Exact confidence intervals for difference between proportions

Confidence intervals Proportions Exact confidence intervals

This describes the ExactCIdiff package, unpacks the arguments to the functions, looks at timings and notes an oddity, a typo I think, in the original paper about the package.

Chris Evans https://www.psyctc.org/R_blog/ (PSYCTC.org)https://www.psyctc.org/psyctc/
2022-04-18
Show code
### this is just the code that creates the "copy to clipboard" function in the code blocks
htmltools::tagList(
  xaringanExtra::use_clipboard(
    button_text = "<i class=\"fa fa-clone fa-2x\" style=\"color: #301e64\"></i>",
    success_text = "<i class=\"fa fa-check fa-2x\" style=\"color: #90BE6D\"></i>",
    error_text = "<i class=\"fa fa-times fa-2x\" style=\"color: #F94144\"></i>"
  ),
  rmarkdown::html_dependency_font_awesome()
)

Warning: this is pretty geeky statistical stuff

Warning2: the package is no longer on CRAN ## Installing package from source

Update added 26.vi.22

The ExactCIdiff package has dropped off CRAN because the authors/maintainers aren’t responding to the CRAN team and a URL in the package is not openning for CRAN. This doesn’t affect the code in the package.

You can download the source package from https://cran.r-project.org/src/contrib/Archive/ExactCIdiff/. It’s the last version, v1.3 you want which is https://cran.r-project.org/src/contrib/Archive/ExactCIdiff/ExactCIdiff_1.3.tar.gz

### you can download it before launching R:
### install.packages("~/Downloads/ExactCIdiff_1.3.tar.gz", repos = NULL, type = "source")
### replace "~/Downloads/" with the directory into which you downloaded it of course!

### or, as I should have remembered, you can omit downloading the package beforehand and just grab it withing the 
### install.packages() call:
install.packages("https://cran.r-project.org/src/contrib/Archive/ExactCIdiff/ExactCIdiff_1.3.tar.gz", repos = NULL, type = "source")

That’s that solved!

Back to the story: background

I came on this as Clara and I need to look at the differences in proportions between two arms of our “elephant” study (we call it that because it’s BIG!) I knew that I wasn’t confident about the best way to get 95% confidence intervals (CIs) for a difference between two proportions. I have been using Hmisc::binconf() for certainly over ten years for CIs around a single proportion but knew this was a bit different. A bit of searching led to the ExactCIdiff package and to the paper about it: Shan, G., & Wang, W. (2013). ExactCIdiff: An R Package for Computing Exact Confidence Intervals for the Difference of Two Proportions. The R Journal, 5(2), 62. https://doi.org/10.32614/RJ-2013-026.

I can’t follow the maths of the method but I do follow the evidence that it does better in terms of coverage probability (the actual probability that it will include the population value) than other methods. It’s a clean package and a nice paper and I say it’s a clean package as it appears to do just two things and the two things it sets out to do, and to do them well.

The two things are to give you a CI around an observed difference in proportions for a paired sample (e.g. proportion above a cut-off at baseline and after therapy) or for the same but for unconnected samples (what we have: students versus non-student people of the same age group).

I will start with the latter.

Difference between proportions in two separate samples

I’ll quote from their paper:

The second data set is from a two-arm randomized clinical trial for testing the effect of tobacco smoking on mice (Essenberg, 1952). In the treatment (smoking) group, the number of mice is n 1 = 23, and the number of mice which developed tumor is x = 21; in the control group, n 2 = 32 and y = 19.

Ugh, not my world but it’s their paper. So here’s their data, rows as smoking/non-smoking groups and columns as whether the poor things developed tumours.

Show code
tribble(~smoking, ~tumour, ~n,
        1, 1, 21,
        1, 0, 2,
        0, 1, 19,
        0, 0, 13) %>%
  uncount(n) -> tibPoorMice

tibPoorMice %>%
  tabyl(smoking, tumour) %>%
  adorn_totals(where = c("row", "col")) 
 smoking  0  1 Total
       0 13 19    32
       1  2 21    23
   Total 15 40    55

And with percentages.

Show code
tibPoorMice %>%
  tabyl(smoking, tumour) %>%
  adorn_totals(where = c("row", "col")) %>%
  adorn_percentages(denominator = "row") %>%
  adorn_pct_formatting(digits = 1)
 smoking     0     1  Total
       0 40.6% 59.4% 100.0%
       1  8.7% 91.3% 100.0%
   Total 27.3% 72.7% 100.0%

Pretty clear that’s going to be statistically significant … and it is.

Show code
tibPoorMice %>%
  tabyl(smoking, tumour) %>%
  chisq.test()

    Pearson's Chi-squared test with Yates' continuity correction

data:  .
X-squared = 5.3625, df = 1, p-value = 0.02057

But what we want is the confidence interval. Here are the separate intervals.

Show code
tibPoorMice %>% 
  group_by(smoking) %>%
  ### bit back to square one given I stared with the raw numbers but 
  ### this is how it would be with real data
  summarise(n = n(),
            nTumour = sum(tumour == 1),
            ### this is a bit messy, Hmisc::binconf returns a matrix
            ### but I just want the first row, hence the "[1,]"
            ### and of course I need the list as that's how dplyr
            ### has to be told that what it's getting is a list or
            ### a vector (as here)
            binconf = list(Hmisc::binconf(nTumour, n)[1,])) %>%
  ### OK, unnest that
  unnest_wider(binconf) %>%
  ### recode smoking to a factor
  mutate(smoking = ordered(smoking,
                           levels = 0:1,
                           labels = c("Smoked mice",
                                      "Lucky ones"))) -> tmpTib 

### print that
tmpTib %>%
  ### do some rounding across the three values PointEst to Upper
  mutate(across(PointEst:Upper, round, 2)) %>%
  pander(justify = "rrrrrr")
smoking n nTumour PointEst Lower Upper
Smoked mice 32 19 0.59 0.42 0.74
Lucky ones 23 21 0.91 0.73 0.98

And here as a plot as I am such a believer in offering both tabulated and plotted data where possible as some people find tables easier to digest (more precision) and others find the plots easier (more, hm, visual impact!)

Show code
### get overall proportion for reference line
tibPoorMice %>%
  summarise(nTumour = sum(tumour == 1),
            n = n(),
            prop = nTumour / n) %>%
  select(prop) %>%
  pull() -> tmpAllProp

ggplot(data = tmpTib,
       aes(x = smoking, y = PointEst)) +
  geom_point() +
  geom_linerange(aes(ymin = Lower, ymax = Upper)) +
  geom_hline(yintercept = tmpAllProp) +
  scale_y_continuous(name = "Proportion", breaks = seq(0, 1, .1), limits = c(0,1)) +
  xlab("Whether the poor mice were smoked or not!") +
  ggtitle("95% confidence intervals for proportions")

The exact CI of the difference

So now (finally) we come to ExactCIdiff! The function is BinomCI and the syntax is that you give the four numbers from the crosstabulation, as n1, n2, count1, count2, so here 23, 32, 21, 19:

BinomCI(23, 32, 21, 19, conf.level = ?, CItype = ?)

I’m going to take things in the order that the authors do in their paper, starting with:

uci <- BinomCI(23, 32, 21, 19, conf.level = 0.95, CItype = "Upper")$ExactCI so the one-sided upper 95% confidence limit.

Show code
Sys.time() -> time1
uci <- BinomCI(23, 32, 21, 19, conf.level = 0.95, CItype = "Upper")$ExactCI
Sys.time() -> time2
elapsedTimeSecs1 <- as.numeric(difftime(time2, time1, units = "secs"))
uci
[1] -1.00000  0.48595

That’s a one-sided interval telling me that the upper 95% confidence limit for the difference is 0.48595: big. The computations are CPU intensive, that took 1 minutes on a fairly powerful laptop. I will generally want a two-sided interval and their next three calls to BinomCI demonstrate the relationship between the two one-sided 97.5% confidence limits and the two-sided CI. First the upper 97.5% limit.

Show code
Sys.time() -> time1
u975 <- BinomCI(23, 32, 21, 19, conf.level = 0.975, CItype = "Upper")$ExactCI
Sys.time() -> time2
u975
[1] -1.00000  0.51259
Show code
elapsedTimeSecs2 <- as.numeric(difftime(time2, time1, units = "secs"))

So upper 97.5% limit 0.51259 (elapsed time 31.6 seconds).

Show code
Sys.time() -> time1
l975 <- BinomCI(23, 32, 21, 19, conf.level = 0.975, CItype = "Lower")$ExactCI
Sys.time() -> time2
elapsedTimeSecs3 <- as.numeric(difftime(time2, time1, units = "secs"))
l975
[1] 0.09468 1.00000
Show code
# [1] 0.09468 1.00000

So lower 97.5% limit 0.09468 (elapsed time 6.7 seconds).

Show code
Sys.time() -> time1
ci95 <- BinomCI(23, 32, 21, 19)$ExactCI
Sys.time() -> time2
elapsedTimeSecs4 <- as.numeric(difftime(time2, time1, units = "secs"))
ci95
[1] 0.09468 0.51259
Show code
# [1] 0.09468 0.51259

And it can be seen there that the two-sided 95% CI is from 0.09468 to 0.51259, i.e. from the lower 97.5% CL to the upper 97.5% CL. (Elapsed time 38.2 seconds).

Paired difference of proportions

Here again I’ll quote from the paper:

We illustrate the usage of the PairedCI() function to calculate the exact smallest lower one-sided confidence interval [LP , 1] for θP in (1) with the data from Karacan et al. (1976). In this study, 32 marijuana users are compared with 32 matched controls with respect to their sleeping difficulties, with n11 = 16, n12 = 9, n21 = 3, and n22 = 4. The second argument in the function is t = n11 + n22 = 20.

The “(1)” refers back to the first equation in the paper which I won’t copy in here as it would need some formatting and doesn’t really matter for our purposes.

Show code
# ```{r makeTable, results='asis'}
# tmpVec <- c("", "Success at t2", "Failure at t2", "",
#             "Success at t1", "N11, p11", "N12, p12", "p1 = p11 + p12",
#             "Failure at t1", "N21, p21", "N22, p22", "",
#             "", "p2 = p11 + p21", "", "Total, p = 1")

tmpVec <- c("", "Success at t2", "Failure at t2", "",
            "Success at t1", "N11, p11", "N12, p12", "",
            "Failure at t1", "N21, p21", "N22, p22", "",
            "", "", "", "Total, p = 1")

tmpMat <- matrix(tmpVec, ncol = 4)
# print(xtable::xtable(tmpMat, type = "html"))
# print(xtable::xtable(tmpMat, getOption("xtable.type", "html")))
# knitr::kable(tmpMat, "html")

tmpMat %>%
  kbl() %>% 
  kable_styling(bootstrap_options = c("striped")) %>%
  # kable_styling() %>%
  row_spec(1, align = "c", bold = TRUE) %>%
  row_spec(2:4, align = "c") %>%
  column_spec(1, bold = TRUE, border_right = TRUE) %>%
  column_spec(2, border_right = TRUE) %>%
  column_spec(3, border_right = TRUE) 
Success at t1 Failure at t1
Success at t2 N11, p11 N21, p21
Failure at t2 N12, p12 N22, p22
Total, p = 1

I do find this way of describing a contingency table pretty counterinuitive!

Digression about tables in R and in Rmarkdown

GRRrrrr!!! I continue to feel that table handling in R is almost its Achilles heel. I’ve just wasted the better part of an hour finding out a way to get that table in an even halfway, no quarterway, decent form. I think I first commented on this perhaps twenty years ago and the R team position has always been, I think, that nice tables are for packages to fix and so we have multiple packages that try to fix this, mostly incompatible and none of them working reliably in Rmarkdown and with all output formats from Rmarkdown. I think the R afficionados all love knocking up tables in LaTeX and I’m sure that’s fine if you are really familiar with LaTeX and I suspect that direct R to LaTeX is the most robust and general way to do things but many of us don’t know TeX/LaTeX and don’t really want to have to learn it. Aarghhhh! OK, flame over!

Back to the data here.

Show code
tmpVec2 <- c("", "Sleep OK, no smokes", "Sleep poor, no smokes", "",
            "Sleep OK, smokes", "16", "9", "",
            "Sleep poor, smokes", "3", "4", "",
            "", "", "", "Total, p = 1")

tmpMat2 <- matrix(tmpVec2, ncol = 4)
# print(xtable::xtable(tmpMat, type = "html"))
# print(xtable::xtable(tmpMat, getOption("xtable.type", "html")))
# knitr::kable(tmpMat, "html")

tmpMat2 %>%
  kbl() %>% 
  kable_styling(bootstrap_options = c("striped")) %>%
  # kable_styling() %>%
  row_spec(1, align = "c", bold = TRUE) %>%
  row_spec(2:4, align = "c") %>%
  column_spec(1, bold = TRUE, border_right = TRUE) %>%
  column_spec(2, border_right = TRUE) %>%
  column_spec(3, border_right = TRUE) 
Sleep OK, smokes Sleep poor, smokes
Sleep OK, no smokes 16 3
Sleep poor, no smokes 9 4
Total, p = 1

Exact CI for a difference between paired proportions

That means that the code is:

PairedCI(9, 20, 3, conf.level = 0.95)

because the syntax is

PairedCI(n12, t, n21, conf.level, CItype, precision, grid.one, grid.two)

where we can ignore grid.one and grid.two for now and leave them at their default values of 30 and 20 and precision is, as the help says:

Precision of the confidence interval, default is 0.00001 rounded to 5 decimals.

OK, so here we go with:

PairedCI(9, 20, 3, conf.level = 0.95)

Show code
Sys.time() -> time1
lciall <- PairedCI(9, 20, 3, conf.level = 0.95) # store relevant quantities
Sys.time() -> time2
elapsedTimeSecs5 <- as.numeric(difftime(time2, time1, units = "secs"))
lciall
$conf.level
[1] 0.95

$CItype
[1] "Two.sided"

$estimate
[1] 0.1875

$ExactCI
[1] -0.03564  0.39521
Show code
# $conf.level
# [1] 0.95
# 
# $CItype
# [1] "Two.sided"
# 
# $estimate
# [1] 0.1875
# 
# $ExactCI
# [1] -0.03564  0.39521

(Elapsed time 15.4 seconds.)

The odd thing here is that this is not what the authors show in the paper:

lciall  # print lciall  
$conf.level  
[1] 0.95    # confidence level  
$CItype  
[1] "Lower" # lower one-sided interval    
$estimate  
[1] 0.1875  # the mle of p1 - p2  
$ExactCI 
[1] 0.00613 1.00000 # the lower one-sided 95% interval  
lci <- lciall$ExactCI # extracting the lower one-sided 95% interval  
lci         # print lci  
[1] 0.00613 1.00000  

The use of marijuana helps sleeping because the interval [ 0.00613, 1 ] for θP is positive.

Which is clearly not what I just got. However, in the paper they go on:

The upper one-sided 95% interval and the two-sided 95% interval for θ~P~ are given below for illustration purpose.

I think that’s a typo. I think what they are showing are the results of

PairedCI(9, 20, 3, conf.level = 0.95, CItype = "lower")

Let’s see:

Show code
Sys.time() -> time1
lciall <- PairedCI(9, 20, 3, conf.level = 0.95, CItype = "Lower") # store relevant quantities
Sys.time() -> time2
elapsedTimeSecs5 <- as.numeric(difftime(time2, time1, units = "secs"))
lciall
$conf.level
[1] 0.95

$CItype
[1] "Lower"

$estimate
[1] 0.1875

$ExactCI
[1] 0.00613 1.00000

Yes! (Elapsed time 3.4 seconds.)

They do go on to give us other things in the paper and that I think confirms that the above call was a typo.

So here is the upper 95% CL.

Show code
Sys.time() -> time1
uci <- PairedCI(9, 20, 3, conf.level = 0.95, CItype = "Upper")$ExactCI
Sys.time() -> time2
elapsedTimeSecs7 <- as.numeric(difftime(time2, time1, units = "secs"))
uci
[1] -1.00000  0.36234
Show code
# [1] -1.00000  0.36234

(Elapsed time 12.4 seconds.)

The upper 97.5% CL.

Show code
Sys.time() -> time1
u975 <- PairedCI(9, 20, 3, conf.level = 0.975, CItype = "Upper")$ExactCI
Sys.time() -> time2
elapsedTimeSecs8 <- as.numeric(difftime(time2, time1, units = "secs"))
u975
[1] -1.00000  0.39521
Show code
# [1] -1.00000  0.39521

(Elapsed time 11.7 seconds.)

The lower 97.5% CL.

Show code
Sys.time() -> time1
l975 <- PairedCI(9, 20, 3, conf.level = 0.975, CItype = "Lower")$ExactCI
Sys.time() -> time2
elapsedTimeSecs9 <- as.numeric(difftime(time2, time1, units = "secs"))
l975
[1] -0.03564  1.00000
Show code
# [1] -0.03564  1.00000

(Elapsed time 3.4 seconds.)

And back to the two-sided 95% CI (and yes, I’m running it again just to be sure I get the same answer as last time!)

Show code
Sys.time() -> time1
ci95 <- PairedCI(9, 20, 3, conf.level = 0.95)$ExactCI
Sys.time() -> time2
elapsedTimeSecs10 <- as.numeric(difftime(time2, time1, units = "secs"))
ci95
[1] -0.03564  0.39521
Show code
# [1] -0.03564  0.39521

(Elapsed time 15.2 seconds.)

Yup, the same again and fits with what they say in the paper:

[1] -0.03564 0.39521 # the two-sided 95% interval
                      # it is equal to the intersection of two one-sided intervals

Computation times

Clearly one issue here is that these are small total sample sizes in their two examples but the process is computationally expensive (though 40x faster than another approach with the same accuracy/coverage).

Show code
vecMultipliers <- 1:4 # check for sample sizes 1 to 8x the example above
vecTimesPaired <- rep(NA, length(vecMultipliers))
matCI95paired <- matrix(rep(NA, length(vecMultipliers) * 2), ncol = 2)

vecTimesPaired[1] <- elapsedTimeSecs10
matCI95paired[1, ] <- ci95

for (mult in vecMultipliers[-1]) {
  Sys.time() -> time1
  matCI95paired[mult, ] <- PairedCI(mult * 9, mult * 20, mult * 3, conf.level = 0.95)$ExactCI
  Sys.time() -> time2
  vecTimesPaired[mult] <- as.numeric(difftime(time2, time1, units = "secs"))
}

Here are those CIs getting tighter as the numbers go up.

Show code
matCI95paired %>%
  as_tibble() %>%
  bind_cols(vecMultipliers) %>%
  rename(nPairs = `...3`,
         LCL = V1,
         UCL = V2) %>%
  mutate(nPairs = 32 * nPairs) %>%
  select(nPairs, everything()) -> tibCI95paired

tibCI95paired %>%
  pander::pander()
nPairs LCL UCL
32 -0.03564 0.3952
64 0.03751 0.3396
96 0.06301 0.3067
128 0.07231 0.2906
Show code
ggplot(data = tibCI95paired,
       aes(x = nPairs)) +
  geom_linerange(aes(ymin = LCL, ymax = UCL)) +
  geom_hline(yintercept = 0.1875) +
  ylab("Difference in proportions") +
  scale_x_continuous(name = "Number of pairs",
                     breaks = vecMultipliers * 32) +
  ggtitle("Two sided 95% CI tightening with increasing sample size",
          subtitle = "Horizontal reference line is observed difference in proportions")

And here are the times (in seconds).

Show code
library(tidyverse)
vecTimesPaired %>%
  as_tibble() %>%
  rename(timeSecs = value) %>%
  bind_cols(vecMultipliers) %>%
  rename(nPairs = `...2`) %>%
  mutate(nPairs = 32 * nPairs) %>% 
  select(nPairs, timeSecs) -> tibTimesPaired

tibTimesPaired %>%
  pander(justify = "rr")
nPairs timeSecs
32 15.21
64 166.6
96 242.9
128 556
Show code
ggplot(data = tibTimesPaired,
       aes(x = nPairs, y = timeSecs)) +
  geom_point() +
  geom_line() +
  ylab("Elapsed time (seconds)") +
  xlab("Number of pairs of participants") +
  ggtitle("Plot of computation time against sample size")

Hm. That doesn’t look that far off linear which is not what I had expected.

Unpaired example: CIs and timings

Show code
ci95 <- BinomCI(23, 32, 21, 19)$ExactCI

vecTimesUnpaired <- rep(NA, length(vecMultipliers))
matCI95unpaired <- matrix(rep(NA, length(vecMultipliers) * 2), ncol = 2)

for (mult in vecMultipliers) {
  Sys.time() -> time1
  matCI95unpaired[mult, ] <- PairedCI(mult * 9, mult * 20, mult * 3, conf.level = 0.95)$ExactCI
  Sys.time() -> time2
  vecTimesUnpaired[mult] <- as.numeric(difftime(time2, time1, units = "secs"))
}
Show code
matCI95unpaired %>%
  as_tibble() %>%
  bind_cols(vecMultipliers) %>%
  rename(nTotal = `...3`,
         LCL = V1,
         UCL = V2) %>%
  mutate(nTotal = 55 * nTotal) %>%
  select(nTotal, everything()) -> tibCI95unpaired

tibCI95unpaired %>%
  pander::pander()
nTotal LCL UCL
55 -0.03564 0.3952
110 0.03751 0.3396
165 0.06301 0.3067
220 0.07231 0.2906
Show code
ggplot(data = tibCI95unpaired,
       aes(x = nTotal)) +
  geom_linerange(aes(ymin = LCL, ymax = UCL)) +
  geom_hline(yintercept = 0.1875) +
  ylab("Difference in proportions") +
  scale_x_continuous(name = "Total number of mice",
                     breaks = vecMultipliers * 32) +
  ggtitle("Two sided 95% CI tightening with increasing sample size",
          subtitle = "Horizontal reference line is observed difference in proportions")

And, again, the times.

Show code
library(tidyverse)
vecTimesUnpaired %>%
  as_tibble() %>%
  rename(timeSecs = value) %>%
  bind_cols(vecMultipliers) %>%
  rename(nTotal = `...2`) %>%
  mutate(nTotal = 32 * nTotal) %>% 
  select(nTotal, timeSecs) -> tibTimesPaired

tibTimesPaired %>%
  pander(justify = "rr")
nTotal timeSecs
32 15.91
64 171.2
96 244.3
128 551.5
Show code
ggplot(data = tibTimesPaired,
       aes(x = nTotal, y = timeSecs)) +
  geom_point() +
  geom_line() +
  ylab("Elapsed time (seconds)") +
  xlab("Number of mice") +
  ggtitle("Plot of computation time against sample size")

Again, looks fairly linear. Interesting.

Summary

The R package ExactCIdiff provides two functions which give what appear to be the best confidence intervals for differences between two proportions, one function, BinomCI() for differences from unpaired samples and the other PairedCI() for paired samples. I think there’s a typo in the paper about the package and the syntax of the arguments isn’t particularly friendly (it’s even case sensitive so CItype = "upper" with throw an error, it has to be CItype = "Upper"). However, it’s not difficult to work those things out (and that’s partly why I’ve created this post) and it does seem that these really are the best ways to get these CIs. They’re faily computationally intensive but from my tiny simulation it looks as if the timing is linear across simple multiples of sample size. Thanks and kudos to Shan and Wang!

Visit count

website counter

Last updated

Show code
cat(paste(format(Sys.time(), "%d/%m/%Y"), "at", format(Sys.time(), "%H:%M")))
16/04/2024 at 10:48

Reuse

Text and figures are licensed under Creative Commons Attribution CC BY-SA 4.0. The figures that have been reused from other sources don't fall under this license and can be recognized by a note in their caption: "Figure from ...".

Citation

For attribution, please cite this work as

Evans (2022, April 18). Chris (Evans) R SAFAQ: Exact confidence intervals for difference between proportions. Retrieved from https://www.psyctc.org/R_blog/posts/2022-04-18-exact-confidence-intervals-for-difference-between-proportions/

BibTeX citation

@misc{evans2022exact,
  author = {Evans, Chris},
  title = {Chris (Evans) R SAFAQ: Exact confidence intervals for difference between proportions},
  url = {https://www.psyctc.org/R_blog/posts/2022-04-18-exact-confidence-intervals-for-difference-between-proportions/},
  year = {2022}
}