The first thing to understand about multiple regression is that it uses several variables to predict a single continuous variable. In your Intro Stats course, you probably learned about simple linear regression, or how a single predictor variable impacts a single outcome variable. Multiple regression is going to be similar to that, but instead of one predictor, it’s going to be many (or at least more than one!).
The second thing to know is how to make good decisions as you build your model. You might be tempted to add in a lot of predictors so you have a really thorough model, but that isn’t always a great choice. You will need to learn about multicollinearity, model overfit, and interpreting findings.
The third thing you will want to understand about multiple regression is that it uses 𝛽 beta weights, which are sort of like the regression coefficients you learned about with simple linear regression, but are standardized so you can more easily compare which variables are having the biggest impact on your outcome.
Let’s dive in.
As psychologists, our goal is to understand and predict human behavior, which often involves lots of complex and interconnected feelings, thoughts, and behaviors (aka, it’s messy!), not just one. So at its core, multiple regression asks: "How much of my outcome variable can be explained by a set of predictor variables?” This means that regression is a helpful tool because it allows us to model more complex (and thus more realistic) relationships between variables.
As with the other analyses in this blog, the APA Dictionary is a good place to start when you spot a statistical analysis that is new to you. In our case, APA defines multiple regression as “a statistical technique for examining the linear relationship between a continuous dependent variable and a set of two or more independent variables. It is often used to predict a single outcome variable from a set of predictor variables.”
Let’s break down what that means.
In simple regression, we use a single predictor to estimate a single outcome. For example, using stress to predict depression. This is a good start and can give us a rough “ballpark” estimate, but humans are complex and behavior is rarely explained or influenced by just one thing. For instance, depression is influenced by many variables beyond just stress, things like social support, sleep, physical health, and coping strategies (just to name a few!).
Thus, multiple regression builds on this idea of human complexity by letting us include several predictors at the same time. In our example of depression, we could assess multiple predictors at once (social support, sleep, physical health, coping strategies). This gives us a stronger and more realistic picture of the outcome we’re trying to understand. In other words, instead of assuming one factor impacts behavior, multiple regression helps us capture how a combination of factors work together to shape the outcome.
Let’s consider an example.
Is social media use related to life satisfaction? In a simple linear regression, we would look at one predictor variable, perhaps the number of hours per day spent on social media and its impact on self-reported life satisfaction. This approach is really helpful for understanding the relationship between social media and life satisfaction. However, we also know several other factors are going to predict how satisfied someone is with their life; it’s not just social media. So if we wanted to more accurately model life satisfaction, we might include additional variables, such as perceived social support,and/or stressful life events.
The next thing to consider about multiple regression (and its use of multiple predictor variables) is that each variable will have a different relationship to your outcome variable; some will have a stronger relationship and some will have a weaker relationship. Some will have a positive relationship (they are associated with an increase in the outcome) and some will have a negative relationship (they are associated with a decrease in the outcome). The cool thing is that multiple regression can actually quantify the relationship in ways that help us understand the relative ‘weight’ of each variable. In our life satisfaction example, perceived social support might have a stronger impact on life satisfaction and thus more ‘weight’ than social media use.
In sum, human behavior is complex, and multiple regression is one tool that can help us model relationships where there are two or more predictors. Below are some examples where multiple regression might be used across disciplines.
Human Resources/Organizational Psychology: How do job satisfaction, leadership satisfaction, and number of hours worked predict employee burnout?
Education: How do school belonging, extracurricular participation, and peer support predict academic persistence?
Marketing: To what extent do product price, customer satisfaction, and perceived product quality predict customer spending?
Political Science: How do education level, political interest, and income predict individuals' likelihood of voting?
Clinical/Counseling Psychology: How do childhood adversity, current social support, and coping skills predict psychological well-being in adulthood?
As you can see, multiple regression is a useful analysis not just in psychology, but in numerous fields where we are looking at multiple predictors! It’s a very handy analysis to know about.
Okay, you’ve got the overview of multiple regression: we get to have multiple predictors for a single continuous outcome. Let’s add some technical knowledge that will help you read through articles that report their results from a multiple regression. You are already familiar with two important statistical concepts on which factor analysis builds: correlations and simple linear regression.
Review Topic #1: Correlations & Correlation Matrices
You learned in your stats class that two variables can be correlated. Or not. You may also have learned about a correlation matrix that looks at how all of the variables in a set are correlated with one another. In multiple regression, you will always start by looking at your bivariate correlations. Knowing about correlations and interpreting correlation matrices will help you build a better regression model.
As a review, look over the correlation matrix (below) that shows relationships between several predictors of health promoting behaviors in college students, including recent stressful events, perceived stress, and hardiness (Cress & Lampman, 2007). As you can see, the majority of correlations among these measures are relatively moderate, though only some are statistically significant (you can tell because there is an asterisk next to them). You’ve probably seen correlation matrices like this one in your intro stats class.
What should you be looking for in the correlation matrix?
At the most general level, you want to make sure your variables are relating to each other in the ways you expected from the literature; it’s an initial check that the pattern of relationships match the expectations from your literature review. But for multiple regression specifically, you are focusing on the strength of relationships because the analysis relies on an assumption we call multicollinearity.
What is multicollinearity? Again, our APA dictionary comes in handy, defining it as “the state that occurs when several independent variables are extremely highly interrelated, making it difficult to determine separate effects on the dependent variable.” That is, when our predictor variables are too similar to each other, we can’t tell their effects apart from one another. So if you tried to use GPA, GRE scores, and SAT scores to predict future job performance, having all three wouldn’t be very useful because GPA, GRE, and SAT scores are too similar to each other, creating redundancy (which we call multicollinearity). When you look at your correlation matrix and find overlapping variables in your research, that is when you want to consider removing or combining them.
Hence, you want to use variables that are related to your dependent variable, but not strongly to each other. So in terms of the job performance example from above, perhaps you use GRE scores, conscientiousness, and years experience with the job. Those three variables might have some overlap, but each is going to provide a unique contribution to your model and therefore give you a better model.
Review Topic #2: Simple Linear Regression (SLR)
You probably learned about SLR in your stats class, but here is a quick review. SLR is sort of like a fancy correlation; it models the relationship between a predictor variable and a dependent variable while assuming the relationship is linear. One of the key elements of SLR is the line of best fit, which is the line that comes as close to as many of the data points as possible with the least amount of error. It won’t model the relationship perfectly (sometimes it will estimate too high, other times too low), but it gives us an estimate with the smallest amount of error. For example, let’s imagine you want to predict how well students will do on their stats exam based on how much they study. The SLR won’t give you a perfect estimate, but it might give you an idea of the scores students will get based on how much they studied.
As you will see in a lot of textbooks, SLR is expressed in a formula to help you estimate your outcome.
Mathematically, the formula is Y = b0 + b1X
But this isn’t a textbook and I kind of hate formulas, so let’s do a quick review of what the different parts tell us.
Y is our outcome variable, the very thing we want to understand and predict. In our hypothetical example, it’s the grade on the stats exam. But Y variables can be anything that is a continuous variable, like how much money people earn, how satisfied they are with their life, or the severity of their depression. It just depends on the research question!
X is our predictor variable, the variable that we think impacts or explains our outcome variable (Y). In our example, our predictor (X) variable is how many hours students studied for their stats exam.
Next, let’s look at the intercept (b0), which is our expected outcome when our predictor (X) is zero. In our stats exam example, what score do we expect if a student does not study at all (zero hours)? For our purposes, let’s say that when students do not study at all, they average a 56% on the exam. That is our intercept.
Finally, the slope (b1; also called the regression coefficient) is the delightful part of the equation because it reflects the rate of change based on an increase in our predictor (X). In our example, how many additional points can students expect to earn for each hour of studying? Here, let’s say that for each additional hour a student studies, their score improves by an average of three percentage points. So if they study for 5 hours, their score will be about 15% higher; if they study 10 hours, it will be about 30% higher.
Regression coefficients are important because they can tell us the direction of the relationship. Similar to correlations, positive coefficients indicate that as the predictor increases, so does our outcome. In our case, as hours studied increases, so too does exam score. But not all coefficients are positive (e.g., you could imagine that as absences increase, exams scores go down; this would have a negative regression coefficient).
Putting it all together. At the end of our stats exam example, our equation might look like this: Y = 56 + 3(X).
Let’s go interpret this together. Y is the exam score we expect students to earn based on a baseline intercept of 56%, plus an additional 3% for every additional hour they study. Using this equation, we see that if a student doesn’t study at all, they would score about 56%. If they study for 5 hours, we expect them to score about 71% (the 56% baseline + and additional 15% from studying for 5 hours, where for each additional hour studied earns 3% more, on average).
If you want to check your knowledge, try and estimate your score if you study for 10 hours.
Review Topic #3: Correlation is not Causation
Similar to what you learned about SLR in your intro stats class, labeling your input variables as independent/predictor and dependent/outcome variables does not actually make them causal. Much SLR and MLR research is correlational, not causal. So don’t be tricked - just because we talk about the variables as predictors or independent variables does not make them causal; only the research design can determine the degree to which we make causal inferences about the model.
Okay, review session complete. Let’s add to it with multiple linear regression.
Before we jump into interpreting MLR in published articles, let’s go over an example and learn some new key terms. For this example, we will stick with trying to predict our score on our stats exam. Except this time, we won’t just look at the number of hours studied, we will also look at class attendance. This gives us two predictors to help us predict exam scores (and ideally, improve the precision of our estimate - two variables will hopefully help us explain exam scores better than just one).
Now, imagine we go around and collect our data (ask students how long they studied, then get attendance records and exam scores from the professor). We analyze our data and need to look over the results. As you learned about, the first thing we need to look at are the biserial correlations. We want to make sure that attendance and studying are both related to our outcome (exam score) but not too strongly related to each other. If attendance and studying are strongly related to each other, they are multicollinear and that hurts the overall quality of our model. For our purposes, let’s assume they are NOT multicollinear and that we can proceed with our MLR model. We run the analysis and need to interpret the output and model. Regardless of which data analysis tool you use (JASP, SPSS, etc.), the output will have consistent pieces of information, including a model summary, standardized and unstandardized coefficients, t-statistics, and p-values for each predictor.
In my opinion, here is where it gets a little hairy. Because the table provides both standardized or unstandardized coefficients, it can be confusing to new folks to tell them apart and successfully interpret them. Let’s talk about each one.
Unstandardized (B) coefficients: This approach keeps the variable in its original units. So if we are talking about how much students study, it would keep the estimate in the unit of hours studied. If we are talking about absences, it would keep the estimate in the number of classes missed. This can be a useful approach because it can be intuitive for people to interpret (e.g., for every hour studied, you can expect a 3% increase in exam score). However, what makes the unstandardized estimates difficult in multiple regression is that the predictors tend to be on different scales, which makes it hard to compare across variables. In turn, different relative scales make it hard to figure out which variable is having a bigger impact on our outcome of interest. That is, you can’t look at the size of the unstandardized coefficients and compare them (it would be like trying to meaningfully compare apples to staplers; they are completely different). Thus, another way to evaluate our predictors is by standardizing them (putting them all on the same scale) so we can see which ones are most important (allowing us to compare apples to apples).
Standardized coefficients (represented by β): We call standardized scores beta weights. Beta weights tell us the relative weight (importance) of each predictor variable in our MLR model. Standardized beta weights are helpful because they measure the relative strength and direction of each predictor variable on our outcome/dependent variable using standard deviation units, with all other predictors held constant. In a nutshell, standardized beta weights put all variables on the same scale and we can now directly compare across predictors (again, apples to apples). Stated simply, the predictor with the largest absolute Beta value is going to have the strongest relative effect on the dependent variable. In our example using the standardized beta weights allows us to compare hours studied to class days missed; two variables on different scales (hours vs. days). Using the standardized coefficients allows us to put them on the same scale and compare which is a stronger predictor of exam scores.
Important sidenote: You can see that unstandardized coefficients are represented with a B and standardized coefficients are represented with a fancy β; this often causes confusion for students because both are represented by “B”s, but you have to pay attention to which is which because we interpret them very differently. With our standardized coefficient (Fancy β), we can directly compare the strength and direction of our predictors. On the other hand, unstandardized coefficients are nice because they keep our predictors in their original units (dollars, hours, pounds, days, etc.), which can make them easier for practical interpretation. So your takeaway should not be that one approach is better than the other, but that standardized and unstandardized coefficients have different interpretive uses.
Statistical significance. In addition to the coefficients, the results will also give you an estimate for the t-statistics and associated p-values for each individual predictor. This should feel familiar to you from Intro Stats, where you learned that when the p-value is less than alpha, results are significant. In this case, there will be a p-value for each predictor, informing you which predictors in your model are significant and which ones are not. In our example, it could be possible that studying significantly predicts exam scores but absences do not (or vice versa), both significantly predict, or neither.
Finally, results will include a model summary, with estimates for the overall model fit, including R² (proportion of variance explained) and Adjusted R².
Understanding R². When we talk about R² and the “fit” of an MLR model, we are talking about variance explained, or how much of the fluctuations in our outcome variable (Y) is explained by all of our predictor variables combined. In our case, how much of the differences in exam scores can be explained by both hours studied and class absences?
The variance explained can range from 0% to 100%; where 0% means our predictors variables are not predicting the outcome variable at all and 100% means that our predictors are perfectly predicting the outcome variable. As a guideline, higher values indicate a better fit (we are better at predicting our outcome) and lower values indicate weaker fit (our model is less effective at predicting our outcome). With our example, it could be that studying and absences don’t predict the differences in exam scores at all (R² = 0%), very little (R² = 10%), quite a bit (R² = 35%), or they perfectly predict exam scores (R² = 100%). As you can imagine, in psychology. we rarely find either end of the spectrum (0% or 100%), but rather degrees of prediction.
Adjusted R². Without going into all the details, looking only at R² has some drawbacks, namely that adding more variables will increase the variance explained, even if the predictors aren’t actually relevant (e.g., adding height to our exam score example). This leads to an inflated value that isn’t necessarily reflective of the quality of our model. To correct for this, we use adjusted R², which takes into account the number and quality of predictors, as well as the sample size. So with adjusted R², you need to think carefully about which variables to include, since the variance accounted for will increase if you have selected strong predictors, but can decrease if you try to include too many unrelated (or poorly related) variables.
That is, the difference is that R² can be manipulated by adding in lots of variables and creating an inflated coefficient, but adjusted R² is going to account for those variables and will actually go down if you are adding predictors to the model that aren’t actually meaningful (i.e., don’t throw in everything but the kitchen sink; be judicious and thoughtful about which variables to include).
We’ll start with a (fictional) Table 2 from our example and interpret:
Interpreting the intercept. In this case, we can see that the intercept is 67, which we would interpret to mean that the average score for a student that studied zero hours and had zero absences would be a 67% on the exam. Note that we interpret the intercept holding all predictor variables at zero.
Interpreting the Unstandardized Coefficients. If we look at the unstandardized coefficients first (B), we see that for every hour increase in studying, exam scores go up by about three percentage points (3.05) holding all other variables constant. Conversely, for every class that a student misses, their exam score goes down by about 2 points (-2.18; the negative sign means the scores decrease), holding all other variables constant.
However, hours studied and number of class absences are not on the same scale (hours vs. days), so it can make it hard to see which one has the bigger impact. Thus, standardized beta (β) helps us because it converts our predictors into standard deviation units, which gives us the relative magnitude of each predictor. Then we can make direct comparisons!
Interpreting the Standardized Coefficients. If we look at the standardized Beta coefficients on our output (look at the fancy β column), the number of hours studied has a β coefficient of .60 and number of days absent has a β coefficient of -.29. This indicates that a one standard deviation increase in hours studied is associated with a .60 standard deviation increase in exam scores (when absences are held constant). Conversely, a one standard deviation increase in class absences is associated with a -.29 standard deviation decrease in exam scores (when hours studied are held constant). So, the key take away is that when holding all other variables constant, studying has a bigger impact on exam scores than absences.
Interpreting Model Significance.Okay, you have your model and the results from the standardized and unstandardized estimates. The next question is, are the results significant?
Let’s look at the p-values for the individual predictors (rows two and three of Table 2) and interpret their statistical significance. To evaluate this, you will follow the same rule you learned in Intro Stats: when our p-value is less than alpha, results are statistically significant.
In this case, the p-value for hours studied is .004. Since .004 is less than alpha (let’s say alpha was .05), then accounting for all other effects in the model, hours studied is a significant predictor of exam score.
Absences are also a significant predictor of exam score, since the p-value for that predictor (.01) is less than alpha.
But does the model as a whole predict our outcome?
The p-value on the first line (intercept) is going to be a reflection of whether or not your model as a whole is significant (as opposed to individual predictors).That is, it tells us that combining these independent variables together predicts our dependent variable better than chance alone. In our example, it means that using hours studied and number of absences predicts exam score significantly better than chance alone. How do we know? In this case, the p-value for our model is less than .001, so results are significant, meaning our model predicts exam scores better than chance alone.
Interpreting Explained Variance. Finally, we look at how well our model explains the variance (fluctuations) in our dependent variable (exam scores), which researchers refer to as model fit. As mentioned above, we can look at the coefficient of determination (R²) as well as the adjusted R².
Let’s look at the R² first, where our results indicate that R² = .68. What does this mean? It means that given our dataset, 68% of the variability in exam scores can be explained by how much students studied and many times they missed class. And we know this is a significant amount because we learned from the previous paragraph that our model was significant (p < .001).
If we look at the adjusted R², the results look similar; adjusted R² = .67, which indicates that 67% of the variability in exam scores can be explained by how much students studied and how many times they missed class.
In this example, the R² and adjusted R² estimates are pretty similar, but that will not always be the case. Remember, adjusted R² accounts for the number of predictors and penalizes the model when there are too many predictors that are not meaningfully contributing to the model. In our example, we only had two predictors, so it is not surprising the R² estimates are similar. However, if we added lots of additional predictors (test anxiety, how much sleep students got, previous math experience, stats anxiety, difficulty of exam, student GPA, etc.), our model would look different. The R² would reward all the predictors and begin to become artificially inflated, which researchers refer to as overfitting the model. However, the adjusted R² would account for the number and quality of predictors (as well as sample size) and produce a more conservative estimate. So, when the R² is much higher than the adjusted R², that might be an indicator that your model has too many or low quality predictors.
So what is your takeaway with this example? That both studying and missing class impact stats exam scores, though our standardized beta weights suggest that studying has a bigger impact than attendance. And don’t forget, based on how data were collected, this is only a correlational finding, not a causal one. Thus, it could be something else entirely - perhaps something like motivation that is the meaningful predictor. So be cautious about your interpretation of multiple regression results. You cannot infer causality simply because we used words like predictor and outcome.
Now that we have reviewed the basics, you’re ready to learn how to interpret multiple regression findings in a published journal article! To do this, let’s walk through an example using an open-access article.
Before continuing, go ahead and download this article about what factors predict who texts during class (Olmsted & Terry, 2014).
Read the article for yourself, but here is the gist: The researchers wanted to understand student texting behavior during their classes, with an emphasis on looking at different factors that might predict texting while in class (Olmsted & Terry, 2014). In this area, they identified several variables they thought would predict student texting behavior: the size of a student’s social network, how frequently they text in other contexts (e.g., while studying or driving), intrusive thoughts when cell phone access is limited, multi-tasking preferences, and mindfulness. To address their research question, they used Multiple Linear Regression (MLR).
Read through their results section. As with most articles, they begin with basic descriptives and other statistics you are familiar with from Intro to Stats. As they begin their key topic, predictors of texting in class (page 187), let’s walk through how to interpret their results. As you read, find the language we reviewed (above) in their article.
The analysis and general findings. Most MLR papers will mention the bivariate correlations and, ideally, a nod to the underlying assumptions of MLR (normality, multicollinearity, linearity, etc.). They should tell you which variables were included in the model and whether or not results (model fit) were significant. They will typically say something like “the model accounted for 45% of the variance in Y” or “45% of the variance in Y was explained by the model.” They will also state which predictor variables were significant and which ones were not, and provide beta weights as evidence.
As for Olmstead & Terry (2014), they state “The multiple regression analysis produced a significant model that accounted for approximately 33% of the variance in self-reported frequency of texting in the classroom, R2 = .33, R2 adj = .30, F(13, 427) = 15.78, p < .001.”
Let’s tackle this.
First, they are telling you their model explains variance in student texting; that their predictors (social network size, texting behaviors, intrusive cell phone thoughts, multi-tasking preferences, and mindfulness) successfully explain variability in how often students text in class. Specifically, their model can explain 33% of the variability in texting habits.
Second, they report an adjusted R2 that is slightly smaller than R2, which suggests they may have had too many variables, but it’s not *that* different, so probably not much to worry about.
Third, they report the significance of the model. They provide the results for the F-test (the specific analysis that tests the overall significance of a regression model; these would look like ANOVA results to you) and we can see that the model is statistically significant because the p value is less than alpha (in this case, p < .001).
To succinctly provide their MLR results to readers, authors will use tables to provide the details of their findings. In general, there are two tables to pay attention to: the correlation matrix and the MLR analysis summary table.
Bivariate Correlation Matrices. In most articles with MLR, the authors will provide a bivariate correlation matrix to show how the variables are related. Remember that we want to avoid multicollinearity (predictors that are too strongly related to each other). Let’s review Table 3 (below), the adapted correlation matrix from Olmstead and Terry (2014). You can see that many of the predictor variables are significantly correlated, though none are above .50. The authors then proceeded with the multiple regression analysis.
Multiple Regression Results Table. If a paper uses MLR, they will report their results in a comprehensive table. Normally, the first column is a list of the predictor variables included in the model. Sometimes (but not always), the first line/row will indicate the intercept, usually labeled as the intercept or the constant. As we learned above, this would be the expected outcome if all predictor variables equal zero. Sometimes, the intercept is not listed (as in this example from Olmstead & Terry, 2014). This is not unusual. If the intercept is not listed, the far left column will list the predictor variables in the model.
The table the lists the unstandardized coefficient (B) for each variable so you can interpret the change in outcome based on the original scale (and with it, the authors typically provide an estimate of the error for B, and sometimes a 95% confidence interval, though this varies quite a bit).
Next comes the standardized coefficient (β) which allows you to compare the strength of each predictor to assess which have the largest and smallest impact on the outcome. Remember that we are looking at absolute values, so the bigger the number (regardless if it is positive or negative), the bigger the impact on the outcome variables.
In terms of assessing whether or not the predictor has a statistically significant impact on the outcome, you will also get t-values and p-values (usually the last two columns). Here, you will follow the rules you learned in Intro Stats (when p < alpha, results are significant).
Let’s review Table 4, the MLR results table from Olmstead and Terry (2014; below).
Look at the list of predictor variables included in the model (first, far left column). In this case, the intercept is not listed. You can see that the authors included several predictors in their model, including sex, age, text habits, intrusive cell phone thoughts, multi-tasking preferences, and mindfulness. These are the variables the authors believed would impact their outcome variable, texting in class.
Next, look at the unstandardized coefficients. Remember, you can’t really compare these values with each other, since they are all on different scales. So you want to interpret them in their original form or scale. For example, holding all the other predictors constant, each additional year of age is associated with a .05 decrease in class texting (notice the negative coefficient value). As you can see, this can be kind of challenging to do for each variable and we still can’t compare them to each other (e.g., how do you compare the impact of age to score on mindfulness? They are on completely different scales). So we need to review the standardized coefficients.
Evaluate standardized coefficients. In reviewing the standardized coefficients (β), we can now compare the relative weight of each predictor. Looking at the the table, you can see that four variables have the largest beta coefficients: texting network size (measured by number of different people they text, β = .18), texting while studying habits (β = .24), texting while driving habits (β = .14), and score on the cellphone intrusive thoughts scale (β = .25). Of these, you can see that the intrusive thoughts around cell phone use has the biggest relative impact on students texting in class.
Determine predictor significance. Once you have reviewed both standardized and unstandardized coefficients, look at the last column, which shows our statistical significance as represented by p-values. Of the 13 predictors in this model, only four were significant, with p-values less than alpha: social network size, texting while studying habits, texting while driving habits, and cellphone intrusive thoughts.
Finally, ask “Does the model explain variability in our outcome?” Once you have reviewed the details of the above model, you can see in the notes of the table that the authors once again report the findings of the model, such that their model is not only statistically significant, it explains 30% (R2 adj) to 33% (R2) of the variability in student texting habits.
You’ve read the article and reviewed the results for their multiple linear regression. Let’s look at some big ideas in the context of Olmstead and Terry (2014).
Results are influenced by the model you choose. We tend to think of results as a direct output of the data, but in fact, the researcher shapes the data through consistent judgments and decisions as they build their model. For example, Olmstead and Terry (2014) included several predictors into their model, some of which were moderately correlated to each other and some were not correlated at all the outcome variable. Other researchers may have made different decisions that would have ultimately impacted their model and results.
There are diminishing returns on too many predictors. Yes, adding more predictors will increase explained variance (R2), but only incrementally and at that point, you are building a model that is overly complicated and not very useful (plus you’d need to look at adjusted R2 anyways and account for the large number of predictors).
We interpret every coefficient as conditional on the other predictors. As you remember from the previous sections, we interpret the regression coefficient for each predictor while holding all other predictors constant. So adding or removing predictors changes the overall calculation, which means it could change the strength, direction, and significance of a predictor. So, in the Olmstead and Terry (2014) article, you can’t just interpret the isolated impact of intrusive cell phone thoughts on in-class texting. The model included several predictors, so we have to interpret the variable within the parameters of the model (and all its predictors).
Correlation is not causation. Just because a model is significant or a predictor explains a lot of variance in the outcome, does not mean the predictor causes the outcome, just that they are related. As always, the research design (including random sample, random assignment, manipulation of the IV, confounding variables, etc.) is going to determine whether results are causal. In this study, the researchers simply collected survey data from participants about their habits (how often they text, how many people they text, self-ratings on intrusive thoughts, etc.) and then correlated the variables (Olmstead & Terry, 2014). So even though they had a significant model, their research was still correlational.
Statistical significance is not the same as practical importance. As with many of the analyses you learned about in Intro Stats, we need to be cautious about overinterpreting statistical significance, which can be heavily influenced by sample size. So to get a more robust understanding of your results, you want to consider other factors like confidence intervals, practical importance, and theoretical relevance.
Okay, we did it! No go look at other MLR articles and see if you can identify the key parts!
Olmsted, N. M., & Terry, C. P. (2014). Who's Texting in Class? A Look at Behavioral and Psychological Predictors. Psi Chi Journal of Psychological Research, 19(4), 183.
-----
Notes from meeting 3.27.26
Beta weights!
Each section mostly standalone
Use real published papers from examples
Send email to Psych to ask about using tables (psichijournal@psichi.org)
---
Finally, we also account for and acknowledge error (e). We won’t go into detail here, but suffice to say that we know our models are imperfect and have errors of one kind or another. The error term helps us understand how well our model fits the data and the difference between the observed values and the predicted values.