adjusting samples for nonresponse bias: pros and cons of surveying nonrespondents compared with...
TRANSCRIPT
Adjusting samples for nonresponse bias: pros and cons of surveying nonrespondents compared with
other approaches in ESS
Jaak Billiet: CeSO - K.U. Leuven Hideko Matsuo: CeSO – K.U. Leuven
The European Social Survey Round 4 launching conference ‘Poland and Europe: continuation and change’.
Institute of Philosophy and Sociology Polish Academy of Sciences, Warsaw 13 Jan 2010
Outline
Very short introduction
Short overview: Approaches to the assessment of bias applied in ESS (Billiet, Matsuo, Beullens & Vehovar, Research & Methods. ASK, vol 18 (1, 2009), pp. 3-43).
The surveys among nonrespondents (post R3): what, how, when, with what…?
Main results of NRS in PL and NO (comparison of results and focus on adjusting samples)
Pros and cons of NRS compared with other 3 approaches
1. Introduction Analysis of nr bias still needed:
WHY? Still large differences in NR rates based on CF R4
Short overview of approaches to the assessment of bias applied in ESS (Billiet, Matsuo, Beullens & Vehovar, Research & Methods. ASK. vol 18 (1, 2009), pp. 3-43).
In all rounds (R1, R2, R3, R4…..)
1. Bias as deviation between obtained sample and population (or ‘Golden standard’ survey) = post-stratification and evaluations of samples before and after weighting
2. Bias as difference between cooperative and converted refusals collected via refusal conversion = comparison of cooperative with reluctant respondents (converted refusals)
3. Bias as difference in ‘observable’ data among all sampling units (collected in contact forms)= sample based comparison between all respondents and all nonrespondents
2. Short overview (1)
In context of R3
4. Bias as difference between respondents and non-respondents collected via post hoc nonresponse survey= surveys among nonrespondents after R3 in PL, NO and CH (real NRS)
in BE (at moment of refusal only among refusals = Doorstep Questions Survey)
Short overview (2)
3. Survey among nonrespondents (1)
New survey among refusals with very small & easy questionnaire (some crucial variables) (Voogt, 2004; Saris)
Implemented in ESS Round 3 : 4 participating countries- Full mail survey (15 questions) months after main survey in NO (medium rr), CH (low rr) & PL (high rr)
- At moment of refusal 7 crucial questions in BE (7 questions)
Response rates
BE (44.7% = 303) response among refusals
NO (30.3% = 342) response among noncontacts & refusals
PL (23.2% = 192)
CH (52.9% = 771)
(cooperative much higher response)
Survey among nonrespondents (2)
1. The questions asked
Key questions procedure (Pedaksi approach)
Short 7 question module (+ at door): work situation, highest level of education, # of members in household, frequency of social activities, feeling (un)safe, interest in politics, attitude towards surveys
Normal 16 questions module: same as short + gender, year of birth, TV watching, voluntary work, trust in people, satisfied with democracy, trust in politics, immigration good/worse for country, (+ reasons for refusal (closed) in one subgroup)
Target group
Timing Mode Use of Incen-tives
Type of questio-naaire
Response
Rates (R/NR)
Sample size
BE ESS3
refusers
Same as ESS
PAPI at door
NO 1 short DQS_R: 44.7%
303
CH ESS3_R
& ESS3_NR
After ESS
Mail/
Web/
CATI
10 Swiss FR.
2 short & long
NRS3_R: 84%
NRS3_NR: 51.8%
1023
NO ESS3_R
& ESS3_NR
After ESS
Mail/
Web/
CATI
NO 1 long NRS3_R: 60.79%
NRS3_NR: 30.25%
487
PL ESS3_R
& ESS3_NR
After ESS
Mail Notepad 2 short & long+
NRS3_R: 59.04%
NRS3_NR: 23.24%
1208
Survey among nonrespondents (3)
2. Overview of the sample design
Survey among nonrespondents (4)
3. Kinds of respondents in NRS decisions to take in view of computing propensity scores for weighting the sample
NRS/(cooperative vs. nrs)
NRS/(cooperative vs. main)
(NRS+reluctant) vs (cooperative (nrs or main?))
NRS/cooperative vs reluctant/cooperative
see Figure next slide
ESS Cooperative Respondent(ESS3_Rco)
ESS Reluctant Respondent(ESS3_Rrel)
ESS Non-Respondent (ESS3_NR)
Kinds of respondents in data analyses [NO, CH & PL]
NRS Cooperative Respondent(NRS3_Rco)
NRS Reluctant Respondent(NRS3_Rrel)
ESS
NRSNRS Non-
Respondent(NRS3_NR)
Survey among nonrespondents (5)
Method used for adjusting the sample for nonresponse bias
1. Identify survey response differences on key explanatory variables between types of respondent (‘nonrespondent vs. cooperative respondent’).
2. Study net effects of key explanatory variables on response probabilities via logistic regression model (dependent variable: prob ratio’s ‘nonrespondent/cooperative’).
3. Obtain propensity scores on all cases on non-response probabilities via logistic regression model (dependent variable: prob ratio’s ‘cooperative/nonrespondent’).
log ( ) / (1 ( )) ' ( )e e f x x x
Survey among nonrespondents (6)
4. Transform propensity scores into weights via stratification
method (Rosenbaum & Rubin 1984; Little 1986; Lee &
Vaillant 2008):
Form 10 strata with equal number of cases after sorting on ps;
Assign each sample unit into correct corresponding sub-strata
Weight = expected probability/observed probability of the
coop. respondent (or nonrespondent) in the corresponding
sub-strata. 5.
Survey among nonrespondents (7)
5. Evaluate effects of propensity weighting via two main
criteria:
1. Tests between unweighted & weighted sample on cooperative
respondents (NRS3_Rco & ESS3_Rco).
1b. In case of significant differences: test differences between
parameters of relevant substantive explanatory models
2. Study differences in distributions on key questions between
types of respondents (NRS3_Rco vs NRS3_NR or ESS3_Rco
vs. NRS3_NR).
Survey among nonrespondents (7)
4. Main results in NO and PL (1)1. Differences between ESS cooperative and NRS !nonrespondents*
1.
* Only single ESS cooperative respondents (not ‘double’ respondents). All tests: ESS resp = expected freq
!
Main results in NO and PL (2)
…differences in distributions
!
Main results in NO and PL (3)
2. Logistic regression parameters nonresp/cooperative
*NRS res are final ESS nonrespondents
NO: NRS res / ESS res (226 vs. 1616)
PL: NRSres / ESS res ( 156 vs. 1434 )
Odds ratio SE Odds ratio SE
Educational level (numeric: 0-8) (NO) Educational level (PL)
ISCED 2 ISCED 3 & 4 ISCED 5 & 6
Reference: ISCED 0 & 1
0.851** 0.045
0.894 1.365*
1.772**
0.146 0.134 0.184
Work status employed
Reference: unemployed
0.742**
0.091
Neighbourhood security after darkness Safe
Unsafe & very unsafe Ref: very safe
1.335* 1.344*
0.129 0.143
Involved in charity organization
(numeric: 0-6: >= once a week – never)
0.996
0.047
Main results in NO and PL (4)
(continued) Logistic regression parameters nonresp/cooperative
NO: NRS res / ESS res (226 vs. 1616)
PL: NRS res / ESS res ( 156 vs. 1434 )
Odds ratio SE Odds ratio
Participation in social activities Less than most
About the same Much more than most
Ref: much less than most
0.829 0.887
0.520**
0.156 0.118 0.194
0.548***
1.287
0.130 0.158
TV watching time per day (numeric: 0-7: no time – 3+hrs per day)
1.072
0.042
Political interest Hardly interested
Not at all interested Ref: very & quite interested
0.936 1.334
0.107 0.154
0.895
0.087
How satisfied with democracy works (numeric 0-10: ex. dissatisfied - ex.satisfied)
0.916*
0.037
Immig. make country worse/better place to live (numeric 0-10: worse place – better place)
0.916*
0.036
R²=0.052; H & L= 9.060
R²=0.042; H & L=3.835
***p-value<0.0001; **p-value<0.01; p < 0.05; H&L stands for Hosmer and Lemeshow
Main results in NO and PL (5)
?
Main net effects on probability ratio coop resp / nonresp (inversed parameters!) In Norway: probability of response INCREASES if • Higher educated• Participate more in social activities then most (subjective…)• More satisfied with democracy• Positive attitude towards ‘consequences’ of immigration
In Poland: probability of nonresponse INCREASES if• Higher educated!!!• Unemployed• Feel safe• Participate less in social activities than most!• (political interested?!!)
Main results in NO and PL (6)
3. Evaluation of the propensity weights
First approach A: is the adjusted sample (weighted) of cooperative ESS respondents significant different from the original sample?
if yes: we may conclude that the adjustment had effect on the sample estimates
conclusion: no significant differences at all example: variable with largest differences = education
Main results in NO and PL (7)
Differences between original sample and adjusted sample even smaller in PL
Not necessary to test a substantive regression model since the univariate distributions do not differ (first approach B)
This is nonetheless checked for model with “consequences of immigration” as relevant dependent variable” and number of predictor variables: age, TV watching, involvement in charity org, trust in politics, social trust, and two value orientations (conservation, self-transcendence)
R² = 0,26 in both models (not weighted & weighted)
all predictors contribute significantly to variance of dept. var
BUT: no differences at all between the two models
Conclusion = was ps weighting useless? Let us see the second approach
Main results in NO and PL (8)
Second approach: do the initial significant differences of belonging to a response category of all key questions between ESS respondents and nonrespondents (NRS res) in first table disappear after adjusting the sample of ESS cooperative respondents?
in other words, did we move from NMAR to MAR
let us see:
Main results in NO and PL (9)
Norway sample (Chisq values or t-values; p-values)
largely successful: all differences disappeared except political interest
(2) NRS res. vs ESS res.
Unweighted prob. Weighted prob.
Education level (df=2) 40.552 <.0001 1.813 0.404
Work status (df=1) 11.594 0.0007 0.470 0.493
Political interest (df=3) 33.014 <.0001 13.247 0.010
Participation in social activities (df=4) 48.105 <.0001 2.301 0.681
T-value prob. T-value prob.
How satisfied democracy (df=1864) 5.31 <.0001 1.01 0.313
Imm. make country worse/better place (df=1871) 5.17 <.0001 0.56 0.577
TV watching time per day (df=1878) -4.38 <.0001 -0.81 0.420
Involved in work for voluntary & charity org.
(df=1885)
-2.69 0.007 -0.69 0.492
* Only key questions with significant differences in distribution (p<.05) in unweighted sample (Table 1) are shown.
Main results in NO and PL (10)
Poland: sample (Chisq values and p-values)
Not completely successful since still sign differences between NRS and ESS for two variables (political interest and social participation)
(2) NRS res. vs ESS res.
Unweighted prob. Weighted prob.
Educational level (df = 3) 14.481 0.002 0.838 0.840
Work status (df = 1) 4.658 0.031 0.340 0.560
Neighbourhood security (AESFDRK) (df=3) 14.295 0.003 2.174 0.537
Political interest (POLINTR) (df = 3) 22.141 <.0001 26.759 <.0001
Social participation (SCLACT) (df = 4) 60.838 <.0001 28.690 <.0001 * Only key questions with significant differences in distribution (p<.05) in unweighted sample (Table 7) are shown. ** Chi² is computed for partial cross-tables: (1) NRS respondent vs double respondent and (2) NRS respondent vs ESS respondent.
5. Pros and cons of NRS compared with 3 other approaches
Approach to study of nonresponse bias in cross-nation research Post-
stratification Comparison coop – reluct
Info from observ data
NRS
Possible in ESS yes yes yes yes
Adjustment of all samples
yes +/- yes +/-
Weakness 1 Small effect Small effect but info on all
variables
Small effect Small effect but info on more relevant vars
Weakness 2 No info on target vars
Differences btw countries
Measurement error in obs
NRS resp differ from ESS resp
Problem Error in Gold standard
(education…)
Final refusals differ from
converted ref.
Only small # of vars with
low cov
mode effects possible
Additional information population Some refusals All samp units Selection of #
How to adjust Ps or prop weights
Not useful Prop weights but low effect
Kinds of prop. weights
Additional costs Low Rather high Rather low Rather high
Conclusions
Future:
- Other methods (contacting sequences using contact forms data) = expect low effect (result of some studies, see Blom)
- More model based method: crucial is what additional information can be used
- Combining different methods
- Info in all methods = view on sensitive variables
- Finally: low effect may mean LOW BIAS