pubh
packageThere has been a long relationship between the disciplines of Epidemiology / Public Health and Biostatistics. Students frequently find introductory textbooks explaining statistical methods and the maths behind them, but how to implement those techniques in a computer is most of the time not explained.
One of the most popular statistical software’s in public health settings is Stata
. Stata
has the advantage of offering a solid interface with functions that can be accessed via the use of commands or by selecting the proper input in the graphical unit interface (GUI). Furthermore, Stata
offers “Grad Plans” to postgraduate students, making it relatively affordable from an economic point of view.
R
is a free alternative to Stata
. The use of R
keeps growing; furthermore, with the relatively high number of packages and textbooks available, it’s popularity is also increasing.
In the case of epidemiology, there are already some good packages available for R
, including: Epi
, epibasix
, epiDisplay
, epiR
and epitools
. The pubh
package does not intend to replace any of them, but to only provide a common syntax for the most frequent statistical analysis in epidemiology.
Most students and professionals from the disciplines of Epidemiology and Public Health analyse variables in terms of outcome, exposure and confounders. The following table shows the most common names used in the literature to characterise variables in a cause-effect relationships:
Response variable | Explanatory variable(s) |
---|---|
Outcome | Exposure and confounders |
Outcome | Predictors |
Dependent variable | Independent variable(s) |
y | x |
In R
, formulas
are used to declare relationships between variables. Formulas are common in classic statistical tests and in regression methods.
Formulas have the following standard syntax:
Where y
is the outcome or response variable, x
is the exposure or predictor of interest and my_data
specifies the name of the data frame where x
and y
can be found.
The symbol ~
is used in R
for formulas. It can be interpreted as depends on. In the most typical scenario, y ~ x
means y
depends on x
or y
is a function of x
:
y = f(x)
Using epidemiology friendly terms:
Outcome = f(Exposure)
Is worth to mention that Stata
requires for variables to be given in the same order as in formulas: outcome first, predictors next.
The pubh
package integrates well with other packages of the tidyverse
wich use formulas and the pipe operator %>%
to add layers over functions. In particular, pubh
uses ggformula
as a graphical interface for plotting and takes advantage of variable labels from sjlabelled
. This versatility allows it to interact also with tables from moonBook
and effect plots from sjPlot
.
One way to control for confounders is the use of stratification. In Stata
stratification is done by including the by
option as part of the command. In the ggformula
package, one way of doing stratification is with a formula like:
Where y
is the outcome or response variable, x
is the exposure or predictor of interest, z
is the stratification variable (a factor) and my_data
specifies the name of the data frame where x
, y
and z
can be found.
pubh
packageThe main contributions of the pubh
package to students and professionals from the disciplines of Epidemiology and Public Health are:
glm_coef
that displays coefficients from most common regression analysis in a way that can be easy interpreted and used for publications.ggformula
package, introducing plotting functions with a relatively simple syntax.Many options currently exists for displaying descriptive statistics. I recommend the function mytable
from the moonBook
package. It does a great job for both continuous and categorical (factors) outcomes.
The estat
function from the pubh
package displays descriptive statistics of continuous outcomes and like mytable
, it can use labels to display nice tables. estat
. As a way to aid in the understanding of the variability, estat
displays also the relative dispersion (coefficient of variation) which is of particular interest for comparing variances between groups (factors).
Some examples. We start by loading packages as suggested in the Template of this package.
library(broom)
library(car)
library(Hmisc)
library(MASS)
library(kableExtra)
library(tidyverse)
library(mosaic)
library(latex2exp)
library(moonBook)
library(pubh)
library(sjlabelled)
library(sjPlot)
theme_set(sjPlot::theme_sjplot2(base_size = 10))
theme_update(legend.position = "top")
options(knitr.table.format = 'pandoc')
We will use a data set about blood counts of T cells from patients in remission from Hodgkin’s disease or in remission from disseminated malignancies. We generate the new variable Ratio
and add labels to the variables.
data(Hodgkin)
Hodgkin <- Hodgkin %>%
mutate(Ratio = CD4/CD8) %>%
var_labels(
Ratio = "CD4+ / CD8+ T-cells ratio"
)
kable(head(Hodgkin))
CD4 | CD8 | Group | Ratio |
---|---|---|---|
396 | 836 | Hodgkin | 0.4736842 |
568 | 978 | Hodgkin | 0.5807771 |
1212 | 1678 | Hodgkin | 0.7222884 |
171 | 212 | Hodgkin | 0.8066038 |
554 | 670 | Hodgkin | 0.8268657 |
1104 | 1335 | Hodgkin | 0.8269663 |
Descriptive statistics for CD4+ T cells:
Hodgkin %>%
estat(~ CD4)
N Min. Max. Mean Median SD CV
1 CD4+ T-cells 40 116 2415 672.62 528.5 470.49 0.7
In the previous code, the left-hand side of the formula is empty as it’s the case when working with only one variable.
We can use kable
to display nice tables.
N | Min. | Max. | Mean | Median | SD | CV | |
---|---|---|---|---|---|---|---|
CD4+ T-cells | 40 | 116 | 2415 | 672.62 | 528.5 | 470.49 | 0.7 |
For stratification, estat
recognises the following two syntaxes:
outcome ~ exposure
~ outcome | exposure
where, outcome
is continuous and exposure
is a categorical (factor
) variable.
For example:
Disease | N | Min. | Max. | Mean | Median | SD | CV | |
---|---|---|---|---|---|---|---|---|
CD4+ / CD8+ T-cells ratio | Non-Hodgkin | 20 | 1.10 | 3.49 | 2.12 | 2.15 | 0.73 | 0.34 |
Hodgkin | 20 | 0.47 | 3.82 | 1.50 | 1.19 | 0.91 | 0.61 |
From the descriptive statistics of Ratio we know that the relative dispersion in the Hodgkin group is almost as double as the relative dispersion in the Non-Hodgkin group.
For new users of R
it helps to have a common syntax in most of the commands, as R
could be challenging and even intimidating at times. We can test if the difference in variance is statistically significant with the var.test
command, which uses can use a formula syntax:
var.test(Ratio ~ Group, data = Hodgkin)
F test to compare two variances
data: Ratio by Group
F = 0.6294, num df = 19, denom df = 19, p-value = 0.3214
alternative hypothesis: true ratio of variances is not equal to 1
95 percent confidence interval:
0.2491238 1.5901459
sample estimates:
ratio of variances
0.6293991
What about normality? We can look at the QQ-plots against the standard Normal distribution.
Let’s say we choose a non-parametric test to compare the groups:
For relatively small samples (for example, less than 30 observations per group) is a standard practice to show the actual data in dot plots with error bars. The pubh
package offers two options to show graphically differences in continuous outcomes among groups:
strip_error
bar_error
For our current example:
In the previous code, axis_labs
applies labels from labelled data to the axis.
The error bars represent 95% confidence intervals around mean values.
Is relatively easy to add a line on top, to show that the two groups are significantly different. The function gf_star
needs the reference point on how to draw an horizontal line to display statistical difference or to annotate a plot; in summary, gf_star
:
Thus:
\[ y1 < y2 < y3 \]
In our current example:
Hodgkin %>%
strip_error(Ratio ~ Group) %>%
axis_labs() %>%
gf_star(x1 = 1, y1 = 4, x2 = 2, y2 = 4.05, y3 = 4.1, "**")
For larger samples we could use bar charts with error bars. For example:
data(birthwt, package = "MASS")
birthwt <- birthwt %>%
mutate(
smoke = factor(smoke, labels = c("Non-smoker", "Smoker")),
Race = factor(race > 1, labels = c("White", "Non-white")),
race = factor(race, labels = c("White", "Afican American", "Other"))
) %>%
var_labels(
bwt = 'Birth weight (g)',
smoke = 'Smoking status'
)
Quick normality check:
The (unadjusted) \(t\)-test:
t.test(bwt ~ smoke, data = birthwt)
Welch Two Sample t-test
data: bwt by smoke
t = 2.7299, df = 170.1, p-value = 0.007003
alternative hypothesis: true difference in means is not equal to 0
95 percent confidence interval:
78.57486 488.97860
sample estimates:
mean in group Non-smoker mean in group Smoker
3055.696 2771.919
The final plot with annotation to highlight statistical difference (unadjusted):
birthwt %>%
bar_error(bwt ~ smoke) %>%
axis_labs() %>%
gf_star(x1 = 1, x2 = 2, y1 = 3400, y2 = 3500, y3 = 3550, "**")
Both strip_error
and bar_error
can generate plots stratified by a third variable, for example:
The pubh
package includes the function glm_coef
for displaying coefficients from regression models and the function multiple
to help in the visualisation of multiple comparisons.
Note: Please read the vignette on Regression Examples for a more comprehensive use of
glm_coef
.
For simplicity, here we show the analysis of the linear model of smoking on birth weight, adjusting by race (and not by other potential confounders).
model_bwt <- lm(bwt ~ smoke + race, data = birthwt)
model_bwt %>%
glm_coef(labels = c("Constant",
"Smoker vs Non smoker",
"African American vs White",
"Other vs White")) %>%
kable(align = "r")
Coefficient | Pr(>|t|) | |
---|---|---|
Constant | 3334.95 (3036.11, 3633.78) | < 0.001 |
Smoker vs Non smoker | -428.73 (-667.02, -190.44) | < 0.001 |
African American vs White | -450.36 (-750.31, -150.4) | 0.003 |
Other vs White | -452.88 (-775.17, -130.58) | 0.006 |
Similar results can be obtained with the function tab_model
from the sjPlot
package.
Birth weight (g) | ||
---|---|---|
Predictors | Estimates | p |
(Intercept) |
3334.95 (3153.89 – 3516.01) |
<0.001 |
Smoking status: Smoker |
-428.73 (-643.86 – -213.60) |
<0.001 |
race: Afican American |
-450.36 (-752.45 – -148.27) |
0.004 |
race: Other |
-452.88 (-682.67 – -223.08) |
<0.001 |
Observations | 189 | |
R^{2} / R^{2} adjusted | 0.123 / 0.109 |
Some adavantages of glm_coef
over tab_model
are:
tab_model
does not recognise.Some advantages of tab_model
over glm_coef
are:
In the previous table of coefficients confidence intervals and p-values for race had not been adjusted for multiple comparisons. We use functions from the emmeans
package to make the corrections.
The pubh
package offers two wrappers to epiR
functions.
contingency
calls epi.2by2
and it’s used to analyse two by two contingency tables.diag_test
calls epi.tests
to compute statistics related with screening tests.Let’s say we want to look at the effect of ibuprofen on preventing death in patients with sepsis.
id | treat | race | fate | apache | o2del | followup | temp0 | temp10 |
---|---|---|---|---|---|---|---|---|
1 | Placebo | White | Dead | 27 | 539.20 | 50 | 35.22222 | 36.55556 |
2 | Ibuprofen | African American | Alive | 14 | NA | 720 | 38.66667 | 37.55555 |
3 | Placebo | African American | Dead | 33 | 551.12 | 33 | 38.33333 | NA |
4 | Ibuprofen | White | Alive | 3 | 1376.14 | 720 | 38.33333 | 36.44444 |
5 | Placebo | White | Alive | 5 | NA | 720 | 38.55556 | 37.55555 |
6 | Ibuprofen | White | Alive | 13 | 1523.39 | 720 | 38.16667 | 38.16667 |
Let’s look at the table:
Treatment | Placebo | Ibuprofen | Total | p |
---|---|---|---|---|
(N=231) | (N=224) | (N=455) | ||
Mortality status | 0.679 | |||
- Alive | 139 (60.2%) | 140 (62.5%) | 279 (61.3%) | |
- Dead | 92 (39.8%) | 84 (37.5%) | 176 (38.7%) |
For epi.2by2
we need to provide the table of counts in the correct order, so we would type something like:
dat <- matrix(c(84, 140 , 92, 139), nrow = 2, byrow = TRUE)
epiR::epi.2by2(as.table(dat))
Outcome + Outcome - Total Inc risk * Odds
Exposed + 84 140 224 37.5 0.600
Exposed - 92 139 231 39.8 0.662
Total 176 279 455 38.7 0.631
Point estimates and 95% CIs:
-------------------------------------------------------------------
Inc risk ratio 0.94 (0.75, 1.19)
Odds ratio 0.91 (0.62, 1.32)
Attrib risk * -2.33 (-11.27, 6.62)
Attrib risk in population * -1.15 (-8.88, 6.59)
Attrib fraction in exposed (%) -6.20 (-33.90, 15.76)
Attrib fraction in population (%) -2.96 (-15.01, 7.82)
-------------------------------------------------------------------
Test that odds ratio = 1: chi2(1) = 0.26 Pr>chi2 = 0.61
Wald confidence limits
CI: confidence interval
* Outcomes per 100 population units
For contingency
we only need to provide the information in a formula:
Bernard %>%
contingency(fate ~ treat)
Outcome
Predictor Dead Alive
Ibuprofen 84 140
Placebo 92 139
Outcome + Outcome - Total Inc risk * Odds
Exposed + 84 140 224 37.5 0.600
Exposed - 92 139 231 39.8 0.662
Total 176 279 455 38.7 0.631
Point estimates and 95% CIs:
-------------------------------------------------------------------
Inc risk ratio 0.94 (0.75, 1.19)
Odds ratio 0.91 (0.62, 1.32)
Attrib risk * -2.33 (-11.27, 6.62)
Attrib risk in population * -1.15 (-8.88, 6.59)
Attrib fraction in exposed (%) -6.20 (-33.90, 15.76)
Attrib fraction in population (%) -2.96 (-15.01, 7.82)
-------------------------------------------------------------------
Test that odds ratio = 1: chi2(1) = 0.26 Pr>chi2 = 0.61
Wald confidence limits
CI: confidence interval
* Outcomes per 100 population units
Pearson's Chi-squared test with Yates' continuity correction
data: dat
X-squared = 0.17076, df = 1, p-value = 0.6794
Advantages of contingency
:
contingency
would show the results of the Fisher exact test at the end of the output.For mhor
the formula has the following syntax:
Thus, mhor
displays the odds ratio of exposure yes against exposure no on outcome yes for different levels of stratum, as well as the Mantel-Haenszel pooled odds ratio.
Example: effect of eating chocolate ice cream on being ill by sex from the oswego
data set.
data(oswego, package = "epitools")
oswego <- oswego %>%
mutate(
ill = factor(ill, labels = c("No", "Yes")),
sex = factor(sex, labels = c("Female", "Male")),
chocolate.ice.cream = factor(chocolate.ice.cream, labels = c("No", "Yes"))
) %>%
var_labels(
ill = "Developed illness",
sex = "Sex",
chocolate.ice.cream = "Consumed chocolate ice cream"
)
First we look at the cross-tabulation:
mytable(ill ~ sex + chocolate.ice.cream, data = oswego, show.total = TRUE)
Descriptive Statistics by 'Developed illness'
____________________________________________________________________
No Yes Total p
(N=29) (N=46) (N=75)
--------------------------------------------------------------------
Sex 0.226
- Female 14 (48.3%) 30 (65.2%) 44 (58.7%)
- Male 15 (51.7%) 16 (34.8%) 31 (41.3%)
Consumed chocolate ice cream 0.127
- No 7 (24.1%) 20 (44.4%) 27 (36.5%)
- Yes 22 (75.9%) 25 (55.6%) 47 (63.5%)
--------------------------------------------------------------------
oswego %>%
mhor(ill ~ sex/chocolate.ice.cream)
OR Lower.CI Upper.CI Pr(>|z|)
sexFemale:chocolate.ice.creamYes 0.47 0.11 2.06 0.318
sexMale:chocolate.ice.creamYes 0.24 0.05 1.13 0.072
Common OR Lower CI Upper CI Pr(>|z|)
Cochran-Mantel-Haenszel: 0.35 0.12 1.01 0.085
Test for effect modification (interaction): p = 0.5434
Example: We compare the use of lung’s X-rays on the screening of TB against the gold standard test.
Freq <- c(1739, 8, 51, 22)
BCG <- gl(2, 1, 4, labels=c("Negative", "Positive"))
Xray <- gl(2, 2, labels=c("Negative", "Positive"))
tb <- data.frame(Freq, BCG, Xray)
tb <- expand_df(tb)
diag_test(BCG ~ Xray, data = tb)
Outcome + Outcome - Total
Test + 22 51 73
Test - 8 1739 1747
Total 30 1790 1820
Point estimates and 95 % CIs:
---------------------------------------------------------
Apparent prevalence 0.04 (0.03, 0.05)
True prevalence 0.02 (0.01, 0.02)
Sensitivity 0.73 (0.54, 0.88)
Specificity 0.97 (0.96, 0.98)
Positive predictive value 0.30 (0.20, 0.42)
Negative predictive value 1.00 (0.99, 1.00)
Positive likelihood ratio 25.74 (18.21, 36.38)
Negative likelihood ratio 0.27 (0.15, 0.50)
---------------------------------------------------------
Using the immediate version (direct input):
diag_test2(22, 51, 8, 1739)
Outcome + Outcome - Total
Test + 22 51 73
Test - 8 1739 1747
Total 30 1790 1820
Point estimates and 95 % CIs:
---------------------------------------------------------
Apparent prevalence 0.04 (0.03, 0.05)
True prevalence 0.02 (0.01, 0.02)
Sensitivity 0.73 (0.54, 0.88)
Specificity 0.97 (0.96, 0.98)
Positive predictive value 0.30 (0.20, 0.42)
Negative predictive value 1.00 (0.99, 1.00)
Positive likelihood ratio 25.74 (18.21, 36.38)
Negative likelihood ratio 0.27 (0.15, 0.50)
---------------------------------------------------------