Group 14: Predict COVID-19 Anxiety Search Trend - Regression

1 Summary

The COVID-19 pandemic led to significant public health measures, such as lockdowns and vaccinations, which have increased stress and anxiety worldwide. This study aims to explore how factors like hospitalizations and new vaccinations correlate with anxiety-related Google searches in the United States (US) during the pandemic.

2 Introduction

In response to the global COVID-19 pandemic, public health measures such as lockdowns, social distancing, and vaccination drives have significantly impacted daily lives. These unprecedented measures, while essential for managing the health crisis, have led to increased levels of stress and anxiety among populations worldwide (Murphy et al., 2021). Several studies have highlighted the psychological impacts of the pandemic, noting increased incidences of anxiety and other mental health disorders as people cope with isolation, uncertainty, and changes to their normal routines (Prati & Mancini, 2021). Furthermore, the rapid dissemination of information and sometimes misinformation through media channels has played a crucial role in shaping public perceptions and reactions to the pandemic (Neill et al., 2023). This phenomenon, often referred to as the “infodemic,” has contributed to heightened anxiety levels, driving more people to seek information online about COVID-19 and its implications for their health and safety (Borges do Nascimento et al., 2022).

This study aims to analyze how certain factors such as hospitalizations, vaccinations, and confirmed cases correlate with the trends in Google searches for anxiety in the United States. More specifically, we want to find out the following: How do COVID-19 vaccination rates, hospitalization rates, and confirmed cases relate to anxiety-related search patterns during the pandemic?

3 Methods:

3.1 Data

To find an answer to this question, we performed an analysis on a dataset of US daily COVID-19 data and Google search trends from January 2020 until now. The raw data set contains 991 rows × 564 columns, within which includs 6 variables relavant to our question. Our variables of interest are as following:

• date: date of data collection in YYYY/MM/DD format
• search_trends_anxiety: the prevalence of anxiety in google searches
• new_persons_vaccinated: count of new persons who have received one or more doses
• new_hospitalized_patients: count of new cases hospitalized after a positive test
• new_confirmed: count of new cases confirmed after a positive test
• new_intensive_care_patients: count of new COVID-19 positive cases admitted to ICU

By studying these variables, we are attempting to observe patterns and correlations that can reveal the mental health response of the public to the pandemic.

3.2 Exploratory Data Analysis (EDA)

During EDA, we will first select only the 6 variables of interest for further exploratory data analysis and regression analysis. This preliminary variable selection is then followed by computation of their summary statistic and visualizations to explore the trends and patterns among the relevant variables:

The following Table 1 displays summary statistics of all 6 variables of interest relavant to our research question. Notably, we see that new_persons_vaccinated and new_intensive_care_patients contain many missing values which would be removed from the data before they enter into further analysis.

Moreover, we see that on average, there are around 80,0000 people getting newly vaccinated across the entirety of the COVID-19 pandemic in the US.

Variable	Min	Q25	Mean	Median	Q75	Max	Observations	Missing
search_trends_anxiety	5.5	7.17	7.736588e+00	7.79	8.29	10.9	986	5
new_persons_vaccinated	4534.0	371081.00	7.822666e+05	798390.00	1157564.00	2005480.0	641	350
new_hospitalized_patients	4332.0	7275.00	1.159166e+04	12988.00	15247.00	18104.0	977	14
new_confirmed	39828.0	69225.00	1.443094e+05	147277.00	200308.00	286196.0	988	3
new_intensive_care_patients	22.0	84.00	1.418353e+02	141.00	186.00	279.0	419	572
date	18609.0	18630.00	1.865100e+04	18651.00	18672.00	18693.0	991	0

Table 1: Summary statistics of selected variables with total observation counts and missing values counts from the original dataset.

From the following Figure 1 pairplot visualization, we see that new vaccination status and time progression (date) has positive moderate association with anxiety search trend. Meanwhile, new hospitalization patient rate is negatively and moderately correlated with anxiety search trend.

The remaining variables such as the number of newly confirmed COVID-19 cases (new_confirmed) and newly admitted intensive care patints (new_intensive_care_patients) either show weak negative correlation or no correlation with anxiety search trend.

Figure 1: Pairs plot for all variables of interest for exploration

new_persons_vaccinated seems to have a moderate positive linear correlation with anxiety serach trend.

new_hospitalized_patients seems to have a weak negative linear correlation with anxiety serach trend.

new_confirmed seems to have a weak negative linear correlation with anxiety serach trend.

date seems to have a weak positive linear correlation with anxiety serach trend.

In Figure 2, we can observe that anxiety searches seem to be consistent across the COVID-19 pandemic except for 3 peaks around winter of 2020, spring of 2021, and spring of 2022.

Figure 2: Anxiety Search Trends Over Time

3.3 Feature Selection: Backward Selection

In this section, we perform feature selection using backward seleciton method. Using backward selection, we will be able to select the number of variables has optimal predictive ability using measures of adjusted \(R^2\). Higher adjusted \(R^2\) indicates better predictive performance of the features.

Using backward selection over the 5 feature variables, Table 2 gives the selection results, showing which feature is being selected into the model each time (the selected variable is denoted as TRUE). Table 3 shows the adjusted \(R^2\) (ADJ.R2) performance of model corresponding to the number of TRUE feature input(s) displayed in Table 2.

(Intercept)	new_persons_vaccinated	new_hospitalized_patients	new_confirmed	new_intensive_care_patients	date
TRUE	TRUE	FALSE	FALSE	FALSE	FALSE
TRUE	TRUE	TRUE	FALSE	FALSE	FALSE
TRUE	TRUE	TRUE	FALSE	FALSE	TRUE
TRUE	TRUE	TRUE	FALSE	TRUE	TRUE
TRUE	TRUE	TRUE	TRUE	TRUE	TRUE

Table 2: The features being selected during each step of backward model selection.

n_input_variables	RSQ	RSS	ADJ.R2
1	0.3213248	16.71652	0.3043580
2	0.3285177	16.53935	0.2940828
3	0.3561980	15.85756	0.3053715
4	0.3752592	15.38806	0.3077197
5	0.3753546	15.38571	0.2885983

Table 3: RSQ, RSS, and Adjusted R^2 of each model with the number of features used.

To select the best-performing features to use to predict Anxiety google searches during the COVID-19 pandemic, we used backward selection to pick the set of features that would produce the best highest adjusted \(R^2\) value. Higher \(R^2\) value greater predicative ability of the feature variables for the response variable, thereby giving us the model that explained our data the best.

As seen in Table 3, a model with 3 features (n_input_variables = 4 due to intercept being counted as an input variable by the selection algorithm) provides the highest adjusted \(R^2\) value at 0.308 based on the validation set. This model contains the variables new_persons_vaccinated, new_hospitalized_patients, and date, which provides some evidence for our hypothesis that higher vaccination rates and greater hospitalization may lead to greater anxiety search trend during COVID-19.

3.4 Creating the final model using selected variables

term	estimate	std.error	statistic	p.value
(Intercept)	337.1161592	257.9924354	1.306690	0.1991712
new_persons_vaccinated	0.0000012	0.0000003	3.562430	0.0010093
new_hospitalized_patients	-0.0000911	0.0000637	-1.430905	0.1606304
date	-0.0176455	0.0138049	-1.278205	0.2089297

Table 4: Coefficients and intercept of the selected linear model.

According to the resultant coefficients from Table 4, we create a linear formula of our regression model:

Anxiety search trend = \(1.16\times 10^{-6}\) New Vaccination \(-9.11\times 10^{-5}\) New Hospitalization \(-0.0176\) Date + \(337.116\)

(note: Date is measured in number of days)

3.5 Model Evaluation

RMSPE	R_square
0.6589689	0.356198

Table 5: Model performance metrics.

The current model has a root mean square prediction error (RMSPE) of 0.659, suggesting a large discrepancy between the actual response variable and the predicted response variable anxiety search.

At the same time, this model has a \(R^2\) of 0.356 as seen in Table 5. This suggests that around 36% of variance in anxiety Google search trend during the COVID-19 pandemic is explained conjointly by new vaccination, new hospitalization, and progression of time.

4 Discussion

4.1 Summary of Findings

Our results discovered that COVID-19 pandemic-anxiety-related search patterns were weakly influenced by the vaccination rate, hospital rates, and the time elapsed. The vaccination rate increase was strongly correlated with anxiety-related search increases, but growing hospital rates were correlated with decreased anxiety searches. The model’s 36% variance explanation suggests the possibility that additional independent variables beyond what was measured within the database may also contribute to a massive amount of this variance.

4.2 Expected vs. Actual Results

The small but positive relationship between vaccination rates and anxiety searches went against our expectations as one would expect higher vaccination rates would reduce anxiety. However, it could be an expression of public concern regarding side effects of vaccinations or the overall impact of vaccination campaigns on society. The negative relationship between hospitalization rates and anxiety searches aligns with the hypothesis that serious COVID-19 complications will generate fewer online searches because of increased stress or disengagement.

4.3 Impact of Findings

These are indicators of the complex interactions between public and mental health interventions during a pandemic. Knowledge of these interactions could allow policymakers to develop improved communication strategies for reducing worry and anxiety in the public.

4.4 Future Questions

• Future research areas might include:
• The role of media coverage in producing anxiety-supportive search trends.
• The impact of other measures of mental health, depression or stress, on searching.

5 References

References

Borges do Nascimento, I. J., Pizarro, A. B., Almeida, J. M., Azzopardi-Muscat, N., Gonçalves, M. A., Björklund, M., & Novillo-Ortiz, D. (2022). Infodemics and health misinformation: A systematic review of reviews. Bulletin of the World Health Organization, 100(9), 544–561. https://doi.org/10.2471/BLT.21.287654

Murphy, L., Markey, K., O’Donnell, C., Moloney, M., & Doody, O. (2021). The impact of the COVID-19 pandemic and its related restrictions on people with pre-existent mental health conditions: A scoping review. Archives of Psychiatric Nursing, 35(4), 375–394. https://doi.org/10.1016/j.apnu.2021.05.002

Neill, R. D., Blair, C., Best, P., McGlinchey, E., & Armour, C. (2023). Media consumption and mental health during COVID-19 lockdown: A UK cross-sectional study across england, wales, scotland, and northern ireland. Z Gesundh Wiss, 31(3), 435–443. https://doi.org/10.1007/s10389-021-01506-0

Prati, G., & Mancini, A. D. (2021). The psychological impact of COVID-19 pandemic lockdowns: A review and meta-analysis of longitudinal studies and natural experiments. Psychological Medicine, 51(2), 201–211. https://doi.org/10.1017/S0033291721000015