Clarifying causality and information flows between time series: Particulate air pollution, temperature, and elderly mortality Authors: Louis Anthony Cox Published: 2024 ABSTRACT AbstractExposure-response associations between fine particulate matter (PM2.5) and mortality have been extensively studied but potential confounding by daily minimum and maximum temperatures in the weeks preceding death has not been carefully investigated. This paper seeks to close that gap by using lagged partial dependence plots (PDPs), sorted by importance, to quantify how mortality risk depends on lagged values of PM2.5, daily minimum and maximum temperatures and other variables in a dataset from the Los Angeles air basin (SCAQMD). We find that daily minimum and maximum temperatures and daily mortality counts two to three weeks ago are important independent predictors of both current daily elderly mortality and current PM2.5 levels. Thus, it is important to control for these variables over a period of at least several weeks preceding death. Such detailed control for lagged confounders has not been performed in influential past papers on PM2.5-mortality associations, but appears to be essential for isolating the potential causal contributions of specific variables to mortality risk, and, therefore, a worthwhile area for future research and risk assessment modeling.Keywords: PM2.5, Mortality, Minimum daily temperature, Confounding, Time series, Partial dependence plots Introduction Numerous studies in recent decades have reported that fine particulate matter (PM2.5) air pollution is statistically significantly associated with mortality rates. These correlations have long been presented to policymaker and the media as reasons to expect that ever tighter regulation of allowed PM2.5 levels will cause ever greater reductions in human mortality rates. They have been used to support successful arguments for reductions in PM2.5 National Ambient Air Quality Standards in the US (Weinhold, 2012; [9]) and for increasingly stringent regulations of PM2.5 internationally [4,12]. Disappointingly, however, dozens of accountability studies have not found that even large reductions in PM2.5 have caused any clear reduction in all-cause mortality rates [2,10,12]. Peer-reviewed published claims that reducing particulate air pollution caused reductions in mortality rates (e.g., [4]) have subsequently turned out to be incorrect, e.g., due to use of unverified modeling assumptions that prove to be wrong in hindsight (e.g., [7,10,23]). The studies that control most carefully for confounders have been found to be least likely to report clear evidence that reducing PM2.5 has reduced mortality rates [2,10]. These findings are consistent with some earlier studies that questioned whether statistical associations between particuate air pollution and mortality risk are causal, e.g., because even large reductions in total suspended particulates do not seem to be clearly and strongly associated with reductions in adult or elderly mortality [3,13]. How can PM2.5 be strongly correlated with mortality if changes in PM2.5 have no detectable effect on changing mortality rates? One way is via confounding by unmeasured or incompletely controlled variables. The vast literature associating PM2.5 with mortality risk has – without exception, to our knowledge – not fully controlled for lagged daily minimum (and maximum) temperatures over the weeks preceding death. This leaves open the question of to what extent, if any, PM2.5-mortality associations have non-causal explanations. Alternative non-causal hypotheses have not been carefully examined or refuted, such as that cold winter days increase elderly mortality over the next one to two weeks, and also predict increased levels of PM2.5 in the same time frame. The needed daily temperature data have sometimes been collected, but investigators have typically averaged it, e.g., using seasonal averages [8], leaving residual confounding intact [5]. Studies that report a causal impact of PM2.5 on mortality rates typically rely on untested (and often untestable) assumptions (ibid). Many of these studies assume that there are no hidden confounders, but without testing for them, and then treat associational exposure-response relations as if they were interventional causal relationships, even while acknowledging that “the true relationship between mortality and PM2.5 has not been established” [17]. Table 1 lists some recent studies that have drawn causal conclusions about the adverse health effects of PM2.5 (left column) based on unverified assumptions (right column). These papers do not fully control for confounding by temperature, as noted in comments in the right column of Table 1. An anonymous reviewer of an earlier version of this paper noted that all of these studies have substantial overlap in authorship and funding, as is typical of much of the literature claiming that PM2.5 causes increased mortality risk; and that “many authors consider omitted covariates a serious problem,” with lagged temperature measurements (and perhaps other meteorological variables) being prominent examples of frequently omitted covariates [16]. Table 1Studies on PM2.5 and Adverse Health Effects Using Unverified Assumptions (Emphases added).Table 1StudyCausal ConclusionsUnverified assumptions[6]“Among the elderly population in the United States, air pollution generally increased the rate of hospital admissions with CVD, MI, stroke, and respiratory disease. PM2.5 and ozone, in particular, were found to increase the risk of admissions across several health comes.”“[O]ur causal approach relies on the parallel trends assumption, which we cannot verify.”[8]“The investigators reported increased risks of all-cause mortality of 6 % to 8 % per 10-μg/m3increase in PM2.5 across the five approaches, with larger effect estimates in a low exposure subcohort. The consistency of the associations across methods provides stronger support than past studies for what is likely a causal effect between long-term exposure to PM2.5 and mortality.”“We believe the threat of unmeasured confounding to our analysis results was minimized, given our use of state-of-the-art statistical methods to control and assess the potential confounding. Yet, because this assumption was not directly verifiable based on the observed data, we cannot rule out the possibility that unmeasured confounding exists.” [Comment: Daily temperatures were collected, but were converted to seasonal averages, allowing residual confounding.]Schwartz et al. [15]“We estimated that 23.5 % of the Medicare population would die before 76 y of age if exposed to PM2.5 at 12 μg/m3compared with 20.1 % if exposed to an annual average of 7.5 μg/m3.”“We also assumed there were no important omitted confounders. This assumption is common to all epidemiology studies whether they use causal modeling techniques or not.” “Finally, causality is a conclusion of humans, not the output of a statistical model. Our model produced causal estimates if certain conditions are met, and it is not possible to verify that those conditions were truly met.” [Comment: The study omitted temperature, an important confounder.][14]“We estimated that reducing PM2.5 and O3 concentrations to levels below current standards would increase life expectancy by substantial amounts”“Nevertheless, inference is still dependent on the assumption that we have captured all confounders. This is why causal inference still requires judgment….” [Comment: The study adjusted for year but omitted the known confounders daily and seasonal temperatures.][18]“Under the assumption that no variable changing differentially over time across census tracts other than seasonal temperatures could confound the association, we found causal associations between PM2.5 and all natural- cause mortality.”The key assumption of no confounding by changes across census tracts and time in variables other than temperature is an “untestable hypothesis, which must be judged on external information.” [Comment: The study adjusted for seasonal average temperatures but not for daily temperatures.][19]“This study indicates long-term residential exposure to PM2.5 could increase risk of MI [myocardial infarction] among the general US population, people with comorbidities, and poorer individuals. The association persists below current [ambient air] standards.”“We assumed no other covariates that were both correlated with exposure and outcome and changed differentially over time across zip codes were left out.” [Comment: Daily, weekly, and monthly temperatures, which are known plausible confounders, were left out.][20]“Assuming the causal framework was valid, we found that higher levels of PM2.5, O3, and NO2 were causally associated with greater risk of mortality and that PM2.5 posed the greatest risk. … lowering each air pollutant concentration from the 70th to 60th percentiles would prevent 65,935 early deaths per year.”“This cannot be verified directly.” “The causal conclusions of this study depend on the key assumption of correct IPW model specification. The validity of this assumption is not testable and relies on outside information.” [Comment: Daily temperatures were collected, but were converted to seasonal averages, allowing residual confounding.][21]“Assuming no omitted predictors changing differently across ZIP codes over time in correlation with PM2.5, we found a causal effect of PM2.5 on mortality incidence rate.”“For a causal interpretation of the DID estimate, [we assume that] all the ZIP codes have parallel trends in mortality rates. … Because this assumption cannot be tested statistically, it must be judged based on knowledge and plausibility from outside the study.” [Comment: Daily, weekly, and monthly temperatures were left out.] Studies on PM2.5 and Adverse Health Effects Using Unverified Assumptions (Emphases added). The following sections seek to close this gap by examining in more detail the temporal relationships between PM2.5 levels, daily minimum temperatures and other meteorological variables, and subsequent daily elderly mortality rates. In doing so, they explore how modern statistical tools can be used to clarify causality and information flows between time series, potential confounding, mediation, and moderation effects in causal analysis of multiple time series including minimum daily temperatures, ambient PM2.5 concentrations, and elderly mortality rates. Data Time series data for California's South Coast Air Quality Management District (SCAQMD), which includes Los Angeles, for 2007–2010 have been made available by Lopiano et al. [11] and Young et al. [22]. The data set is included in the online supplementary material. It has 1461 rows (days) of data, from January 1, 2007, to December 31, 2010. Each row gives the year, month, and day on which data were collected; the number of daily deaths (excluding accidental deaths) among individuals aged 75 or older, denoted by AllCause75, as reported by the California Department of Health; the daily average ambient concentration of fine particulate matter (PM2.5), measured in micrograms per cubic meter, as reported by the California Air Resources Board (CARB); and meteorological variables: tmin (minimum daily temperature), tmax (maximum daily temperature), and MAXRH (maximum relative humidity), obtained from the Oak Ridge National Laboratory and the U.S. Environmental Protection Agency (EPA). To help analyze long-term trends, we also add a derived “time” variable, defined as the number of days since January 1, 2007. Table 2 shows the layout of the data for the first few rows (days). Fig. 1 plots the time series for PM2.5, AllCause75, tmin, and tmax. The data and sources are discussed further by Young et al. [22]. Table 2Layout of data.Table 2timeyearmonthdayAllCause75PM2.5tmintmaxMAXRH120071115138.4367268.8220071215817.4367548.9320071313919.9447561.3420071416464.6376887.952007151366.1406147.5620071615218.8396939 Layout of data. Time series for daily PM2.5; All Cause Mortality (excluding accidents) counts among people 75 or older (AllCause75); and minimum and maximum temperatures (tmin and tmax) in the Los Angeles air basin (SCAQMD). Methods The key research questions that we address are as follows: •Qualitative dependence:What variables in this data set does mortality (AllCause75) directly depend on? The other variables considered are the same-day values of tmin, tmax, MAXRH, and PM2.5; their lagged values at various times over the past 30 days; and the year, month, and time. Qualitative dependence:What variables in this data set does mortality (AllCause75) directly depend on? The other variables considered are the same-day values of tmin, tmax, MAXRH, and PM2.5; their lagged values at various times over the past 30 days; and the year, month, and time. •Quantitative dependence:How does mortality depend on each of the variables that significantly helps to predict it? Specifically, what time lags for the other variables are relevant for predicting current mortality counts, and what is the shape of the dependence of AllCause75 on each variable that significantly helps to predict it? Quantitative dependence:How does mortality depend on each of the variables that significantly helps to predict it? Specifically, what time lags for the other variables are relevant for predicting current mortality counts, and what is the shape of the dependence of AllCause75 on each variable that significantly helps to predict it? •Confounding: How does PM2.5 depend on other variables on which AllCause75 also depends, and over what time windows? To what extent do other variables explain a PM2.5-mortality association? Confounding: How does PM2.5 depend on other variables on which AllCause75 also depends, and over what time windows? To what extent do other variables explain a PM2.5-mortality association? To answer these questions, we use partial dependence plots (PDPs) [5] and Random Forest (RF, a popular non-parametric machine learning predictor) to investigate how the expected death count per day (AllCause75) varies as one independent variable at a time is varied, holding all other variables fixed at their current levels. For example, the PDP for AllCause75 vs. tmin_lag14 shows how different values of tmin two weeks ago are predicted to affect AllCause75 today if all other variables (including other lagged values of tmin) are held constant at their observed levels, with only tmin_lag14 being varied. In the terminology of mediation analysis, the resulting PDP is an estimate of the natural direct effect of the independent variable that is varied (e.g., tmin_lag14) on current mortality, AllCause75, if all relevant variables are measured and all except tmin_lag14 and AllCause75 are held fixed at their observed levels. Methods for calculating and interpreting PDPs are now well-developed in machine learning and are increasingly being applied in epidemiology and health risk assessment (ibid). A useful feature of Random Forest is that it ranks independent variables in order of decreasing importance for predicting the dependent variable. “Importance” of a variable can be defined in several different ways. A standard approach in Random Forest, adopted here, defines it as the estimated percentage increase in mean squared prediction error (MSE) if a variable is excluded as a predictor. Cox [5] provides additional details and discussion of PDPs applied to exposure-response modeling. To get a sense of dependencies among variables on time scales of days and weeks, we consider PDPs between dependent variables (AllCause75 and PM2.5) and independent variables with time lags of 0, 1, 2, 4, 7, 14, 21, and 28 days. Results Inspection of Fig. 1 suggests that daily death counts tend to be relatively low when daily temperatures (both tmin and tmax) are relatively high. Associations between PM2.5 and mortality are harder to discern visually, but Table 3 shows that multiple linear regression of AllCause75 against all other variables evaluated on the same day (i.e., without lags) gives a significant positive regression coefficient (p = 0.0076) for the association between same-day levels of PM2.5 and AllCause75. (The same is true when stepwise variable selection is used: all of the independent variables remain in the model as significant predictors of AllCause75.) The next step is to investigate how different values of potential causes on one day predict different natural direct effects on mortality over the days that follow. Table 3Multiple linear regression model for predicting AllCause75 from other same-day variables.Table 3VariableEstimated slope coefficientStd. Errort valuep-value)(Intercept)1,653,000417,0003.9647.75e-05 ***time2.2540.56873.9647.73e-05 ***year−823.3207.7−3.9637.76e-05 ***month−69.9717.32−4.045.64e-05 ***day−2.2630.5692−3.9767.36e-05 ***PM2.50.092550.034632.6730.007609 **tmin−0.46970.08514−5.5174.10e-08 ***tmax−0.22820.06075−3.7560.000179 ***MAXRH−0.12890.03276−3.9358.71e-05 ***Significance Codes:***: p-value ≤0.001, **: p-value ≤0.01, *: p-value ≤0.05.: p-value ≤0.1. Multiple linear regression model for predicting AllCause75 from other same-day variables. Significance Codes:***: p-value ≤0.001, **: p-value ≤0.01, *: p-value ≤0.05.: p-value ≤0.1. Fig. 2 shows a family of PDP curves for (a) AllCause75 vs. same-day (unlagged) values of tmin (red curve, relatively flat but slightly increasing for tmin values above about 65 degrees, with predicted conditional expected AllCause75 values between about 134.1 and 134.3 deaths per day); and (b) AllCause75 vs. tmin at various lags, from one day (orange curve) to 1 or 2 weeks (the relatively steep greenish curves) to a month (the relatively flat dark blue curve). The natural direct effect of tmin on AllCause75 is greatest for lags of about 1–2 weeks. This indicates that a low minimum temperature today has a relatively large natural direct effect on mortality risk on days one to two weeks in the future. Partial dependence plots (PDPs) for mortality vs. tmin lagged by 0–4 weeks (0–28 days). Fig. 3 explores further the temporal and meteorological predictors of daily elderly mortality. It shows the PDPs for the 20 most important of these predictors of AllCause75. They are sorted in order of decreasing importance (defined as % increment in MSE if a variable is not used to help predict AllCause75) from the upper left to the lower right. The most important predictor of mortality is month. Partial dependence plots (PDPs) for mortality (AllCause75) vs. its 20 most important temporal and meteorological predictors. Source: Appendix A gives R code for generating Fig. 2. Source: Appendix B gives R code for generating Fig. 3. Source: Appendix C gives R code for generating Fig. 4. Partial dependence plots (PDPs) for PM2.5 vs. its 20 most important predictors. January, February, and March are relatively high-risk months for elderly mortality, other variables being held constant. Time is also an important predictor (4th most important). This indicates that variables not in the data set are important in determining elderly mortality risk, since the PDP for time would be flat if time acted only through the other observed variables, such as same-day and lagged values of tmin, tmax, MAXRH, and PM2.5, because they are held fixed in calculating the PDP for time. The second most important predictor of mortality is recent mortality (specifically, AllCause75_lag 2, with AllCause75 at lags 1, 4, and 14 also appearing in the array with less importance). This suggests that factors not included in the data set that affect mortality may persist for at least two weeks. The third most important variable is tmin_lag4. The minimum daily temperatures at lags 14 and 21 days and maximum daily temperatures at lags 1 and 14 days are also identified as important predictors of AllCause75. The remaining of the top 20 most important variables are all various lagged values of tmin, tmax, and MAXRH, but not PM2.5. Higher values of tmin and tmax at various lags over the interval from lag 1 day to lag 28 days generally predict lower values of elderly mortality risk. Although the estimated natural direct effects of each individual day's temperatures on future mortality are small, these effects are found for lags extending back several weeks, so the total effects of daily temperature extremes accumulated over several weeks might plausibly be at least an order of magnitude larger than the effects from a single day, shown in Fig. 3. Fig. 4 presents analogous information for PM2.5 as the dependent variable. The most important predictor of ambient PM2.5 concentration today is ambient PM2.5 concentration yesterday. The second and third most important predictors are same-day maximum relative humidity, MAXRH, and maximum temperature, tmax. Lagged values of these variables with lags of up to 4 days for tmax and 14 days for tmin and MAXRH help to predict current PM2.5, with higher tmin values at lags 1–14 days typically being associated with lower PM2.5 values. Unexpectedly, AllCause75 also helps to predict future PM2.5 values 2, 3, and 4 weeks later (right column of Fig. 4), as well as future values of AllCause75 up to 2 weeks later (upward-sloping PDPs in Fig. 3). This suggests that noncausal sources of PM2.5-mortality associations (e.g., unobserved confounders with effects distributed over time) are important in this data set. Time and month also help to predict PM2.5, with ambient concentrations of PM2.5 increasing in November and December. If lagged values of other variables explain the significant positive association (regression coefficient) between same-day PM2.5 and AllCause75 in simple regression models such as the one in Table 3, then it might be expected that the regression coefficient for PM2.5 as a predictor of AllCause75 would no longer be significant after conditioning on past values of other variables. Table 4 shows that this is indeed the case: conditioning on variables over the past week (for lags 1, 4, and 7) increases the p-value for the PM2.5 coefficient from its previous value of 0.0076 in Table 1 (no lagged variables) to 0.674 in Table 2. Of course, including lagged values of variables does not by itself address other limitations of regression modeling. Performing regression diagnostics on the model in Table 2 (using the gvlma linear model assumption validation package in R) flags the assumption of a linear link function as being not satisfied by the data (p = 4.167e-07), as expected from the conspicuous nonlinear patterns in Fig. 3. Table 4Multiple linear regression model for predicting AllCause75 from same-day and lagged variables.Table 4VariableEstimated slope coefficientStd. Errort valuep-value(Intercept)777,900374,3002.0780.037847 *time1.0620.51042.080.037741 *year−387.5186.5−2.0780.037879 *month−32.7615.55−2.1070.035323 *day−1.0720.5109−2.0970.036149 *PM2.50.016650.039630.420.674524tmin0.31050.1062.9290.003457 **tmax0.17090.075242.2720.023248 *MAXRH0.048870.036231.3490.177565PM2.5_lag10.039990.040270.9930.320846PM2.5_lag4−0.054080.03355−1.6120.107214PM2.5_lag70.026240.032690.8030.422223tmin_lag1−0.11550.1028−1.1230.261598tmin_lag4−0.20410.09161−2.2280.026011 *tmin_lag7−0.22270.08427−2.6430.008307 **tmax_lag1−0.28790.08293−3.4710.000534 ***tmax_lag4−0.091440.06178−1.480.139043tmax_lag7−0.074980.06006−1.2490.212039MAXRH_lag1−0.170.03714−4.5775.14e-06 ***MAXRH_lag4−0.031740.03018−1.0520.293095MAXRH_lag7−0.042390.02986−1.420.155891AllCause75_lag10.15730.026126.0232.19e-09 ***AllCause75_lag40.12080.026074.6333.94e-06 ***AllCause75_lag70.1490.025625.8157.47e-09 ***Significance Codes:***: p-value ≤0.001, **: p-value ≤0.01, *: p-value ≤0.05.: p-value ≤0.1. Multiple linear regression model for predicting AllCause75 from same-day and lagged variables. Significance Codes:***: p-value ≤0.001, **: p-value ≤0.01, *: p-value ≤0.05.: p-value ≤0.1. Discussion The patterns of information flow between time series variables in Fig. 1, Fig. 2, Fig. 3, Fig. 4 have potentially striking implications for PM2.5-mortality exposure-response modeling and epidemiology. Although PM2.5 does not appear as a predictor of same-day or future elderly mortality risk (AllCause75) values in Fig. 3, AllCause75 values at various lags are found to be important predictors of future PM2.5 values (Fig. 4), as well as of future mortality (upward-sloping curves for AllCause75 vs. lagged AllCause75 values in Fig. 3). This highlights an important conceptual ambiguity in the static exposure-response curves that dominate the literature on PM2.5 and mortality. Exposure-response associations may reflect unmeasured confounders and distributed lags that mimic reverse causality, with present and future (but not past) values of PM2.5 being predicted by hidden variables that also affect present mortality risk. A static association between same-day PM2.5 and mortality counts does not reveal (a) the direction of causation; (b) the presence of hidden confounders; or (c) the role of recent history of meteorological variables in explaining current same-day PM2.5-mortality associations. Proposed causal interpretations of such associations based on expert judgment, unverified assumptions, and weight-of-evidence considerations as in Table 1, typically do not consider the distributed lag structure of the associations. Instead, they typically assume that there are no hidden confounders without actively checking for the appearance of reverse causation such as that revealed by Fig. 3, Fig. 4. The method illustrated in Fig. 3, Fig. 4, of systematically evaluating PDPs for current and lagged values of variables, complements static exposure-response association analysis. It builds on the ideas used in Granger causality and transfer entropy studies of directed information flows among time series [1], but applies the well-developed non-parametric framework of PDPs [5] to achieve readily visualizable descriptions of how the conditional expected value of a dependent variable depends on current and lagged values of independent variables. These PDPs can be interpreted causally as the estimated natural direct effects of independent variables on the dependent variable if all relevant variables are measured. Otherwise, if unmeasured variables confound the observed PDPs, the curves can still be interpreted statistically (but not causally) as predictions of how the conditional expected value of a dependent variable depends on observed current and lagged values of independent variables. A major implication of the lagged PDP analysis, illustrated by the results in Fig. 2, Fig. 4, is that the dynamic processes that determine daily temperatures, PM2.5 exposure concentrations, and mortality counts have memories on the order of at least days to weeks. Exposure-response modeling for PM2.5 and mortality should, therefore, consider the effects of lagged daily temperature extremes (tmin and tmax) over an interval of not just one or a few days, but over an interval of at least 2–4 weeks in order to adequately control for the effects of potential dynamic confounding. For example, tmin_lag14, the daily minimum temperature two weeks ago, is a predictor of both PM2.5 and AllCause75, other variables being held constant. A possible interpretation is that colder days increase the probability of contracting an illness that results in death a week or two later, and also contribute to atmospheric chemistry conditions (perhaps in part from increased consumption of fuels for heating) that eventuate in increased levels of PM2.5 a fortnight later. The voluminous literature on quantitative PM2.5-mortality associations has generally ignored such lagged confounders. Temperature is seldom lagged by more than a day or two if it is considered at all. It is common practice to simply ignore temperature over the one to two weeks preceding death, even when the data have been collected, as illustrated by the papers in Table 1. Fig. 2, Fig. 3, Fig. 4 suggest that seasonal averages of daily temperatures are relatively useless as predictors of mortality. Instead, lags on the order of 1–3 weeks are likely to be the most important for understanding and controlling for temperature (tmin and tmax) as a confounder of same-day PM2.5-mortality associations. To our knowledge, none of the influential studies and models used to inform regulatory risk assessments and decision-making for PM2.5 and mortality has controlled for confounding by tmin and tmax (or MAXRH) over this relevant interval. A third implication of Fig. 2, Fig. 3 is that nonlinearity matters. For example, elderly mortality is increased for tmax_lag1 values both above about 90 degrees and below about 65 degrees compared to values in between those extremes (Fig. 3). The PDP for AllCause75 vs. tmin_lag4 has plateaus below about 40 degrees and above 57 degrees, with a decline in risk in between. The dependence of PM2.5 on tmin_lag4 and tmin_lag14 is also conspicuously nonlinear. This contrasts with the widespread use of generalized linear models and Cox Proportional Hazards models in much of the literature on PM2.5-mortality exposure-response associations (Table 1). In these models, each independent variable has only one coefficient to represent its main effect on mortality risk. Nonlinear effects for most of the independent variables, such as thresholds and plateaus, are not modeled. Regression diagnostics (e.g., careful analyses of residuals) for these models are seldom presented in the literature that adopts them. The non-parametric PDPs in Fig. 2, Fig. 3, Fig. 4 show that many dependencies between AllCause75 and independent variables cannot be well approximated by such models. Flexible non-parametric methods are needed to capture these nonlinear dependence relationships. A frequently expressed concern in the literature on PM2.5-mortality exposure-response modeling is that unmeasured variables may confound observed exposure-response associations. Many recent papers argue that the assumption that no important variables have been omitted requires untestable assumptions that are best assessed by expert human judgment (Table 1). The lagged PDP methodology illustrated in Fig. 3, Fig. 4 provides a more objective, data-driven, approach to detecting important unmeasured variables. The key idea is to include time itself as an independent variable and to see whether the response variable is conditionally independent of time (flat PDP), given the values of the observed variables. This should be the case if the response variable depends only on the observed variables. In Figs. 3 (upper right corner) and 4 (row 2,column 3), however, time is found to be an important independent predictor of both AllCause75 and PM2.5 even after conditioning on the levels of the observed variables. This indicates that unobserved variables play a role in predicting these variables. The findings that daily mortality risk depends on lagged values of meteorological variables in the days (and weeks) preceding death (Fig. 3), but not on PM2.5 once these lagged variables are conditioned on (Fig. 3 and Table 4), are consistent with some previous research. An anonymous reviewer pointed out a useful paper by Smith [16] that investigated “the sensitivity of short-term associations between mortality in the Medicare population and fine particulate matter (PM) to various statistical modeling assumptions” using mortality data from Medicare, particulate matter data from EPA, and temperature and dewpoint data from NOAA. The main result was that “Significant associations were found when a linear concentration response function was fitted to the full range of PM [analogous to our Tabke 3], or in a broken-stick model above 12 μg/m3. No significant association was found below 12 μg/m3. However when lagged meteorology was omitted from the model, the estimated coefficients greatly increased and were significant at all levels of PM. Conclusions: It is important to take lagged meteorology into account in investigating short-term associations between PM and mortality.” This reinforces the importance of lagged values in interpreting concurrent PM2.5-mortality associations. Despite the potential advantages of the lagged PDP methodology in clarifying (a) the direction of information flows between time series over time (e.g., from AllCause75, which presumably serves as a marker for other conditions, to present and future PM2.5 values; but not from past PM2.5 values to present AllCause75 values); (b) the length of relevant time intervals over which to control for potential lagged confounders (e.g., weeks instead of days for tmax and tmin); (c) the shapes of nonlinear dependencies; and (d) the presence of important omitted variables, this approach has several limitations. One is that the PDPs we have used address only one independent variable at a time, holding all other variables fixed. It might be useful to allow pairs or larger sets of variables to vary at the same time to clarify their joint effects and possible interactions in affecting AllCause75. For example, a cold day followed by a second cold day in the middle of a warm series of days might have a different effect on mortality risk in the weeks ahead than would two cold days separated by several warm days. The methodology of 2-dimensional PDPs can potentially be used to investigate pairwise interactions and joint effects between two independent variables [5]. A second limitation, intrinsic to PDPs, is that they only quantify average effects of an independent variable on a dependent variable, averaged over the joint distribution of the remaining variables. By contrast, more informative Individual Conditional Expectation (ICE) plots “de-average” the PDPs, allowing heterogeneity in exposure-response relationships across days in the data set to be characterized; and accumulated local effect (ALE) plots consider only realistic combinations of values for multiple variables (ibid). Applying these methods to air pollution health effects data from multiple populations and locations may add important details to the research reported here and help to show the extent to which its conclusions for the SCAQMD area apply elsewhere. Conclusions Although the exposure-response association between PM2.5 and mortality is one of the most extensively studied associations in public health and air pollution epidemiology, the extent to which it depends on lagged values of daily minimum and maximum temperatures in the weeks preceding death has not, to our knowledge, been carefully evaluated in previous influential papers (Table 1). This paper has introduced a way to use lagged partial dependence plots (PDPs), sorted by a measure of importance, to visualize the dependence of mortality risk (or other dependent variables, such as current ambient PM2.5) on lagged values of PM2.5 and other independent variables. In these plots, one independent variable at a time is varied, holding all other observed variables constant, and the effects on predicted values of the dependent variable are quantified (Fig. 3, Fig. 4). Applying this lagged PDP analysis to a dataset from the Los Angeles air basin (SCAQMD) uncovers several dynamic relationships that are not obvious from studies of static PM2.5-mortality exposure-response associations alone. One is that AllCause75 is identified as an important predictor of present and future values of PM2.5, but not conversely: PM2.5 is not identified as an important predictor of current or future values of AllCause75, given the values of other variables, such as lagged tmin, tmax, and MAXRH values, that help to predict both. Since mortality presumably does not cause PM2.5, this pattern suggests that there are one or more unmeasured variables for which AllCause75 is a marker that help to predict future values of PM2.5. A second major finding is that daily minimum and maximum temperatures and daily mortality counts two to three weeks ago are important independent predictors of both current daily mortality (Fig. 3) and current PM2.5 levels (Fig. 4), holding all other variables constant. Thus, in contrast to the large literature on PM2.5-mortality associations, it appears to be important to control for these variables over a period of at least several weeks preceding death. Including seasonal average temperatures [8] or same-day and previous-day temperatures in models for predicting mortality risk is inadequate to control for confounding by temperature. This implies that research on PM2.5-mortality associations can potentially be improved by better controlling for lagged values of variables. Meanwhile, associations that have been interpreted causally without such detailed control for lagged confounding (e.g., Table 1) should probably be viewed as provisional at best. The absence of any apparent effect on all-cause mortality of large-scale, very substantial reductions in PM2.5 in recent decades (e.g., [12]) is consistent with the possibility that PM2.5-mortality associations might be explained largely (or perhaps even wholly) by uncontrolled confounding by lagged temperatures and other variables. This possibility appears to be worth investigating further in other datasets. Finally, the method of lagged PDPs illustrated in Fig. 2, Fig. 3, Fig. 4 appears to have many potential applications in epidemiology when time series data are available for exposure variables, response variables, and covariates. This method builds on ideas used in Granger causality testing for time series, but its use of non-parametric PDPs sorted by importance (as estimated by Random Forest) appears to be novel. We hope that this method may help to clarify information flows and dependencies among time series variables in many other public health applications. CRediT authorship contribution statement Louis Anthony Cox: Writing – review & editing, Visualization, Software, Formal analysis, Conceptualization. Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Footnotes Supplementary data to this article can be found online at https://doi.org/10.1016/j.gloepi.2024.100176. R Code for Fig. 2 # Load necessary libraries. library(randomForest). library(pdp). library(dplyr). library(ggplot2). # Load the original dataset (ladata.csv). data <− read.csv(“ladata.csv”). # Function to create lagged variables. create_lagged_data <− function(data, vars, max_lag) {. for (var in vars) {. for (lag in 0:max_lag) {. data[[paste0(var, “_lag”, lag)]] < − dplyr::lag(data[[var]], lag). } } return(data). } # Variables to lag. vars_to_lag <− c(“tmin”, “tmax”, “MAXRH”, “PM2.5”, “AllCause75”). # Create lagged data with lags from 0 to 30 days. lagged_data <− create_lagged_data(data, vars_to_lag, 30). # Remove rows with NA values (due to lagging). lagged_data_clean <− na.omit(lagged_data). # Set seed for reproducibility. set.seed(1). # Fit the Random Forest model to predict AllCause75. rf_model <− randomForest(AllCause75 ∼ ., data = lagged_data_clean, ntree = 100). # Specify the variables and lags for PDPs: 0 (same day), and lags 1, 7, 14, 21, 28. lags_to_plot <− c(0, 1, 7, 14, 21, 28). vars_to_plot <− c(“tmin”, paste0(“tmin_lag”, lags_to_plot)) # Include same-day tmin and lagged versions. # Create a data frame to store the PDP values for all lags. pdp_data <− data.frame(). # Generate PDPs and store them in the data frame. for (var in vars_to_plot) {. pdp_values <− pdp::partial(rf_model, pred.var. = var., plot = FALSE, grid.resolution = 100) # Increase grid resolution. pdp_values$lag <− var. # Add a column for lag identifier. pdp_values$tmin <− pdp_values[[var]] # Make sure x-axis uses the same variable name for all. pdp_values <− pdp_values % > % select(tmin, yhat, lag) # Keep relevant columns. pdp_data <− rbind(pdp_data, pdp_values). } # Export the PDP dataset to a CSV file. write.csv(pdp_data, “pdp_data_export.csv”, row.names = FALSE). # Create a color palette using ROYGBIV scheme (red for short lags, blue for long lags). colors <− rainbow(length(lags_to_plot) + 1, start = 0, end = 0.7) # Add one more color to match the number of groups. # Sort lags in order for the plot and legend. pdp_data$lag <− factor(pdp_data$lag, levels = c(“tmin”, paste0(“tmin_lag”, lags_to_plot))). # Plot all PDPs on a single graph with sorted legend, larger fonts, and heavier curves. ggplot(pdp_data, aes(x = tmin, y = yhat, color = lag, group = lag)) +. geom_line(linewidth = 1.2) + # Heavier lines. scale_color_manual(values = colors) + # ROYGBIV color scheme. labs(title = “Superimposed PDPs for tmin and its Lags (Random Forest)”, x = “tmin”, y = “Predicted AllCause75”) +. theme_minimal() +. theme(. text = element_text(size = 16), # Larger font size for axes. axis.title = element_text(size = 18), legend.title = element_text(size = 18), legend.text = element_text(size = 14), axis.line = element_line(linewidth = 1), # Display axes more clearly. axis.ticks = element_line(linewidth = 1), panel.grid.major = element_line(linewidth = 0.5), panel.grid.minor = element_line(linewidth = 0.25). ) R Code for Fig. 3 # Load necessary libraries. library(randomForest). library(pdp). library(dplyr). library(ggplot2). library(gridExtra). # Load the original dataset (ladata.csv). data <− read.csv(“ladata.csv”). # Function to create specific lagged variables. create_selected_lagged_data <− function(data, vars, lags) {. for (var in vars) {. for (lag in lags) {. data[[paste0(var, “_lag”, lag)]] < − dplyr::lag(data[[var]], lag). } } return(data). } # Variables to lag. vars_to_lag <− c(“tmin”, “tmax”, “MAXRH”). # Lags to apply. lags_to_apply <− c(1, 2, 4, 7, 14, 21, 28). # Create lagged data. lagged_data <− create_selected_lagged_data(data, vars_to_lag, lags_to_apply). # Remove rows with NA values (due to lagging). lagged_data_clean <− na.omit(lagged_data). # Set seed for reproducibility. set.seed(123). # Fit the Random Forest model using the specified variables as predictors. rf_model <− randomForest(AllCause75 ∼ month + day + year + time + tmin + tmax + MAXRH +. tmin_lag1 + tmin_lag2 + tmin_lag4 + tmin_lag7 + tmin_lag14 + tmin_lag21 + tmin_lag28 +. tmax_lag1 + tmax_lag2 + tmax_lag4 + tmax_lag7 + tmax_lag14 + tmax_lag21 + tmax_lag28 +. MAXRH_lag1 + MAXRH_lag2 + MAXRH_lag4 + MAXRH_lag7 + MAXRH_lag14 + MAXRH_lag21 + MAXRH_lag28, data = lagged_data_clean, ntree = 100, importance = TRUE). # Get the importance measures (this includes both %IncMSE and IncNodePurity). importance_df < − as.data.frame(importance(rf_model)). # Use %IncMSE for ranking. importance_df < − data.frame(Variable = rownames(importance_df), Importance = importance_df[, “%IncMSE”]). # Select the top 20 important variables. top_20_vars <− importance_df % > % arrange(desc(Importance)) % > % head(20) % > % pull(Variable). # Create a list to store PDP plots. pdp_plots <− list(). # Loop over the top 20 variables and generate PDPs for each. for (i in seq_along(top_20_vars)) {. var. < − top_20_vars[i]. pdp_var < − partial(rf_model, pred.var. = var., plot = FALSE) # Generate partial dependence values. # Create PDP plot using ggplot2 (with shortened titles). pdp_plot <− ggplot(pdp_var, aes(x =!!sym(var), y = yhat)) +. geom_line(size = 1.2) +. geom_point(size = 2) +. labs(title = paste(“PDP: AllCause75 vs”, var), x = var) +. # Only add y-axis labels to the first column, but keep tick marks. theme_minimal() +. theme(. text = element_text(size = 12), axis.title.x = element_text(size = 14), axis.title.y = if (i %% 4 == 1) element_text(size = 14) else element_blank(), # Show y-axis title only for first column. axis.text.x = element_text(size = 10), axis.text.y = element_text(size = 10) # Keep y-axis tick labels for all. ) # Append the plot to the list. pdp_plots[[var]] < − pdp_plot. } # Arrange PDP plots in a grid (e.g., 5 rows × 4 columns). grid.arrange(grobs = pdp_plots, nrow = 5, ncol = 4). R code for Fig. 4 # Load necessary libraries. library(randomForest). library(pdp). library(dplyr). library(ggplot2). library(gridExtra). # Load the original dataset (ladata.csv). data <− read.csv(“ladata.csv”). # Function to create specific lagged variables. create_selected_lagged_data <− function(data, vars, lags) {. for (var in vars) {. for (lag in lags) {. data[[paste0(var, “_lag”, lag)]] < − dplyr::lag(data[[var]], lag). } } return(data). } # Variables to lag. vars_to_lag <− c(“tmin”, “tmax”, “MAXRH”, “PM2.5”, “AllCause75”). # Lags to apply. lags_to_apply <− c(1, 2, 4, 7, 14, 21, 28). # Create lagged data. lagged_data <− create_selected_lagged_data(data, vars_to_lag, lags_to_apply). # Remove rows with NA values (due to lagging). lagged_data_clean <− na.omit(lagged_data). # Set seed for reproducibility. set.seed(123). # Fit the Random Forest model using the specified variables as predictors and PM2.5 as the dependent variable. rf_model <− randomForest(PM2.5 ∼ month + day + year + time + tmin + tmax + MAXRH + AllCause75 +. tmin_lag1 + tmin_lag2 + tmin_lag4 + tmin_lag7 + tmin_lag14 + tmin_lag21 + tmin_lag28 +. tmax_lag1 + tmax_lag2 + tmax_lag4 + tmax_lag7 + tmax_lag14 + tmax_lag21 + tmax_lag28 +. MAXRH_lag1 + MAXRH_lag2 + MAXRH_lag4 + MAXRH_lag7 + MAXRH_lag14 + MAXRH_lag21 + MAXRH_lag28 +. PM2.5_lag1 + PM2.5_lag2 + PM2.5_lag4 + PM2.5_lag7 + PM2.5_lag14 + PM2.5_lag21 + PM2.5_lag28 +. AllCause75_lag1 + AllCause75_lag2 + AllCause75_lag4 + AllCause75_lag7 + AllCause75_lag14 + AllCause75_lag21 + AllCause75_lag28, data = lagged_data_clean, ntree = 100, importance = TRUE). # Get the importance measures (this includes both %IncMSE and IncNodePurity). importance_df < − as.data.frame(importance(rf_model)). # Use %IncMSE for ranking. importance_df < − data.frame(Variable = rownames(importance_df), Importance = importance_df[, “%IncMSE”]). # Select the top 20 important variables. top_20_vars <− importance_df % > % arrange(desc(Importance)) % > % head(20) % > % pull(Variable). # Create a list to store PDP plots. pdp_plots <− list(). # Loop over the top 20 variables and generate PDPs for each. for (i in seq_along(top_20_vars)) {. var. < − top_20_vars[i]. pdp_var < − partial(rf_model, pred.var. = var., plot = FALSE) # Generate partial dependence values. # Create PDP plot using ggplot2 (with shortened titles). pdp_plot <− ggplot(pdp_var, aes(x =!!sym(var), y = yhat)) +. geom_line(size = 1.2) +. geom_point(size = 2) +. labs(title = paste(“PDP: PM2.5 vs”, var), x = var) +. # Only add y-axis labels to the first column, but keep tick marks. theme_minimal() +. theme(. text = element_text(size = 12), axis.title.x = element_text(size = 14), axis.title.y = if (i %% 4 == 1) element_text(size = 14) else element_blank(), # Show y-axis title only for first column. axis.text.x = element_text(size = 10), axis.text.y = element_text(size = 10) # Keep y-axis tick labels for all. ) # Append the plot to the list. pdp_plots[[var]] < − pdp_plot. } # Arrange PDP plots in a grid (e.g., 5 rows × 4 columns). grid.arrange(grobs = pdp_plots, nrow = 5, ncol = 4). Supplementary data Supplementary material References Associated Data Supplementary Materials Supplementary material