Examination of Trust Items in UNICEF data

This provides further analysis of the trust-related items in the UNICEF survey. The survey asks about trust in the following types of institutions/professions/organizations, including:

  1. International news
  2. National news
  3. Friends
  4. Social media
  5. National government
  6. Doctors
  7. Scientists
  8. Religious organizations
  9. Police

All questions are asked on a 3-point scale, which has been recoded here in the following way: 1=No trust at all; 2=A little and 3=A lot.

We now consider the inter-correlation of these items. Due to the nature of the data, a polychoric correlation is used. While this is a different type of calculation from the more familiar Pearson’s correlation to account for the ordinal nature of the data, the interepretation is similar: a score closer to 1 indicates a strong (positive) relationship between any two question-items (and a score close to -1 indicates a close negative relationship), while a score close to 0 indicates no relationship.

This analysis looks at the polychoric correlations of all three items in the following ways:

  1. Results for all respondents (pooled across cohorts and countries)
  2. Results for 15-24 year olds ONLY (pooled across all countries)
  3. Results for 40+ year olds ONLY (pooled across all countries)

While the cross-country data is weighted, it is not projection weighted. This could be considered a limitation of the analysis – however the advantage of the current approach is that it allows for broadly equal representation across the countries.

Towards a Factor Analysis Solution: Why a general (non-cohort specific) seems appropriate

The correlation results are broadly similar, suggesting a general (rather than a cohort specific approach) may be appropriate for this analysis. The below analysis confirms this, showing that 4 factors are suggested when we look at the all respondent data, only 15-24 year olds and only 40+ year olds.

## Parallel analysis suggests that the number of factors =  4  and the number of components =  NA

## Parallel analysis suggests that the number of factors =  4  and the number of components =  NA

## Parallel analysis suggests that the number of factors =  4  and the number of components =  NA

Consequently, the factor analysis will focus on the cross-country, cross-cohort data.

We first apply Bartlett’s and KNMO test for correlation and sampling adequacy on the pooled data and find no issues for either test.

## $chisq
## [1] 41663.01
## 
## $p.value
## [1] 0
## 
## $df
## [1] 36
## Kaiser-Meyer-Olkin factor adequacy
## Call: KMO(r = poly_cor_all$rho)
## Overall MSA =  0.83
## MSA for each item = 
##  trust_international_news       trust_national_news             trust_friends 
##                      0.82                      0.83                      0.85 
##        trust_social_media trust_national_government             trust_doctors 
##                      0.84                      0.84                      0.83 
##          trust_scientists       trust_religious_org              trust_police 
##                      0.83                      0.83                      0.82

Factor Analysis Solution

We now move into the actual factor analysis. As can be seen, 4 factors are discovered. The first factor, titled “Media” contains both trust in international and national news.

The second factor (called Govt) measures trust in the major governmental/powerful institutions included in the battery of questions – the national government and police.

The third factor, called “Science” encapsulates trust in both doctors and scientists.

The fourth factor may seem to include the most diverse institutions – including religious organizations, friends and social media. However, we might consider this as a measure of ‘community’.

## 
## Loadings:
##                           ML1   ML2   ML3   ML4  
## trust_international_news  0.697 0.125 0.269 0.146
## trust_national_news       0.534 0.402 0.215 0.234
## trust_friends                         0.227 0.441
## trust_social_media        0.368             0.471
## trust_national_government 0.319 0.641 0.179 0.238
## trust_doctors             0.151 0.265 0.628 0.171
## trust_scientists          0.250 0.164 0.581 0.113
## trust_religious_org       0.109 0.240       0.477
## trust_police                    0.491 0.207 0.103
## 
##                  ML1   ML2   ML3   ML4
## SS loadings    1.110 0.993 0.982 0.829
## Proportion Var 0.123 0.110 0.109 0.092
## Cumulative Var 0.123 0.234 0.343 0.435

Creating Scores for Each Sub-Index

To create an index score, a simple weighted approach seems sensible (i.e. taking the average of each item for each factor). This is a reflection of the fact that the factor loading scores are somewhat similar.

##  [1] "WP5"                  "RESPONDENT_NUM_TOTAL" "iso3c"               
##  [4] "wbi3"                 "WP22140"              "WT"                  
##  [7] "WP1219"               "WP3117"               "factor_media"        
## [10] "factor_govt"          "factor_science"       "factor_community"
## `summarise()` ungrouping output (override with `.groups` argument)
## Warning: The `x` argument of `as_tibble.matrix()` must have unique column names if `.name_repair` is omitted as of tibble 2.0.0.
## Using compatibility `.name_repair`.
## This warning is displayed once every 8 hours.
## Call `lifecycle::last_warnings()` to see where this warning was generated.
Factor Cohort_15.24 Cohort_40plus
Media 2.285 2.211
Government 2.196 2.203
Science 2.522 2.492
Community 2.200 2.175 
## `summarise()` regrouping output by 'wbi3' (override with `.groups` argument)
wbi3 WP22140 Media Government Science Community
LILMC 15-24 2.307 2.251 2.499 2.393
LILMC 40+ 2.274 2.327 2.462 2.361
UMC 15-24 2.084 1.912 2.563 2.129
UMC 40+ 2.030 1.839 2.470 2.154
HIC 15-24 2.211 2.211 2.750 1.960
HIC 40+ 2.148 2.247 2.554 1.946

Next Steps

Country/cohort data will be provided (using the average of each factor, which will fall between 1 and 3).

For next step, we could recode the trust factors such that a score of nearly 3 is equal to 1 and all other scores (i.e. lower than nearly 3, so maybe below 2.5) is equal to 0 and run logit regressions for each factor. Any other thoughts?