slides_codealong.qmd

---
title: "code-along"
format: html
editor: visual
---

## Example Dataset: A Brief Overview

We'll explore the relationship between physical activity and pulse in the National Health and Nutrition Examination Survey (NHANES) dataset, which has been modified for this course with simulated variables.

To try this out in practice, we load the NHANES dataset, only keep the variables we need for the example.

The variables used in the dataset:

-   exposure of interest (a): physical activity (continuous variable)

-   outcome (y): Pulse (continuous variable)

-   outcome (y2): Diabetes(binary variable)

-   mediator (m): BMI (continuous variable)

Note that the example is created to demonstrate the methods, and there is no clinical relevance of the results. For the convenience of analysis, we have transformed the mediator (m) and outcome variable (y) so that they follow a normal distribution.

Four demographic variables are considered in the analysis: age, gender, education, and smoking. We assume that adjusting for these four confounding is sufficient to block the backdoor paths.

## 

```{r}
# load the library
library(readr)
# load the data 
nhanes <- read_csv(here::here("data/NHANES.csv")) 
library(dplyr) 
# keep the variables used in the analyses 
data <- nhanes %>% 
  dplyr::select(ID, w1 = Age, w2 = Gender, w3 = Education, w4 = Smoke100, a = PhysActive, m = BMI, y = Pulse, y2 = Diabetes ) %>% 
  na.omit()
```

## Continuous outcome

```{r}
# Physical activity is a binary variable, dichotomized as 'Yes' or 'No'. 
table(data$a)
# check the distribution of outcome 
summary(data$y)
hist(data$y)
# check the distribution of mediation
summary(data$m)
hist(data$m)
# check the education categories and relevel 
table(data$w3)
data$w3 <- relevel(factor(data$w3), ref = "College Grad")
```

i\) run a model for the mediator only adjusting for confouding factors.

```{r}
lm_m <- lm(m ~ a + w1 + w2 + w3 + w4, data = data)
summary(lm_m)
```

We see that physical activity (a) is associated with lower BMI.

ii\) run the model for the outcome, including an interaction term between exposure and mediator.

```{r}
lm_y <- lm(y ~ a + m + a:m + w1 + w2 + w3 + w4, data = data)
summary(lm_y)
```

In this model we can see:

-   physical activity reduces pulse rate, independent of BMI (m);

-   BMI is positively associated with pulse rate (y);

-   There is interaction between physical activity (a) and BMI (m)

    ### We will use the CMAverse R package to conduct the mediation analyses.

First, we load the package and then we setup the model object. We have to specify:

-   model: this is the type of model, or the approach for the causal mediation. We used the regression-based approach, so we will also do that here. However, one can also use weighting-based approach marginal structural models or the gformula.

-   outcome: your outcome variable

-   exposure: your exposure/intervention/treatment

-   mediator: your mediator

-   mreg: the type of model for the mediator. It is a list because you can have multiple mediators

-   yreg: type of regression for the outcome. E.g., linear, logistic, cox.

-   astar: Control/comparison intervention (value from)

-   a: Intervention (value to)

-   mval: the value for the mediator when estimating the CDE

-   EMin0t: indicator of whether there should be exposure and mediator interaction in the models.

-   estimation: method for estimating causal effects. Can use parametric functions or imputation.

-   inference: how to calculate confidence intervals. For regression-based we use the delta method. But for the other methods we use a bootstrap.

Load the library:

```{r}
## install.packages("CMAverse")  #reinstall the package if the library is not loaded
library(CMAverse)
```

Build the models:

-   Since both the mediator(BMI) and outcome(pulse rate) are continuous, we use linear regression models.

-   

-   We set mediator (BMI9 values to 25kg/m2, suppose mediator could be intervened to normal weight

-   I will first pretent there is no interaction, set EMint to FALSE.

```{r}
res_rb <- cmest(
  data = data, model = "rb", outcome = "y", exposure = "a",
  mediator = "m", basec = c("w1","w2","w3","w4"), EMint = FALSE,
  mreg = list("linear"), yreg = "linear", #specifiy model types
  astar = 0, a = 1, mval = list(25),
  estimation = "paramfunc", inference = "delta"
)

summary(res_rb)
```

Now change EMint to TRUE to allow for interaction.

```{r}
res_rb <- cmest(
  data = data, model = "rb", outcome = "y", exposure = "a",
  mediator = "m", basec = c("w1","w2","w3","w4"), EMint = TRUE,
  mreg = list("linear"), yreg = "linear", #specifiy model types
  astar = 0, a = 1, mval = list(25),
  estimation = "paramfunc", inference = "delta"
)

summary(res_rb)
```

## Binary outcome

Suppose now i am interested in how physical activity influences the risk of diabetes. I will first check the proportion of diabetes in the sample.

`{# y2 is diaebtes} table(data$y2)`

```{r}
# recode diabetes to 1(Yes) and 0(No)
data$y2 <- case_when(
  data$y2 == "Yes" ~ 1,
  data$y2 == "No" ~ 0
)
```

```{r}
res_rb <- cmest( data = data, model = "rb", outcome = "y2", exposure = "a", mediator = "m", basec = c("w1","w2","w3","w4"), EMint = TRUE, 
mreg = list("linear"), 
yreg = "logistic",  mval = list(25), estimation = "paramfunc", inference = "delta" )

summary(res_rb)
```

We can see that a great proportion of the effect of physical activity on diabetes risk is mediated through BMI (65%). undefinedThe remaining \~35% would represent physical activity's direct effect on diabetes risk through other pathways