Tutorial_lookup.Rmd

---
title: "Regression Commands"
author: "Chris Mainey"
date: "14/10/2019"
output:
  pdf_document
---

```{r setup, include=FALSE}
knitr::opts_chunk$set(echo = TRUE,
  comment = "#>",
  collapse = TRUE,
  fig.align = "center"
)
```

# Tutorial Cheatsheet

The tutorial on regression model uses many different commands, but most of them are simple and based on default `R` objects and methods.  Never-the-less, it takes time to learn, search and try these commands out, so this document is a summary of the commands shown in the slides.  The will use the `LOS_model` dataset from the `NHSRdatasets` package, but will not usually show the output of each command.

You can install/load the package and data with:
```{r pckload, eval=FALSE, message=FALSE, warning=FALSE}
install.packages("NHSRdatasets")

library(NHSRdatasets)
data(LOS_model)
```

# Correlation

Although we are not using correlation in the practical, it is mentioned in the slides.

```{r cor, eval=FALSE}
# Correlation coefficient for Age and LOS
cor(LOS_model$Age, LOS_model$LOS)

# With a significane test:
cor.test(LOS_model$Age, LOS_model$LOS)

# Correaltion matrix plot of the data. frame (numeric columns 3 - 5)
library(corrplot)

M <-cor(LOS_model[3:5])
corrplot(M, type="upper", order="hclust")

```

# Linear regression

Linear regression is the building block for what we are looking at, and we will treat regressions as 'object', i.e. assign each regression to a variable.

```{r reg1, eval=FALSE}
#Linear model
lm(LOS ~ Age, data=LOS_model)

# Linear model assigned to variable, this includes two parameters, 
# age and death, with death as a categorical
mod1 <- lm(LOS ~ Age + factor(Death), data=LOS_model)

# View model summary
summary(mod1)

# Exctract coefficients
coef(mod1)

# calculate confidence interval
confint(model1)

# plot the residuals
plot(mod1)


# interaction model
mod2 <- lm(LOS ~ Age * factor(Death), data=LOS_model)

# Predict from your model
predict(mod2, newdata=my_newdata)

```

# Generalized Linear Models (GLM)

GLMs share most of the architecture of linear models, with the additional 'family' argument

```{r glm, eval=FALSE}
# GLM to nodel death as binary (binomial).  We'll also mean-centre and scale then
mod3 <- glm(Death ~ scale(Age) + scale(LOS), data=LOS_model, family="binomial")

# Extract AIC
aic(mod3)

# predict on scasle of the data
predict(mod3, newdata=my_data, type = "response")

```