Interactions

library(vigicaen)
library(rlang)
library(dplyr)

Introduction

It is possible to explore interactions between drugs on an adr reporting.

This tutorial does not aim at covering the concepts underlying interactions in pharmacovigilance. It is about running them in practice.

In particular, we will not cover the differences between additive interactions, statistical/synergistic interactions.

compute_interaction() use is shown at the end of the vignette.

We use built-in example dataset.


# ---- Tables ---- ####

demo <- demo_
drug <- drug_

# ---- Dictionary step ---- ####

d_drecno <- ex_$d_drecno
a_llt <- ex_$a_llt

# #### Data management #### ####

# ---- Drugs ---- ####

demo <-
    demo |>
    add_drug(
      d_code = d_drecno,
      drug_data = drug_
    ) 
#> ℹ `.data` detected as `demo` table.

# ---- Adrs ---- ####

demo <- 
  demo |>
    add_adr(
      a_code = a_llt,
      adr_data = adr_
    )
#> ℹ `.data` detected as `demo` table.

# ---- Sex ---- ####

demo <- 
  demo |> 
  mutate(
    sex = case_when(Gender == "1" ~ 1,
                    Gender == "2" ~ 2,
                    TRUE ~ NA_real_)
    )

Additive interactions

Multivariate analysis

Additive effect of two covariates can be obtained by multiplying the Odds-Ratio of each.

mod3 <- glm(a_colitis ~ ipilimumab + sex,
            data = demo,
            family = "binomial")

mod_or <- 
  compute_or_mod(
  summary(mod3)$coefficients,
      estimate = Estimate,
    std_er = Std..Error
  ) |> select(rn, orl, ci, up_ci)

mod_or
#>             rn    orl          ci     up_ci
#>         <char> <char>      <char>     <num>
#> 1: (Intercept)   0.15 (0.08-0.28) 0.2830974
#> 2:  ipilimumab   2.00 (1.14-3.53) 3.5274694
#> 3:         sex   1.06 (0.69-1.62) 1.6167786

With reporting Odds-Ratio of ipilimumab being 2.00 and the reporting Odds-Ratio of sex being 1.06, the additive effect of both is 2.00 * 1.06.

Subgroup comparisons

Some way to approach multiplicative interactions is to compare the disproportionality signal in subgroups.

The compute_dispro() function can be used for these analyses, assuming the initial dataset is filtered on the appropriate subgroup.

Say we want to investigate the interaction between ipilimumab and nivolumab and colitis reporting.

demo |> 
  filter(nivolumab == 1) |> 
  compute_dispro(
    y = "a_colitis",
    x = "ipilimumab"
    )
#> # A tibble: 1 × 9
#>   y         x          n_obs n_exp or    or_ci          ic ic_tail ci_level
#>   <chr>     <chr>      <dbl> <dbl> <chr> <chr>       <dbl>   <dbl> <chr>   
#> 1 a_colitis ipilimumab    18  11.5 2.36  (1.18-4.73) 0.620  -0.123 95%

The overall analysis implies to perform additional analysis in different settings.

In our example:

Both IC and ROR can be used here.

Statistical interactions

Logistic regression model

The true statistical interaction is obtained with the following model

mod4 <- glm(a_colitis ~ ipilimumab + sex + ipilimumab * sex,
            data = demo,
            family = "binomial")

compute_or_mod(
  summary(mod4)$coefficients,
  estimate = Estimate,
  std_er = Std..Error
)
#>                rn   Estimate Std..Error   z.value     Pr...z..         or
#>            <char>      <num>      <num>     <num>        <num>      <num>
#> 1:    (Intercept) -2.2464633  0.3632657 -6.184077 6.246672e-10  0.1057727
#> 2:     ipilimumab  2.7892022  0.8927900  3.124141 1.783247e-03 16.2680357
#> 3:            sex  0.2910744  0.2363515  1.231532 2.181241e-01  1.3378641
#> 4: ipilimumab:sex -1.4905928  0.6289493 -2.369973 1.778940e-02  0.2252391
#>        low_ci      up_ci    orl           ci ci_level signif_ror
#>         <num>      <num> <char>       <char>   <char>      <num>
#> 1: 0.05189925  0.2155687   0.11  (0.05-0.22)      95%          0
#> 2: 2.82742376 93.6007499  16.27 (2.83-93.60)      95%          1
#> 3: 0.84183934  2.1261543   1.34  (0.84-2.13)      95%          0
#> 4: 0.06565701  0.7726920   0.23  (0.07-0.77)      95%          0

Bayesian Information Component, compute_interaction()

The formula for the interaction between 3 variables (\(y\), the event of interest, \(x1\) and \(x2\), two potential explanatory factors) in information component is

\(log_2\frac{n_{y, x1, x2}}{n.expected_{interaction}}\)

with \(n.expected_{interaction}\) equal to

\(\frac{n_{x1, x2} * n_{y, x1} * n_{y, x2} * n.total}{n_{x1} * n_{x2} * n_y}\)

The parameters are read as follows

Parameter Case
\(n_{x1}\) number of cases reporting x1
\(n_{x1, x2}\) number of cases reporting x1 AND x2
\(n_{y, x1, x2}\) number of cases reporting x1 AND x2 AND y
\(n.total\) total number of cases in the study population

The credibility interval is calculated as for the usual IC.

compute_interaction() produces this interaction statistic.


demo |>
  compute_interaction(
    y = "a_colitis",
    x = "ipilimumab",
    z = "nivolumab"
  )
#> # A tibble: 1 × 8
#>   y         x          z         n_obs n_exp    ic ic_tail ci_level
#>   <chr>     <chr>      <chr>     <dbl> <dbl> <dbl>   <dbl> <chr>   
#> 1 a_colitis ipilimumab nivolumab    18  7.88  1.14   0.400 95%