Lecture 11_Measurement Error Flashcards
Measurement Error
failure of the recorded responses to reflect the true characteristics of the respondents
Measurement error variance
Variance across respondents- social desirability- men diff answers than women-but ind person makes same errors
(y_i=μ+ε_i)
Variance across trials -diff answer each time for same survey
(y_it=μ+ε_it)
Measurement error bias
When response deviation is non zero over conceptual replications
(y_i=μ+ε_i)
Ways to Study Measurement error (5)
Repeated Measures Lab investigations -non exp and exp Interpenetration/random ass. To diff methods/modes record checks; external sources collecting correlates of error
Repeated measures TYPES (4)
reinterview survey with subsample of survey sample
single int with repeated measures
combo of survey and admin rec’s
multiple rec’s collected by panel survey that refer to diff time pds
Lab investigations - Non experimental
Expanded interviews (Think aloud, Comprehension probes, Debriefing)
Methods to target source of error (vignettes, response timing)
Confidence measuresBehavior coding
Issues with lab investigations
Attention level of Respondents
Motivation in response process
Type of participants
Reality
random ass. To diff methods/modes
Split ballot; split sample within surveys
Difference between T groups used to est biasMost focus on bias not Measurement error variance
Inference is to the sampling frame
Assumptions of lab investigations
Reality
Heterogeneity of errors represented
Interpenetration definition
Randomly assign subsamples to different production units (interviewers)
Many diff treatments are applied and compared- more common to focus on variance
All survey attributes are replicated for diff int. Estimation of variance due to production units (interviewers)
Interviewer variance
Interpenetration model
y_ij=χ_i+M_ij+ε_ij
χ_i:true
M_ij:effect of jth method
Issues with interpenetration/Interviewer variance models
Changes survey conditions
Logistical issues -(Attrition, Refusal conversion ass. Of interviewers)
interpreted to mean that reduction in assigned cases (m) leads to lower int effects
Record checks; external sources Assumptions
Most interested in bias NOT measurement error variance
Record is assumed true value
Record checks; external sources TYPES
Reverse record check
Forward record check
*Full design record check
Reverse record check
Sample people with known characteristic covered in record
Measures failure to report on something E.g., abortions, crime incidents
Record check model
y_i=R_i+ε_ι
R: record value for reporter
Issues with Reverse record check
Cant measure overreporting (only a subset of events sampled)
Forward record check
Samples reporters from a frame; then compares survey answers to recorded value Measures overreports (when survey report not found in record)E.g., medical events
Issues with forward record check
If no report is given; no record search conducted
*Full design record check
Requires full population covered by one or more record systems; full access
All events captured
Asking about variables related to the events that are available for population of interest
Measure both under and over reporting E.g., official divorce records
Issues with record checks
Assumes all records are complete
Linking errors
Using Benchmarks
Universe data available to compute the same statistic as the sample
e.g., voting
correlated response variance
interviewer variance
answers from one respondent are correlated with answers interviewed by the same interviewer
simple response variance; A.K.A. reliability in psychometrics
respondent vary in their response deviations over conceptual trials anyway- even without interviewer effects
Index of consistency
proportion of the element variance of y that’s due to the variance of the response deviations
Var(ε)/Var(y)
Measure response variance
Y_i1=χ_i+ε_i1
Y_i2=χ_i+ε_i2
Variance in errors
E=(ε_i1-ε_i2)
psychometric “reliability”
p_y=1-index of consistency
Assumptions to measure Simple Response Variance
No change in true values, over X trials (threatened by long time between trials)
No memory effect of first trial on second (threatened by short time between trials)
Survey conditions the same between trials
(threatened by using supervisors to conduct int’s)
Parallel measures
survey items that measure the same underlying construct with different response deviations
Both measures are asked of each respondent as indicators of same underlying attribute
Correlation should be high
Generalizing parallel measures/multiple indicators
if multiple constructs measured and different methods are used for each- then separate out variance into
True values
Method effects
random measurement error variance
Challenges of using multiple indicators
Context effects- similar questions
estimation issues with discrete variables
“just-identified” models
Paradata in surveys
data collected as a byproduct of field activities
- response latency
- keystroke data
- probng behavior
- edit failure records