This report is written for a public health policy team that needs a clear, non-technical summary of how COVID-19 outcomes differ across countries and what factors are most closely tied to those differences. The intended audience does not need advanced statistics knowledge, so each section starts with a plain-language explanation and then adds the supporting analysis.
The main objective of this analysis is to test whether country-level vaccination coverage, demographics, economic conditions, and policy stringency are associated with COVID-19 case fatality rate. That central problem is examined through three related questions:
The analysis uses the Our World in Data (OWID) COVID-19 dataset, which compiles global COVID-19 information from official public health sources such as national health agencies, the WHO, and other government reporting systems. For this project, the data was cleaned and organized into country-date observations, making it suitable for trend analysis and cross-country comparison. OWID is widely used in academic and policy research, which makes it a credible source for this analysis.
| Variable | What it means | Why it matters |
|---|---|---|
case_fatality_rate |
Share of cases that ended in death | Main outcome for severity |
vax_coverage |
Vaccination coverage level | Main outcome for vaccine progress |
continent |
Broad geographic region | Used for group comparison |
median_age |
Median age of the population | Captures population structure |
gdp_per_capita |
Economic capacity per person | Captures resource differences |
stringency_index |
How strict policy responses were | Captures response intensity |
reproduction_rate |
Transmission pressure | Captures outbreak momentum |
Read the EDA to see the patterns in the data, the hypothesis test to see whether group differences are statistically meaningful, and the regression sections to see which factors matter after controlling for other variables. Read the TL;DR first if you only want the main conclusions.
This report follows the same structure as the weekly assignments, but with stronger interpretation and more direct recommendations:
ggplot(monthly_trend, aes(x = year_month, y = mean_cfr)) +
geom_line(linewidth = 1) +
geom_point(size = 1.5) +
geom_vline(xintercept = as.Date(c("2020-03-01", "2021-06-01", "2021-11-01")),
linetype = "dashed", alpha = 0.5) +
annotate("text", x = as.Date("2020-03-01"), y = plot_cfr_max * 0.98,
label = "Early pandemic waves", angle = 90, vjust = -0.4, size = 3) +
annotate("text", x = as.Date("2021-06-01"), y = plot_cfr_max * 0.98,
label = "Delta becomes dominant", angle = 90, vjust = -0.4, size = 3) +
annotate("text", x = as.Date("2021-11-01"), y = plot_cfr_max * 0.98,
label = "Omicron emerges", angle = 90, vjust = -0.4, size = 3) +
labs(
title = "Average COVID-19 Case Fatality Rate Over Time",
subtitle = "Annotations highlight major pandemic waves that help explain the peaks and valleys",
x = "Year-Month",
y = "Mean Case Fatality Rate"
) +
theme_minimal()
This figure gives a time-based view of how fatality rates changed across the full data set. The annotated peaks and valleys help connect the trend to major pandemic phases, such as the Delta wave in mid-2021 and the emergence of Omicron later in 2021. That context makes the graph easier to interpret because the line is not just moving randomly; it reflects changes in variants, treatment, testing, and vaccination over time.
p_continent_full <- ggplot(covid_model, aes(x = continent, y = case_fatality_rate, fill = continent)) +
geom_boxplot(alpha = 0.75) +
labs(
title = "Case Fatality Rate by Continent",
subtitle = "Full distribution view",
x = "Continent",
y = "Case Fatality Rate"
) +
theme_minimal() +
theme(legend.position = "none")
p_continent_zoom <- ggplot(covid_model, aes(x = continent, y = case_fatality_rate, fill = continent)) +
geom_boxplot(alpha = 0.75) +
coord_cartesian(ylim = c(0, plot_cfr_q95)) +
labs(
title = "Case Fatality Rate by Continent",
subtitle = "Zoomed-in view to make the box ranges easier to compare",
x = "Continent",
y = "Case Fatality Rate"
) +
theme_minimal() +
theme(legend.position = "none")
p_continent_full + p_continent_zoom + plot_layout(ncol = 2)
This boxplot compares the distribution of fatality rates across continents. The zoomed-in view makes the box ranges and median lines easier to compare, especially when a few outliers stretch the full scale. Together, the two panels show both the overall spread and the more readable center of each regional distribution.
ggplot(scatter_data, aes(x = vax_coverage, y = case_fatality_rate)) +
geom_point(alpha = 0.35) +
geom_smooth(method = "lm", se = FALSE) +
coord_cartesian(xlim = c(plot_vax_q05, plot_vax_q95), ylim = c(plot_cfr_q05, plot_cfr_q95)) +
labs(
title = "Vaccination Coverage and Case Fatality Rate",
subtitle = "Zoomed-in view to make the slope easier to see",
x = "Vaccination Coverage",
y = "Case Fatality Rate"
) +
theme_minimal()
This scatterplot shows whether higher vaccination coverage tends to be associated with lower fatality rates. Each dot represents one country-date observation, and the fitted line gives a quick summary of the overall trend. The zoomed-in axis makes the slope easier to read, which is important because the relationship is subtle in the full-scale version.
ggplot(scatter_data, aes(x = median_age, y = vax_coverage)) +
geom_point(alpha = 0.35) +
geom_smooth(method = "lm", se = FALSE) +
labs(
title = "Vaccination Coverage and Median Age",
subtitle = "Population age structure appears to matter for vaccination uptake",
x = "Median Age",
y = "Vaccination Coverage"
) +
theme_minimal()
This graph examines whether older countries tended to have higher vaccination coverage. That relationship is important because age structure can affect vaccine priority, public risk perception, and the ability to reach large coverage levels. If the line trends upward, then median age is a plausible predictor of vaccination success.
This plot helps motivate the logistic regression section. It also shows a relationship that a general audience can understand quickly without needing statistical background.
A continent-level comparison gives a simple way to test whether fatality rates differ across broad geographic regions. That is useful because a client audience often wants a high-level answer first. If the regions are meaningfully different, then the rest of the report should not treat the world as one uniform population.
Null hypothesis (H0): Mean case fatality rate is the same across continents.
Alternative hypothesis (H1): At least one continent has a different mean case fatality rate.
## Df Sum Sq Mean Sq F value Pr(>F)
## continent 5 0.61 0.12283 27.41 <2e-16 ***
## Residuals 39573 177.31 0.00448
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1## Tukey multiple comparisons of means
## 95% family-wise confidence level
##
## Fit: aov(formula = case_fatality_rate ~ continent, data = anova_data)
##
## $continent
## diff lwr upr p adj
## Asia-Africa -0.010170982 -0.0141393942 -0.006202570 0.0000000
## Europe-Africa -0.005321778 -0.0089669780 -0.001676578 0.0004553
## North America-Africa 0.001072502 -0.0039615256 0.006106529 0.9905634
## Oceania-Africa -0.018652966 -0.0249545485 -0.012351383 0.0000000
## South America-Africa -0.006497737 -0.0115463951 -0.001449078 0.0033414
## Europe-Asia 0.004849204 0.0024208304 0.007277577 0.0000002
## North America-Asia 0.011243484 0.0070066278 0.015480340 0.0000000
## Oceania-Asia -0.008481984 -0.0141670046 -0.002796964 0.0003058
## South America-Asia 0.003673245 -0.0005809841 0.007927475 0.1359942
## North America-Europe 0.006394280 0.0024585294 0.010330031 0.0000537
## Oceania-Europe -0.013331188 -0.0187954941 -0.007866882 0.0000000
## South America-Europe -0.001175958 -0.0051304057 0.002778489 0.9585259
## Oceania-North America -0.019725468 -0.0261994614 -0.013251474 0.0000000
## South America-North America -0.007570238 -0.0128325196 -0.002307957 0.0005906
## South America-Oceania 0.012155229 0.0056698525 0.018640606 0.0000014The ANOVA test checks whether the differences seen in the boxplot are large enough to be unlikely under random variation alone. If the p-value is below 0.05, we reject the null hypothesis and conclude that fatality rates are not equal across all continents. In practical terms, that means geography is associated with meaningful differences in COVID severity or reporting patterns.
The post-hoc Tukey test then helps identify which continent pairs differ from one another. That matters because a significant ANOVA result only tells us that at least one group differs; it does not say exactly which ones. The pairwise results make the conclusion more actionable.
This section gives the audience a clear yes/no answer to a simple question. It also shows that regional differences, if present, are worth taking seriously in later modeling and recommendations.
The ANOVA gives a broad comparison, but it does not explain the drivers behind fatality rates. A multiple linear regression model helps answer a more practical question: when vaccination, demographics, and economic conditions are considered together, which factors are most closely associated with case fatality rate?
case_fatality_ratevax_coverage, median_age,
gdp_per_capita, stringency_indexThese predictors were selected because they represent vaccination progress, age structure, economic capacity, and policy response.
##
## Call:
## lm(formula = case_fatality_rate ~ vax_coverage + median_age +
## gdp_per_capita + stringency_index, data = lm_data)
##
## Residuals:
## Min 1Q Median 3Q Max
## -0.04686 -0.01995 -0.01151 0.00183 2.25340
##
## Coefficients:
## Estimate Std. Error t value Pr(>|t|)
## (Intercept) -6.038e-03 2.327e-03 -2.595 0.00946 **
## vax_coverage -1.508e-04 1.651e-05 -9.136 < 2e-16 ***
## median_age 7.986e-04 5.438e-05 14.687 < 2e-16 ***
## gdp_per_capita -4.104e-07 2.262e-08 -18.147 < 2e-16 ***
## stringency_index 3.052e-04 2.108e-05 14.480 < 2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## Residual standard error: 0.06639 on 39574 degrees of freedom
## Multiple R-squared: 0.01968, Adjusted R-squared: 0.01958
## F-statistic: 198.6 on 4 and 39574 DF, p-value: < 2.2e-16ggplot(lm_data, aes(x = vax_coverage, y = case_fatality_rate)) +
geom_point(alpha = 0.3) +
geom_smooth(method = "lm", se = FALSE) +
coord_cartesian(xlim = c(plot_vax_q05, plot_vax_q95), ylim = c(plot_cfr_q05, plot_cfr_q95)) +
labs(
title = "Higher Vaccination Coverage Is Associated With Lower Fatality Rates",
subtitle = "Zoomed-in view to make the negative slope easier to see",
x = "Vaccination Coverage",
y = "Case Fatality Rate"
) +
theme_minimal()
This plot gives a direct visual check of the regression relationship. The goal here is not just to see whether the line slopes up or down, but also to see whether the points are roughly centered around a straight-line pattern. The zoomed-in axis makes the negative slope easier to read, which addresses the fact that the full-scale graph can look nearly flat.
This plot is the simplest way to explain the key regression idea to a general audience. It shows the main relationship without requiring anyone to read coefficient tables first.
knitr::kable(
lm_tidy,
digits = 4,
caption = "Linear Regression Coefficients With Confidence Intervals"
)| term | estimate | std.error | statistic | p.value | conf.low | conf.high |
|---|---|---|---|---|---|---|
| (Intercept) | -6e-03 | 0.0023 | -2.5950 | 0.0095 | -0.0106 | -0.0015 |
| vax_coverage | -2e-04 | 0.0000 | -9.1359 | 0.0000 | -0.0002 | -0.0001 |
| median_age | 8e-04 | 0.0001 | 14.6866 | 0.0000 | 0.0007 | 0.0009 |
| gdp_per_capita | 0e+00 | 0.0000 | -18.1468 | 0.0000 | 0.0000 | 0.0000 |
| stringency_index | 3e-04 | 0.0000 | 14.4798 | 0.0000 | 0.0003 | 0.0003 |
## vax_coverage median_age gdp_per_capita stringency_index
## 1.097356 1.474312 1.464795 1.125965vax_coverage means that
higher vaccination coverage is associated with lower fatality
rates.median_age suggests that
older populations experience higher fatality rates, which fits the
expectation that older adults are at greater risk of severe COVID
outcomes.gdp_per_capita and stringency_index help
capture economic capacity and policy response. Even if one coefficient
is not statistically significant, it still matters conceptually because
it may help reduce omitted variable bias.A 1-unit increase in each predictor changes fatality rate by the amount shown in the coefficient table, holding the other predictors constant. For a non-technical reader, the most important thing is the direction of the effect. If the vaccination coefficient is negative, that means more vaccination is associated with lower fatality. If the median-age coefficient is positive, that means older populations tend to have higher fatality rates.
This is the core analytic section of the report. It turns the visual pattern into a controlled comparison and identifies which factors still matter after accounting for the others.
The previous model focused on fatality rates. This section flips the question and asks what helps explain whether a country reaches high vaccination coverage. That is useful because vaccination success is itself an important policy outcome.
##
## 0 1
## 35441 4138We define high vaccination coverage as greater than 50 percent. That cutoff is a practical midpoint for comparison, not a claim that 50 percent is enough to stop COVID-19 spread. The purpose of the threshold is to split the sample into lower- and higher-coverage groups so the logistic model can estimate which factors make crossing that midpoint more likely.
ggplot(logit_data, aes(x = median_age, y = vax_coverage, color = factor(high_vax))) +
geom_point(alpha = 0.6) +
geom_hline(yintercept = 50, linetype = "dashed") +
labs(
title = "Median Age and Vaccination Coverage",
subtitle = "The dashed line marks the 50 percent threshold used to define the binary outcome",
x = "Median Age",
y = "Vaccination Coverage",
color = "High Vax"
) +
theme_minimal()
This scatterplot shows why median age is a plausible predictor. Countries with higher vaccination coverage are concentrated at higher median ages, while lower vaccination coverage appears more common across younger populations. That pattern suggests that demographic structure may be tied to vaccine uptake.
This plot helps a non-technical reader understand why the logistic model was built in the first place. It also connects the model to a policy question: which countries are more likely to cross a meaningful vaccination threshold?
##
## Call:
## glm(formula = high_vax ~ reproduction_rate + stringency_index +
## median_age, family = binomial, data = logit_data)
##
## Coefficients:
## Estimate Std. Error z value Pr(>|z|)
## (Intercept) -0.836460 0.143363 -5.835 5.39e-09 ***
## reproduction_rate -0.213001 0.055685 -3.825 0.000131 ***
## stringency_index -0.049699 0.001151 -43.174 < 2e-16 ***
## median_age 0.040139 0.002725 14.727 < 2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## (Dispersion parameter for binomial family taken to be 1)
##
## Null deviance: 26515 on 39578 degrees of freedom
## Residual deviance: 23808 on 39575 degrees of freedom
## AIC: 23816
##
## Number of Fisher Scoring iterations: 6knitr::kable(
logit_tidy,
digits = 4,
caption = "Logistic Regression Odds Ratios and Confidence Intervals"
)| term | estimate | std.error | statistic | p.value | conf.low | conf.high |
|---|---|---|---|---|---|---|
| (Intercept) | 0.4332 | 0.1434 | -5.8346 | 0e+00 | 0.3268 | 0.5733 |
| reproduction_rate | 0.8082 | 0.0557 | -3.8251 | 1e-04 | 0.7241 | 0.9008 |
| stringency_index | 0.9515 | 0.0012 | -43.1736 | 0e+00 | 0.9494 | 0.9537 |
| median_age | 1.0410 | 0.0027 | 14.7273 | 0e+00 | 1.0354 | 1.0466 |
## (Intercept) reproduction_rate stringency_index median_age
## 0.4332415 0.8081556 0.9515157 1.0409555median_age has an odds ratio above 1,
then each additional year in median age increases the odds of high
vaccination coverage.reproduction_rate has an odds ratio below 1, then
more transmission pressure is associated with lower odds of high
vaccination coverage.This model is useful because it turns a continuous policy question into a simple decision-oriented outcome: which countries are more likely to cross the 50 percent vaccination threshold?
This section translates the vaccination story into a simple yes/no outcome that is easy to communicate to decision-makers. It also gives the report a second model that checks whether the same broad patterns still appear when the question is reframed.
## Confusion Matrix and Statistics
##
## Reference
## Prediction 0 1
## 0 35391 4138
## 1 50 0
##
## Accuracy : 0.8942
## 95% CI : (0.8911, 0.8972)
## No Information Rate : 0.8954
## P-Value [Acc > NIR] : 0.7968
##
## Kappa : -0.0025
##
## Mcnemar's Test P-Value : <2e-16
##
## Sensitivity : 0.9986
## Specificity : 0.0000
## Pos Pred Value : 0.8953
## Neg Pred Value : 0.0000
## Prevalence : 0.8954
## Detection Rate : 0.8942
## Detection Prevalence : 0.9987
## Balanced Accuracy : 0.4993
##
## 'Positive' Class : 0
## The confusion matrix checks whether the model classifies countries correctly. This matters because a model can have statistically significant coefficients but still perform poorly in prediction. Accuracy, sensitivity, and specificity give a fuller view of whether the logistic model is actually useful for classification.
A model is more convincing when it does not only look significant on paper but also performs reasonably well in classification. This helps the audience trust the practical usefulness of the results.
A reasonable policy recommendation is to prioritize vaccination rollout and maintain targeted public health responses in countries with older populations, higher transmission pressure, and weaker economic capacity.
The model results should be used as guidance rather than as proof of causation, but they are still valuable because they identify which factors are repeatedly associated with worse outcomes and lower vaccination success.
This project analyzes global COVID-19 data to identify which country-level factors are most closely associated with case fatality rate and vaccination coverage. The goal is to turn a large and messy public-health dataset into a small set of practical insights that a non-technical audience can use.
The most useful public-health strategy is to focus on measurable factors that move outcomes in a better direction, especially vaccination coverage and population structure. The report shows where those relationships are strongest and gives a practical basis for decision-making.
The strongest message from the report is that a small number of measurable factors, especially vaccination coverage and demographic structure, explain a meaningful share of the differences in COVID outcomes. That makes targeted, data-driven action more useful than broad assumptions.