Load and process all data

library(dplyr)
library(tidyr)
library(stringr)
library(yaml)
library(purrr)
library(pander)
library(ggplot2)
library(ggrepel)
library(scales)
library(lubridate)
library(plotly)

panderOptions('table.split.table', Inf)
panderOptions('table.split.cells', Inf)
panderOptions('keep.line.breaks', TRUE)
panderOptions('table.style', 'multiline')
panderOptions('table.alignment.default', 'left')

Dead addresses, domains, and bounces

# Load data from tracking database
db.email <- src_sqlite(path=file.path(PROJHOME, "Data", "Survey", "list", 
                                      "final_list.db"))
email.full <- tbl(db.email, "full_list") %>% collect() %>%
  separate(id_org, c("db", "id.in.db"))
removed <- tbl(db.email, "remove") %>% collect()
bounced.raw <- tbl(db.email, "bounces") %>% collect()
completed <- tbl(db.email, "survey_completed") %>% collect()
sending.groups <- tbl(db.email, "groups") %>% collect()

email.by.db <- email.full %>%
  group_by(db) %>%
  summarise(num.apparently.valid = n())

dead.searches <- paste(c("Hippo", "Invalid", "Dead", "weird opt"), collapse="|")
dead <- removed %>%
  filter(str_detect(remove_notes, regex(dead.searches, ignore_case=TRUE))) %>%
  select(fk_org, notes = remove_notes)

bounced <- bounced.raw %>%
  gather(notes, value, hard_bounce, email_dead, email_bounce) %>%
  filter(value != 0) %>%
  select(fk_org, notes)
  
dead.and.bounced <- bind_rows(dead, bounced) %>%
  group_by(fk_org) %>%
  slice(1) %>%  # Get rid of duplicate removal entries
  ungroup() %>%
  left_join(select(email.full, fk_org = index_org, db), by="fk_org")

dead.and.bounced.by.db <- dead.and.bounced %>%
  group_by(db) %>%
  summarise(num.dead.bounced = n())

# Load full survey data (minus the Q4\* loop for simplicity)
survey.orgs.all <- readRDS(file.path(PROJHOME, "Data", "data_processed", 
                                     "survey_orgs_all.rds"))

# Load cleaned, country-based survey data (*with* the Q4\* loop)
survey.clean.all <- readRDS(file.path(PROJHOME, "Data", "data_processed", 
                                      "survey_clean_all.rds"))

# Load cleaned, organization-based data (without the Q4 loop)
survey.orgs.clean <- readRDS(file.path(PROJHOME, "Data", "data_processed", 
                                       "survey_orgs_clean.rds"))

# Load cleaned, country-based data (only the Q4 loop)
survey.countries.clean <- readRDS(file.path(PROJHOME, "Data", "data_processed", 
                                            "survey_countries_clean.rds"))

# Load valid partial responses
survey.partials <- readRDS(file.path(PROJHOME, "Data", "data_processed", 
                                     "survey_partials.rds"))

Summarize response rates from each database

# Load YAML metadata for survey lists
raw.lists <- yaml.load_file(file.path(PROJHOME, "data", "data_raw",
                                      "NGO lists", "ngo_lists.yml"),
                            as.named.list=TRUE)

# Convert to nice dataframe with purrr::map_df()
list.details <- seq(1:length(raw.lists$lists)) %>%
  map_df(function(x) raw.lists$lists[[x]][c("title", 'name',
                                            "num_rows_raw", "description")]) %>%
  arrange(desc(num_rows_raw))

response.summary <- list.details %>%
  left_join(email.by.db, by=c("name" = "db")) %>%
  left_join(dead.and.bounced.by.db, by=c("name" = "db")) %>%
  mutate(num.invited = num.apparently.valid - num.dead.bounced) %>%
  select(-c(description))

response.summary.total <- response.summary %>%
  summarise_each(funs(sum), -title, -name) %>%
  mutate(title = "**Total**")

response.summary.with.total <- bind_rows(response.summary,
                                         response.summary.total) %>%
  select(-name) %>%
  mutate(perc.valid = num.apparently.valid / num_rows_raw,
         perc.bounced.from.valid = num.dead.bounced / num.apparently.valid,
         perc.invited.from.raw = num.invited / num_rows_raw,
         perc.invited.from.valid = num.invited / num.apparently.valid)

response.summary.display <- response.summary.with.total %>%
  mutate_each(funs(comma), starts_with("num")) %>%
  mutate_each(funs(percent), starts_with("perc"))

Full technical details of how I ran the survey are available at my research notebook.

The complete database of NGOs to receive a survey invitation came from 5 different sources. After collecting the details of each organization listed at each source, I cleaned the raw lists by removing all organizations without valid e-mail addresses and by attempting to filter out obviously domestic NGOs. I filtered out domestic NGOs either by not collecting them in the first place (in the case of the Yearbook of International Organizations), or using information from the database to identify them. For example, the UN iCSCO database includes a field for an organization’s geographic scope: local, national, regional, and international. I omitted local and national.

I filtered out invalid e-mail addresses using Email Hippo, which pings each address to verify (1) that the domain exists, and (2) that the address exists at the domain.

pandoc.table(response.summary.display)
title num_rows_raw num.apparently.valid num.dead.bounced num.invited perc.valid perc.bounced.from.valid perc.invited.from.raw perc.invited.from.valid
UN Integrated Civil Society Organizations System (iCSO) 27,028 7,498 1,101 6,397 27.7% 14.7% 23.7% 85.3%
Directory of Development Organizations 14,834 14,540 7,457 7,083 98.0% 51.3% 47.7% 48.7%
Yearbook of International Organizations (YBIO) 9,325 3,065 322 2,743 32.9% 10.5% 29.4% 89.5%
Global Anti-Human Trafficking (TIP) NGOs 1,421 1,063 93 970 74.8% 8.7% 68.3% 91.3%
Arab Institute of Human Rights NGO directory 761 606 209 397 79.6% 34.5% 52.2% 65.5%
Total 53,369 26,772 9,182 17,590 50.2% 34.3% 33.0% 65.7%

Some databases were more responsive than others (though this is hardly accurate; half of the responses aren’t linked to specific databases):

survey.dbs <- survey.orgs.clean %>%
  group_by(database) %>%
  summarise(num.responses = n()) %>% 
  ungroup()

response.summary.actual <- survey.dbs %>%
  left_join(response.summary, by=c("database"="name")) %>%
  mutate(pct.responded.from.invited = num.responses / num.invited,
         pct.responded.clean = percent(pct.responded.from.invited)) %>%
  mutate(database = ifelse(database == "unknown", "zzzunknown", database),
         title = ifelse(is.na(title), "Unknown", title)) %>%
  arrange(database) %>%
  select(title, num.responses, num.invited, pct.responded.clean)
pandoc.table(response.summary.actual)
title num.responses num.invited pct.responded.clean
Arab Institute of Human Rights NGO directory 4 397 1.01%
Directory of Development Organizations 59 7083 0.83%
UN Integrated Civil Society Organizations System (iCSO) 321 6397 5.02%
Global Anti-Human Trafficking (TIP) NGOs 30 970 3.09%
Yearbook of International Organizations (YBIO) 80 2743 2.92%
Unknown 147 NA NA%

Figure out partials, incompletes, completes

Determine the best cutoff point for partially answered questions based on the number of questions answered.

What was the minimum number of questions answered by an INGO that finished the survey?

complete.ingos <- survey.orgs.all %>%
  filter(Finished == 1, Q2.4 == "Yes") %>%
  # Count of Q* questions answered
  mutate(num.answered = rowSums(!is.na(select(., starts_with("Q")))))

min(complete.ingos$num.answered)
## [1] 27

Thus, my rough cut-off point for partials = 20. But with Q3.3 and Q3.8, missing values are converted to “Not applicable”, so even if a respondent didn’t ever see it, they’ll get a recorded response for that question. This means that there are 11 ghost responses that have to be ignored (5 in Q3.3, 6 in Q3.8), so the technical cut-off point is 31.

However, some respondents quit before answering any questions about the countries they work in. I count any respondent that answered more than six questions in the loop of country questions. I use six because of how the Q4 variables are generated and cleaned. If an organization answered Q4.1 (the country name), the script converted it to COW and ISO codes, resulting in 3 valid Q4.1 questions. Additionally, the script converts text-based Q4 questions into characters and will sometimes yield NULL instead of NA, which then gets counted in the number of questions (so it’s possible for a respondent to answer just the country name and have that count as 6 questions). Also, Q4.16 creates 6 ghost questions, so the limit is technically 12 (6 + 6).

So, I use a combination of factors to determine partiality. A respondent has to answer at least 20 (31) questions, and at least 6 have to come from the Q4 loop. This is a better, more robust cutoff than simply using a 20-question minimum arbitrarily.

Thus, there are this many valid complete and partial responses:

nrow(survey.orgs.clean)
## [1] 641
table(survey.orgs.clean$complete)
## 
## FALSE  TRUE 
##   104   537

Survey meta-metrics

Absorption rate

The absorption rate [measures] the ability of the survey company to manage and keep up-to-date their database of email addresses and communications with panel members (Callegaro and DiSogra 2008, 1026).

\[ \frac{EI - BB - NET}{EI} \]

EI <- response.summary.total$num.apparently.valid
BB.NET <- response.summary.total$num.dead.bounced

absorption.rate <- (EI - BB.NET) / EI
  • EI = e-mail invitations sent: 26,772
  • BB = bounced: 9,182
  • NET = network undeliverable: included in BB
  • Absorption rate: 65.7%

Break-off rate

The break-off rate is a possible indicator of problems in the design of the questionnaire (e.g., too long, boring…) or struggle with technical problems during the survey administration (e.g., streaming media or animations that may “break” a survey at some point) (Callegaro and DiSogra 2008, 1026).

\[ \frac{BO}{I + P + BO} \]

# Only consider the organizations that were not screened out
survey.orgs.ingos <- survey.orgs.all %>%
  filter(Q2.4 == "Yes")

BO <- survey.orgs.ingos %>%
  filter(!(ResponseID %in% unique(c(survey.partials$ResponseID,
                                    survey.clean.all$ResponseID)))) %>%
  select(ResponseID) %>% unique() %>% nrow() %>% unlist()

I.survey <- survey.clean.all %>%
  filter(Finished == 1) %>% select(ResponseID) %>% 
  unique() %>% nrow() %>% unlist()

P.survey <- length(unique(survey.partials$ResponseID))

break.off.rate <- BO / (I.survey + P.survey + BO)
  • BO = number of surveys broken off (i.e. incomplete and not partial): 309
  • I = complete: 537
  • P = partial: 104
  • Break-off rate: 32.5%

Completion rate (participation rate)

The most intuitive response metric is the survey’s completion rate. It is also the one metric most often mislabeled as a response rate. The completion rate is the proportion of those who completed the Web survey among all the eligible panel members who were invited to take the survey (Callegaro and DiSogra 2008, 1021–22).

Using such a rate as an indicator of possible nonresponse error makes little sense; however, the participation rate may serve as a useful indicator of panel efficiency (The American Association for Public Opinion Research 2016, 49).

\[ \frac{I + P}{(I + P) + (R + NC + O)} \]

# Number who refused (i.e. explicitly did not give consent)
R.survey <- survey.orgs.all %>%
  select(Q6.1) %>%
  filter(!is.na(Q6.1)) %>%
  nrow() %>% unlist()

Non-contact is impossible to determine, since I don’t know how many organizations self-screened without taking the survey or e-mailing me. So, this participation rate is not accurate, but no participation rate ever is.

NC <- response.summary.total$num.invited - nrow(survey.orgs.all)

participation.rate <- (I.survey + P.survey) / 
  (I.survey + P.survey + BO + R.survey + NC)
  • I = complete: 537
  • P = partial: 104
  • R = refusal and break-off: 309 (break-off) and 2 (refusal)
  • NC = non-contact: 15,757
  • O = other: None
  • Participation rate: 3.84%

Study-specific screening completion rates

Study-specific screening completion rates and eligibility rates measure the incidence of a particular phenomenon among panel members. When these rates are significantly different from an external “gold standard,” they may indicate issues of question wording in the screener module or respondents purposively self-selecting themselves for a particular study (e.g., to gain rewards) even if they do not really qualify.… These rates may also reveal a skew in the panel membership along a particular dimension that may raise concerns regarding bias (Callegaro and DiSogra 2008, 1026).

\[ \frac{SCQ + SCNQ}{INV} \]

SCQ <- survey.orgs.all %>%
  filter(Q2.4 == "Yes") %>%
  select(ResponseID) %>%
  unique() %>% nrow() %>% unlist()

SCNQ <- survey.orgs.all %>%
  filter(Q2.4 == "No") %>%
  select(ResponseID) %>%
  unique() %>% nrow() %>% unlist()

INV <- response.summary.total$num.invited

screening.completion.rate <- (SCQ + SCNQ) / INV
  • SCQ = screening completed and qualified: 950
  • SCNQ = screening completed and not qualified (i.e. screened out): 482
  • INV = survey invitations sent out: 17,590
  • Study-specific screening completion rate: 8.14%

Study-specific eligibility rate

The problem with a screening rate is that nonresponse is confounded with the screening. In fact, we do not know if a person qualifies unless they provide that information by answering the screening questions. For this reason, we talk about screening completion rate and not screening rate (Callegaro and DiSogra 2008, 1023).

\[ \frac{SCQ}{SCQ + SCNQ} \]

study.eligibility.rate <- SCQ / (SCQ + SCNQ)
  • SCQ = screening completed and qualified: 950
  • SCNQ = screening completed and not qualified (i.e. screened out): 482
  • Study-specific eligibility rate: 66.3%

Other details

In addition to these rates, we also believe that it is the best practice to report the length of the field period with its start and close dates, the number of reminders sent and their form (email, letter, IVR call, or personal call), and the use of any incentive (Callegaro and DiSogra 2008, 1026).

Timeline of e-mail invitations

invited.groups.summary <- email.full %>%
  filter(!(index_org %in% dead.and.bounced$fk_org)) %>%
  mutate(id_group = as.integer(group)) %>%
  group_by(id_group) %>%
  summarise(num.in.group = n())

sending.groups.summary <- sending.groups %>%
  left_join(invited.groups.summary, by="id_group") %>%
  # Because I stupidly didn't include a final reminder column, I put the
  # timestamp of the final reminder in the notes column. This extracts the
  # timestamp with a regex.
  mutate(email_final_reminder = 
           str_extract(notes, "\\d{4}-\\d{2}-\\d{2} \\d{2}:\\d{2}:\\d{2}")) %>%
  mutate_each(funs(ymd_hms), starts_with("email")) %>%
  gather(email_type, email_date, starts_with("email")) %>%
  # Make Group 9's reminder be the final reminder
  mutate(email_type = ifelse(email_type == "email_reminder" & id_group == 9,
                             "email_final_reminder", email_type))


make_range <- function(x) {
  if (length(x) == 1) {
    return(paste("Group", as.character(x)))
  } else {
    return(paste0("Groups ", min(x), "-", max(x)))
  }
}

sending.groups.plot <- sending.groups.summary %>%
  mutate(email_day = ceiling_date(email_date, unit="day")) %>%
  filter(!is.na(email_day)) %>%
  group_by(email_day, email_type) %>%
  summarise(emails_sent = sum(num.in.group),
            group_names = make_range(id_group)) %>%
  ungroup() %>%
  mutate(total = cumsum(emails_sent),
         email_type = factor(email_type, 
                             levels=c("email_invitation", "email_reminder", 
                                      "email_final_reminder"),
                             label=c("Invitation  ", "Reminder  ", "Final reminder"),
                             ordered=TRUE))

plot.timeline <- ggplot(sending.groups.plot, aes(x=email_day, y=total)) +
  geom_step(size=0.5, colour="grey50") + 
  scale_y_continuous(labels=comma) +
  scale_x_datetime(date_labels="%B %e", date_breaks="1 week") +
  guides(fill=FALSE, colour=guide_legend(title=NULL)) +
  labs(x=NULL, y="Approximate total number of emails") +
  theme_light()

plot.timeline.static <- plot.timeline + 
  geom_point(aes(color=email_type)) + 
  geom_label_repel(aes(label=group_names, fill=email_type),
                   size=2.5, colour="white") +
  theme(legend.position="bottom", 
        legend.key.size=unit(0.65, "lines"),
        legend.key=element_blank(),
        panel.grid.minor=element_blank())

plot.timeline.interactive <- plot.timeline +
  geom_point(aes(color=email_type, text=group_names))

ggplotly(plot.timeline.interactive)

Timeline of survey responses

survey.time.plot <- survey.orgs.clean %>%
  select(EndDate) %>%
  arrange(EndDate) %>%
  mutate(done = 1,
         num.completed.cum = cumsum(done))

plot.responses.timeline <- ggplot() + 
  geom_step(data=survey.time.plot,
            aes(x=EndDate, y=num.completed.cum),
            size=0.5, colour="grey50") + 
  scale_y_continuous(labels=comma) +
  scale_x_datetime(date_labels="%B %e", date_breaks="1 week") +
  guides(fill=FALSE, colour=guide_legend(title=NULL)) +
  labs(x=NULL, y="Cumulative number of responses") +
  theme_light()

plot.responses.timeline.static <- plot.responses.timeline +
  geom_vline(data=sending.groups.plot, 
             aes(xintercept=as.numeric(email_day), 
                 colour=email_type),
             size=0.5) + 
  geom_label_repel(data=sending.groups.plot,
                   aes(x=email_day, y=400, 
                       label=group_names, fill=email_type),
                   size=2.5, colour="white") +
  theme(legend.position="bottom", 
        legend.key.size=unit(0.65, "lines"),
        legend.key=element_blank(),
        panel.grid.minor=element_blank())

plot.responses.timeline.interactive <- plot.responses.timeline +
  geom_vline(data=sending.groups.plot, 
             aes(xintercept=as.numeric(email_day), 
                 colour=email_type, text=group_names),
             size=0.5)

ggplotly(plot.responses.timeline.interactive)

References

Callegaro, Mario, and Charles DiSogra. 2008. “Computing Response Metrics for Online Panels.” Public Opinion Quarterly 72(5): 1008–32.

The American Association for Public Opinion Research. 2016. Standard Definitions: Final Dispositions of Case Codes and Outcome Rates for Surveys. 9th ed. AAPOR.