Epidemiological data

From the recording

Points the lecturer made in the lecture recording on Canvas that are not (fully) on the slides. Times refer to the recording.

• 02:04 — The WHO 1948 definition was praised because it was the first to define health positively (complete well-being), not only negatively as the absence of disease. Its critique: 'complete' well-being is utopian and cannot last (any flu or chronic disease would make you 'unhealthy'), hence the newer definition of health as the ability to adapt and self-manage. 'Environment' in environmental health includes the built environment (houses, neighbourhood design, nearby schools and facilities), not only the natural environment (air, green and blue spaces). (Slides p. 4–6 (definitions and the 'utopian' critique; the positive-definition point and the built-environment examples are verbal))
• 06:14 — Environmental hazards are classed as physical (e.g. heat, noise, particulate matter), chemical (e.g. pesticides, microplastics, gases such as NO₂) or biological (bioaerosols such as viruses, bacteria, moulds and pollen). Air pollution itself can be chemical, physical or biological, and most pollutants come from human activity. (Slides p. 7, 10 (lists the problems but does not sort them into physical/chemical/biological))
• 06:14 — The approaches map onto the risk-assessment paradigm: toxicology characterises the hazard itself (its toxic properties), physiology studies how the hazard affects the body, and environmental epidemiology (and One Health) covers the second half of the paradigm: assessing population exposure and characterising risk, i.e. how disease is distributed and what its environmental determinants are. (Slides p. 9, 11 (paradigm and approaches shown separately; the mapping between them is verbal))
• 10:18 — Short-term effects = health effects of today's exposure (e.g. what happens in the following days); long-term effects = effects of exposure sustained over years (here a 10-year average). Before analysis you narrow the question by fixing study area, population, one or a few exposures (there are hundreds of pollutants), an outcome that the exposure could plausibly affect, and short- vs long-term. (Slides p. 16–17 ('Short vs. long-term' listed without definition))
• 12:22 — Why the residential address is the proxy: people visit many micro-environments each day, so true personal exposure cannot be measured for a large population; researchers assume people spend most of their time at home. NO₂ is very high next to roads and drops sharply with distance, so each person needs an address-specific estimate rather than an area average. (Slides p. 21–22 (state the proxy and 'large spatial variability'; the 'most time at home' assumption and the steep road gradient are verbal))
• 16:24 — Remote sensing means satellite data; together with surface (ground) monitoring it is a type of routine monitoring network. Routine network data are free/open access, go back decades (useful for historical/long-term exposure) and each station measures several pollutants at once; satellite data have coarser spatial resolution. The Dutch network (Luchtmeetnet) gives hourly concentrations. (Slides p. 25–27 (lists surface monitoring and remote sensing under routine networks; 'satellite', 'historical' and 'coarser resolution' are verbal))
• 18:25 — Two ways to predict concentrations where nothing is measured: deterministic models (dispersion or chemical transport) start from the pollution sources and their emissions and use physical/chemical laws to predict how pollutants travel away from the source; stochastic/statistical models such as land-use regression ignore sources and emissions and start from measured concentrations, interpolating them to unmeasured locations. (Slides p. 30–32 (titles only: 'Deterministic – physical laws' vs 'Stochastic – statistical (LUR)'))
• 22:33 — Always inspect downloaded data before manipulating it (summarise and plot). The hourly boxplots show weekday NO₂ peaks at the morning rush hour and a smaller evening peak (traffic), lower values at weekends, and higher values in the cold season (heating adds NO₂). The prediction maps show the same: e.g. 8 a.m. on weekdays higher than 12:00. Plotting only the medians is a simpler alternative to crowded boxplots. (Slides p. 35–36, 40 (plots and maps shown without interpretation))
• 24:33 — How LUR learns: the outcome (Y) is the measured concentration at each station and the predictors are the land-use characteristics around it. If stations surrounded by more roads have higher NO₂, roads get a positive coefficient; if stations surrounded by green areas have low NO₂, green area gets a negative coefficient. The fitted equation is then applied to any location whose land use is known. The speaker said 'the more predictors, the better'; strictly, more predictors can overfit a model built on only 69 stations, which is one reason validation is needed. (Slides p. 33, 37–39 (steps and predictor table; the sign reasoning is verbal))
• 26:37 — Validating the LUR model: you can measure yourself at predicted locations (time-consuming), but the usual approach is a hold-out check: fit the model on part of the stations (e.g. 40 of 69), predict at the remaining stations and compare predictions with measurements; a good correlation suggests it will also predict well at unmeasured addresses (an assumption, not a proof). Also check input data quality first: a model built on bad data makes no sense. (Slides p. 41 (lists 'model verification and validation' and 'quality of routine monitoring networks' only))
• 30:43 — Of the four health-data categories (Galetsi et al. 2020), real-time patient data are e.g. smartwatch heart rate, sleep and step counts, and administrative data (healthcare costs/reimbursements) can serve as a proxy for clinical data, which are often too sensitive to access: you may not know the diagnosis but know that someone paid for certain therapies. (Slides p. 44 (categories and examples; the proxy use is verbal))
• 32:46 — Routine databases depend on administrative factors: the mortality registry shows more deaths on Mondays than on Sundays, not because people die more on Mondays but because of registration practices and medical shifts. Questionnaires: bias from wording and from mode of administration (interview vs written), and pooling questionnaires from different sources is hard when wording differs or answers are open. (Slides p. 46–47 ('depends on administrative factors', 'wording', 'mode of administration' listed without examples))
• 39:03 — Data processing is usually the most time-consuming step: a registry gives a date of death, not 'outcome yes/no', so you derive a 0/1 outcome (1 if the person died within the 2009–2019 follow-up). Choose the endpoint to match the exposure (e.g. for smoking you would look at lung cancer), and check there are enough events to analyse. (Slides p. 52–53 (analysis table and 'data processing' challenge; the derivation and endpoint example are verbal))
• 56:43 — Parametric models assume one fixed functional form (linear, polynomial); non-parametric models assume none and infer the shape from the data (more flexible); semi-parametric models assume part of the form and leave the rest flexible. Example: splines are a chain of polynomials whose shape changes with the exposure level, e.g. temperature–mortality is roughly linear at low temperatures but rises steeply at high temperatures. Model choice: look at what others did and at your own data. (Slides p. 57 (table of model families; splines and the temperature example are verbal))
• 60:48 — Cox regression was chosen because it is semi-parametric and the standard model for long-term survival (time-to-event) studies: the time until a person experiences the outcome. The HR is computed for a fixed increment in concentration; her example was 10 µg/m³ (she said 'micrograms per meter square'; air concentrations are per cubic metre). HR = 1: no association; HR > 1: positive; HR < 1: protective. She said 'if the hazard ratio is negative, so it's lower than one'. An HR cannot be negative; it is the coefficient (log HR) that is negative when HR < 1. (Slides p. 63 (HR definition and the 1/>1/<1 rule; the survival rationale and 10 µg/m³ example are verbal))
• 64:50 — Reading the Cox output: the model calculates exp(coef), which is the HR (1.04 for NO₂); coef is the log HR, followed by its SE, z and p-value. The baseline hazard h0(t) is the hazard when all covariates are 0, because exp(0) = 1. In R, Surv(age_b, age_end, mort) sets the time scale (age at start and end of follow-up) and the death indicator. She calls this 'the baseline hazard', which is loose: Surv() defines the outcome, not h0(t). Note the printed NO₂ coef (0.00392) actually gives exp = 1.004 per µg/m³; 1.04 corresponds to a 10 µg/m³ step, matching her 10 µg/m³ example. The exam will likely follow the slide's '4% per unit'. (Slides p. 67–69 (formula, R code and output table; how to read the columns is verbal))
• 66:59 — Significant vs substantial: yes, significant (HR > 1, p < 0.001). Substantial? 4% per unit looks small, but it is a typical size for a very large population study, and an HR is not a measure of health burden: to judge public-health impact you need to know how many people are exposed and to which levels. A student read 1.04 as '104% risk'; the correct reading is 4% higher hazard per increment. (Slides p. 69 (poses both questions without answering them))
• 69:07 — The exposure–response curve applies the HR across the whole exposure distribution: risk rises with concentration, and the study population is exposed to potentially harmful levels but not at the extreme end of the curve. She did not explain that the left panel of the results slide is per interquartile range (HR about 1.065, CI about 1.05–1.08), a different increment from the 1.04 per unit, or that the curve is non-linear (dipping below 1 around 10 µg/m³). (Slides p. 70 (left: 'HR per interquartile range'; right: spline HR curve over residential NO₂ with histogram))
• 71:11 — The course lecturer (not the guest) said he hoped everyone had recognised several steps of the data science lifecycle in the talk: question → data gathering (exposure and health) → processing and checking → exploration → modelling → interpretation and communication. Expect to map this case study onto the lifecycle from Lecture 1. (not on slides)
• 73:18 — Long-term exposure assessment must account for residential moves (residential history). Daily commuting within a region (e.g. living in Utrecht, studying in Amsterdam; within the Randstad) makes little difference to the estimate, but moving between very different regions (e.g. from Groningen to Utrecht) does. Research is moving towards better estimates of daily personal exposure. (not on slides)
• 75:27 — Limitations: predicted residential concentration is a proxy, not actual personal exposure, and the data are heavily modelled and processed, so such studies always list many limitations. Confidence comes from replication: similar results in other populations or countries (e.g. Belgium) and in subgroups (e.g. low-income groups, to help rule out confounding by income). One study alone rarely justifies strong conclusions. The course lecturer added that such studies were used as evidence in Volkswagen emissions-scandal court cases (he said 'CO2'; the scandal concerned NOx/NO₂ emissions). (not on slides)
• 79:39 — Privacy and data-science responsibility (the course lecturer's discussion): removing names does not make data anonymous, because address, age and other attributes can re-identify people. CBS microdata (individual-level data on every Dutch resident) is accessed only via an approved research proposal, after a short test on the usage rules, inside a secure environment that nobody else may watch. You see only the variables your project was approved for. Outputs are checked before export, with a minimum number of subjects per reported cell (she thought about 10) so that, for example, a single rare-disease case cannot be located. Results must be published because the data are public. (Slides p. 50–51, 53 (CBS microdata link and 'privacy regulations' only))