HSCI 341 — Lesson 15

Validity in Observational Studies

Fundamental Epidemiological Concepts and Approaches

Kiffer G. Card, PhD, Faculty of Health Sciences, Simon Fraser University

Learning objectives for this lesson:

  • Identify different types of selection bias and assess whether a study is likely to suffer from it
  • Determine the likely direction and magnitude of selection bias using sampling fractions or sampling odds
  • Apply principles of bias prevention in study design, including secondary-base studies
  • Explain differences between non-differential and differential misclassification bias in terms of sensitivity and specificity
  • Evaluate misclassification of exposure, disease, or both in 2×2 tables
  • Evaluate the likely impact of misclassification using sensitivity analysis
  • Apply validation studies and regression calibration to adjust observed data
  • Modify sample-size estimates to account for misclassification

This course was developed by Kiffer G. Card, PhD, as a companion to Dohoo, I. R., Martin, S. W., & Stryhn, H. (2012). Methods in Epidemiologic Research. VER Inc.

Section 1 of 5

Introduction & Selection Bias

⏱ Estimated reading time: 20 minutes

12.1 Introduction to Validity

An awareness of the key features of study design, implementation, and analysis should help ensure we obtain valid results from research. The term validity relates to the absence of a systematic bias in results — a valid measure of association in the study group will have the same value as the true measure in the source population (except for variation due to sampling error).

To the extent that the study group and source population measures differ systematically, the result is said to be biased.

Key Concept: Internal vs. External Validity

Internal validity means the study allows unbiased inferences about associations in the source population.

External validity relates to making correct inferences to populations beyond the source population (the target population).

Generalisability is an inferential step beyond external validity — extending valid scientific theories to broadly defined populations (e.g., across populations and/or species).

The Three Major Types of Bias

Click each card to learn more about the three major categories of bias that threaten the validity of observational studies:

🎯
Selection Bias
Click to explore
🔍
Information Bias
Click to explore
🔀
Confounding Bias
Click to explore

12.2 Selection Bias

Selection bias results from the fact that the composition of the study group differs from that in the source population, and this biases the association observed between the exposure and the outcome of interest. Selection bias can affect study results significantly.

From a sampling and study-design perspective, each study will have an objective that relates to a defined target population. Ideally, the study group would completely reflect the source population, which in turn would reflect the target population. In practice, this is rarely the case.

Bias Variables and DAGs

Bias variables influence participation in a study in a way that causes the initial or final composition of the study group to differ from the source population, thus biasing the observed association.

The basic conditions for selection bias can be shown using directed acyclic graphs (DAGs):

Scenario 1: No association in source E D Selection Scenario 2: Bias variable present E D Selection Bias Variable

In Scenario 1, both E and D independently affect selection. When we condition on selection (study only the responders), E and D become associated even though they are independent in the source population. In Scenario 2, a bias variable related to both exposure and disease directly affects selection, creating a spurious association.

Sampling Fractions & Sampling Odds

We can understand selection bias by examining sampling fractions. The source population and study group follow the structure shown below:

Source PopulationE+E−
D+A1A0M1
D−B1B0M0
N1N0N

The four sampling fractions (sf) represent the proportion selected from each cell:

Eq 12.1 — Sampling Fractions
sf11 = a1/A1    sf12 = a0/A0
sf21 = b1/B1    sf22 = b0/B0

If subjects were selected by random sampling, all four fractions would be equal — no selection bias. If the sampling fractions are equal, the OR of the sampling fractions (ORsf) equals 1, and there is no bias in the observed OR.

Key Insight

The four sampling fractions can be unequal and still produce no bias in the observed OR, provided ORsf = 1. Also, if ORsf = 1, there is no bias to the risk ratio (RR) if disease is infrequent.

Sampling Odds

In practice, sampling odds may be easier to conceptualise than individual sampling fractions. For a cohort study, we compare the sampling odds of disease among exposed versus non-exposed subjects:

Eq 12.2 — Sampling Odds
soD+|E+ = sf11 / sf21
soD+|E− = sf12 / sf22

If these selection odds are equal, there is no bias. If the ratio of sampling odds is greater than 1, bias is away from the null; if less than 1, bias is toward the null.

Example 12.1: Selection Bias Due to Non-Response

Consider a source population where 10% are exposed, with disease risk of 25% in the exposed and 12% in the non-exposed. If non-response is related to exposure only (30% non-response in exposed, 10% in non-exposed) and unrelated to outcome, the study group RR (2.04) matches the source population RR (2.08) and OR (2.49 vs 2.44) — no bias.

However, if non-response is related to both exposure and outcome (disease risk twice as high in non-responders), then:

  • Study group RR = 1.73 vs true RR = 2.04 (biased toward the null)
  • Study group OR = 1.90 vs true OR = 2.38 (biased toward the null)
  • ORsf = 0.8, so observed OR = true OR × 0.8 = 2.38 × 0.8 = 1.90

Reflection

Think about a study you have encountered (from previous lessons or your own reading). Could selection bias have affected the results? How would you assess whether the study group was representative of the source population?

Minimum 20 characters required.

✓ Reflection saved

Knowledge Check: Section 1

1. Which of the following best describes internal validity?

Internal validity means the study allows unbiased inferences about associations in the source population, free from systematic error. External validity refers to generalising beyond the source population.

2. If all four sampling fractions (sf) are equal, what can we conclude?

Equal sampling fractions indicate that selection into the study group is independent of both exposure and disease status, meaning there is no selection bias. This does not address confounding or the actual exposure-outcome association.

3. A bias variable in the context of selection bias is one that:

Bias variables influence participation in the study (selection) in a way related to both exposure and/or disease, causing the study group composition to differ from the source population. The direction of bias depends on the specific relationships involved.
Section 2 of 5

Examples & Reduction of Selection Bias

⏱ Estimated reading time: 25 minutes

12.3 Examples of Selection Bias

Selection bias can manifest in many ways across different study designs. Understanding these patterns helps researchers anticipate and prevent bias during the design phase.

12.3.1 Choice of Comparison Groups

A general principle is that study groups should be selected from the same source population. In cohort studies, it is important that the non-exposed group be comparable with respect to other risk factors for the outcome. In case-control studies, the control group should reflect either the prevalence of exposure in the ‘non-case’ members of the population from which the cases arose.

Design Principle

A single-cohort design (where exposed and non-exposed come from the same population) is generally less susceptible to selection bias than a two-group cohort design, since both groups come from the same population by definition.

Types of Selection Bias

Click each card to explore different types of selection bias:

📬
Non-Response Bias
Click to explore
💪
Healthy Worker Effect
Click to explore
🏫
Berkson’s Fallacy
Click to explore
🚶
Loss to Follow-Up
Click to explore
🔬
Detection Bias
Click to explore
❓
Missing Data Bias
Click to explore

12.4 Reducing Selection Bias

Prevention Strategies
  • Be aware of potential pitfalls in selecting study subjects from the proposed source population
  • In cohort studies, take care when selecting the comparison group and ensure equal follow-up of both exposed and non-exposed groups
  • Minimise non-response bias, missing data, and detection bias
  • Case-control studies are particularly susceptible; minimise differential response to study participation between cases and potential controls
  • Where possible, use only incident cases and obtain controls from the same source population as the cases
Evaluating and Correcting Selection Bias

For valid control of selection bias, one of two conditions must be met:

  1. The factors associated with selection must be antecedents of both exposure and disease, and the distributions must be known in the source population — allowing the bias to be controlled like confounding.
  2. A bias breaker (a variable strongly related to selection and study participation that produces the bias) can be identified. Unbiased estimates of its population distribution can then be obtained, and the ‘corrected’ estimates are not associated with ‘selection’.

Additionally, the potential impact of selection bias can be assessed by examining sampling fractions using deterministic or stochastic sensitivity analysis (as in Example 12.2).

Example 12.2: Evaluating Potential Selection Bias

In a study of childhood respiratory disease (CRD) and regular daycare attendance, the observed OR was 2.33 (95% CI: 1.04–5.19). Using deterministic sampling fractions (sf) to assess the impact of possible selection bias:

CellDeterministic sf
Exposed cases (E+D+)0.5
Non-exposed cases (E−D+)0.6
Exposed controls (E+D−)0.05
Non-exposed controls (E−D−)0.1

The ‘adjusted’ OR (after accounting for the sampling fractions) was 1.40 — a 67% reduction from the observed OR. The true association would be considerably weaker than what was observed if this selection bias were present.

Reflection

Consider Berkson’s fallacy in the context of hospital-based case-control studies. Why might using hospital controls lead to biased estimates of the exposure-disease association? Can you think of an example where this might occur?

Minimum 20 characters required.

✓ Reflection saved

Knowledge Check: Section 2

1. The “healthy worker effect” is an example of:

The healthy worker effect is a form of selection bias where workers tend to be healthier than the general population because only those healthy enough to work are included in the study. This selective entry biases comparisons between workers and the general population.

2. Berkson’s fallacy is most likely to occur in:

Berkson’s fallacy arises specifically in hospital-based (secondary-base) case-control studies where admission risk is related to both the exposure and the disease, creating a spurious association in the hospitalised population.

3. Which of the following is true about non-response and selection bias?

Low response rates are not uncommon and do not automatically produce selection bias. The key determinant is whether non-response is differential with respect to both exposure and outcome simultaneously. A high response rate does not guarantee absence of bias either.
Section 3 of 5

Information Bias & Misclassification

⏱ Estimated reading time: 25 minutes

12.5 Information Bias

The previous discussion was concerned with whether study subjects had the same exposure-disease association as that in the source population. Now we review the effects of incorrectly classifying or measuring the study subjects’ exposure, extraneous factors, and/or outcome status.

When describing errors in classification of categorical variables, the resultant bias is called misclassification bias. The errors can be described in terms of sensitivity (Se) and specificity (Sp):

  • Sensitivity (Se): the probability that an individual with the event (e.g., exposed) will be correctly classified as having it
  • Specificity (Sp): the probability that an individual without the event will be correctly classified as not having it

When variables of interest are continuous, classification errors are termed measurement error or bias. The bias can arise from:

  • A lack of accuracy (systematic bias in the measurement)
  • A lack of precision (variability in repeated measurements)

Non-differential measurement error tends to bias the dose-response curve towards the null.

12.6 Bias from Misclassification

Misclassification bias results from a rearrangement of study individuals into incorrect categories because of errors in classifying exposure, outcome, or both.

12.6.1 Non-Differential Misclassification of Exposure

If misclassification of the exposure and outcome are independent (i.e., errors in classifying exposure are the same in diseased and non-diseased subjects, and vice versa), the misclassification is called non-differential.

Non-Differential Exposure Misclassification
SeE|D+ = SeE|D− = SeE     and/or     SpE|D+ = SpE|D− = SpE

With dichotomous exposures and outcomes, non-differential errors will bias measures of association toward the null (given SeE + SpE > 1). The observed cell values are a mixture of correctly and incorrectly classified subjects:

True NumberObserved (Incorrectly Classified)
a1a1′ = SeE·a1 + (1−SpE)·a0
a0a0′ = (1−SeE)·a1 + SpE·a0
b1b1′ = SeE·b1 + (1−SpE)·b0
b0b0′ = (1−SeE)·b1 + SpE·b0
Important

Exposure misclassification does not affect the disease status totals. Only the exposure category totals change. Relatively small errors (10–20%) can have sizable effects on relative risks.

Example 12.3: Impact of Non-Differential Exposure Misclassification

Consider a study with a true OR of 3.86 (90 exposed cases, 70 non-exposed cases, 210 exposed non-diseased, 630 non-exposed non-diseased). If we assume SeE = 0.80 and SpE = 0.90:

  • Exposed cases: 90×0.8 + 70×0.1 = 79
  • Non-exposed cases: 90×0.2 + 70×0.9 = 81
  • Observed OR = (79 × 690) / (81 × 310) = 2.17

As predicted, the non-differential errors have reduced the OR from 3.86 to 2.57 — bias toward the null.

12.6.2 Evaluating Non-Differential Exposure Misclassification

If the most likely values of SeE and SpE are known, we can correct the observed classifications. Since b1′ + b0′ = b1 + b0 = m0, we can solve for the true cell values:

Eq 12.3 — Correcting for Exposure Misclassification
b1 = [b1′ − (1 − SpE) × m0] / (SeE + SpE − 1)
Eq 12.4
a1 = [a1′ − (1 − SpE) × m1] / (SeE + SpE − 1)

12.6.3 Non-Differential Misclassification of Disease

In cohort studies, with non-differential misclassification of disease:

Non-Differential Disease Misclassification
SeD|E+ = SeD|E− = SeD    and/or    SpD|E+ = SpD|E− = SpD

There are two components: establishing initial health status (to exclude prevalent cases) and identifying new cases during follow-up. Imperfect sensitivity fails to exclude subjects with the outcome at the study outset; imperfect specificity has less impact.

For binary outcomes, non-differential errors bias the association measure toward the null.

In case-control studies, diagnostic errors applicable to cohort studies do not apply unless SpD = 1.00. This is because imperfect disease sensitivity does not bias the RR or IR, and only biases the OR if disease frequency is common.

The key is to verify diagnoses so there are no false positive cases. When SpD < 1, non-cases will be included as cases. The case-control sensitivity and specificity differ from the population values:

Eq 12.5 & 12.6 — Case-Control Se & Sp
Secc = SeD / [(SeD + sf·(1 − SpD))]
Spcc = sf·SpD / [(1 − SpD) + sf·SpD]

Thus, external estimates of SeD and SpD cannot be used to correct misclassification in case-control studies.

12.6.5 Misclassification of Both Exposure and Disease

When both exposure and disease are misclassified, we need to pay close attention to reducing these errors whenever possible. Most researchers prefer to evaluate errors for the more important one first, conducting a “what if?” analysis one set of errors at a time.

12.6.6 Differential Misclassification

If the errors in exposure classification are related to the status of the outcome under study, the errors are called differential:

Differential Exposure Misclassification
SeE|D+ ≠ SeE|D−    and/or    SpE|D+ ≠ SpE|D−

The resulting bias may be in any direction — either exaggerating or underestimating the true association. In case-control studies, recall bias is one common illustration: ‘affected’ subjects (cases) may have increased sensitivity, and perhaps lower specificity, than non-affected subjects in recalling previous exposures.

12.6.7 Reducing Misclassification Errors

Strategies for Reducing Misclassification
  • Use clear and explicit guidelines for classification
  • Have well-trained, consistent research personnel
  • Double-check exposure and disease status when possible (e.g., lab confirmations, confirmatory records)
  • Validate the test or survey instrument prior to widespread use
  • Collect specific rather than general exposure data (to reduce attenuation)
  • Use blinding techniques so survey personnel cannot equalise errors
  • Reduce misclassification of extraneous variables (confounders) as well, since poorly measured confounders cannot be fully controlled

Reflection

Why is non-differential misclassification generally considered less “dangerous” than differential misclassification? Under what circumstances might non-differential misclassification still be problematic?

Minimum 20 characters required.

✓ Reflection saved

Knowledge Check: Section 3

1. Non-differential misclassification of a dichotomous exposure typically biases the odds ratio:

Non-differential misclassification of a dichotomous exposure (where Se + Sp > 1) biases the association measure toward the null. This is because the errors mix exposed and non-exposed subjects equally across disease groups, diluting the true association.

2. Recall bias is an example of:

Recall bias occurs when cases recall their past exposures differently from controls (typically with higher sensitivity but lower specificity). Since the errors differ by disease status, this is differential misclassification. The direction of resulting bias is unpredictable.

3. Why can population SeD and SpD estimates not be used to correct disease misclassification in case-control studies?

In case-control studies, the sampling fractions for cases and controls change the effective case-control sensitivity and specificity. The Se and Sp observed at the case-control level differ from population-level values (Eq 12.5 and 12.6), so external population estimates cannot be directly applied for correction.
Section 4 of 5

Validation, Measurement Error & Correction

⏱ Estimated reading time: 20 minutes

12.7 Validation Studies to Correct Misclassification

A thorough review of validation studies to correct misclassification identified four main approaches: regression calibration, maximum likelihood, semi-parametric, and Bayesian methods. One key finding is that the more advanced methods are not user-friendly, while ‘simple’ approaches have important limitations.

For validation, we select a subsample of study subjects and verify their exposure and/or disease status. For direct estimates of sensitivity and specificity, we are determining:

Validation: Observed → True
p(D′ = 1 | D = 1) — probability of observed state given true state

Whereas when correcting for misclassification, we attempt to determine the reverse:

Correction: True → Observed
p(D = 1 | D′ = 1) — probability of true state given observed state

Two-stage samples (Chapter 10) are useful for validation. We select a subsample and verify their true status to obtain direct estimates of Se and Sp.

ApproachDescriptionLimitations
Regression CalibrationUse a validation subsample to calibrate measurement errors; regress true values on observed valuesAssumes non-differential errors; needs modification for differential errors
Maximum LikelihoodJointly model the true and observed data using likelihood functionsComplex; not user-friendly
Semi-parametricFewer distributional assumptions than maximum likelihoodStill technically demanding
BayesianIncorporate prior information about error rates; can use hidden Markov modelsRequires specification of priors; can be sensitive to prior choices
Caution: Sensitivity to Error Rate Estimates

Post-hoc adjustments for misclassification are very sensitive to changes in the error rate estimates used. Unless there is an extremely thorough validation procedure, different ‘corrected’ results could arise from a range of apparently sensible choices of the correction factor.

It is very important for the sensitivity and specificity of misclassification to be equivalent (‘transportable’) in the two datasets (validation and study) before attempting to adjust for errors.

12.8 Measurement Error

Errors in measuring quantitative factors can lead to biased measures of association. The bias can arise because the variable is not measured accurately (systematic bias) or due to a lack of precision (variability).

Regression Calibration Estimate (RCE)

To introduce the concepts of correcting measurement errors, suppose we have 2 quantitative exposure factors (X1 and X2) and a binary or continuous outcome (Y). The uncorrected ‘naive’ model is:

Eq 12.7 — Naïve Model
Y = β0u + β1uX1′ + β2uX2′

where the subscript ‘u’ indicates the coefficients are biased because the predictor variables (X′) are measured with error. The regression calibration estimate (RCE) involves:

Step 1: Perform a Validation Study

Take a random subset of study subjects and obtain the true values for X1 and X2. Regress each true X variable on the set of observed predictor variables:

Eq 12.8 & 12.9
X1 = β0 + λ11X1′ + λ12X2′
X2 = β0 + λ21X1′ + λ22X2′
Step 2: Predict and Regress

Calculate the estimated (predicted) X values for all study subjects (X1rc and X2rc) using the calibration equations. Then regress Y on these estimated values:

Eq 12.10 — Calibrated Model
Y = β1rc + β1rcX1rc + β2rcX2rc

The coefficients β1rc should provide less biased estimates of the true X–Y association than the naïve estimates. Standard errors need to be adjusted for the calibration process.

12.9 Errors in Surrogate Measures of Exposure

Often, epidemiologists study the effects of a complex exposure using surrogate measures. For example, in air pollution studies, what is the ‘appropriate’ measure? It could be a complex mixture of agents, doses, and durations.

Key Considerations for Surrogate Measures
  • Should exposure be measured on a continuous scale (preferred) or categorised as dichotomous/ordinal?
  • If specific agents are highly correlated, which one should be analysed, or should a composite variable be created?
  • Even if variables are measured “without error,” they may still be surrogates that fail to reflect true exposure
  • One solution: ask about the effects of measurable components (e.g., sulphur dioxide) rather than the broad concept (“air pollution”)

12.10 Impact of Information Bias on Sample Size

Classification and measurement errors can have a serious impact on measures of association. With non-differential misclassification, measures are biased toward the null; with classical measurement error models, the same is true for continuous variables. This leads to an important conclusion:

Sample Size Implications

The projected loss of power due to information errors should be considered and the sample size increased accordingly. The formulae for sample size estimation assumed that p1 and p2 were true population levels. However, with an imperfect test, the observed disease frequencies would be:

p1′ = Se·p1 + (1 − Sp)(1 − p1)
p2′ = Se·p2 + (1 − Sp)(1 − p2)

The difference p1′ − p2′ is usually less than p1 − p2, and it is the adjusted estimates that should be used to calculate sample size. Obuchowski (2008) generalises sample-size estimation to account for misclassification, response bias, and other features of clinical trials.

Summary: Types of Information Bias

TypeDefinitionDirection of BiasExample
Non-differential misclassification Classification errors are equal across comparison groups Toward the null (for dichotomous variables) Self-reported smoking with same error rate in cases and controls
Differential misclassification Classification errors differ by disease or exposure status Any direction (unpredictable) Recall bias in case-control studies
Non-differential measurement error Errors in continuous variables are equal across groups Toward the null Random variability in blood pressure readings
Misclassification of confounders Errors in measuring extraneous variables Incomplete control of confounding Poorly categorised socioeconomic status

Reflection

Consider the practical challenges of conducting a validation study to correct for misclassification. Why might it be difficult to obtain “true” values, and how could the sensitivity of corrections to error rate estimates affect your confidence in the adjusted results?

Minimum 20 characters required.

✓ Reflection saved

Knowledge Check: Section 4

1. Regression calibration is a method for:

Regression calibration uses a validation subsample (where true values are obtained) to calibrate the measurement errors. The true values are regressed on the observed values to create prediction equations, which are then applied to all subjects.

2. Why should sample sizes be increased when information bias is expected?

Non-differential misclassification and measurement error bias associations toward the null, reducing the effective difference between groups. Since the observed difference (p1′ − p2′) is smaller than the true difference (p1 − p2), a larger sample is needed to detect it with adequate power.

3. A major caution with post-hoc adjustments for misclassification is that:

A major limitation is that corrections for misclassification are very sensitive to the error rate estimates used. Small changes in assumed Se and Sp can produce large changes in the ‘corrected’ results. Without extremely thorough validation, different ‘corrected’ results could arise from apparently sensible choices.
Section 5 of 5

Lesson 15 — Final Review & Assessment

⏱ Estimated time: 20 minutes

Lesson Summary

In this lesson, you have explored the major threats to validity in observational studies, including selection bias and information bias, and learned strategies for detecting, preventing, and correcting these biases.

Core Concepts Reviewed

Section 1: Concepts of internal and external validity, selection bias mechanisms, DAG representations of selection bias, sampling fractions and their relationship to the odds ratio (Eqs 12.1–12.2), and Example 12.1 illustrating how selection bias arises.

Section 2: Specific types of selection bias (non-response, healthy worker effect, Berkson’s fallacy, loss to follow-up, detection bias, missing data), strategies for prevention and reduction, and Example 12.2 demonstrating bias in practice.

Section 3: Information bias and misclassification (non-differential vs. differential), sensitivity and specificity of exposure and disease measurement, correction formulae for non-differential misclassification (Eqs 12.3–12.6), and the predictable direction of bias under non-differential errors.

Section 4: Validation studies and their design, regression calibration for measurement error correction (Eqs 12.7–12.10), surrogate measures, the impact of misclassification on sample size requirements, and summary of information bias types.

Lesson 15 Comprehensive Assessment

This final assessment covers all topics from Lesson 15: Validity in Observational Studies. You must score 100% to complete this lesson. Review the feedback for any incorrect answers before retrying.

Final Reflection

Reflecting on this entire lesson, how would you design a study to minimise both selection bias and information bias? What are the key trade-offs between achieving high internal validity and maintaining generalisability (external validity)?

Minimum 20 characters required.

✓ Reflection saved

Final Assessment

1. Validity in epidemiological studies primarily refers to:

Validity relates to the absence of systematic bias — a valid measure of association in the study group will have the same value as the true measure in the source population (except for random sampling variation).

2. Which of the following is not one of the three major types of bias discussed in this chapter?

The three major types of bias discussed are selection bias, information bias, and confounding bias. Publication bias, while important in meta-analysis, is not one of the three categories covered in this chapter on validity.

3. If the OR of the sampling fractions (ORsf) equals 1, this indicates:

When ORsf = 1, the four sampling fractions, while potentially unequal, do not produce bias in the observed OR. Note that the sampling fractions do not need to all be equal — just their odds ratio needs to be 1.

4. In a DAG representation of selection bias, the bias typically occurs because:

When both exposure and disease (or a bias variable related to both) affect selection, conditioning on selection (studying only those selected) creates a spurious association between E and D even if none exists in the source population.

5. Berkson’s fallacy requires which condition for bias to occur?

Berkson’s fallacy requires that the exposure of interest has an independent risk of admission to the hospital or registry (i.e., p(H|E+) > 0). A differential admission risk between cases and control diseases is also needed to produce the bias.

6. The Hawthorne effect is best described as:

The Hawthorne effect refers to behavioural changes in study subjects as a result of being studied. The act of observation or enquiry into specific factors could lead participants to modify their behaviour, potentially causing differential management by exposure status.

7. Sensitivity (Se) of exposure classification is defined as:

Sensitivity for exposure classification is the probability that an individual who truly has the event (e.g., is exposed) will be correctly classified as having it. Specificity is the converse — the probability that a truly unexposed individual is correctly classified as unexposed.

8. Non-differential misclassification of a dichotomous exposure (with SeE + SpE > 1) biases the OR:

With dichotomous exposures and outcomes, non-differential errors (where Se + Sp > 1) bias measures of association toward the null. This occurs because the errors mix exposed and non-exposed subjects equally across disease groups, diluting the true association.

9. What is unique about non-differential exposure misclassification in terms of its effect on the 2×2 table?

Exposure misclassification only rearranges subjects between exposure categories within the same disease status row. Therefore, only the exposure category totals (column margins) change; the disease status totals (row margins) remain the same.

10. Differential misclassification differs from non-differential because:

In differential misclassification, the Se and/or Sp differ between comparison groups (e.g., SeE|D+ ≠ SeE|D−). The resulting bias can be in either direction — toward or away from the null — making it unpredictable and more problematic than non-differential misclassification.

11. Why can population-level SeD and SpD values not be used to correct disease misclassification in case-control studies?

In case-control studies, cases and controls are sampled separately at different rates (sampling fractions). This sampling process transforms the population-level SeD and SpD into different case-control Se and Sp values (Eq 12.5 and 12.6), making direct correction with population estimates invalid.

12. Which strategy is most effective for reducing recall bias?

Recall bias is a differential misclassification that arises when cases recall exposures differently from controls. Using objective, independently verified records (lab results, medical records, administrative data) bypasses the reliance on subjective recall and reduces this bias.

13. The regression calibration estimate (RCE) approach involves:

RCE involves: (1) obtaining true values for a validation subsample, (2) regressing true values on observed values, (3) using these calibration equations to predict corrected values for all study subjects, and (4) using the predicted values in the final analysis. This assumes non-differential measurement errors.

14. A “bias breaker” in the context of selection bias is:

A bias breaker is a variable that is strongly related to selection and study participation and produces the bias. If unbiased estimates of its population distribution can be obtained, the ‘corrected’ estimates are not associated with ‘selection,’ allowing the selection bias to be addressed.

15. When information bias (non-differential misclassification) is expected, the sample size should be:

Non-differential misclassification biases the association toward the null, effectively reducing the detectable difference between groups (p1′ − p2′ < p1 − p2). To maintain adequate statistical power to detect this smaller observed difference, a larger sample size is needed.

🏆 Congratulations!

You have successfully completed Lesson 15: Validity in Observational Studies.

You now understand the key concepts of selection bias, information bias, misclassification, measurement error, and methods for evaluating and correcting these biases in epidemiological research.