Class Notes
(810,290)

Canada
(494,047)

University of Toronto Mississauga
(23,388)

Sociology
(3,990)

SOC350H5
(9)

Lecture 9

Unlock Document

University of Toronto Mississauga

Sociology

SOC350H5

David Pettinicchio

Winter

Description

Lecture 9
Omitted variable bias
Left out variables that are negatively affecting outcomes
What are known predictors influencing outcome
Something destabilizing model (lack of it)
It’s a problem with x highly related to each other
Because cannot decide what variable has the effect cuz you can t separate the
two
Model specification
Parameters that you decide are boundaries of outcome
Part of understanding outcome is first am I putting in irrelevant things
Not affecting outcome but also screwing up outcome
Jeopardize ability to predict outcome
When understanding multivariate model – saying stat sig are interpreted in
relation to all x
Standard error
When you have omitted variable bias
Can make things be non sig when they may be sig but fact that model is on
shaky grounds because of variables that you select
This happens when you add irrelevant x or miss important ones
Omitted Variable bias
Solution – add variable that you think are important
You cannot substitute something else in there if data doesn’t contain it
Find another way to deal with problem
Adjusted r2 very low – problem
People rely on comparing r2 to adj r2 to see is variable bias
You saying x is increasing y in a linear way but non linear relationship could
be problem
A lot of concern about what you do when you don’t know how to construct
model based on unavailability of data
Look at r2 with caution
SPSS OVB
Requires extra step
Create a variable –
Create a model
Ask spss to sq outcome of predicted values
It’ll create a new variable
This is a method that is more reliable than just looking at adjusted r sq
But first have to produce original model first
If you suspect OVB have to sq predicted values as new variable Re run original dep regression – with predicted value and sq value
Two outcomes
DON’T WANT SQUARED PREDICTED TERM TO BE SIGNIFICANT
When values are sq, relationships are not linear
When you run reg with predicted values and they are sig – there is a problem
with structure of model – variables causing model to violate assumption
This is an indication that you must readjust or fix kind of variables you have
in model
If its not doing anything, why is it in there
Basically to look at omitted variable bias – squared predicted values
shouldn’t be sig
Have sq predict value and unstand predicted value – 0.05 (not a problem)
Does not reach significance at alpha of .05
What if spv is sig
Adding variables
Enter variables that were omitted
What is an important predictor that might be missing
Is it causing multicolinearity
Outliers?
In project – have to be honest about what you did – what you thought was
the problem
Multicollinearity
Independent variables not detached from each other in social world
Problem is not whether x are correlated but whether its just too high
Pearson correlation matrix
Are your x variables correlated to each other
MC problem with x so closely related that you cant separate what is closely
related
If you have something super correlated that SPSS takes it out for you
Diagnostic to look at whether you should worry about ti
Cant figure out what effects are when cant untangle – PROBLEM
Interaction term is variable term
Lecture 4 looked at pearson correlation – see how

More
Less
Related notes for SOC350H5