NR-503 · Week 6 of 8 · Bias, confounding and causal inference

NR-503 Week 6 Bias, Confounding and Causal Inference: How to Write It

The short answer

NR-503 Week 6 asks the hardest question in the discipline: when is an association worth believing. Three explanations compete with a real effect every time a study reports one. Chance, which the interval speaks to. Bias, which is a systematic error built into how people were selected or how things were measured. And confounding, where a third factor is tied to both the exposure and the outcome and produces an association that is not what it appears to be. Against those sit the considerations for causal reasoning: strength, consistency, sequence in time, dose response, plausibility and reversibility. Your section may print this as NR 503 or NR503; it is the same course. Chamberlain publishes no syllabi outside Canvas. The placement here is our teaching judgment from the course's catalog arc; your section's rubric decides what your week actually asks.

NR-503 Week 6 grading scale at Chamberlain, the criterion levels this assessment is scored on, from Chamberlain Tutors
How Chamberlain grades NR-503 Week 6, visualized by Chamberlain Tutors.

What NR-503 Week 6 asks for

Precision matters in the vocabulary here, because the terms are often used loosely in conversation and strictly in grading. Bias is not partiality on the part of a researcher; it is a systematic distortion introduced by the design. Selection bias occurs when the people studied differ from the population in a way related to both exposure and outcome. Information bias occurs when exposure or outcome is measured differently between groups, and recall bias is the version of it that haunts case control studies, where people who have the outcome search their memories harder.

Confounding is different in kind and it has three conditions worth stating in any paper: the factor is associated with the exposure, it is independently related to the outcome, and it does not sit on the causal path between them. That last condition is what separates a confounder from a mediator, and mislabelling a mediator as a confounder is a mistake graders look for.

The causal considerations are a framework for judgment, not a checklist to be ticked. Sequence in time is the only one that is strictly required; the rest raise or lower confidence. Deliverables at this point in an eight-week session are often a critique memo of three to five pages applying these ideas to a study or to a claim. If your section runs a discussion this week, draft it separately, since posts do not reopen once submitted in Canvas.

The NR-503 Week 6 method, step by step

Six moves that take an association apart before deciding what to believe about it.

  1. State the association in one plain sentence

    Exposure, outcome, population, direction and size. Everything that follows is an examination of this sentence, and papers that never state it cleanly end up examining a claim that keeps shifting shape.

  2. Deal with chance first, briefly

    Read the interval and say what it leaves open. This takes two sentences and it clears the ground, because there is no point analysing bias in an estimate whose interval comfortably includes no association at all.

  3. Work through selection, then measurement

    Ask who got into the study and how, then ask how exposure and outcome were each recorded and whether the recording could have differed between groups. Name the specific bias rather than saying the study may have been biased.

  4. Nominate confounders and test them against the three conditions

    For each candidate, say whether it is related to the exposure, whether it is independently related to the outcome, and whether it lies on the causal path. Anything that fails the third test is a mediator and belongs in a different part of your argument.

  5. Check what the researchers did about it

    Restriction, matching, stratification and adjustment in the analysis are the usual tools, and randomization is the one that handles unknown confounders as well as known ones. Say which was used, and say what remains uncontrolled, since residual confounding is almost always present in observational work.

  6. Apply the causal considerations as a weighing, not a tally

    Go through the considerations that your evidence can speak to, say which are met and which are not, and give sequence in time its special place. Then write a judgment with a confidence level attached rather than a verdict.

A layout and word budget for a critique memo

Sized for a memo of roughly 1,100 to 1,400 words. This is our own drafting frame rather than a university-issued template, and where your week's rubric names its own sections, follow the rubric and keep only the proportions.

SectionWhat belongs in itWord target
The association under examinationExposure, outcome, population, direction and size, in one paragraph with the source cited.120 to 150
Chance consideredThe interval read, with a sentence on what it leaves open and what it rules out.100 to 130
Bias examinedSelection and information problems named specifically, each tied to something in the study's methods.280 to 340
Confounding examinedTwo or three candidates tested against the three conditions, with mediators separated out.280 to 340
Control and what remainsWhat the researchers did about confounding, and what is left uncontrolled after it.180 to 220
Causal judgmentThe considerations weighed, sequence in time given its place, and a conclusion with a confidence level.250 to 300

Evidence craft for a critique

Point at something specific for every bias you claim. A named bias with no mechanism is an accusation. Say how the recruitment, the response rate, the instrument or the follow-up could have produced it, and say which direction it would push the estimate.

State the direction of the distortion. A bias that would move the estimate toward the null is a different problem from one that inflates it. Saying which way the error runs turns an objection into an analysis, and it is the sentence most student critiques leave out.

Look for dose response in the tables, not in the abstract. A gradient across exposure categories strengthens a causal reading, and it is usually visible in a stratified table even when the abstract reports only a single comparison. Report the gradient with its categories rather than asserting that one exists.

Keep consistency claims sourced. Saying that other studies found the same thing requires those studies. Give two, with their designs and populations, and note where a finding failed to replicate, since selective consistency is its own form of distortion.

Five mistakes that cost points in this week's territory

  • Bias used to mean prejudice. The technical sense is a systematic error in design or measurement, and using the everyday sense makes the whole section unscoreable.
  • Confounders asserted without the three conditions. Listing plausible-sounding variables is not analysis until each is tested for association with exposure, independent relation to outcome, and position off the causal path.
  • A mediator treated as a confounder. Adjusting for something on the causal path removes part of the effect you were trying to measure, and calling for that adjustment shows the distinction was missed.
  • Considerations ticked like a checklist. Announcing that six of nine are met is not judgment, and it also misrepresents how the considerations were meant to be used.
  • Sequence in time assumed. In a cross sectional study nothing establishes which came first, and a causal claim built on one is the clearest overreach at this stage.

Before you submit

  • The association is stated once, plainly, with its size and direction
  • Chance is addressed through the interval before bias is discussed
  • Every bias named is tied to a specific feature of the study and given a direction
  • Each proposed confounder is tested against all three conditions
  • What the researchers controlled and what remains uncontrolled are both stated
  • The causal judgment carries a confidence level rather than a verdict

Critique due this week?

Send the study or the claim, plus the prompt and rubric. A premium original critique comes back in 24 to 48 hours with each bias tied to a mechanism and every confounder tested properly.

Questions students ask about this stage

How do I tell a confounder from a mediator in my own topic?
Ask whether the variable sits on the road between the exposure and the outcome or beside it. If the exposure causes the variable and the variable then causes the outcome, it is a mediator and it is part of how the effect happens; adjusting for it would remove some of the effect you are trying to see. If the variable causes or accompanies the exposure and independently affects the outcome without being produced by the exposure, it is a confounder and leaving it alone distorts the estimate. Drawing the arrows on paper settles most cases in a minute. Where the direction is genuinely uncertain, say so in the paper and describe what the estimate would look like under each reading, since an author who can see the ambiguity and reason about both branches is doing precisely what the analysis row rewards.
Can an observational study ever support a causal claim?
It can contribute to one, and a body of observational work sometimes supports a causal conclusion that no single study could. What does the work is the accumulation: a strong association found repeatedly in different populations with different designs and different sources of error, a clear sequence in time, a gradient across levels of exposure, a plausible mechanism, and a decline in the outcome where exposure was reduced. That is how several major public health conclusions were established without a randomized trial, since trials of harmful exposures are not ethically available. What one observational study supports on its own is a well-characterized association with named limitations. Write your conclusion at the level your evidence reaches, and say what further evidence would raise your confidence, which is the sentence that shows the reasoning was understood rather than recited.
What do I write if the study controlled for almost everything?
Say what was controlled, then examine how well rather than whether. Adjustment depends on the confounder being measured accurately, and a variable captured crudely leaves residual confounding behind even after it appears in the model. Income recorded in three broad bands, or smoking recorded as ever and never, will not fully account for either. Ask also what could not be measured at all, since observational studies cannot adjust for unknown factors and randomization is the only tool that handles them. Then state your conclusion with the residual uncertainty attached: the association survived adjustment for these factors, measured this way, and the most likely remaining explanation is this. That is a stronger paper than either accepting a heavily adjusted result at face value or dismissing it because adjustment can never be perfect.

Keep going

Online now