Select Procedures in Health Screening Studies

Start With the Question and the Outcome

Health studies can ask whether a disease rate is higher than a known benchmark, whether one treatment has a higher success rate than another, or whether a measured quantity differs from a clinical reference value. These questions may sound similar because each asks about a health-related difference. The right procedure depends instead on what is recorded, how many groups are compared, and how the observations are related.

A disease indicator such as “disease present” or “disease absent” is categorical and binary. Its population summary is a proportion. A measurement such as blood pressure is quantitative, and its population summary is a mean. Then the design determines whether the procedure concerns one group or a comparison of two groups. Before calculating anything, name the outcome, the parameter, the comparison, and the design.

Procedure-selection rule: For a binary outcome in one group, use a one-proportion procedure. For a binary outcome in two independent groups, use a two-proportion procedure. For a quantitative outcome in one group compared with a benchmark, use a one-sample t procedure when the population standard deviation is unknown. For a quantitative outcome in two independent groups, use a two-sample t procedure, commonly Welch’s procedure when equal population variances are not assumed.

A Health-Study Decision Map

Health question	Outcome and design	Parameter(s)	Procedure family
Does the disease rate in one screened population exceed a benchmark?	Binary disease status; one group	One population proportion $p$	One-proportion test or interval
Do two treatments have different success rates?	Binary success status; two independent groups	Difference $p_1-p_2$	Two-proportion test or interval
Is a group’s mean measurement above a clinical reference value?	Quantitative measurement; one group	One population mean $\mu$	One-sample t test or interval
Do two independent treatments have different mean outcomes?	Quantitative measurement; two independent groups	Difference $\mu_1-\mu_2$	Two-sample t test or interval

A benchmark does not automatically mean a one-sample procedure: first identify the kind of outcome. A disease rate compared with a percentage benchmark concerns $p$; a measured quantity compared with a numerical reference concerns $\mu$. Likewise, having two treatment groups does not automatically mean a t procedure: the outcome may be binary, in which case the comparison is between proportions.

One Disease Rate: A One-Proportion Procedure

Suppose each person is classified as having or not having a particular disease, and the study asks about the fraction with the disease in a target population. Let $p$ be that population proportion. If the question compares $p$ with a specified benchmark $p_0$, a significance test uses $H_0:p=p_0$ and an alternative determined by the question. The one-proportion z test uses the null standard error $\sqrt{p_0(1-p_0)/n}$, provided the sampling and large-count conditions are appropriate.

For a test, the expected numbers under the null model are $np_0$ with disease and $n(1-p_0)$ without disease. These should be sufficiently large for the normal approximation, commonly at least 10 each. The observations should also be independent; a random sample should be a small fraction of the population when sampling without replacement. For an interval estimating $p$, the success-failure check instead uses the observed counts $n\hat{p}$ and $n(1-\hat{p})$.

Worked Example: Screening Rate Above a Benchmark

In a hypothetical random sample of 200 adults from a large region, 30 are found to have a specified disease. The health team asks whether the disease rate in the region exceeds 10%. Let $p$ be the regional disease proportion. The outcome is binary and there is one sampled group, so this calls for a one-proportion z test with $H_0:p=0.10$ and $H_a:p>0.10$.

The sample proportion is $\hat{p}=30/200=0.150$. Under the null, the expected disease and non-disease counts are $200(0.10)=20$ and $200(0.90)=180$, respectively, both at least 10. If the sampling was genuinely random and the region contains far more than 2,000 adults, treating the observations as independent is reasonable.

The test statistic uses the null proportion in its standard error:

z=\frac{\hat{p}-p_0}{\sqrt{p_0(1-p_0)/n}} =\frac{0.150-0.100}{\sqrt{0.100(0.900)/200}} =\frac{0.050}{0.02121} \approx 2.357.

The one-sided p-value is approximately $P(Z\ge 2.357)=0.0092$, where $Z$ has a standard normal distribution under the null model. At the 5% significance level, this is evidence that the regional disease rate is greater than 10%. It does not establish why the rate is higher or show that a screening program caused the difference.

Two Treatment Success Rates: A Two-Proportion Procedure

When success is recorded as a binary outcome for each patient in two independent treatment groups, the target is a difference between population proportions. If group 1 receives treatment 1 and group 2 receives treatment 2, define the direction explicitly as $p_1-p_2$. A test of equal success rates has $H_0:p_1-p_2=0$. Under that null hypothesis, the usual two-proportion z test estimates the common rate by pooling the successes across both groups.

The groups must be independent: an individual in one group should not also contribute an outcome to the other group, and the group assignments or samples should support the intended inference. For a pooled test, check that the expected success and failure counts in each group under the null are sufficiently large. If the study is randomized, random assignment supports a causal comparison of assigned treatments for the study participants; it does not by itself guarantee that the results generalize to every patient population.

Worked Example: Comparing Two Treatment Success Rates

In a hypothetical randomized trial, 68 of 100 patients assigned to treatment A meet a prespecified success criterion, compared with 52 of 100 assigned to treatment B. The question is whether the success rates differ. Set $p_A$ and $p_B$ as the population success proportions for the two treatments and define the difference as $p_A-p_B$. The alternatives are $H_0:p_A-p_B=0$ and $H_a:p_A-p_B\ne0$.

The groups contain different patients, so the design is independent rather than paired. The pooled success proportion under the null is $\hat{p}_{\text{pool}}=(68+52)/(100+100)=120/200=0.600$. Expected successes are 60 in each group and expected failures are 40 in each group, satisfying the large-count check. The randomized assignment supports comparison of the treatments, assuming outcomes are recorded consistently and the trial is carried out as planned.

The observed difference is $68/100-52/100=0.160$, or 16 percentage points. The pooled standard error and test statistic are:

SE=\sqrt{0.600(1-0.600)\left(\frac{1}{100}+\frac{1}{100}\right)} =\sqrt{0.0048} \approx 0.06928,

z=\frac{0.160-0}{0.06928}\approx 2.309.

The two-sided p-value is approximately 0.0209. At the 5% level, the data provide evidence of different treatment success rates. The observed difference favors A by 16 percentage points in this trial, but the test alone does not establish that every patient would benefit more from A.

One Measurement Against a Benchmark: A One-Sample t Procedure

For a quantitative measurement from one group, a comparison with a reference value concerns the population mean $\mu$. If the population standard deviation $\sigma$ is unknown, use a one-sample t procedure based on the sample mean $\bar{x}$, sample standard deviation $s$, and sample size $n$. For testing $H_0:\mu=\mu_0$, the statistic is $t=(\bar{x}-\mu_0)/(s/\sqrt{n})$, with $df=n-1$.

The observations should be independent and arise from an appropriate sampling or assignment design. For a large sample, the sampling distribution of $\bar{x}$ is often adequately approximated by a normal distribution; for a small sample, the measurement distribution should be roughly normal without severe outliers or strong skew. A clinical reference value supplies the comparison value, not proof that the value is an appropriate threshold for every clinical decision.

Worked Example: Mean Blood Pressure Above a Reference

In a hypothetical sample of 16 adults receiving a screening assessment, the mean systolic blood pressure is 126 mmHg and the sample standard deviation is 12 mmHg. The question is whether the target population mean exceeds the reference value of 120 mmHg. This is one quantitative sample compared with one benchmark, so use a one-sample t test: $H_0:\mu=120$ mmHg and $H_a:\mu>120$ mmHg.

Assume the adults were independently sampled from the target population. With only 16 observations, the t procedure requires the measurements to be reasonably consistent with a normal population, without severe skewness or extreme outliers; this should be checked in the data before proceeding. The population standard deviation is not supplied, so $s$, rather than a known $\sigma$, belongs in the standard error.

Here $SE=s/\sqrt{n}=12/\sqrt{16}=12/4=3$ mmHg. Therefore,

t=\frac{\bar{x}-\mu_0}{s/\sqrt{n}} =\frac{126-120}{12/\sqrt{16}} =\frac{6}{3} =2.000,\qquad df=16-1=15.

For $t=2.000$ with 15 degrees of freedom, the one-sided p-value is approximately 0.032. At the 5% level, these data give evidence that the population mean systolic blood pressure is above 120 mmHg. The conclusion concerns the population mean in context, not the blood pressure of each person in the sample.

Independent Treatment Groups: A Two-Sample t Procedure

When two independent treatment groups produce a quantitative outcome, the target is a difference in population means. Define the order before interpreting the result, such as $\mu_A-\mu_B$. The two-sample t procedure compares $\bar{x}_A-\bar{x}_B$ with zero or estimates that difference with an interval. Welch’s procedure uses separate sample variances and does not require equal population variances. A pooled t procedure requires the stronger assumption that the two population variances are equal; do not select it merely because two sample standard deviations happen to look similar.

Both groups should have independent observations, with no patient contributing to both groups. The outcome should be quantitative, and each group’s data should be suitable for t inference: with small samples, check for strong skewness and outliers; larger samples can tolerate moderate non-normality more readily. Random assignment or appropriate sampling is needed for the desired interpretation. If the same patients are measured under both treatments or at two times, the observations are paired and an independent two-sample procedure does not fit.

Why the Welch standard error has this form: For independent groups, the estimated variance of $\bar{x}_A$ is $s_A^2/n_A$, and that of $\bar{x}_B$ is $s_B^2/n_B$. Independence makes the covariance zero, so the variance of their difference is the sum: $$ \widehat{\operatorname{Var}}(\bar{x}_A-\bar{x}_B) =\frac{s_A^2}{n_A}+\frac{s_B^2}{n_B}. $$ Taking the square root gives $SE=\sqrt{s_A^2/n_A+s_B^2/n_B}$. Welch’s degrees of freedom use these two estimated variance contributions and need not be an integer.

Worked Example: Comparing Mean Recovery Times

In a hypothetical study, different patients are randomly assigned to treatment A or B. Time to meet a defined recovery criterion is measured in days. For A, $n_A=12$, $\bar{x}_A=5.4$ days, and $s_A=1.2$ days; for B, $n_B=10$, $\bar{x}_B=4.6$ days, and $s_B=1.5$ days. The question asks whether mean recovery time differs. The outcome is quantitative and the groups are independent, so choose a two-sample t procedure. Without a justified equal-variance assumption, choose Welch’s procedure and define the contrast as $\mu_A-\mu_B$.

The random assignment creates independent treatment groups, and each patient contributes one outcome. Since the sample sizes are modest, the group distributions should be checked for severe skewness and influential outliers. Assuming those checks are satisfactory, the observed difference is $5.4-4.6=0.800$ days. Its Welch standard error is:

SE=\sqrt{\frac{1.2^2}{12}+\frac{1.5^2}{10}} =\sqrt{0.120+0.225} =\sqrt{0.345} \approx 0.5874\text{ days}.

The test statistic for $H_0:\mu_A-\mu_B=0$ against $H_a:\mu_A-\mu_B\ne0$ is $t=0.800/0.5874\approx1.361$. Welch’s degrees of freedom are approximately 17.16, giving a two-sided p-value of about 0.191. At the 5% level, the study does not provide sufficient evidence of a difference in mean recovery time. The estimated difference is 0.800 days for A minus B, but the test result does not show that the true difference is zero.

A Short Selection Workflow

Name the outcome.
Decide whether each observation is a binary category, such as success or disease status, or a quantitative measurement, such as blood pressure or recovery time.

Name the target parameter.
A binary outcome leads to a population proportion $p$; a quantitative outcome leads to a population mean $\mu$. A comparison of two groups requires a difference, with the group order stated.

Read the design.
Identify one group or two, and determine whether two groups are independent or paired. A benchmark value is not itself a second sample.

Match the goal and verify conditions.
Use a test when the question evaluates a specified claim and an interval when it requests plausible parameter values. Check sampling or assignment, independence, and the relevant count or distribution conditions.

Common Selection Errors

Using a t procedure for a disease rate. Disease status is binary, so the parameter is a proportion, not a mean. Use a one-proportion procedure for one group or a two-proportion procedure for two independent groups.
Using a two-sample procedure just because a benchmark is mentioned. One group compared with a fixed benchmark is a one-sample question; the outcome type determines whether it concerns a proportion or a mean.
Ignoring how observations are connected. If each patient supplies outcomes under both conditions, the measurements are paired. Treating them as independent discards the design and uses the wrong standard error.
Reversing a group difference after calculating it. If the reported estimate is A minus B, a positive value means the sample mean or proportion for A is larger. Keep that order consistent in the hypotheses and conclusion.
Interpreting a non-significant result as proof of equality. A large p-value means the data do not provide sufficient evidence against the null at the chosen level; it does not establish that the population parameters are identical.

Procedure choice is not a label chosen from the subject area. It follows from the parameter and design: one binary outcome, two binary groups, one quantitative group, or two independent quantitative groups. Once that match is made, checking conditions and stating the comparison in context help ensure that the result answers the health question actually asked.

Check Your Understanding

For each scenario, identify the outcome type, parameter or contrast, procedure family, and one relevant condition or design check.

A random sample of patients is classified by whether a condition is present. The question asks whether its population rate is below a stated benchmark. What procedure fits?
Patients are randomly assigned to one of two vaccines, and each is classified as protected or not protected. Which population quantities should be compared, and which procedure family fits?
A sample of patients has a quantitative laboratory measurement, and its mean is compared with a clinical reference value. What procedure fits if the population standard deviation is unknown?
Two distinct treatment groups have a quantitative outcome, with no reason to assume equal population variances. What procedure should be considered, and what feature makes the groups independent?
The same patients have a quantitative measurement before and after treatment. Why would an independent two-sample t procedure be a poor selection?

Health question	Outcome and design	Parameter(s)	Procedure family
Does the disease rate in one screened population exceed a benchmark?	Binary disease status; one group	One population proportion \(p\)	One-proportion test or interval
Do two treatments have different success rates?	Binary success status; two independent groups	Difference \(p_1-p_2\)	Two-proportion test or interval
Is a group’s mean measurement above a clinical reference value?	Quantitative measurement; one group	One population mean \(\mu\)	One-sample t test or interval
Do two independent treatments have different mean outcomes?	Quantitative measurement; two independent groups	Difference \(\mu_1-\mu_2\)	Two-sample t test or interval