Tuesday, October 2, 2018 From rOpenSci (https://ropensci.org/blog/2018/10/02/outcomerate/). Except where otherwise noted, content on this site is licensed under the CC-BY license.
Surveys are ubiquitous in the social sciences, and the best of them are meticulously planned out. Statisticians often decide on a sample size based on a theoretical design, and then proceed to inflate this number to account for “sample losses”. This ensures that the desired sample size is achieved, even in the presence of non-response. Factors that reduce the pool of interviews include participant refusals, inability to contact respondents, deaths, and frame inaccuracies.^{1} The more non-response, the more a study becomes open to criticism about its veracity.
To simplify communication about non-response, a set of “outcome rates” have emerged over time. These rates are essentially quality declarations about the operational elements in a survey. Broadly speaking, they relate how well an actual respondent set r matches its intended sample s.
Some of popular outcome rates include:
Sounds straightforward. But for a long time, it was common to find disparate definitions of these rates. This made it hard to judge the quality of one survey against another.
The package calculates standard outcome rates in R in order to encourage transparent research, open methods, and scientific comparability. It does so using the American Association of Public Opinion Research’s (AAPOR) definitions of the rates, which are the industry standard. By collecting them in one package, it saves you the time of repetitively looking up each rate and calculating them separately.
As a contrived example, let’s assume we draw a sample of 10 people. Five individuals provide a complete interview (I
), one provides a partial interview (P
), one refuses to participate (R
), and two are unreachable but known to be eligible (NC
). Our 10th case is also unreachable but we don’t know if they are illegible to begin with (UH
).
library(outcomerate)
x <- c(I = 5, P = 1, R = 1, NC = 2, UH = 1)
outcomerate(x)
#> RR1 RR2 RR5 RR6 COOP1 COOP2 COOP3 COOP4 REF1 REF3 CON1 CON3 LOC1
#> 0.50 0.60 0.56 0.67 0.71 0.86 0.71 0.86 0.10 0.11 0.70 0.78 0.90
By default, the function returns only the rates that are defined based on the input. Here, we see that the example case achieved a Response Rate 2 (RR2) of 60%, and a Cooperation Rate 1 (COOP1) of 71%. Other rates, such as Response Rate 3 (RR3), are not returned as they only become available if you specify additional parameters (see ?outcomerate
for details).
The same output can be obtained by specifying the input as a vector of cases. This format may be more natural if you have a dataframe of interviews, or a stream of daily data coming from a server:
y <- c("I", "I", "I", "I", "I", "P", "R", "NC", "NC", "UH")
outcomerate(y)
#> RR1 RR2 RR5 RR6 COOP1 COOP2 COOP3 COOP4 REF1 REF3 CON1 CON3 LOC1
#> 0.50 0.60 0.56 0.67 0.71 0.86 0.71 0.86 0.10 0.11 0.70 0.78 0.90
outcomerate()
replicates some of the functionality that AAPOR already makes available in their Excel-based calculator. Theirs is a wonderful free tool, but has some limitations. For one, the Excel calculator is not designed to compute many rates by sub-domain. In my own work, I often need rates in real time by day, by researcher/enumerator, and by geography. Doing each analysis in a separate Excel workbook is impractical. Instead, outcomerate()
makes it easy to achieve this with dplyr
, aggregate()
, or data.table
.
Another problem is that Excel is notorious for failing silently. It will give you a result even if the input contains errors or inconsistencies. Especially when I am updating this information daily, those mistakes are quick to slip in. outcomerate
helps overcome this by failing fast and failing helpfully:
outcomerate(c("I", "I", "P", NA))
#> Error in assert_disposition(x): The input 'x' contains NA values. Consider converting them to
#> NE (known inelligibles) or UO / UH (unknown elligibility)
The package also implements weighted calculations of rates. Weighted figures further advance reproducible research because they typically estimate the outcome rate at the population level rather than for the sample. This is useful information to share with the world, as it helps others plan for sample losses in future studies. Unless you plan to replicate the sampling design exactly, it is often more useful to know the (weighted) population rate and/or rates by sub-domain.
outcomerate(y, weight = 10:1) # fake weights
#> RR1w RR2w RR5w RR6w COOP1w COOP2w COOP3w COOP4w REF1w REF3w
#> 0.73 0.82 0.74 0.83 0.82 0.92 0.82 0.92 0.07 0.07
#> CON1w CON3w LOC1w
#> 0.89 0.91 0.98
I welcome contributions to outcomerate
(#hacktoberfest). Some ideas for enhancements include:
My thanks goes to Neil Richardson, Carl Ganz, and Karthik Ram for their peer review and editorial support at rOpenSci. outcomerate
is better for it, and my programming style and knowledge has improved. Naturally, I also wish to acknowledge AAPOR for advancing the practice of survey research— this package is built on the definitions they helped standardize.
Frame inaccuracies are mistakes in the list of population elements from which the sample is drawn. Their post-facto discovery has implications for design-based inference, because the original sample then no longer relates cleanly to the target population. Such inaccuracies impact outcome rate calculation, depending on whether or not you use your initial mistaken frame or make an adjustment to correct for ineligible cases. ↩︎