Recent PMA surveys from Côte d’Ivoire and Nigeria include questions about the abortion experiences of both respondents and other women of reproductive age with whom they share personal information.
In our last post, we demonstrated that there are multiple ways to estimate abortion incidence with data provided by respondents to the 2018 PMA surveys from Côte d’Ivoire and Nigeria. With help from an interactive shiny dashboard, we saw that estimates may vary depending on how the researcher incorporates responses to one or both of both of these survey questions:
PMA surveys include multiple ways of asking about a woman’s experiences with abortion as a way of mitigating social desirability bias inherent in this sensitive topic. In this post, we’ll explore an additional set of questions that can help researchers assess the validity of incidence measures like those we’ve discussed.
Prior to the questions related to her own abortion experiences, each woman in the 2018 PMA surveys from Côte d’Ivoire and Nigeria are asked to consider other women aged 15-49 with whom they share personal information. If the respondent indicates that she knows at least one such person, the questionnaire goes on to ask about that woman’s age, educational background, and experiences with pregnancy removal and period regulation.
These questions mirror those that are later posed to the respondent about her own abortion history, so we can construct the same incidence measures shown in our last post with data from each woman’s closest confidante.1 Comparing these measures can help give us a sense of the extent to which social social desirability bias or other factors limit the information respondents share about their own experiences. However, we should note that the confidante data have their own limitations:
Both of these issues are addressed by Bell et al. (2020a; 2020b), whose work we highlighted in our last post. In both publications, they report separate incidence estimates from data derived from respondents and from information provided about their confidantes. Overall, they find higher abortion incidence rates using the confidante data compared with the respondent data, ultimately suggesting that the real-world target may be closer to the former than the latter.
We’ll follow their steps again in this post, focusing on the 2018 sample from Nigeria. Notably: these data are available in longitudinal format combined with data from a 2020 follow-up. To avoid confusing the numeric suffix assigned to variables from each round of the survey and an additional numeric suffix assigned to each of the respondent’s closest confidantes, we’ll simply download a cross-sectional extract for the 2018 survey (we’ll revisit the longitudinal data in an upcoming post).
We’ve downloaded a cross-sectional extract for the 2018 Nigeria sample (female respondents only) that includes parallel sets of variables for respondents and their confidantes, plus a few that are only available for respondents. As a reference, here is a table showing the corresponding variable names for each group (where available).
| Measure | Respondent Version | Confidante Version (Friend 1) |
|---|---|---|
| Age | AGE | AGEFRND1 |
| Highest level of school attended | EDUCATTGEN | EDUCATTFRND1 |
| Ever terminated a pregnancy | ABOREV | ABOREV_FRND1 |
| Year of last terminated pregnancy | ABORYR | ABORYR_FRND1 |
| Used multiple termination methods | ABORMULT | ABORMULT_FRND1 |
| Only termination method used | ABORONLYMETH | ABORONLYMETH_FRND1 |
| First of multiple termination methods used | ABORFIRSTMETH | ABORFIRSTMETH_FRND1 |
| Last of multiple termination methods used | ABORLASTMETH | ABORLASTMETH_FRND1 |
| Sought termination care at a facility | ABORCARE | ABORCARE_FRND1 |
| Place where termination medication was obtained (if first / only method) | ABORFIRSTMEDLOC | ABORFIRSTMEDLOC_FRND1 |
| Place where termination medication was obtained (if last method) | ABORLASTMEDLOC | ABORLASTMEDLOC_FRND1 |
| Place where termination surgery was obtained (if first / only method) | ABORFIRSTSURGLOC | ABORFIRSTSURGLOC_FRND1 |
| Place where termination surgery was obtained (if last method) | ABORLASTSURGLOC | ABORLASTSURGLOC_FRND1 |
| Ever regulated a period (if never terminated a pregnancy) | REGPREGEV | REGPREGEVFRND1 |
| Ever regulated a period (besides terminated a pregnancy) | REGPREGEV_ABOR | REGPREGEV_ABORFRND1 |
| Year of last regulated period | REGYR | REGYRFRND1 |
| Used multiple regulation methods | REGMULT | REGMULTFRND1 |
| Only regulation method used | REGMETH | REGMETHFRND1 |
| First of multiple regulation methods used | REG1ST | REG1STFRND1 |
| Last of multiple regulation methods used | REGLAST | REGLASTFRND1 |
| Sought regulation care at a facility | REGTREAT | REGTREATFRND1 |
| Place where regulation medication was obtained (if first / only method) | REGMEDLOC | REGMEDLOCFRND1 |
| Place where regulation medication was obtained (if last method) | REGLASTMEDLOC | REGLASTMEDLOCFRND1 |
| Place where regulation surgery was obtained (if first / only method) | REGSURGLOC | REGSURGLOCFRND1 |
| Place where regulation surgery was obtained (if last method) | REGLASTSURGLOC | REGLASTSURGLOCFRND1 |
| Number of close female friends aged 15-49 | FRIENDNUM | – |
| Household wealth quintile | WEALTHQ | – |
| Marital status | MARSTAT | – |
| Religion of household head | RELIGION | – |
| Ethnicity | ETHNICITYNG | – |
| Parity (live birth events) | BIRTHEVENT | – |
| Urban or rural residence | URBAN | – |
| Nigerian State | GEONG | – |
| Female Questionnaire Sampling Weight | FQWEIGHT | – |
| Sampling cluster (enumeration area) | EAID | – |
| Sampling strata (Nigerian state) | STRATA | – |
After you’ve downloaded an extract containing these variables, load it together with the following packages in R.
Our analysis will focus only on members of the de facto population, identifiable as all cases where FQWEIGHT is not 0.
We’ll also create a short ID for each woman by row index.
ng <- ng %>% rowid_to_column("id")Next, we’ll build an analytic dataset dat with recoded variables matching the categories shown throughout Bell et al. (2020a; 2020b). First, we’ll handle the variables for each respondent (largely following the workflow shown in our last post).
dat <- ng %>%
mutate(
.keep = "none",
id,
age = AGE %>%
case_match(
15:19 ~ "15-19", 20:24 ~ "20-24", 25:29 ~ "25-29", 30:34 ~ "30-34",
35:39 ~ "35-39", 40:44 ~ "40-44", 45:50 ~ "45+"
) %>%
fct_relevel("15-19","20-24","25-29","30-34","35-39", "40-44"),
edu = EDUCATTGEN %>%
case_match(
1 ~ "Never", 2 ~ "Primary", 3 ~ "Secondary", 4 ~ "Higher"
) %>%
fct_relevel("Never", "Primary", "Secondary"),
wealth = WEALTHQ %>%
as_factor() %>%
str_remove(" quintile") %>%
fct_relevel("Lowest", "Lower", "Middle", "Higher"),
marstat = MARSTAT %>%
case_match(
10 ~ "Never Married", 21:22 ~ "Partnered", 31:32 ~ "Separated / Widowed"
) %>%
fct_relevel("Never Married", "Partnered", "Separated / Widowed"),
relig = RELIGION %>%
case_match(
210 ~ "Catholic", 290 ~ "Other Christian", 100 ~ "Muslim",
.default = "Other"
) %>%
fct_relevel("Catholic", "Other Christian", "Muslim"),
ethnicity = ETHNICITYNG %>%
case_match(
5 ~ "Hausa", 6 ~ "Igbo", 15 ~ "Yoruba",
.default = "Other"
) %>%
fct_relevel("Hausa", "Igbo", "Yoruba"),
parity = BIRTHEVENT %>%
case_match(
0 ~ "0", 1:2 ~ "1-2", 3:4 ~ "3-4", 5:90 ~ "5+"
) %>%
fct_relevel("0", "1-2", "3-4", "5+"),
urban = URBAN %>% as_factor()
)
dat # A tibble: 11,106 × 9
id age edu wealth marstat relig ethnicity parity urban
<int> <fct> <fct> <fct> <fct> <fct> <fct> <fct> <fct>
1 1 35-39 Never Lowest Partnered Other Christian Other 0 Rural
2 2 35-39 Never Lowest Partnered Muslim Hausa 3-4 Rural
3 3 15-19 Secondary Lower Partnered Catholic Other 0 Rural
4 4 40-44 Never Lowest Partnered Muslim Hausa 5+ Rural
5 5 30-34 Never Lowest Partnered Other Christian Other 1-2 Rural
6 6 45+ Never Lowest Partnered Muslim Hausa 5+ Rural
7 7 30-34 Higher Higher Partnered Other Christian Other 1-2 Urban
8 8 25-29 Never Lower Partnered Muslim Hausa 3-4 Rural
9 9 35-39 Higher Highest Partnered Other Christian Igbo 5+ Urban
10 10 15-19 Secondary Middle Separated / Widowed Other Christian Other 0 Urban
# … with 11,096 more rowsAs in our last post, we’ll create four separate incidence measures from the following recoded variables:
term indicating whether the respondent terminated a pregnancy after January 1, 2017reg indicating whether the respondent regulated a period after January 1, 2017any indicating whether the respondent terminated a pregnancy or regulated a period after January 1, 2017avg the mean value of term and any, as reported by Bell et al. (2020a; 2020b)yrs - the number of months between January 1, 2017 and the date of a woman’s interview divided by 12 (e.g. 1.5 represents 18 months)dat <- ng %>%
mutate(
.keep = "none",
id,
term = ABORYR %in% 2017:2018,
reg = REGYR %in% 2017:2018,
any = reg | term,
avg = pick(term, any) %>% rowMeans,
yrs = (INTFQCMC - 1405)/12 # Jan 2017 as CMC = 1405
) %>%
print() %>% # preview columns to be joined with `dat`
full_join(dat, by = "id")# A tibble: 11,106 × 6
id term reg any avg yrs
<int> <lgl> <lgl> <lgl> <dbl> <dbl>
1 1 FALSE FALSE FALSE 0 1.25
2 2 FALSE FALSE FALSE 0 1.25
3 3 FALSE TRUE TRUE 0.5 1.25
4 4 FALSE FALSE FALSE 0 1.25
5 5 FALSE FALSE FALSE 0 1.25
6 6 FALSE FALSE FALSE 0 1.33
7 7 FALSE TRUE TRUE 0.5 1.25
8 8 FALSE FALSE FALSE 0 1.25
9 9 FALSE FALSE FALSE 0 1.33
10 10 FALSE FALSE FALSE 0 1.25
# … with 11,096 more rowsFinally, we’ll attach simpmle numeric versions of a few technical variables:
dat <- ng %>%
mutate(
.keep = "none",
id,
weight = FQWEIGHT %>% zap_labels,
eaid = EAID %>% zap_labels,
strata = STRATA %>% zap_labels,
) %>%
print() %>% # preview columns to be joined with `dat`
full_join(dat, by = "id")# A tibble: 11,106 × 4
id weight eaid strata
<int> <dbl> <dbl> <int>
1 1 0.220 113 56604
2 2 0.109 134 56604
3 3 0.527 616 56607
4 4 0.109 134 56604
5 5 1.37 384 56609
6 6 0.759 124 56604
7 7 2.26 356 56609
8 8 1.64 496 56605
9 9 0.913 155 56606
10 10 0.959 556 56608
# … with 11,096 more rowsAs shown in the table above, IPUMS PMA uses the shorthand “Friend 1” to reference variables pertaining to the respondent’s closest confidante. PMA surveys collect a limited amount of demographic information about Friend 1: we’ll use her age and education level.
dat <- ng %>%
mutate(
.keep = "none",
id,
age_f1 = AGEFRND1 %>%
case_match(
15:19 ~ "15-19",
20:24 ~ "20-24",
25:29 ~ "25-29",
30:34 ~ "30-34",
35:39 ~ "35-39",
40:44 ~ "40-44",
45:50 ~ "45+"
) %>%
fct_relevel("15-19","20-24","25-29","30-34","35-39", "40-44"),
edu_f1 = EDUCATTFRND1 %>%
case_match(
100 ~ "Never",
200 ~ "Primary",
400 ~ "Secondary",
600 ~ "Higher"
) %>%
fct_relevel("Never", "Primary", "Secondary")
) %>%
print() %>% # preview columns to be joined with `dat`
full_join(dat, by = "id")# A tibble: 11,106 × 3
id age_f1 edu_f1
<int> <fct> <fct>
1 1 40-44 Never
2 2 <NA> <NA>
3 3 <NA> <NA>
4 4 <NA> <NA>
5 5 <NA> <NA>
6 6 <NA> <NA>
7 7 30-34 Higher
8 8 25-29 Never
9 9 35-39 Higher
10 10 <NA> <NA>
# … with 11,096 more rowsNotice that many of the values in our modified variables age_f1 and edu_f1 are NA? These values are produced by case_match for any original values that are not explicitly specified: here, NA covers codes for “don’t know”, “no response or missing”, and “NIU (not in universe)”. Cases marked NIU represent respondents who reported that they had no confidante women aged 15-49, so these questions were skipped.
We’ll create a helper variable has_f1 to mark women who indicated that they did know at least one such person.
# A tibble: 2 × 2
has_f1 n
<lgl> <int>
1 FALSE 5223
2 TRUE 5883As we’ve mentioned, respondents could report their knowledge about the abortion experiences of Friend 1 with different degrees of certainty. For example, the question about pregnancy termination for Friend 1 looks like this:
712a.i. Now I want to ask some more questions about {friend1_name}. Has she ever
done something to remove a pregnancy when she was pregnant or worried she was
pregnant?
Probe to confirm whether the pregnancy removal was successful. If not, select
'no.'
[] Yes, I am certain
[] Yes, I think so
[] No
[] Do not know
[] No responseBell et al. (2020a; 2020b) use all of the “Yes, I am certain” responses to estimate abortion incidence, and they also use cases where the respondent answered “Yes, I think so” only if they could specify at least one method for the procedure. This information can be found in multiple variables depending on the number of methods that were ultimately needed to terminate the pregnancy.
ng <- ng %>%
mutate(
termmethod_f1 = if_any(
c(ABORONLYMETH_FRND1, ABORFIRSTMETH_FRND1, ABORLASTMETH_FRND1),
~.x < 97 & .x != 4
),
regmethod_f1 = if_any(
c(REGMETHFRND1, REG1STFRND1, REGLASTFRND1),
~.x < 97 & .x != 4
),
termev_f1 = case_when(
has_f1 ~ ABOREV_FRND1 == 2 | {ABOREV_FRND1 == 1 & termmethod_f1}
),
regev_f1 = case_when(
has_f1 ~ REGPREGEVFRND1 == 2 | {REGPREGEVFRND1 == 1 & regmethod_f1}
)
)Finally, we’ll construct the same four abortion measures we made from the respondent data. In this case, we’ll use termev_f1 and regev_f1 to mark some cases FALSE where the respondent was uncertain and could not identify an abortion method for Friend 1.
term_f1 indicating whether Friend 1 terminated a pregnancy after January 1, 2017reg_f1 indicating whether Friend 1 regulated a period after January 1, 2017any_f1 indicating whether Friend 1 terminated a pregnancy or regulated a period after January 1, 2017avg_f1 - the mean value of aboryr_f1 and anyyr_f1, as reported by Bell et al. (2020a; 2020b)dat <- ng %>%
mutate(
.keep = "none",
id, has_f1,
term_f1 = if_else(termev_f1, ABORYR_FRND1 %in% 2017:2018, FALSE),
reg_f1 = if_else(regev_f1, REGYRFRND1 %in% 2017:2018, FALSE),
any_f1 = term_f1 | reg_f1,
avg_f1 = pick(term_f1, any_f1) %>% rowMeans,
) %>%
print() %>% # preview columns to be joined with `dat`
full_join(dat, by = "id")# A tibble: 11,106 × 6
id has_f1 term_f1 reg_f1 any_f1 avg_f1
<int> <lgl> <lgl> <lgl> <lgl> <dbl>
1 1 TRUE FALSE FALSE FALSE 0
2 2 FALSE NA NA NA NA
3 3 FALSE NA NA NA NA
4 4 FALSE NA NA NA NA
5 5 FALSE NA NA NA NA
6 6 FALSE NA NA NA NA
7 7 TRUE FALSE FALSE FALSE 0
8 8 TRUE FALSE FALSE FALSE 0
9 9 TRUE FALSE FALSE FALSE 0
10 10 FALSE NA NA NA NA
# … with 11,096 more rowsWe could now proceed to calculate annualized abortion incidence with the each of the measures we’ve constructed for respondents and confidantes. As a reminder, our previous post demonstrated that incidence rates more than doubled if we included period regulation or pregnancy termination (41.9 women per 1,000) compared with incidence constructed from pregnancy termination alone (19.8 women per 1,000). Bell et al. (2020a; 2020b) report the mean of these estimates (30.8 women per 1,000), represented by the variable avg.
dat %>%
as_survey_design(weight = weight, id = eaid, strata = strata, nest = TRUE) %>%
summarise(across(
c(term, reg, any, avg),
~pick(everything()) %>%
summarise(1000 * survey_mean(.x / yrs, vartype = "ci"))
)) %>%
pivot_longer(everything()) %>%
unnest(value)# A tibble: 4 × 4
name coef `_low` `_upp`
<chr> <dbl> <dbl> <dbl>
1 term 19.8 15.3 24.4
2 reg 25.2 18.0 32.5
3 any 41.9 33.4 50.4
4 avg 30.8 24.9 36.8We can use the same calculation for Friend 1 where information about Friend 1 was reported. However, we’ll have to include na.rm = TRUE to ignore cases where this information is missing.
dat %>%
as_survey_design(weight = weight, id = eaid, strata = strata, nest = TRUE) %>%
summarise(across(
c(term_f1, reg_f1, any_f1, avg_f1),
~pick(everything()) %>%
summarise(1000 * survey_mean(.x / yrs, vartype = "ci", na.rm = TRUE, proportion = TRUE))
)) %>%
pivot_longer(everything()) %>%
unnest(value)# A tibble: 4 × 4
name coef `_low` `_upp`
<chr> <dbl> <dbl> <dbl>
1 term_f1 40.2 33.4 48.3
2 reg_f1 21.3 15.7 28.8
3 any_f1 59.6 50.5 70.1
4 avg_f1 49.9 42.3 58.8At least initially, it seems that each of these four incidence measures is higher than those derived from data about the respondents’ own experiences! This is true for every measure except if we estimate incidence with period regulation alone - in that case, the respondent incidence is higher! This suggests that respondents may be more willing to report their own abortion experiences in circumstances where their pregnancy status was unknown or ambiguous, as discussed in Bell et al. (2021).
Above, we mentioned that an important limitation on the data from Friend 1 is that only around half of the respondents to the Nigeria 2018 sample identified such a person.
To help correct for selection bias in the Friend 1 data, Bell et al. (2020a; 2020b) replace the missing NA values for unreported confidantes with probable values derived from information about the confidantes we know about.
For example, let’s model the probability that a known confidante removed a pregnancy after January 1, 2017 (setting aside period regulation for now). Predictors could include any of the respondent characteristics, but not the age or education of Friend 1 (those factors may be very predictive, but they aren’t available for the missing confidante values we want to impute).
Moreover, we won’t be able to impute any values if the respondent has missing NA values for any of the covariates in our model. First, we’ll specify a Poisson model with svyglm; then, we’ll use that model for all cases without missing NA values for any of the model covariates.
mod_term <- dat %>%
as_survey_design(weight = weight, id = eaid, strata = strata, nest = TRUE) %>%
svyglm(
term_f1 ~ age + edu + wealth + marstat + relig + ethnicity + parity + urban,
design= .,
family = "poisson"
)
result <- dat %>%
filter(!if_any(
c(age, edu, wealth, marstat, relig, ethnicity, parity, urban),
~is.na(.x)
)) %>%
mutate(predicted = predict(mod_term, newdata = pick(everything())) %>% exp) Let’s compare the predicted values with those in term_f1 for the first 20 rows in the resulting data frame.
# A tibble: 11,084 × 3
predicted has_f1 term_f1
<svystat> <lgl> <lgl>
1 0.021639958 TRUE FALSE
2 0.016593557 FALSE NA
3 0.059412618 FALSE NA
4 0.012548495 FALSE NA
5 0.031222636 FALSE NA
6 0.006188911 FALSE NA
7 0.068066845 TRUE FALSE
8 0.019658020 TRUE FALSE
9 0.023480061 TRUE FALSE
10 0.096823637 FALSE NA
11 0.021164499 TRUE FALSE
12 0.087638625 FALSE NA
13 0.046756815 FALSE NA
14 0.162033942 TRUE FALSE
15 0.044651003 TRUE FALSE
16 0.033077578 FALSE NA
17 0.066055775 TRUE TRUE
18 0.035565384 FALSE NA
19 0.055831264 TRUE FALSE
20 0.068178847 FALSE NA
# … with 11,064 more rowsYou can see in has_f1 that only 9 of the 20 respondents above were able to identify someone as Friend 1; in term_f1 you see that, of those 9, only one (row 17) indicated that her Friend 1 had removed a pregnancy after January 1, 2017.
The NA values in term_f1 appear for respondents who identified no person as Friend 1. That’s where the predicted values come in: we’ll now substitute the predicted value in place of NA values. Because these values are numeric, we’ll coerce the existing values in term_f1 to 0 (for FALSE) and 1 (for TRUE).
result <- result %>%
mutate(final_f1 = if_else(
is.na(term_f1),
as.double(predicted),
as.double(term_f1)
))
result %>%
select(predicted, has_f1, term_f1, final_f1) %>%
print(n = 20)# A tibble: 11,084 × 4
predicted has_f1 term_f1 final_f1
<svystat> <lgl> <lgl> <dbl>
1 0.021639958 TRUE FALSE 0
2 0.016593557 FALSE NA 0.0166
3 0.059412618 FALSE NA 0.0594
4 0.012548495 FALSE NA 0.0125
5 0.031222636 FALSE NA 0.0312
6 0.006188911 FALSE NA 0.00619
7 0.068066845 TRUE FALSE 0
8 0.019658020 TRUE FALSE 0
9 0.023480061 TRUE FALSE 0
10 0.096823637 FALSE NA 0.0968
11 0.021164499 TRUE FALSE 0
12 0.087638625 FALSE NA 0.0876
13 0.046756815 FALSE NA 0.0468
14 0.162033942 TRUE FALSE 0
15 0.044651003 TRUE FALSE 0
16 0.033077578 FALSE NA 0.0331
17 0.066055775 TRUE TRUE 1
18 0.035565384 FALSE NA 0.0356
19 0.055831264 TRUE FALSE 0
20 0.068178847 FALSE NA 0.0682
# … with 11,064 more rowsWith these substitutions in place, we can now calculate an adjusted confidante incidence estimate similar to the estimate shown in Bell et al. (2020a; 2020b)2
result %>%
as_survey_design(weight = weight, id = eaid, strata = strata, nest = TRUE) %>%
summarise(1000 * survey_mean(final_f1 / yrs, vartype = "ci", proportion = TRUE))# A tibble: 1 × 3
coef `_low` `_upp`
<dbl> <dbl> <dbl>
1 38.3 34.4 42.7Compared with our previous estimate (40.2), we now find 2 fewer women per 1,000 (38.3) with a pregnancy termination within one year. Here’s a way to repeat that estimation with all four abortion measures via across:
dat %>%
summarise(across(
c(term_f1, reg_f1, any_f1, avg_f1),
function(y){
model <- pick(everything()) %>%
as_survey_design(weight = weight, id = eaid, strata = strata, nest = TRUE) %>%
svyglm(
y ~ age + edu + wealth + marstat + relig + ethnicity + parity + urban,
design = .,
family = "poisson"
)
dat %>%
filter(!if_any(
c(age, edu, wealth, marstat, relig, ethnicity, parity, urban),
~is.na(.x)
)) %>%
mutate(
predicted = predict(model, newdata = pick(everything())) %>% exp,
z = if_else(is.na(y), as.double(predicted), as.double(y))
) %>%
as_survey_design(weight = weight, id = eaid, strata = strata, nest = TRUE) %>%
summarise(1000 * survey_mean(z / yrs, vartype = "ci", proportion = TRUE))
}
)) %>%
pivot_longer(everything()) %>%
unnest(value)# A tibble: 4 × 4
name coef `_low` `_upp`
<chr> <dbl> <dbl> <dbl>
1 term_f1 38.3 34.4 42.7
2 reg_f1 21.0 17.7 24.8
3 any_f1 57.0 51.7 62.9
4 avg_f1 47.7 43.2 52.6Overall, whether our estimates include adjustments for missing confidantes or not, we see that the estimated annualized abortion incidence derived from confidantes is higher than the incidence estimates derived from respondents. This holds true whether we use pregnancy termination responses alone, or if we combine them with responses for period regulation; on the other hand, when our estimate includes responses for period regulation only, the respondent data matches or exceeds the incidence derived from confidantes! In light of these findings, it’s clear why abortion researchers benefit from consulting multiple measures like those included in PMA surveys: social desirability bias and other related factors likely influence respondent reporting on their own experiences with pregnancy removal.
Respondents were invited to provide this information for multiple close confidantes, as indicated by a numeric suffix attached to the variable name for each. For example, ABOREV_FRND1 references the woman’s closest confidante, while , ABOREV_FRND2 references the confidante she listed second. In this post, we’ll focus only on the responses provided about the confidante listed first.↩︎
Bell et al. (2020a; 2020b) further adjust confidante data with post-stratification weights modeled on respondent values. The survey package provides tools for this in rake and calibrate↩︎
If you see mistakes or want to suggest changes, please create an issue on the source repository.