Adjusted and unadjusted quitting-study estimates: what changed?

Keep the estimate attached to its model

An article may present a simple group comparison and several adjusted models. Identify which result the headline quotes, its effect measure, outcome, follow-up, and uncertainty. The word “adjusted” should lead you to the model description, not function as a synonym for corrected or clinically accurate.

Cochrane’s nonrandomized-study guidance emphasizes reporting how an estimate was derived, including the confounders controlled for. This reading exercise helps locate that information. It does not select covariates, evaluate a named study, or claim that a particular model removed every source of bias.

Cochrane: Cochrane Handbook, Chapter 24: Including non-randomized studies on intervention effects

Read the variables rather than assuming their purpose

Look for the authors’ rationale, the timing of measurement, and whether the adjustment was planned. A variable measured after an intervention may have a different role from a preexisting group difference. Do not assume that adding more variables always improves a result; determining an appropriate causal model requires methodological expertise.

For a hypothetical example, adults choose between two support programs, and an analysis adjusts for recorded age and baseline use. That adjustment does not establish that motivation, access, or other relevant differences were fully measured. Nor does it prove those unmeasured factors explain the finding. It identifies a question the report needs to address.

Compare models without selecting the favorable number

If the adjusted and unadjusted results differ, retain the difference and the stated rationale. It can reveal that the model matters to the interpretation. Do not automatically treat the largest apparent benefit as the final answer or declare that a smaller estimate proves the intervention has no effect.

Check whether the outcome, analysis population, and effect scale are the same before comparing the estimates. A model excluding participants with missing covariates may use a different set of people. A numerical change can therefore reflect more than one alteration. Keep unavailable details unresolved rather than attributing the whole change to a single variable.

Write a causal claim only as far as the design supports

A bounded note is: “This estimate adjusted for these stated variables using this analysis; these remaining limitations affect interpretation.” Nonrandomized data can provide useful evidence without establishing the same protections as randomized assignment. A model name alone does not resolve selection, measurement error, or omitted information.

If a treatment decision depends on the result, ask a qualified professional about the full analysis. Personal tracking cannot recreate the unobserved outcome you would have had under another support choice. You can continue describing your own observations while keeping research attribution separate. The reading task is complete when “adjusted” points to documented choices and limits rather than an automatic proof badge.