When a clean-air policy shrinks the fine particulate matter (PM2.5) readings at a monitor, researchers usually read that drop as the policy's effect. This proposal tests a specific way the reading can mislead. The instruments that weigh PM2.5 do not retain every chemical component equally. Filters are dried in the lab before weighing, and some real-time mass monitors heat the sample. Both steps drive off particle-bound water and the volatile species, mainly ammonium nitrate and semi-volatile organics. The chemical-speciation instruments add a second blind spot: they cannot vaporize mineral dust or black carbon at all.
China's coordinated "2+26" regional controls moved Beijing's pollution away from coal, whose products are sulfate and refractory soot, and toward traffic and secondary formation, which are nitrate- and organic-rich. Under that shift, the share of true airborne mass that instruments fail to record may itself have changed. If so, part of the reported improvement is a change in what the instruments miss, not a change in what people breathe.
The proposal builds a station-level difference-in-differences design, meaning a comparison of before-versus-after changes at treated stations against the same changes at untreated stations. Its outcome is the full mass budget: each measured species, the unexplained residual, and a closure-corrected estimate of ambient mass. The design asks whether the measured policy effect depends on which instrument did the weighing and on each station's starting chemistry. If there were no artifact, the policy effect on bulk mass would not depend on whether a heated microbalance or a beta gauge recorded it. That invariance gives a clean test.
Core Claim
The claim is that composition-shifting policies change the chemical mass closure gap. That gap is the part of weighed PM2.5 that the sum of measured species does not explain, plus the ambient mass that weighing never captures. When the gap changes, bulk-mass policy estimates are biased in a predictable direction. The bias is concealed twice: once by treating total mass as the outcome, and again by averaging stations into city means, which removes the station-level link between composition and instrument retention.
A one-line accounting makes this concrete. Write measured mass as
$$M_{st} = \sum_k r_{k}(T_{st}, RH_{st}, I_s)\, m_{k,st} + \epsilon_{st},$$
where $m_{k,st}$ is the true ambient mass of component $k$ (sulfate, nitrate, ammonium, organic matter, black carbon, dust, water) at station $s$ and time $t$. The term $r_k$ is the fraction of that component the instrument retains, which depends on temperature $T$, relative humidity $RH$, and instrument type $I_s$. The term $\epsilon$ is noise.
A policy that lowers high-retention sulfate while raising the share of low-retention nitrate and water changes $M$ by more than it changes $\sum_k m_k$. The resulting bias has a known sign given the direction of the composition shift. What is not known is how large it is, and that is the empirical question.
What Prior Work Establishes, and What It Leaves Open
The motivation comes from two literatures that have not been joined.
On the measurement side, several field studies document the retention problem directly:
- The PM2.5 Aerosol Chemical Speciation Monitor deployment in Nanjing, Field characterization of the PM 2.5 Aerosol Chemical Speciation Monitor, documents that thermal vaporization at 600 °C misses dust, sea salt and refractory chloride.
- The Beijing 2013 study, Characterization of submicron aerosols during a month of serious pollution in Beijing, 2013, notes the same exclusion for refractory black carbon.
- The Beijing–Tianjin–Hebei multi-site filter campaign, Chemical characterization and source identification of PM 2.5 at multiple sites in the Beijing–Tianjin–Hebei region, explicitly attributes its unclosed mass to evaporation of semi-volatiles and loss of bound water, and notes the loss is worse at seasonal temperature extremes.
- Tropical filter work, Fine particulate matter in the tropical environment, left nearly half the mass unidentified under high humidity.
None of these studies asks how the gap moves when a policy moves the chemistry.
On the policy side, Evaluating the "2+26" regional strategy is the closest existing work. It compares four Beijing pollution episodes using a coupled weather and chemical transport simulation (WRF-CAMx) with a source-tagging module (PSAT). It reports mean PM2.5 16.43% lower in March 2018 than March 2013, and 32.70% lower in November 2017 than November 2016. Over the same episodes, coal's share fell from 40% to 11%, nitrate became the dominant component, and vehicle exhaust reached 40–54% of mass.
That paper supplies the composition shift our argument depends on. It does not ask whether the measured mass it compares against is itself composition-dependent. It also views the region largely from Beijing receptors. Analysis of the air pollution reduction and climate change mitigation effects of the Three-Year Action Plan for Blue Skies models 28-prefecture emission cuts: 17% primary PM2.5, 25% SO2, 21% NOx, but only 3% NH3. That stoichiometry favors ammonium nitrate formation, which is exactly the low-retention direction. The model omits fugitive dust and is never reconciled against measured closure.
The econometric heterogeneity literature shows that uniform policies land differently depending on local context:
- Does an Emissions Trading Policy Improve Environmental Efficiency? finds larger gains in central and western provinces.
- Identifying the Impact of Climate Policy on Urban Carbon Emissions finds effects only in high-innovation cities.
- Carbon Emission Trading Policy and Carbon Emission Efficiency finds stronger effects in resource-dependent cities.
These studies model context as economic structure measured at the city or province level. We add physical context, namely station chemistry and instrument type, and test whether it explains heterogeneity that economic moderators cannot.
The aggregation lens is methodological. Predicting benefit from adjuvant therapy with corticosteroids in community-acquired pneumonia shows a pooled null hiding benefit in patients with high baseline C-reactive protein and harm in those with low baseline levels, which emerged only from patient-level data. Integrative prescreening in analysis of multiple cancer genomic studies shows that summary-statistic pooling loses the covariance structure needed to recover weak signals. We borrow the logic: baseline station composition plays the role that baseline C-reactive protein plays in the trial. A cautionary result also constrains the design. A critical assessment of matching-adjusted indirect comparisons shows that reweighting to aggregate targets breaks down under unmeasured moderators. This is why we do not try to repair city means after the fact.
Research Program
The work has three linked pieces.
Retention functions. First, we estimate how much of each component each instrument retains. The CARE-China network, Characteristics of PM 2.5 mass concentrations and chemical species … CARE-China network, mixes two mass monitors: tapered element oscillating microbalances (heated microbalances) and beta gauges. It pairs them with 48-hour weekly filter speciation at 11 sites. Wherever an instrument's mass coincides with filter speciation and meteorology, we regress the closure residual on species loadings interacted with temperature, relative humidity and instrument type. The fitted coefficients give empirical retention factors, $\hat r_k$.
Ambient water is not measured. We treat it as a latent term with a humidity-dependent prior tied to measured inorganic loading. We report policy conclusions across the range of that prior rather than committing to a single number. Organic mass enters with an organic-matter-to-organic-carbon ratio that is allowed to vary with season and oxidation proxies, rather than being fixed.
Station-level policy design. Second, we set up the policy comparison at the station level. Treated units are stations inside the 2+26 prefectures after the 2017 rollout. Controls are CARE-China and other network stations outside the cluster, matched on pre-period composition and meteorology. We estimate effects on five outcomes, all in one hierarchical model with station-specific effects that are partially pooled toward shared means:
- bulk measured mass;
- each identified species;
- the unexplained residual;
- a closure-corrected ambient mass, $\sum_k \hat m_k$, reconstructed with the estimated retention factors;
- an ACSM-visible (non-refractory) mass, separating what speciation monitors can and cannot see.
The model lets us read off both the pooled effect and the distribution of station effects. Moderators are baseline nitrate share, baseline dust share, instrument type, and distance to major point sources.
Transport bounding. Third, the 2+26 episode study reports local contributions ranging from 49.46% to 89.35%. That range means untreated nearby stations may be contaminated by transported pollution. We therefore run the analysis with a buffer that excludes near-cluster controls, and separately with the transported fraction as a moderator. The goal is to bound spillover rather than to assume it is absent.
The main risk is data access. The file documents CARE-China design and pre-policy BTH speciation, but I found no record on Lacuna of station-level speciated data published after 2017. Securing raw post-2017 records from the network operators is a prerequisite. If only city-month summaries can be obtained, the aggregation half of the proposal cannot be tested. That limitation would itself illustrate the aggregation problem, but it would still be a limitation.
Baselines
The naive reference. This is a standard difference-in-differences on bulk measured PM2.5 at the city-month level. It is the default in environmental economics. We run it on exactly the same stations and period, with the same matching, fixed effects and clustered errors. Every other estimate is compared to it. The file contains no reported difference-in-differences estimate of the 2+26 effect on PM2.5 (no reported result found on Lacuna for that quantity), so we produce this baseline ourselves.
The strongest existing evaluation. This is the chemical-transport-model episode comparison in Evaluating the "2+26" regional strategy, which reports mean PM2.5 16.43% lower in March 2018 versus March 2013 and 32.70% lower in November 2017 versus November 2016. It is the strongest competitor because it already resolves species and transport. For a fair comparison we restrict our estimates to the same Beijing receptors and the same calendar windows. We then check whether the closure-corrected ambient reduction falls inside or outside the range implied by those percentages.
Beating this baseline means two things. First, the closure-corrected ambient reduction differs from the bulk-mass reduction by more than its uncertainty interval, in the direction the retention model predicts. Second, that difference is reproduced by the instrument-type test described below, which a transport model cannot produce because it never sees the instruments.
The economic-heterogeneity baseline. This follows Identifying the Impact of Climate Policy on Urban Carbon Emissions and Carbon Emission Trading Policy and Carbon Emission Efficiency: a city-level difference-in-differences with interactions for innovation capacity and resource dependence. We give it the same outcome and sample. Station chemistry and instrument moderators beat it if they explain additional between-station variance in the effect, measured on held-out stations, after the economic moderators are already included.
The ablations. Two ablations isolate the key ingredients. The first replaces the estimated, condition-dependent retention factors with a single constant correction per instrument. This tests whether the dependence on composition and meteorology matters or whether a flat scaling suffices. The second fits the full model on city-month means instead of station records. This tests whether aggregation alone erases the effect heterogeneity.
Why not just…?
Why not just run difference-in-differences on total gravimetric mass? Total mass is what regulators target, so its change is a legitimate estimand. The problem is that the file's own evidence shows the composition moved sharply during the 2+26 period, while the BTH campaign shows that volatile and water loss differ by species and by season. A bulk-mass estimate therefore mixes the real change with a change in how much the instruments retain. Our instrument-invariance test settles whether this matters. If heated-microbalance and beta-gauge stations yield the same policy effect after conditioning on composition, the naive estimate is vindicated and we will report that result.
Why not just simulate species with a chemical transport model? Transport models predict ambient mass. That is useful, but they are calibrated against the same monitors whose retention is in question. The Blue Skies GAINS analysis also omitted fugitive dust for lack of local emission factors, and the dust fraction is exactly what speciation monitors cannot see. We use model output as a comparison, not as ground truth. Where model output and closure-corrected estimates diverge, that divergence is itself a result.
Why not just apply a fixed correction factor for volatile loss? This is the real threat. If one scaling per instrument explains most of the gap, the proposal reduces to a simple adjustment. The first ablation exists for this reason. If the constant-correction model matches the full model's policy estimate within uncertainty on held-out stations, we will say so and recommend the simple fix.
Why not just add economic moderators at the city level? The city-level heterogeneity studies do find context effects. However, their moderators are economic structure, and they cannot see a station whose nitrate share tripled. The held-out variance comparison in the Baselines section measures whether physical moderators add information beyond the economic ones.
Proposed Experiments and Decisive Checks
Instrument invariance. This is the decisive check. Within the treated cluster, we compare policy effects on bulk mass between heated-microbalance and beta-gauge stations, holding baseline composition fixed. The prediction is that heated stations show larger apparent reductions, and that the gap between the two grows with each station's post-policy nitrate share. A null here, with adequate power, falsifies the core claim.
Sign heterogeneity. We check whether any stations show a positive effect on closure-corrected mass, particularly peri-urban or industrial sites such as those near Tangshan, while the cluster mean is negative. We also check whether city-mean aggregation hides these stations. This is the environmental counterpart of the corticosteroid result.
Residual as an outcome. We test whether the unexplained residual responds to treatment. A significant treatment effect on the residual is direct evidence that the policy moved the closure gap.
Refractory split. We compare policy effects on non-refractory mass (the portion speciation monitors can measure) against filter-based dust and elemental carbon. This checks whether monitor-based evaluations miss the dust and soot reductions that land-use and coal controls should produce.
Sensitivity across the water prior. We report every headline estimate across the full range of the water prior. We will claim a measurement artifact only if its sign survives that entire range.
References
- Field characterization of the PM 2.5 Aerosol Chemical Speciation Monitor: insights into the composition, sources, and processes of fine particles in eastern China
- Characterization of submicron aerosols during a month of serious pollution in Beijing, 2013
- Chemical characterization and source identification of PM 2.5 at multiple sites in the Beijing–Tianjin–Hebei region, China
- Fine particulate matter in the tropical environment: monsoonal effects, source apportionment, and health risk assessment
- Evaluating the “2+26” regional strategy for air quality improvement during two air pollution alerts in Beijing: variations in PM 2.5 concentrations, source apportionment, and the relative contribution of local emission and regional transport
- Analysis of the air pollution reduction and climate change mitigation effects of the Three-Year Action Plan for Blue Skies on the “2+26” Cities in China
- Does an Emissions Trading Policy Improve Environmental Efficiency? Evidence from China
- Identifying the Impact of Climate Policy on Urban Carbon Emissions: New Insights from China’s Environmental Protection Tax Reform
- Carbon Emission Trading Policy and Carbon Emission Efficiency: An Empirical Analysis of China’s Prefecture-Level Cities
- Predicting benefit from adjuvant therapy with corticosteroids in community-acquired pneumonia: a data-driven analysis of randomised trials
- Integrative prescreening in analysis of multiple cancer genomic studies
- A critical assessment of matching-adjusted indirect comparisons in relation to target populations
- Characteristics of PM 2.5 mass concentrations and chemical species in urban and background areas of China: emerging results from the CARE-China network