Você está na página 1de 7

The Craft of Statistical Analysis Webinar

Series: A free program of The Analysis


Factor

Approaches to Dealing with Missing Data:


The Good, the Bad, and the Unthinkable
Karen Grace-Martin

Missing Data Structure


V1 V2

Observed

Copyright 2013 The Analysis Factor


www.TheAnalysisFactor.com

Missing

The Craft of Statistical Analysis Webinar


Series: A free program of The Analysis
Factor

Analysis Strategies

Listwise Deletion
Imputation
Multiple Imputation
Full Information Maximum Likelihood

Decision Factors
Missing Data Mechanism
Missing Completely at Random (MCAR)
Missing at Random (MAR)
Non-Ignorable (NI)

Percentage and Distribution of Missing Data


How much?
How is missing data distributed?

What analysis will you be using?

Copyright 2013 The Analysis Factor


www.TheAnalysisFactor.com

The Craft of Statistical Analysis Webinar


Series: A free program of The Analysis
Factor

Listwise Deletion
V1 V2

V1 V2

Listwise Deletion
Works when:
Data are MCAR and loss of power is tolerable
Data are MAR and percentage missing is small
Missing Data is concentrated on a few variables

Copyright 2013 The Analysis Factor


www.TheAnalysisFactor.com

The Craft of Statistical Analysis Webinar


Series: A free program of The Analysis
Factor

Imputation
V1 V2

V1 V2

Example: Imputing Means


100

80

60

40

INCOME

20

EDUCAT
INC_IMP
EDUCAT

0
6

Copyright 2013 The Analysis Factor


www.TheAnalysisFactor.com

10

12

14

16

18

20

22

The Craft of Statistical Analysis Webinar


Series: A free program of The Analysis
Factor

Imputation
Works when:

Data are MAR


Percentage missing is very small
Missing Data is spread across variables
Imputation method is based on other variables
P-values are not involved

Multiple Imputation
V1 V2

V1 V2

V1 V2
+

Copyright 2013 The Analysis Factor


www.TheAnalysisFactor.com

V1 V2
+

The Craft of Statistical Analysis Webinar


Series: A free program of The Analysis
Factor

Multiple Imputation
Works when:
Data are MAR
Any percentage is missing
Missing Data is concentrated or spread across
variables
Imputation method is based on other variables, and
other variables are good predictors

Full Information
Maximum Likelihood
V1 V2

V1
=

V1 V2
x

L(Parameters|Data)
= L(Parameters|Complete Variables) x
L(Parameters|Complete Cases)

Copyright 2013 The Analysis Factor


www.TheAnalysisFactor.com

The Craft of Statistical Analysis Webinar


Series: A free program of The Analysis
Factor

Full Information Maximum Likelihood


Works when:
Data are MAR
Any percentage is missing
Missing Data is concentrated or spread across
variables
Model is linear or log-linear

More Information
Collection of resources on missing data:
http://www.TheAnalysisFactor.com/missing-data/

Copyright 2013 The Analysis Factor


www.TheAnalysisFactor.com

Você também pode gostar