Most Master of Public Health (MPH) programs put epidemiology and biostatistics at the center of the core curriculum. Students learn to measure how often disease occurs in a population. Then they learn to judge whether an association between an exposure and an outcome could be explained by bias.
Rates, ratios, confidence intervals, and logistic regression cover most of the work, and an introductory statistics course teaches the basics of each. Public health studies are mostly observational, though. Every estimate comes with a question attached about who was measured and how, and most of the effort in an MPH goes into answering it.
Descriptive Epidemiology: Rates Before Comparisons
Before comparing anything, students have to count properly. Prevalence is the share of people with a condition at one point in time. Incidence counts new cases over a period, and the incidence rate divides those cases by the person-time at risk.
A cohort followed for 12,000 person-years with 30 new cases of type 2 diabetes has an incidence rate of 30 / 12,000, or 2.5 per 1,000 person-years. For the interval, treat the count as Poisson. Its standard error is √30 = 5.48, so an approximate 95% confidence interval for the count is 30 ± 1.96 × 5.48 = [19.3, 40.7]. Divide by person-time, and the rate interval is [1.6, 3.4] per 1,000 person-years. With 30 cases, the normal approximation is acceptable. Below about 10 cases, use an exact interval.
Age standardization trips up more students than it should. Two regions can report the same crude death rate even though one has higher death rates in every age group, because its population is younger. Apply each region’s age-specific rates to a common standard population, then compare.
Assessments then ask for the result in plain words. A city council will not act on “2.5 per 1,000 person-years”. It will understand about one new diabetes case a year for every 400 adults followed. Understanding how different metrics handle uncertainty in statistics is vital for communicating risk accurately to non-technical stakeholders.
Analytic Study Designs and the Measures They Produce
Once rates are in hand, the next question is whether an exposure is associated with an outcome. Each design produces its own measure of association, and examiners expect students to match them.
| Design | How participants are selected | Typical measure |
|---|---|---|
| Ecological | Groups such as countries, not individuals | Correlation or rate ratio |
| Cross-sectional | A sample measured at one point in time | Prevalence ratio or prevalence odds ratio |
| Case-control | People with and without the outcome | Odds ratio |
| Cohort | People with and without the exposure, followed over time | Risk ratio or rate ratio |
Cohort studies are the main longitudinal design. Cross-sectional studies are quicker and cheaper, but they cannot show whether the exposure came before the outcome. When handling large datasets for these studies, researchers often rely on specialized statsmodels functions to inspect the underlying statistical outputs and parameters.
This next calculation turns up in almost every first assessment. A case-control study has 100 cases and 100 controls. Forty of the cases were exposed, against 20 of the controls. The odds ratio is (40 × 80) / (60 × 20) = 2.67.
The confidence interval is built on the log scale, where ln(2.67) = 0.981 and the standard error is √(1/40 + 1/60 + 1/20 + 1/80) = 0.323:
0.981 ± 1.96 × 0.323 = [0.348, 1.614]
Exponentiate both limits, and the 95% interval is [1.42, 5.02]. It excludes 1. Look at where the width comes from: the 1/20 term, from the 20 exposed controls, is the largest piece of that standard error.
Bias and Sensitivity Analysis
A confidence interval reflects random error only. Systematic error needs separate treatment, and MPH coursework gives it a lot of time.
Courses usually split bias into two kinds. Selection bias is about who ends up in the data. Information bias is about how exposure or outcome was measured, for example, asking cases and controls to recall their diet from years ago. Confounding is taught on its own, mostly through stratification and regression adjustment. In modern digital research, avoiding statistical pitfalls in AI is equally important when machine learning models are used to evaluate complex health exposures.
Sensitivity analysis asks how far a plausible bias could move the result. For unmeasured confounding, a common tool is the E-value (VanderWeele and Ding, 2017). For a risk ratio above 1, E-value = RR + √(RR × (RR − 1)).
Take a cohort study reporting RR = 1.8, 95% CI [1.3, 2.5]. The E-value for the point estimate is 1.8 + √(1.8 × 0.8) = 1.8 + 1.2 = 3.0. For the lower limit, it is 1.3 + √(1.3 × 0.3) = 1.92. To explain away the estimate entirely, an unmeasured confounder would need a risk ratio of 3.0 with both exposure and outcome. About 1.9 with both would be enough to pull the interval down to include 1.
The most common mistake in first assignments is reporting the point-estimate E-value and stopping there. The lower-limit figure says more about how fragile the conclusion is. Neither number shows that such a confounder exists, so markers usually want a simpler check alongside it, such as rerunning the model without participants lost to follow-up and saying whether the conclusion changed.
What Two Biostatistics Units Cover
One program shows how this material is usually split across a degree. In Victoria University’s online master in public health, the statistics sit in two units, Epidemiology and Biostatistics 1 and 2. The first covers descriptive epidemiology: analyzing observational data and reporting it to different audiences. The second moves to cross-sectional and longitudinal designs, where students identify sources of bias and run sensitivity analyses to see how far a result could shift.
The units use SPSS, so the SPSS guides on this site apply directly. Entry is open to graduates from any discipline, and some students meet a 2×2 table for the first time here. The menu-driven interface helps them. Anyone who moves to R or Stata later will need new syntax for the same models. You can also draw parallels to financial services software trends, where robust analytics tools are critical for processing heavy operational streams.
Preparing with Statology’s Concept Pages
A student starting from scratch can work through Statology in roughly this order:
- Confidence intervals for a mean and a proportion. Every epidemiological estimate is reported with one.
- Relative risk and odds ratios, plus the rare-outcome condition under which the odds ratio approximates the risk ratio.
- The chi-square test of independence, which is the standard test for a 2×2 table.
- Simple then multiple logistic regression, since adjusted odds ratios come from it.
Read the SPSS tutorials too. Run a crosstab with risk estimates, then a binary logistic regression, and go through each output table line by line. Many assessments are built on that exact output.
Putting It Into Practice
Before the first epidemiology unit, test yourself on one task. Take a 2×2 table, get the odds ratio and its 95% interval by hand, then match it in SPSS. If the numbers disagree, rework the case-control example above until they don’t.
For bias, pick a published cohort study. Compute the E-value for its main risk ratio and its lower limit, then write two sentences on whether a confounder that strong is plausible.
Mastering the quantitative foundations of public health requires patience, careful notation, and a solid understanding of how observational data behaves under pressure. By tackling calculation exercises systematically and verifying analytical steps through structured software guides, students can bridge the gap between abstract statistical theory and real-world epidemiological insights.
