How to Easily Generate Predictions from Linear Models in R Using predict() and lm()

The lm() function in R is specifically designed to fit linear regression models.

Once we’ve fit a model using the observed data, we can then employ the predict() function to calculate the estimated response value of a new observation. This process is commonly referred to as scoring the model.

Syntax and Core Arguments of predict()

To effectively leverage the prediction capabilities within R, it is necessary to understand the standard syntax used by the function when applied to linear models. While the function has many optional parameters, a basic call requires the model object and, usually, a new data structure containing the input features for prediction.

This function uses the following general structure:

predict(object, newdata, type=”response”, interval=”none”, level=0.95)

The core parameters dictate how the prediction is executed and what auxiliary information is returned. Mastering these arguments is key to tailoring the prediction output to specific analytical needs.

The essential arguments are defined as follows:

• object: This is the resulting object (typically of class ‘lm’) generated by fitting the linear model using the lm() function. This object contains all the necessary coefficients and model parameters required for the prediction calculation.
• newdata: This highly crucial argument specifies a data frame containing the values of the predictor variables at which predictions are required. If newdata is omitted, the function defaults to predicting the values used to fit the original model.
• type: This determines the type of prediction output. For lm() models, the default is “response”, which provides the predicted mean value of the response variable. This is the output most frequently utilized in standard linear regression.

Practical Demonstration: Setting Up the Data

To illustrate the application of predict() with lm(), we will use a dataset detailing basketball player statistics. Our goal is to predict the number of points a player scores based on minutes played and the number of fouls committed. This requires defining a specific dataset in R as a data frame.

Suppose we generate the following sample data frame in R, which captures information across ten different basketball players. This step ensures we have the necessary input variables (predictors) and output variable (response) before modeling begins.

#create data frame
df <- data.frame(minutes=c(5, 10, 13, 14, 20, 22, 26, 34, 38, 40),
                 fouls=c(5, 5, 3, 4, 2, 1, 3, 2, 1, 1),
                 points=c(6, 8, 8, 7, 14, 10, 22, 24, 28, 30))

#view data frame
df

   minutes fouls points
1        5     5      6
2       10     5      8
3       1       3      8
4       14     4      7
5       20     2     14
6       22     1     10
7       26     3     22
8       34     2     24
9       38     1     28
10      40     1     30

This data frame, named df, contains our response variable, points, and our two predictor variables, minutes and fouls. The next step involves specifying the structure of the linear model we wish to estimate from this data.

Fitting the Multiple Linear Regression Model

We aim to model the relationship where points is a linear function of minutes and fouls. This is expressed mathematically as a multiple linear regression equation:

points = β₀ + β₁(minutes) + β₂(fouls) + ε

Here, $beta_0$ represents the intercept, and $beta_1$ and $beta_2$ are the coefficients associated with minutes and fouls, respectively. We use the lm() function to estimate the optimal values for these coefficients that best fit our observed data, thereby minimizing the sum of squared residuals.

We execute the model fitting using lm() and store the result in an object named fit:

#fit multiple linear regression model
fit <- lm(points ~ minutes + fouls, data=df)

#view summary of model
summary(fit)

Call:
lm(formula = points ~ minutes + fouls, data = df)

Residuals:
    Min      1Q  Median      3Q     Max 
-3.5241 -1.4782  0.5918  1.6073  2.0889 

Coefficients:
            Estimate Std. Error t value Pr(>|t|)    
(Intercept) -11.8949     4.5375  -2.621   0.0343 *  
minutes       0.9774     0.1086   9.000 4.26e-05 ***
fouls         2.1838     0.8398   2.600   0.0354 *  
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Residual standard error: 2.148 on 7 degrees of freedom
Multiple R-squared:  0.959,	Adjusted R-squared:  0.9473 
F-statistic: 81.93 on 2 and 7 DF,  p-value: 1.392e-05

The summary output confirms the strength of the relationship, highlighted by the high Multiple R-squared value of 0.959. Using the coefficients estimated by lm(), we can formally write the derived regression equation that forms the basis of our predictions:

points = -11.8949 + 0.9774(minutes) + 2.1838(fouls)

This equation is now parameterized and ready to be used by the predict() function to estimate outcomes for players not included in the original training data.

Generating Predictions for New Observations

The true utility of a regression model lies in its predictive capability. We now apply the fitted model object (fit) to a hypothetical new player using the predict() function. Suppose we want to determine the expected points for a player who logs 15 minutes and commits 3 fouls.

To perform this prediction, we must construct a newdata data frame containing this specific observation. Crucially, the column names in this new data frame must exactly match the predictor names (‘minutes’ and ‘fouls’) used in the original lm() call.

We define the new observation and then pass it to the predict() function along with the fitted model object:

#define new observation
newdata = data.frame(minutes=15, fouls=3)

#use model to predict points value
predict(fit, newdata)

       1 
9.317731

The output indicates that the model predicts this player, with 15 minutes played and 3 total fouls, will score 9.317731 points. This result represents the point estimate of the mean predicted response based on the linear relationship established by the training data.

Predicting Multiple Observations Simultaneously

The predict() function is highly versatile and optimized to handle batch predictions efficiently. Instead of predicting one observation at a time, we can provide a newdata data frame containing multiple rows, where each row represents a distinct new observation for which we require a prediction.

For instance, we can predict the points scored for three different hypothetical players: Player A (15 minutes, 3 fouls), Player B (20 minutes, 2 fouls), and Player C (25 minutes, 1 foul). This scenario showcases how the predict() function handles vectorized inputs seamlessly.

The following code demonstrates how to define the batch of new observations and execute the prediction using the fitted lm() object:

#define new data frame of three players
newdata = data.frame(minutes=c(15, 20, 25),
                     fouls=c(3, 2, 1))

#view data frame
newdata

  minutes fouls
1      15     3
2      20     2
3      25     1

#use model to predict points for all three players
predict(fit, newdata)

        1         2         3 
 9.317731 12.021032 14.724334 

The result is a vector of predicted values, corresponding positionally to the rows in the newdata data frame. Here is the interpretation of the output for each hypothetical player, rounding to two decimal places for practical use:

• The predicted points for the player with 15 minutes and 3 fouls is 9.32.
• The predicted points for the player with 20 minutes and 2 fouls is 12.02.
• The predicted points for the player with 25 minutes and 1 foul is 14.72.

Advanced Usage: Incorporating Prediction and Confidence Intervals

While point predictions provide an estimated mean value, statistical rigor often demands quantifying the uncertainty around these estimates. The predict() function facilitates this by allowing the computation of confidence intervals and prediction intervals using the optional interval argument.

A confidence interval quantifies the uncertainty in the estimate of the mean response for a given set of predictors. These intervals are suitable when estimating the average outcome for a large group of individuals who share the same characteristics. We compute this by setting interval="confidence", and the confidence level can be adjusted using the level argument (default is 0.95).

In contrast, a prediction interval quantifies the uncertainty around a single future observation. Because single observations include the model’s residual error, prediction intervals are inherently wider than confidence intervals for the same input values. This is the preferred interval type when forecasting an individual outcome. To calculate this, we set interval="prediction".

Critical Note on Data Structure: The newdata Requirement

One of the most frequent sources of errors when using the predict() function is a structural mismatch between the newdata data frame and the structure of the data used to train the original lm() model. Strict adherence to variable naming, data types, and ordering is mandatory.

Specifically, the names of the columns in the new data frame must exactly match the names of the predictor variables used in the formula passed to lm(). In our earlier example, the model fitting relied on the predictor variables:

• minutes
• fouls

Thus, when we created the new data frame called newdata we made sure to also name the columns:

• minutes
• fouls

If the column names are misspelled, miscapitalized, or if a required variable is missing, R will fail to evaluate the model equation correctly. This frequently results in a cryptic error message:

Error in eval(predvars, data, env)

This error message indicates that R cannot find the variables it needs within the provided environment (the newdata object) to apply the model coefficients. Always keep this structural requirement in mind when preparing data for prediction using the predict() function.

Summary of Key Predict() Concepts

The predict() function is an indispensable component of statistical modeling in R, especially when working with objects created by lm(). It provides the mechanism to extrapolate the findings of the training data onto new, unobserved inputs.

Successful implementation relies on three key principles: first, ensuring the model object is correctly fitted using lm(); second, providing the prediction input via the newdata argument, structured identically to the training data; and third, selecting the appropriate prediction type (e.g., simple response, confidence interval, or prediction interval). Mastering these elements ensures powerful and reliable forecasting capabilities within the R environment.