Research Paper First Draft
Raeann Cruz
2026-11-11
library(tidyverse)## ── Attaching core tidyverse packages ──────────────────────── tidyverse 2.0.0 ──
## ✔ dplyr 1.2.1 ✔ readr 2.2.0
## ✔ forcats 1.0.1 ✔ stringr 1.6.0
## ✔ ggplot2 4.0.3 ✔ tibble 3.3.1
## ✔ lubridate 1.9.5 ✔ tidyr 1.3.2
## ✔ purrr 1.2.2
## ── Conflicts ────────────────────────────────────────── tidyverse_conflicts() ──
## ✖ dplyr::filter() masks stats::filter()
## ✖ dplyr::lag() masks stats::lag()
## ℹ Use the conflicted package (<http://conflicted.r-lib.org/>) to force all conflicts to become errorslibrary(readxl)
library(car)## Loading required package: carData
##
## Attaching package: 'car'
##
## The following object is masked from 'package:dplyr':
##
## recode
##
## The following object is masked from 'package:purrr':
##
## somelibrary(lmtest)## Loading required package: zoo
##
## Attaching package: 'zoo'
##
## The following objects are masked from 'package:base':
##
## as.Date, as.Date.numericloaded_objects <- load("NSDUH_2023.Rdata")
print(loaded_objects)## [1] ".Random.seed" "data"nsduh_2023 <- get(loaded_objects[1])nsduh_2023<-data
nsduh_clean <- nsduh_2023 %>%
select(IRAMDEYR, BNGDRMDAYS, AGE3) %>%
drop_na(IRAMDEYR, BNGDRMDAYS, AGE3)Binge Drinking and Major Depressive Episode Among U.S. Adults: An Analysis of the 2023 National Survey on Drug Use and Health
Raeann Cruz
Karlerik Naslund
PAD 6833
University of Teas at San Antonio
Fall 2026
Abstract
Binge drinking and depression are two concerns that can affect people’s health, relationships, and every day lives. This paper asks whether adults who experienced a major depressive episode in the past year reported different patterns of binge drinking than adults who did not, while accounting for age. I used the 2023 National Survey on Drug and Health (NSDUH) and reviewed research suggesting that depression and heavy drinking are connected, although studies do not agree on which comes first. My preliminary linear regression included 45,133 observations. Depression and age were statistically associated with the code binge drinking variable, but together they explained only about 0.55% of its variation. More importantly, the outcomes is made up of categories rather than exact drinking day counts, and it’s highest code means no past month binge drinking. For that reason, these results cannot tell us whether depression is linked to more frequent binge drinking. The findings point to the need for a better suited statisticul model and careful interpretation of behavioral health data.
Introduction
Binge drinking is often discussed as a personal choice, but it effects can reach far beyond one night of drinking. Heavy drinking can affect physical health, emotional well being, relationships, and a person’s ability to manage daily responsibilities. Depression can create its own challenges, including difficulties with motivation, relationships and everyday functioning. When these experiences overlap, it becomes especially important to understand what the relationship actually looks like rather than assuming that one causes the other. Researchers have approached this question from different directions. Some have asked wehther people experiencing depressive symptoms are more likely to drink heavily as a way of coping. Others have explored whether repeated episodes of heavy drinking contribute to later depression. Paljarvi et al. (2009), found an association between binge drinking and later depressive symptoms, while Kim et al. (2022) examined whether depressive symptoms preceded changes in heavy episodic drinking. Together, these studies suggest that the relationship deserves attention, but they do not provide one simple explanation that applies to everyone. Age also matters. People’s drinking habits, responsibilites, and experiences of stress can change over time. Research by Keyes et al. (2020) suggests that the relationship between binge drinlng and depressive symptioms have also shifted across times and groups.Including age in this study allows me to consier one part of that variation. My research question is: Among U.S. adults, is experiencing a major depressive episode in the past year associated with more frequent binge drinking after accounting for age. I expected adults who experienced a major depressive episode to report more frequent binge drinking. This hypothesis reflects research linking emotional distress and alcohol use, but it is not a claim that depression causes someone to drink.
Literature Review
Heavy Drinking and Depressive Symptoms
One reason I chose this topic is that binge drinking and depression do not always fit into separate categories in people’s lives. Paljärvi et al. (2009) followed adults over five years and found that binge drinking was associated with later depressive symptoms. Their study is useful here because it looked at drinking patterns rather than simply whether someone consumed alcohol. It also reminds us that a person’s pattern of drinking may matter as much as the overall amount. Other longitudinal research involving people with and without HIV has likewise examined how risky drinking and depressive symptoms change together over time. Because that study focused on a particular clinical population, its findings provide context rather than a direct estimate for all U.S. adults. An important distinction in this literature is the difference between drinking alcohol, drinking frequently, and binge drinking. These behaviors are related, but they do not measure the same thing. Someone might drink on many days without consuming enough alcohol on any single occasion to meet the definition of binge drinking. Another person might drink only occasionally but consume a large amount in one sitting. This distinction matters for my research because I am interested in episodes of heavy consumption, not simply whether someone drinks. Paljärvi et al. (2009) help make this point by examining binge-drinking patterns in relation to later depressive symptoms. Their approach suggests that the way alcohol is consumed deserves attention when studying mental health.The longitudinal work involving participants with and without HIV also raises questions about how broader circumstances shape this relationship. A person managing health concerns, treatment responsibilities, or limited support may experience depression and alcohol use differently from someone without those circumstances. I do not have enough information in my current model to account for all of those experiences. As a result, this literature helps me recognize that a relationship observed in a statistical model may reflect more than just the two variables included in it.
Depression as a Possible Antecendent of Drinking
Another possibility is that depression comes before heavier drinking. Kim et al. (2022) followed adults across three waves of data collection and found evidence that depressive symptoms predicted later heavy episodic drinking. This finding fits the idea that some people may turn to alcohol when they are struggling emotionally. Still, that explanation will not fit everyone, and the study’s smaller sample limits how widely the results can be applied. Research with high school students has also considered whether changes in depressive symptoms come before changes in drinking frequency. These studies help explain why there may be a relationship in the NSDUH data, even though the cross-sectional analysis cannot determine which condition came first.The possibility that depression precedes heavier drinking particularly relevant to social work practice. Emotional distress can affect how people respond to stressful events, and drinking may be one response among many. However, the research does not justify assuming that everyone experiencing depression drinks to cope. Kim et al. (2022) followed people over time, which gives their study an advantage over a single survey: the researchers could observe whether one measure was recorded before a later change in another measure. My analysis uses information collected at one point in time, so it cannot provide that same sequence.There is also a difference between depressive symptoms and a major depressive episode. Some studies measure symptoms on a scale, capturing a range of emotional experiences. My study uses IRAMDEYR, which indicates whether respondents met the survey criteria for a past-year major depressive episode. These measures overlap conceptually but are not interchangeable. That difference is important when comparing my findings with studies that use symptom scores rather than a past-year episode indicator. A weak association in my analysis would not necessarily contradict a stronger association reported using another definition of depression.
Differences by Age, Sex, and Study Design
The relationship may also look different depending on a person’s age, sex, and social circumstances. Keyes et al. (2020) found that the connection between binge drinking and depressive symptoms among U.S. adolescents changed over time. Other studies have examined women’s binge-drinking patterns, young people followed over several years, and adults moving through middle age. Their findings are not identical. Some report stronger associations in particular groups, while others find that associations weaken after accounting for additional experiences and circumstances. These differences matter for this study because a result from adolescents or a specialized clinical sample should not automatically be treated as a result for all adults.The age-related research is useful because it shows why one overall association may hide important differences. The experiences of a person in early adulthood may not resemble those of someone approaching retirement. Social expectations around drinking, responsibilities at home or work, and access to supportive relationships may differ. Keyes et al. (2020) also show that relationships between drinking and depressive symptoms can change over historical time, not just as people grow older. These findings support my decision to include age, even though my initial model handles the age codes too simply.Research on Australian women and research following participants in the NCANDA study also offer a reminder about generalizability. Their samples and follow-up periods differ from the adult U.S. population represented in NSDUH. Longitudinal studies may be better suited to exploring timing, while large surveys can offer a broader picture of patterns across groups. Neither design answers every question. I see these approaches as complementary: the longitudinal literature helps frame possible explanations, and the NSDUH analysis allows me to explore whether an association appears in a large adult sample.
Synthesis and Research Gap
Overall, the literature provides a reason to study binge drinking and depression together, but it also makes me cautious about drawing conclusions. Some research suggests that depressive symptoms may come before heavier drinking; other work suggests that drinking may come before later depressive symptoms. Differences in age, population, measurement, and study design may help explain these mixed findings. My analysis adds a look at a large U.S. survey, while recognizing that one year of survey data cannot establish cause and effect. It also shows how much the way a variable is coded can shape what a statistical result means.The most important gap for my project is not that no one has studied depression and alcohol use. Many researchers have. Instead, I want to understand what can be learned when past-year major depressive episode status and binge-drinking frequency are examined together in recent U.S. survey data while accounting for age. I also want to be careful about the limits of that comparison. Studies differ in how they define binge drinking, how they measure depression, and whether they can follow participants over time. These choices affect the kinds of conclusions researchers can make. My preliminary findings need to be read in that context, rather than presented as a definitive test of whether depression leads to heavier drinking.
Methods
Data and Sample
I used the 2023 National Survey on Drug Use and Health (NSDUH), a national survey administered by SAMHSA that collects information about substance use, mental health, and related experiences among the civilian, noninstitutionalized population (SAMHSA, 2025). This was a secondary analysis, meaning I worked with information that had already been collected rather than interviewing participants myself. My analysis focused on adults with available responses for binge-drinking frequency, past-year major depressive episode status, and age. After excluding records with missing values on the selected variables, the preliminary regression included 45,133 observations. Because the analysis used complete cases, people without usable responses were not represented in the final model.
NSDUH is especially useful for this question because it asks respondents about both behavioral health and substance use. Rather than drawing only from people receiving treatment, it includes people with different levels of contact with health services. At the same time, the survey does not include everyone living in the United States. People who are institutionalized, for example, are outside its civilian noninstitutionalized target population. That matters when interpreting the findings, particularly because some people experiencing serious mental health or substance use concerns may be in settings that the survey does not cover (SAMHSA, 2025).
The data are also based on survey responses, so they may be affected by recall or willingness to disclose sensitive experiences. Respondents may not remember the exact timing of drinking episodes, and some may feel uncomfortable reporting heavy drinking or mental health symptoms. These concerns do not make the survey unusable, but they should be considered when discussing measurement limitations. Because I used existing data, I could not change the questions or collect follow-up information to clarify responses.
Measures
I used BNGDRMDAYS as the dependent variable. It describes past-month binge-drinking frequency in categories: 1 for one to two days, 2 for three to five days, 3 for six to nineteen days, 4 for twenty to thirty days, and 5 for no past-month binge drinking (SAMHSA, 2025). This last category is especially important. A value of 5 does not mean more binge drinking; it means none. The main independent variable was IRAMDEYR, indicating whether an adult experienced a major depressive episode in the past year. I also included AGE3 to account for age. These measures let me explore the relationship I was interested in, but the category coding creates limits for linear regression.
Before running the regression, I selected the three variables and excluded records with missing values on any of them. This is commonly called complete-case analysis. It makes the model easier to run because each included respondent has information for every variable, but it may change who remains in the sample. For example, if people with certain experiences are more likely to have missing responses, the analyzed group may differ from the group that was originally surveyed. I did not conduct a separate missing-data analysis, so I cannot determine how much this affected the results.
I also need to distinguish between a variable’s numeric storage format and what its values actually mean. R can treat the codes 1 through 5 as numbers, but that does not make them measurements on a continuous scale. In BNGDRMDAYS, the gaps between the frequency ranges are unequal, and the last category means no binge drinking at all. Similarly, AGE3 represents age groups rather than exact years of age. These details are central to choosing a statistical model, not just technical issues in data preparation (SAMHSA, 2025).
Statistical Procedures and Assumptions
For the preliminary analysis, I used ordinary least squares linear regression in R.The model was BNGDRMDAYS = Age3 +IRAMDEYR. The model tested whether age and past year depression status were associated with differences in the numerical codes assigned to binge drinking categories. I reviewed residual plots to examine the model’s fit, a Normal Q–Q plot to assess residual normality, and a Scale–Location plot to look for changes in residual spread. I also used the Durbin–Watson test to check for first-order autocorrelation and variance inflation factors to check for multicollinearity (Fox & Weisberg, 2019). Because the outcome is categorical, however, ordinary least squares is not the best final model. Ordinal regression, with the categories placed in their correct order, would better match the measurement scale (Agresti, 2010). The preliminary analysis also did not account for the NSDUH complex survey design, so its standard errors should not be treated as survey-adjusted estimates.In an ideal ordinary least square model, the outcome is measured on a scale for which differences between numerical values are meaningful. The errors should also behave in ways that make the estimated uncertainty reasonable. My diagnostic plots suggested that the residuals did not meet some of those expectations. The repeated horizontal bands in the Residuals vs. Fitted plot are especially understandable once the outcome’s category structure is considered. The Normal Q–Q plot also showed a pronounced departure from a straight line, while the Scale–Location plot raised questions about whether the spread of errors was consistent across fitted values.I considered whether a transformation might address these patterns. However, a log transformation would not solve the underlying problem that the outcome consists of categories with unequal intervals and a non-use category coded as 5. A more defensible next step would be to reorder the categories from no binge drinking to increasingly frequent binge drinking and estimate an ordinal regression model, checking its assumptions as well (Agresti, 2010). Another option would be to define a binary outcome indicating any past-month binge drinking and use logistic regression. That approach would answer a different, but still useful, question: whether depression is associated with reporting any binge drinking rather than how often it occurred. Finally, NSDUH uses a complex sampling design. My preliminary model did not apply survey weights or design-based variance estimation. This limits how confidently I can generalize the estimates to the U.S. adult population. A revised analysis should use the survey documentation to identify the appropriate weight and design variables and should report survey-adjusted estimates where possible (SAMHSA, 2025).
Results
Descriptive Characteristics
The regression used 45,133 observations with usable values for all three selected variables. This is a large sample, but sample size alone does not guarantee that a model answers the research question correctly. The current output does not provide a complete descriptive table for the final sample, such as the percentage reporting a past-year major depressive episode or the percentage in each binge-drinking category. Those percentages should be calculated from the same cleaned dataset before submission. I have not estimated them from the overall NSDUH sample because that would describe a different group of respondents.
Regression Results
In the preliminary regression, the age coefficient was 0.0502 (p < .001), and the coefficient for past-year major depressive episode status was −0.0479 (p = .0346). Both were statistically significant. The model’s R² was 0.005534, meaning that age and depression together explained about 0.55% of the variation in the coded outcome. This is a very small amount. More importantly, the negative depression coefficient should not be read as proof of more or less binge drinking. Because the outcome’s code 5 represents no binge drinking and the remaining codes represent unequal frequency ranges, the numerical direction of the coefficient does not translate cleanly into actual drinking frequency.
Diagnostic Figures
Figures 1–5 are the original diagnostic plots from my RStudio analysis. I included them because they help show not only what the regression produced, but also why its results need to be treated carefully.
cordata<-data%>%select(IRAMDEYR,BNGDRMDAYS,AGE3) %>% drop_na(IRAMDEYR,BNGDRMDAYS,AGE3)
cor(cordata)## IRAMDEYR BNGDRMDAYS AGE3
## IRAMDEYR 1.00000000 -0.02039946 -0.14357527
## BNGDRMDAYS -0.02039946 1.00000000 0.07372612
## AGE3 -0.14357527 0.07372612 1.00000000nsduh_2023<- lm(BNGDRMDAYS ~AGE3+IRAMDEYR, data = data)
summary(nsduh_2023)##
## Call:
## lm(formula = BNGDRMDAYS ~ AGE3 + IRAMDEYR, data = data)
##
## Residuals:
## Min 1Q Median 3Q Max
## -3.3006 -0.2503 0.7999 0.9004 1.0990
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) 3.747937 0.027227 137.655 <2e-16 ***
## AGE3 0.050239 0.003297 15.240 <2e-16 ***
## IRAMDEYR -0.047917 0.022682 -2.113 0.0346 *
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 1.521 on 45130 degrees of freedom
## (11572 observations deleted due to missingness)
## Multiple R-squared: 0.005534, Adjusted R-squared: 0.00549
## F-statistic: 125.6 on 2 and 45130 DF, p-value: < 2.2e-16Regression Assumption Tests
Figure 1. Linearity: Residuals vs. Fitted
plot(nsduh_2023, which = 1)
The Residuals vs. Fitted plot shows distinct bands rather than a random scatter of points. This suggests that linear regression may not be appropriate for the categorical binge-drinking outcome.
raintest(nsduh_2023)##
## Rainbow test
##
## data: nsduh_2023
## Rain = 0.99131, df1 = 22567, df2 = 22563, p-value = 0.744bptest(nsduh_2023)##
## studentized Breusch-Pagan test
##
## data: nsduh_2023
## BP = 294.37, df = 2, p-value < 2.2e-16Normality: Normal Q_Q Plot
plot(nsduh_2023, which = 2)
The Normal Q–Q plot shows that the points do not follow the diagonal reference line. This suggests that the residuals are not normally distributed, meaning the normality assumption is not met.
dwtest(nsduh_2023)##
## Durbin-Watson test
##
## data: nsduh_2023
## DW = 2.0031, p-value = 0.6306
## alternative hypothesis: true autocorrelation is greater than 0Figure 3: Homoscedascticity; Scale Location Plot
plot(nsduh_2023, which = 3)
The Scale-Location plot suggests that the residual variance may not be constant across fitted values.
The graphs show important limitations. In Figure 1, the residuals appear in distinct bands rather than a random cloud, which is consistent with the categorical outcome. Figure 2 shows a slight change in residual spread across fitted values, raising a concern about equal error variance. In Figure 3, the points depart sharply from the reference line, indicating that the residuals are not approximately normally distributed. The Durbin–Watson statistic was 2.003 (p = .716), so the test did not detect first-order autocorrelation. Both variance inflation factors were 1.021, which does not suggest problematic multicollinearity (Fox & Weisberg, 2019). These favorable tests do not resolve the more basic issue that the outcome is not a continuous measure.
##Answer to the research question My original question was whether adults who experienced a major depressive episode reported more frequent binge drinking after accounting for age. The preliminary analysis found a statistically significant relationship with the coded outcome, but it does not give me a reliable yes-or-no answer about actual binge-drinking frequency. I therefore cannot say that the hypothesis was supported or rejected based on this OLS model. A revised analysis using an appropriate categorical outcome model is needed before making that conclusion. Discussion and Conclusion This project helped me see why statistical significance is only one part of a research finding. The literature gives good reasons to examine depression and binge drinking together, and my preliminary analysis found an association in a large sample. At the same time, the model explained very little variation, and the diagnostic plots revealed clear problems with its assumptions. Most importantly, the coding of BNGDRMDAYS means that the regression coefficients cannot be interpreted as changes in the number of days someone binge drank.
These limitations do not make the research question unimportant. Instead, they point to what I would do differently in the next stage of the project: put the binge-drinking categories in their correct order, use an ordinal or appropriately defined binary regression model, account for the NSDUH survey design, and report descriptive statistics for the final sample. I would also consider other factors that may shape drinking behaviors, such as social stressors and access to mental health care. For social work and public health, the broader lesson is that depression and alcohol use deserve attention together, but research findings need to be interpreted in ways that match what the data can actually tell us.
One part of the project that stood out to me was how easy it would have been to focus only on the p-values. The depression coefficient had a p-value below .05, and that could sound like a clear answer at first. But the model explained less than one percent of the variation in the coded outcome. That small R-squared value means that age and depression, as entered here, provide very limited information about why people fall into different binge-drinking categories. Other factors that I did not include could be important, and some of the unexplained variation may also reflect how the outcome was measured. Statistical significance is worth reporting, but it does not automatically mean that an association is large, clinically meaningful, or useful for making decisions.
The graphs made those limitations more visible. Instead of relying only on the regression table, I could see that the residuals did not resemble the patterns expected from a well-fitting linear model. This is an important lesson for my own research process. Choosing a model should begin with the research question and the way the variables are measured. It should not be based only on whether a command runs successfully or produces a statistically significant result. The diagnostic findings do not erase the association in the preliminary output, but they do change how confidently I can interpret it.
There are several additional limitations. The analysis is cross-sectional, so I cannot determine whether depression came before binge drinking or vice versa. The measures rely on self-reported survey information, which may be affected by recall and disclosure. Complete-case analysis excluded respondents with missing information, and I did not test whether those exclusions were related to drinking or mental health. My model also treated age-category codes as a single numeric predictor and did not incorporate the survey’s complex sampling design. Each of these decisions could affect the estimates or the conclusions drawn from them.
If I continued this study, I would first create a descriptive table showing how many respondents fell into each binge-drinking category, how many reported a past-year major depressive episode, and how the categories differed by age. I would then choose a model that matched the outcome, report confidence intervals alongside p-values, and use appropriate survey-design procedures. I would also compare the results with the literature rather than treating any one estimate as the final answer. This would make the study more useful for understanding whether an association is present and how substantial it may be.
From a social work perspective, I think the value of this topic is that depression and alcohol use can affect the same person’s daily life, even when a dataset cannot fully explain their relationship. The research supports paying attention to both concerns without assuming that everyone with depression drinks heavily or that binge drinking always reflects an underlying mental health condition. A thoughtful assessment should leave room for people’s different experiences, sources of stress, and support needs. My preliminary analysis is not strong enough to support a clinical recommendation on its own, but it reinforces why careful, integrated behavioral health research matters.