Skip to content
PaperFren

Research method

Logistic Regression

8 studies

Ask PaperFrenWatch this

Logistic regression models the log-odds of a binary outcome — preterm versus term, high cholesterol versus not, preeclampsia versus not — as a function of predictors. Results are usually odds ratios per IQR or per standardised unit. Close cousins in this set include conditional logistic models in matched case-control designs, elastic-net screening followed by unpenalized logistic, and modified Poisson models that report risk ratios instead of odds. An OR is not automatically a risk ratio, especially when the outcome is common.

Environmental papers use logistic (and related binary) models when the scientific question is a yes/no event under a pollution or PFAS contrast. The method answers 'how do the odds of this outcome change with this assigned exposure after measured confounders?' Its main limitation is that a selected AOR cannot prove the chemical caused the event, correlated traffic or PFAS layers remain, and swapping OR for RR, or PAH stations for a city-wide surface, changes what the number means.

Studies

8

Findings

6

7 supporting · 0 challenging · 0 qualifying citations

Open tensions

2

Currently

What we know

  1. A Los Angeles case-control of births near MATES monitors (preterm prevalence 9.2%) used stations, land-use regression and source-apportioned PM in conditional logistic models. Total PAHs and several PAH species showed ~30% odds increases (benzo(a)pyrene 13%), but PAH samples were small and limited to two stations. IQR increases in ammonium nitrate PM₂.₅ and 9–13% odds rises appeared for benzene, diesel/EC/OC and biomass-burning PM₂.₅.
  2. PROTECT also studied preterm birth (~9.1% of 1,092 pregnancies) but with modified Poisson risk ratios, not logistic ORs. Mean pregnancy PM₂.₅ was 6.8 μg/m³; an IQR increase gave RR 1.014 (individual-adjusted) and 1.012 (also municipal-adjusted). Phthalate metabolites did not modify the association. Monitor-based municipality assignment likely biases toward the null.
  3. In Ronneby versus Karlshamn, regression on ln-PFAS found cholesterol and LDL 7–9% higher in the exposed town and odds of high cholesterol about 1.3 per ln-PFAS unit. PFOS (median 157 ng/ml) and PFHxS (median 136 ng/ml) were correlated at r ≈ 0.9, so the OR is not compound-specific.
  4. Shanghai deliveries quantified eight PFAS in cord blood. HDP, preeclampsia and gestational hypertension rates were 6.1%, 2.8% and 3.3%. PFOA had the highest median (6.98 ng/ml), but elastic-net selected PFBS (median 0.047 ng/ml) for HDP/preeclampsia. Unpenalized logistic AORs were 1.64 for HDP and 1.81 for preeclampsia per standardised PFBS unit. Cord blood is not a full pregnancy-long exposure history.
  5. An administrative neurologic cohort reported 13,170 non-Alzheimer dementia, 4,201 Parkinson, 1,277 Alzheimer and 658 MS cases. Road proximity raised risks; greenness attenuated NAD/PD HRs by 0.3–6.2% and AD ORs by 11–28%. Air pollution associated with NAD and PD, not AD or MS; noise effects were generally null — diagnosis-specific binary/survival models, not one 'neurologic OR'.

Largest unresolved question

Two preterm-birth analyses both sit near a 9% outcome and do not report the same contrast. Los Angeles finds ~30% PAH odds increases (benzo(a)pyrene 13%) plus 9–13% for several PM sources, with PAH data from only two stations. PROTECT reports RR 1.014 per IQR of pregnancy PM₂.₅ at a mean of 6.8 μg/m³. Odds ratios and risk ratios are not interchangeable, and the exposure products (PAH stations versus municipality monitors) are not the same assignment.

Procedure

How it works

The main stages of the method. Specific protocols and analysis choices vary between studies.

  1. Code a binary environmental outcome

    The event is yes/no — preterm versus term, high cholesterol versus not, preeclampsia versus not. When the outcome is common, an odds ratio is not automatically a risk ratio.

  2. Assign an exposure

    Pollution, PFAS, or another contrast is taken from monitors, a land-use surface, or a biospecimen. Two stations, a city-wide surface, and a mixture of correlated PFAS are not interchangeable assignments.

  3. Fit log-odds or a close cousin

    Logistic, conditional logistic in matched case-control, elastic-net then unpenalized logistic, or modified Poisson that reports a risk ratio instead of odds. The software family is a choice about the number's meaning, not a default.

  4. Report an odds ratio after measured confounders

    Results are usually ORs per IQR or per standardised unit. Elastic-net selection among correlated PFAS is a prediction step, not a toxicological ranking of which compound caused the event.

  5. Do not treat the OR as a proven cause or as a risk ratio

    The model answers how the odds of this outcome change with this assigned exposure after measured confounders. Correlated traffic or PFAS layers remain, and swapping OR for RR, or PAH stations for a city-wide surface, changes what the number means.

Common misconceptions

  • An odds ratio of 1.3 is essentially a 30% increase in risk.

    OR approximates RR only when the outcome is rare. Preterm prevalence was 9.2% in Los Angeles and 9.1% in PROTECT; high cholesterol in Ronneby is not a rare event. PROTECT reports RRs near 1.01 per IQR — a different scale from a 30% PAH OR.

  • Elastic-net selected PFBS, so PFOA is unimportant in that Shanghai sample.

    PFOA had the highest median (6.98 ng/ml); PFBS (0.047 ng/ml) was the selected HDP/preeclampsia predictor after penalisation. Selection is a prediction step among correlated PFAS, not a toxicological clearance of PFOA, and cord blood is not pregnancy-long exposure.

  • Logistic occurrence models of canopy lichens measure biomass loss in the upper canopy.

    Swedish NFI presence on host trees does not measure upper-canopy biomass, and interacting drivers (logging, N deposition, climate) mean no single lever is shown to reverse the ~1.70%/year Alectoria loss.

Related