bad? –  Evan Defend May 11 at 21:44
    
According to where on the web? Those plots look fine to me. I would caution you however, to search for responses to some of the many questions on stepwise regression here (the issues are basically the same for any form of stepwise regression). –  Glen_b May 11 at 23:19

1 Answer 1

up vote 3 down vote accepted

The residual versus fitted values plot is useful to illustrate if a model presents heteroskedasticity.

According to the Wikipedia article:

...the presence of heteroscedasticity can invalidate statistical tests of significance that assume that the modelling errors are uncorrelated and normally distributed and that their variances do not vary with the effects being modelled.

If one had observed heteroskedasticity, the parameters' standard errors (calculated through t tests) would not make much sense.

Your plot seems to be OK, though.

share|improve this answer
    
It also identifies unusual observations (#s 7, 10, & 28) and displays a locally weighted regression line. To reiterate the OP, do you think it can be used to infer anything about missing variables or the curvilinearity of the relationship between predictor and outcome? –  Nick Stauner May 11 at 22:00
    
@NickStauner, I guess just under observation of autocorrelation. –  Andre Silva May 11 at 22:14
1  
The residuals vs fitted plot is usually more for identifying non-linearity: the scale-location plot is slightly better tuned for diagnosing heteroscedasticity, although these two plots are admittedly pretty similar. –  Ben Bolker May 11 at 23:10

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.