Presentation is loading. Please wait.

Presentation is loading. Please wait.

Adjust Survey Response Distributions Using Multiple Imputation: A Simulation with External Validation Frank Liu & Yu-Sung Su.

Similar presentations


Presentation on theme: "Adjust Survey Response Distributions Using Multiple Imputation: A Simulation with External Validation Frank Liu & Yu-Sung Su."— Presentation transcript:

1 Adjust Survey Response Distributions Using Multiple Imputation: A Simulation with External Validation Frank Liu & Yu-Sung Su

2 We Want a Better Guess of Respondents' Preferences.

3 How to Utilize Multiple Imputation to make better guesses about the respondent’s preferences?

4 Calculating proportions based on raw data and omitting the non-response data result in biased proportion of interested variables (Bernhagen & Marsh, 2007). Multiple Imputation (MI) is a cost-efficient and methodological sound approach for better use of raw survey and poll data (Barzi, 2004).

5 Picture Source: Kyle F. Edwards, Christopher A. Klausmeier, and Elena Litchman (2011)

6 MI Modeling vs. MI for DV

7 While most studies use MI for modeling, there is room to examine to which MI can be applied to electoral forecast (King et al., 2001; Snijders & Bosker, 2011).

8 Joint MI (Amelia II) vs. Conditional MI (mi & MICE)

9 Joint MI assumes a multivariate normal distribution. vs. Conditional MI does NOT.

10 mi takes advantage of existing regression models to handle various kinds of variables types (Su, Gelman, Hill, & Yajima, 2011): using a logistical regression model to predict a binary outcome, Using an ordered logit regression model to predict an ordinal outcome, Using a multinomial logit regression model to predict an unordered categorical outcome.

11 mi results can be very biased when data include extreme values (He and Raghunathan, 2009)

12 Data RDD telephone survey data about Taiwanese voters’ national identification (N=1,078), collected from Jan. 23 to Feb. 4, 2013. (AAPOR #3: 21.56%)

13 Sensitive Question (DV) and Item Non-Response VariablesQuestion Wording DistributionMissing (%) VariablesQuestion Wording DistributionMissing (%) camp (v33) Do you support for any political party? Blue camp: 215 Green camp: 204 658 (61.04)

14

15 Auxiliary Variables (1) v Wording v6Do you agree that we can influence the government with our votes? v7Do you agree that we have little influence on what the government plans to do? v9Do you agree that we should use Taiwanese as the major language in Taiwan? v10Do you agree that Taiwanese children perform better than those in mainland China? v17Do you agree that those identifying with Taiwan can be titled as Taiwanese? v18Do you agree that Chinese from mainland China have more money than sense? v20Do you agree that those people should not be called Taiwanese if they don't know Matsu(name of a sea goddess widely worshipped on the SE China coast and in SE Asia)?

16 v22Do you agree that our government should give more restrict policy on mainland China tourists? v27Some call themselves Taiwanese, some Chinese, and some Both, how about you? v28Do you agree that “Taiwan” is the formal name of our country? v29Do you agree that Taiwanese people and those in mainland China belongs to a same nation? v32Do you agree that we should seek for unification with mainland China if it becomes a democracy? v37Do you agree that the two sides of Taiwan Straight will eventually becomes one country? v38Do you ever go to mainland China in the past two years? ageWhich year were you born? (re-coded to real age) eduWhat is your education level? sex(coded by interviewer)

17 Study 1: Check Missingness Patterns Using conditional MI algorithm to check its missingness patterns and to conduct MI. the data should be at least missing at random (MAR), i.e., the missingness mechanism does not depend on the unobserved data. MCAR > MAR > MNAR

18

19 Check the Missingness Patterns 1: Specify a conditional MI model include many valid auxiliary variables (17); 2: Contrast Simulated Data with MI Data Set the imputed datasets as the baseline for comparison create three copies of the MI data randomly remove values from the completely imputed data according to the original missing rate of the data, i.e., 61%.

20 We create simulated datasets by modeling the missingness of each variable conditional on a linear combination of the rest of variables with logistic regressions. Then we use the predictive missingness to create missing values on the three imputed data. In short, we compare simulated datasets with the original MI dataset.

21 Results of the Check for MCAR (1)

22 Results of the Check for MCAR (2) MeanSEBlueGreen Original Data0.4880.50051.1%48.8% Imputed Original Data0.474 + 0.500 + 52.6%47.4% #1 Imputed MCAR Data0.464 + 0.499 + 53.5%46.5% #2 Imputed MCAR Data0.464 + 0.500 + 53.6%46.4% #3 Imputed MCAR Data0.453 + 0.500 + 54.7%45.3% Table 2 : Summary of the camp variable between the original data and imputed MCAR datasets. Note: + The mean’s and SE’s reported here are pooled mean and SE’s for three chains of MI.

23 Results of the Check for MCAR (3) Figure 3: Plots of Camp Variable against Other Variable Using the Imputed MCAR data.

24

25

26 Results of the Check for MAR (1)

27 Results of the Check for MAR (2) Mean*SE*BlueGreen Original Data0.4880.50051.1%48.8% Imputed Original Data0.474+0.500+52.6%47.4% #1 Imputed MAR Data0.477+0.500+50.1%49.9% #2 Imputed MAR Data0.497+0.500+50.3%49.7% #3 Imputed MAR Data0.499+0.500+50.1%49.9% Table 3 : Summary of the camp variable between the original data and imputed MAR datasets. Note: + The mean’s and SE’s reported here are pooled mean and SE’s for three chains of MI.

28 Results of the Check for MAR (3)

29

30

31 Summary: the assumption of MAR holds

32 Study 2: External Validation 1: compare respondents' answers with MI guesses figure out how well the MI prediction works 2: understand why prediction performs not so well, if this is the case. follow-up telephone surveys with identity check Face-to-face one-by-one interviews for explanations

33 Follow-up telephone surveys 1.Called out for the 658 respondents’ camp identification whose political camp id is NA. (April 13-15, 2013) Forced to choose between Green and Blue. N=143 2.Identity check Answers must be consistent with the first survey regarding two questions: (1) whether ever going to mainland China in the past two years; (2) the frequency of watching political news

34 3. In-depth Interview: 45 out of 143 respondents were contacted for face-to- face one-by-one Interview. (incentive: cash $70) 5 out of 45 respondents were interviewed between April 20 to May 6, 2013.

35 Findings IDMICamppred.correct 50.161B1 80.400B1 90.804G1 150.728G1 240.348B1 250.259B1 360.300B1 440.574G1 580.587B0 590.730G1 690.382B1 the total number of the 145 respondents whose values fall between.45 and.55: 19

36 Rate of Successful Prediction using MI: 70% (100/145).

37 Case #1 IDSexAgeEduDate- Time PlaceMICamp ID Causes of Inconsistency 905F3432013.4.27 10:00 Taipei Main Station, Taipei City 0.42 (B) GDisappointed by the KMT's reform on domestic policies. Feeling lost some reasons to keep supporting KMT.

38 Case #2 IDSexAgeEduDate- Time PlaceMICamp ID Causes of Inconsistency 206M43213.4.30 10:00 NSYSU campus, Kaohsiung City 0.50 (B) BDisappointed by both political camps but felt more concerned about DPP than KMT regarding DPP's ideology of seeking Taiwan independence.

39 Case #3 IDSexAgeEduDate- Time PlaceMICamp ID Causes of Inconsistency 140M29313.4.30 14:00 Kaohsiung City 0.54 (G) BGrowing up with KMT supporter mother and have been ok with KMT. Turning to like DPP for a growing Taiwanese national identification.

40 Case #4 IDSexAgeEduDate- Time PlaceMICamp ID Causes of Inconsistency 384M25413.5.4 14:00 Taipei Main Station, Taipei City 0.47 (B) GFeeling cross- pressured because parents support for KMT but friends support for DPP. The first vote is for DPP in 2008 presidential election. Feeling disappointed by KMT's performance but not aware the core ideology of DPP.

41 Case #5 IDSexAgeEduDate- Time PlaceMICamp ID Causes of Inconsistency 286M3732013.5.4 16:00 Taipei Main Station, Taipei City 0.40 (B) GFeeling cross- pressured because his family have been supporting for KMT but wife's family support for DPP. Feeling disappoint about KMT's leadership.

42 Summary 1.These “danglers” are politically aware. 2.Explanations for the ambivalence: 1.potential blue camp supporters (ID 905, 206, 384, and 286) have become unsatisfied with the incumbent's performance and public policy. 2.potential green camp supporters (ID 905, 206, 384, and 286) are affected by nationalism. 3. Cross-pressured (ID 140, 384, and 286).

43 Conclusion 1.MI scores reflect respondents’ partisan orientation, including the level of their ambivalence. It seems reasonable to adopt this method to reconstruct the distribution of partisan orientation of the electorate. 2.By face-to-face interviews inconsistency of their answers can be explained.

44 Implications & Suggestions Encourage respondents to give up answering those they could not answer Explore auxiliary questions for partisanship (e.g., policy and performance ones) and encourage respondents to answer sincerely that seem less sensitive to them. Use these easier auxiliary variables for MI.

45 Continue to test the reliability of this method using panel data. Be aware of the limit of forcing respondents to choose an answer. Alternative methods or question wording of probing true answers about partisanship are still needed.

46 Thank you. request for the paper, data, and/or R code frankcsliu@gmail.com


Download ppt "Adjust Survey Response Distributions Using Multiple Imputation: A Simulation with External Validation Frank Liu & Yu-Sung Su."

Similar presentations


Ads by Google