Clyde Tucker and Brian Meekins U.S. Bureau of Labor Statistics and Paul Biemer

21
www.bls.gov Using Substantive Diagnostics to Evaluate the Validity of Micro-level Latent Class Indicators of Measurement Error Clyde Tucker and Brian Meekins U.S. Bureau of Labor Statistics and Paul Biemer Research Triangle Institute

description

Using Substantive Diagnostics to Evaluate the Validity of Micro-level Latent Class Indicators of Measurement Error. Clyde Tucker and Brian Meekins U.S. Bureau of Labor Statistics and Paul Biemer Research Triangle Institute. Background. - PowerPoint PPT Presentation

Transcript of Clyde Tucker and Brian Meekins U.S. Bureau of Labor Statistics and Paul Biemer

Page 1: Clyde Tucker and Brian Meekins U.S. Bureau of Labor Statistics and Paul Biemer

www.bls.gov

Using Substantive Diagnostics to Evaluate the Validity of Micro-level

Latent Class Indicators of Measurement Error

Clyde Tucker and Brian MeekinsU.S. Bureau of Labor Statistics

and Paul BiemerResearch Triangle Institute

Page 2: Clyde Tucker and Brian Meekins U.S. Bureau of Labor Statistics and Paul Biemer

Background Developed by Lazarsfeld (1950)—

unobserved or “latent” variable drawn from relationships between two or more “manifest” variables

Lazarsfeld and Henry (1968) and Goodman (1974) extended mathematics of theory

Software for latent class analysis (LCA) developed (MLLSA, lEM, M-PLUS)

LCA used to study measurement or response error (VandePol and deLeeuw 1986; Tucker 1992; Van de Pol and Langeheine 1997; Bassi et al. 2000; Biemer and Bushery 2000; Tucker, et al. 2002, 2003, 2004, 2005, 2006, and 2008)

Page 3: Clyde Tucker and Brian Meekins U.S. Bureau of Labor Statistics and Paul Biemer

Creation of Manifest Variables

Try to create at least three Try to avoid direct relationships

with outcome variable (expenditures, in this case)

Use LCA to “triangulate” them to produce a latent variable with more information than any one of them alone

Page 4: Clyde Tucker and Brian Meekins U.S. Bureau of Labor Statistics and Paul Biemer

Statistical Logic

In mathematical terms, when manifest variables A and B are not independent, the following relationship will not hold:

where i indexes the classes of A, j indexes the classes of B, πij

AB is the probability an individual is in cell ij, πi

A is the probability an individual is in class i, and πj

B is the probability an individual is in class j.

Bj

Ai

ABij

Page 5: Clyde Tucker and Brian Meekins U.S. Bureau of Labor Statistics and Paul Biemer

Statistical Logic For the above expression to be true, A and B must

be independent. The purpose of the latent variable X is to achieve that independence. Thus, the following latent class model is desired:

where t indexes the classes of X, πijtABX is the

probability of being in cell ijt of the unobserved ABX table, πt

X is the probability that an individual is in one of the mutually exclusive and exhaustive classes of X, πit

AX and πjtBX are the conditional

probabilities that an individual is in a particular class of A and B, respectively, given that a person is in a certain class of X. Equation (2) indicates that, within a class of X, A and B are independent.

BXjt

AXit

Xt

ABXijt

Page 6: Clyde Tucker and Brian Meekins U.S. Bureau of Labor Statistics and Paul Biemer

Purpose of Paper

Concept of LCA relatively straightforward—create a variable to account for common variance among observed variables

Issues: What is the new variable? What do its classes mean? Does it really tell us anything useful?

Statistical diagnostics don’t help us here. We need substantive ones.

Paper explores some of this type of diagnostics

Page 7: Clyde Tucker and Brian Meekins U.S. Bureau of Labor Statistics and Paul Biemer

Data Sources

CED 2 week diaries All expenditures Small items and grocery expenditures Used for CPI cost weights

CEQ 5 quarters (first for bounding) PV All consumer expenditures 2 hours Larger consumer items Used for CPI cost weights

Page 8: Clyde Tucker and Brian Meekins U.S. Bureau of Labor Statistics and Paul Biemer

Three Examples

1985 CED Operational Test (micro level) 3 treatments—specific, nonspecific, control 800 households in each Latent response error measure of underreporting of grocery

expenditures using manifest performance indicators CEQ (1996-2001) (micro level)

Only analyzed the 2nd wave 43,000 completed 2nd wave interviews Latent response error measure of underreporting for 7

expenditure categories for purchasers using manifest performance indicators

CEQ (1996-2001) (micro level) Analyzed all four waves 14,877 remained in sample throughout Latent response error measure of underreporting for almost

30 expenditure categories for all households (purchasers and nonpurchasers) using manifest performance indicators and indicators of pattern of wave nonresponse

Page 9: Clyde Tucker and Brian Meekins U.S. Bureau of Labor Statistics and Paul Biemer

Critical Assumption

Response errors in CE only come from underreporting of expenditures and not overreporting Tedious Time-consuming Recall problems Lack of knowledge

Page 10: Clyde Tucker and Brian Meekins U.S. Bureau of Labor Statistics and Paul Biemer

Methodological Issues

Weighted vs. unweighted Variances for complex sample design

vs. SRS Local vs. global maxima Sparse cells (too many manifest

variables) Restricted vs. unrestricted models Boundary problems (no overreporting)

Page 11: Clyde Tucker and Brian Meekins U.S. Bureau of Labor Statistics and Paul Biemer

1985 Diary Test

Manifest variables Difference in first and second week grocery

expenditures Difference in usual and average weekly grocery

expenditures Amount of expenditure information collected by

recall Respondent’s attitudes and behavior with

respect to diarykeeping Latent variable

3 classes (low, moderate, high response error)

Page 12: Clyde Tucker and Brian Meekins U.S. Bureau of Labor Statistics and Paul Biemer

CEQ Micro-level Manifest Indicators for First Study

Interview level indicators considered:1. Number of contacts2. Ratio of respondents/household members3. Missing income data4. Type and frequency of records used5. Length of interview6. Ratio of expenditures in last month to

quarter7. Combination of type of record and

interview length

Page 13: Clyde Tucker and Brian Meekins U.S. Bureau of Labor Statistics and Paul Biemer

Indicator Coding

#contacts (1=0-2; 2=3-5; 3=6+) Resp/hh size (1= <.5; 2= .5+) Income missing (1=present; 2=missing) Records use (1=never; 2=single type or

sometimes; 3=multiple types and always) Interview length (1= <45; 2=45-90; 3= 90+) Month3 expn/all (1= <.25; 2= .25-.5; 3= +.5) Combined records and length (1= poor; 2=

fair; 3=good)

Page 14: Clyde Tucker and Brian Meekins U.S. Bureau of Labor Statistics and Paul Biemer

Latent Variables

Three-class latent variables (poor, fair, good reporting) for Kid’s Clothing Women’s Clothing Men’s Clothing Furniture Electricity Minor Vehicle Repairs Kitchen Accessories

Page 15: Clyde Tucker and Brian Meekins U.S. Bureau of Labor Statistics and Paul Biemer

Second CEQ Micro-level Study

Based on results of first CEQ study, analysis of purchasers and nonpurchasers together

Used Interviews 2-5 data. Not limited to within-interview indicators Developed model using all Interview 2

respondents Latent variable is still intended to

represent quality of reporting

Page 16: Clyde Tucker and Brian Meekins U.S. Bureau of Labor Statistics and Paul Biemer

New Manifest Indicators

Overall Panel level indicators considered1. Number of completed interviews (1-4)2. Attrition combined with # of complete

interviews3. Average number of commodity categories for

which CU had expenditure4. Number of interviews the ratio of third month

expenditure to quarter was between .25 - .55. Panel averages of interview level indicators

from first CEQ study

Page 17: Clyde Tucker and Brian Meekins U.S. Bureau of Labor Statistics and Paul Biemer

Model Selection Ran both ordered (fixed or restricted ordinal

constraints) latent class models and unordered. Order was determined based on theoretical

relationship between values of indicators and level of underreporting.

Ran all combinations of indicators in groups of 3 & 4, using 3 or 4 category LC variable for each commodity category & overall

Multiple iterations to avoid local maxima Best model candidates were selected based on

fit From those candidates, models selected based

on relationship of indicators to latent construct

Page 18: Clyde Tucker and Brian Meekins U.S. Bureau of Labor Statistics and Paul Biemer

Application of Model For the final models for each commodity:

Each combination of indicators was assigned to a latent class based on probability of being in that class given the value of the indicators

Ran demographic analysis to identify characteristics of members of each latent class

Expenditure means were found for each latent class

Examined the pattern of mean expenditure and the contribution of the latent variable in predicting these expenditures

Page 19: Clyde Tucker and Brian Meekins U.S. Bureau of Labor Statistics and Paul Biemer

Second CEQ Micro-level Study– Expenditure Categories

Cable/satellite TV Men’s apparel Women’s apparel Men’s clothing only Women’s clothing only Men’s accessories Women’s accessories Men’s shoes Women’s shoes Kid’s apparel Kid’s clothing only Kid’s Accessories Kid’s shoes Dental care Drugs and medical supplies Electricity Gas (household) Eye care

Sports equipment Televisions, video, & sound

equip. Vehicle service, major Vehicle service, minor Vehicle service, oil changes only Vehicle expenses, other Pets and pet supplies Sports equipment Trash collection Televisions, video, & sound

equip. Vehicle service, major Vehicle service, minor Vehicle service, oil changes only Vehicle expenses, other Pets and pet supplies Kitchen accessories Other household items

Page 20: Clyde Tucker and Brian Meekins U.S. Bureau of Labor Statistics and Paul Biemer

Conclusions

When doing LCA for measuring response error, one cannot rely on statistical diagnostics alone. Substantive diagnostics are needed to judge the meaningfulness of the results.

Sometimes the models work and sometimes they don’t. Unfortunately, this is likely to depend on the characteristic you’re analyzing.

We need better manifest variables to explain more variance. We have been unable to develop meaningful latent variables

with more than three or four categories, and, in some cases, we could only identify two. LCA software really does work best with large sample sizes.

Besides only defining a few latent classes, we certainly will not progress beyond the most rudimentary ordinal rankings any time soon.

LCA problems are likely to be multiplied many times for response error measures for non-factual items such as attitudes or opinions.

Page 21: Clyde Tucker and Brian Meekins U.S. Bureau of Labor Statistics and Paul Biemer

Contact Information

www.bls.gov

Clyde TuckerSenior Survey Methodologist

OSMR202-691-7371

[email protected]