基于中国城市面板数据的研究：空气质量改善显著降低公众对抑郁情绪的关注

摘要

抑郁情绪已成为一个重要的公共卫生问题，其病因与自然环境尤其是空气质量密切相关。本研究以环境压力理论和社会支持理论为基础，利用中国277个地级市2011—2023年的面板数据，从理论上分析并实证检验了空气质量对公众抑郁情绪关注度的影响及其内在机制。研究发现：第一，空气质量的改善显著降低了公众对抑郁情绪的注意力，这一结论在多项稳健性检验中依然成立。第二，失业率在空气质量与公众抑郁情绪关注度之间起部分中介作用，而医疗人力资源配置和公共财政支出有效缓冲了空气质量恶化带来的不良心理影响。第三，空气质量对公众抑郁情绪关注度的影响存在显著的城市异质性，在东部和中部地区、人口净流出城市以及成熟型和衰退型资源型城市中更为明显。本研究通过考察环境公共产品和社会支持体系，为理解空气质量与公众心理健康之间的关系提供了新的经验证据，为促进环境政策与社会保障政策的协同提供了有价值的启示。

1 引言

在全球快速城市化和工业化的背景下，保持良好的空气质量已成为公共卫生领域的核心挑战。大量流行病学证据表明，长期暴露于细颗粒物（PM2.5）等大气污染物与呼吸系统和心血管疾病的发病率和死亡率显著相关（Zhang et al., 2020），其生理健康危害已在国际学术界达成广泛共识。然而，作为一项基本环境条件，空气质量的影响超越了传统的身体健康领域，同样影响着公众的心理状态，并与抑郁、焦虑等常见精神障碍密切相关。近期研究表明，大气污染物可能通过两条路径影响心理健康：一方面，污染物通过诱导神经炎症、引发氧化应激和扰乱神经内分泌调节等生物学机制直接损害大脑结构和功能（Braithwaite et al., 2019）；另一方面，持续暴露于不良空气环境会限制户外活动、减少社会交往并引发健康相关担忧，从而加剧情绪困扰和心理负担（Beemer et al., 2021）。

这一背景有两个特征促成了本研究，二者都涉及容易被低估的量级。公众对抑郁情绪的关注意是第一个。据估计，抑郁症影响全球约2.8亿人，是导致伤残损失健康生命年的主要原因之一（GBD 2019 Mental Disorders Collaborators, 2022）；在中国，抑郁障碍的终生患病率接近7%，但仅有少数受影响个体能够获得专科诊疗（Huang et al., 2019）。因此，大多数情绪困扰从未进入临床记录，而正是这一亚临床群体最先以信息寻求的形式显现出来。在人群层面测量的关注度因而发挥着诊断率无法发挥的作用：它先于正式就诊作出反应，将求助意向与症状一并记录，并且对每个城市都可连续观测，而不仅限于被调查的样本。厘清哪些因素会推动它，对预防规划以及心理健康服务供给的时机与靶向具有直接价值（图1）。

图1

空气质量是第二个。在研究窗口期内，中国的空气质量发生了剧烈且不均衡的变化，这正是它富有信息价值而不仅仅是重要的原因。继2013年《大气污染防治行动计划》和2018年蓝天保卫战之后，自2011年以来有连续监测记录的120个城市中，达到二级标准的天数平均占比（图2）从2014年的约67%上升至2023年的约82%，而年均AQI从约95降至约75。这一改善并未伴随趋同。中部地区城市的达标天数平均仍接近78%，比东北和西部平均值低约七至九个百分点，且北方平原每年冬季仍会重污染天气频发。强劲的长期改善与持续存在的城市间差异相结合，正是使空气质量能够在面板设计中与国家趋势相分离的配置。

图2

尽管现有文献已初步探讨了空气质量与公众对抑郁情绪关注之间的关联（Bhui et al., 2023），但仍存在若干方法学局限。第一，在数据时效性与测量视角方面，传统分析主要依赖污染物浓度指标或低频健康调查数据，未能从空气质量达标天数的维度揭示清洁空气可及性对公众抑郁情绪关注的持续影响。第二，在测量工具与样本覆盖方面，基于临床诊断或自评量表的研究往往受制于高成本和样本量有限，易受报告偏倚和社会赞许效应影响，难以对大规模人群的心理状态进行连续、客观的监测。

数字时代为突破这些瓶颈提供了新路径。互联网搜索引擎已成为公众表达健康关切、探索症状信息的重要渠道（Barros et al., 2020）。个体内在心理状态，尤其是尚未达到临床识别阈值的情感困扰，往往敏感地反映在特定的在线搜索行为中。百度作为中国市场份额最高的搜索引擎，其生成的“百度指数”能够实时、匿名地捕捉海量用户对特定关键词的集体关注，为测量群体心理状态提供了高频、客观的“数字足迹”。已有研究表明，此类搜索数据能够有效追踪公共卫生事件、经济不确定性和季节性情感障碍的动态变化（Liang et al., 2025），为以“抑郁”及相关词汇的搜索量作为公众抑郁情绪关注度的代理指标奠定了方法学基础。

在此背景下，本研究聚焦以下核心问题：在中国城市层面，以“空气质量达标天数比例”为代表的空气质量是否与公众抑郁情绪关注度存在系统性关联？这一影响通过何种机制传导？该影响在不同类型城市间是否存在异质性？为回答上述问题，本研究以环境压力理论和社会支持理论为分析框架，利用2011至2023年中国277个地级市的面板数据，从理论和实证两方面考察空气质量对公众抑郁情绪关注度的影响及其机制。主要发现如下：第一，空气质量改善显著降低公众抑郁情绪关注度，该效应在控制多项城市特征和气象条件后依然稳健。第二，空气质量通过城市失业率间接影响公众心理健康，而充足的医疗人力资源配置和公共财政支出能够有效缓冲空气质量不佳带来的负面心理影响。第三，空气质量对公众抑郁情绪关注度的影响存在显著的城市异质性，在东部和中部地区、人口净流出城市以及资源依赖型城市（尤其是成熟型和衰退型）中更为明显。

本研究的边际贡献体现在三个方面。第一，方法创新：运用百度指数在中国城市层面构建“抑郁情绪公众关注度”指标，实现群体心理状态的动态、客观测量，为环境心理健康研究提供新的数据来源与测量范式。第二，理论拓展：构建“环境—社会—心理”一体化分析框架，将失业率、公共卫生服务、财政能力等宏观因素纳入模型，分析其在空气质量与抑郁情绪公众关注度关系中的中介与调节作用，从而厘清宏观环境风险向个体心理健康传导的路径。第三，异质性深化：系统揭示空气质量对抑郁情绪公众关注度的影响在区域发展阶段、人口流动结构和城市资源类型上的差异化表现，识别环境心理风险的高脆弱区域，为精准化、差异化的环境健康治理提供实证依据。

2 文献综述

空气质量是衡量大气环境状况的综合指标，通常依据PM2.5、PM10、二氧化硫、二氧化氮等主要污染物的浓度水平进行评价（World Health Organization, 2021）。空气污染作为空气质量的负面表征，反映大气环境偏离健康基准的程度，是环境风险评估的核心维度。这一从环境暴露到健康效应的逻辑链条，使空气质量及其污染状况影响公众心理健康的机制成为环境科学、流行病学与心理学交叉研究的关键议题。

大量观察性研究为空气污染与抑郁情绪之间的关联提供了实证基础。早期研究主要基于特定区域或人群的横断面调查，揭示了空气污染物暴露与抑郁症状评分或患病率之间的正向关联（Khosrorad et al., 2022）。随着纵向队列研究的积累，长期暴露于细颗粒物（PM2.5）与抑郁发病风险升高之间的关系得到进一步证实（Braithwaite et al., 2019）。Zeng et al. (2019) 整合多项观察性研究的荟萃分析表明，长期PM2.5暴露以及短期暴露于PM10、二氧化氮（NO2）、二氧化硫（SO2）和一氧化碳（CO）与抑郁风险升高显著相关。然而，不同污染物的效应存在异质性；例如，臭氧（O3）对抑郁的影响在不同研究中仍不一致，提示污染物类型、暴露时间窗口和研究设计可能调节最终效应（Zhang et al., 2026）。

深入探索潜在机制对于理解这些关联至关重要，当前研究沿两条主要路径展开。第一，神经生物学路径：研究表明，空气污染物可通过诱导全身性炎症和氧化应激损害血脑屏障完整性，激活小胶质细胞并引发神经炎症，从而损伤与情绪调节密切相关的脑区（如海马体和前额叶皮层），最终增加抑郁风险（Block and Calderón-Garcidueñas, 2009）。近期研究通过生物标志物进一步验证了该路径：Zhang et al. (2026) 发现，PM2.5暴露与抑郁症状之间的关联部分由神经炎症标志物（如脑脊液sTREM2）介导，可能涉及阿尔茨海默病相关病理过程。第二，社会心理与行为路径：该路径关注空气污染如何通过影响认知、行为和社会互动间接损害心理健康（Ventriglio et al., 2021）。理论认为，持续的空气污染限制居民户外活动，减少社会互动，引发持续的健康担忧，并造成感官不适（如能见度降低），这些因素可能诱发无助感，加剧心理压力，降低感知生活质量，从而促成或加重抑郁情绪（Hahad et al., 2020）。

现有研究一致发现，空气污染的心理健康影响呈现显著异质性（Cao et al., 2023），主要受个体特征和宏观情境因素调节。在个体层面，遗传构成是关键调节因素。基因-环境交互研究发现，携带特定高风险基因型的个体在污染暴露下抑郁风险显著放大（Sun et al., 2025）。此外，生命周期阶段（如儿童/青少年、中年/老年）、性别（女性通常风险更高）以及既往健康状况是重要的脆弱性因素（Qiu et al., 2023）。在宏观层面，社会经济地位和社区环境资源发挥缓冲或加剧作用。研究表明，面临环境不公的群体，如低收入人群和少数族裔社区，往往承受更高的污染暴露和心理健康风险（Burrows et al., 2025）；而高质量的社区资源（如充足的绿地）可有效缓解空气污染对心理健康的负面影响，其保护效应对遗传易感个体尤为显著（Sun et al., 2025）。

研究方法的演进也深刻影响了这一领域的发展。传统上，抑郁结果测量主要依赖临床诊断记录、结构化心理量表或处方数据（Fried et al., 2022）。这些方法虽然具有较高的效度，但也存在明显局限：医疗数据覆盖不完整且存在固有滞后，量表调查成本高昂且难以频繁实施，这些都制约了对大规模人群心理状态进行动态、实时监测的能力。为克服这些瓶颈，研究前沿正在积极引入创新数据与方法。在因果识别方面，工具变量方法和准实验设计（如利用政策冲击或气象条件变化）被广泛采用，以增强因果推断的稳健性（Zhao et al., 2025）。在测量工具方面，利用数字足迹数据（如社交媒体情绪和在线搜索行为）作为群体心理状态的实时代理指标，已成为一个有前景的研究方向。此类数据具有高频、客观和覆盖广泛等优势，能有效捕捉公众对健康问题的集体关注和情绪波动，为在较大时空尺度上探索环境与心理的动态关系提供了新的可能。

综上所述，现有文献已确立了空气污染作为抑郁情绪风险因素的地位，并在生物机制和个体异质性方面取得了重要进展。然而，以下研究空白仍有待填补：第一，在测量层面，利用实时行为大数据（如搜索引擎指数）对公众抑郁情绪关注进行高频、宏观监测的研究仍处于起步阶段，尤其需要在中国情境下加强应用与验证。第二，在机制层面，现有研究主要聚焦于微观生物通路，从宏观城市社会经济结构视角（如失业率、公共财政和医疗服务配置）系统检验机制的实证研究相对不足。第三，在异质性分析层面，现有研究强调个体层面差异，缺乏对城市层面结构异质性（如区域发展阶段、人口流动模式和资源依赖）的系统考察。

3 理论分析与研究假设

3.1 空气质量对公众抑郁情绪关注的直接影响

直接影响是从两种理论而非一系列视角中发展而来的。环境压力理论具体说明了持续的环境状况如何成为一种心理负担；社会支持理论则具体说明了当失去的资源是人人共享的资源时，为何这种负担无法被吸收。两者是互补而非并列的。前者确定了产生痛苦的机制，后者确定了痛苦被缓冲或放大的条件。两者结合起来，为直接影响提供了一个单一的预测，并为后续检验的调节变量提供了明确的理由，且二者都不是仅仅为了添加一个标签而被援引。

从环境压力理论的视角来看，空气污染构成了一种典型的慢性、弥散性环境压力源。根据经典压力理论（Folkman et al., 1986），压力源于个体对环境需求与自身应对资源之间失衡的认知评估。空气污染的物理特征（如PM2.5浓度升高）和感官信号（如能见度降低和弥漫的气味），作为持续的环境刺激，首先被公众感知并启动认知评估过程。在初级评估阶段，个体将污染天气判断为对其健康、安全和自主活动构成威胁；在次级评估阶段，由于污染的区域性和持续性特征，个体普遍产生“应对无力感”，认为自己无法有效改变或逃离这一环境。这种持续性的威胁评估和低控制感认知，构成了诱发焦虑、无助和抑郁等负面情绪的核心心理机制（Arnsten, 2015）。从神经生物学角度来看，慢性环境压力导致下丘脑—垂体—肾上腺（HPA）轴失调，抑制前额叶皮层功能（负责高级认知和情绪调节），同时过度激活杏仁核（负责恐惧和焦虑反应）——这些神经内分泌变化共同构成了抑郁情绪产生的生理基础。此外，空气污染物可直接穿透血脑屏障，通过诱导神经炎症和氧化应激等途径，损害海马体和前额叶皮层等情绪调节脑区，进一步加剧抑郁风险。因此，空气污染不仅是物理环境危害，更是通过持续激活压力反应直接损害心理健康的重要风险因素（Thomson, 2019）。

空气污染与这一文献所依赖的大多数急性压力源相比有三个区别性特征，这些特征解释了为何其心理代价会不断累积。它是慢性的而非偶发性的，因此暴露之后没有恢复间歇。它在个体层面是不可控的，因此以问题为中心的应对方式基本不可用，居民转而依赖以情绪为中心的应对方式，而后者与抑郁结果（而非焦虑结果）的关联最为一致。它是环境性的，因此回避需要从日常活动中退出，而非回避某个离散事件。其结果是，对应对资源形成低强度但持续不断的需求，而正是在这种配置下，评估模型预测的是抑郁情绪而非急性警觉。

从社会支持理论的视角来看，清洁空气的缺失更应被理解为一种支持资源的丧失，而非又多了一种压力源（Ntontis et al., 2020）。该理论区分了两条路径：直接路径，即资源无论压力暴露与否都能提升福祉；以及缓冲路径，即资源削弱压力源与其心理后果之间的联系（Thoits, 1995）。清洁空气在两条路径上均发挥作用。它是城市所提供的最广泛共享的环境资源，无需申请、无需资格认定、也无需个人支出，并且它支撑着普通积极情感所依赖的生理舒适感与感知安全感。长期匮乏削弱了直接路径，因为一项基线福利投入被同时从每位居民身上抽走。

缓冲路径同时被削弱，而这是较为不明显的损失。支持在很大程度上通过互动产生，而互动正是污染空气所抑制的。当居民待在室内时，户外休闲、锻炼、非正式的邻里接触以及日常共处——工具性支持与情感性支持正是从中汲取的——都会减少，因此，本可吸收环境压力的资源恰在其需求上升之时收缩（Hazlehurst et al., 2018）。由于暴露是区域性的，这种收缩是集体性的：邻居、家人和同事同时受到约束，支持无法像在个体层面的冲击之后那样，从同一网络中未受影响的成员那里重新分配。环境剥夺与支持萎缩因此相互强化，而正是这一相互过程，将一种物理环境状况转化为集体情绪状态的持久恶化。

基于以上分析，我们提出核心研究假设H1：

H1：空气质量与公众对抑郁情绪的关注呈显著负相关；空气质量越高，公众对抑郁情绪的关注水平越低。

3.2 失业率的中介效应

空气质量对公众心理健康的影响可能通过城市劳动力市场状况间接传导，其中失业率构成重要的中介变量。

首先，关于空气质量影响城市失业率的机制（Li et al., 2020）：从劳动力供给端看，长期暴露于空气污染环境会损害劳动者的呼吸系统和心血管健康，导致因病缺勤增加、工作能力下降，部分劳动者可能因健康原因被迫退出劳动力市场（Hanna and Oliva, 2015）。从劳动力需求端看，空气质量是城市宜居性和营商环境的重要组成部分；持续的重污染天气降低了城市对高素质人才和知识密集型企业的吸引力，制约了服务业和新兴产业的发展，从而影响就业创造能力（Khanna et al., 2025）。供需两端相互强化，因此，尽管供给端和需求端的反应在时间上有所不同，空气质量的持续恶化预计会推高当地失业率。

关于第一条链接，有三点值得明确说明。各渠道的速度不同。健康驱动的劳动力供给减少会在当年内通过缺勤和有效工时减少而显现，而企业的选址和投资决策则滞后一年或更长时间才作出反应，这与表1中报告的滞后估计的持续性相一致。各渠道的影响分布也不同。与污染相关的健康损失不成比例地落在户外、体力及非正规劳动者身上，他们拥有的合同保护最少，再就业前景也最弱，因此城市失业率的某一上升幅度所承载的心理分量比总体数字所显示的更重。而且这两方面相互作用。随着受污染城市失去其最具流动性的高技能劳动者，当地的技能构成变得稀薄，这进一步阻碍了知识密集型雇主的进入，并将环境劣势转变为一种自我强化的劳动力市场劣势。

表1

考虑核心变量滞后效应的稳健性检验。

括号内为t统计量。***p < 0.01，**p < 0.05，*p < 0.1。

第二，关于失业率影响公众对抑郁情绪关注度的机制：失业对个人心理健康的损害已在众多研究中得到证实（Paul and Moser, 2009）。失业不仅意味着经济收入的损失和物质生活水平的下降，还伴随着心理影响，包括社会角色的丧失、日常生活结构的崩塌以及自我价值的动摇。Jahoda（1982）的潜在剥夺理论指出，就业除了提供收入外，还满足个体对时间结构、社会互动、集体目标、身份认同和规律活动的心理需求——失业意味着同时失去这些“潜在功能”。在城市层面，更高的失业率反映了区域经济停滞和积累的社会压力；即使是在业者也可能因担心失业风险而经历焦虑。

对于此处考察的机制，有两点细化很重要。失业的心理成本并不局限于失业者：仍在工作者的工作不安全感所产生的痛苦与失业本身相当，而城市失业率上升是这种不安全感最明显的信号，因此城市层面的失业率所捕捉的暴露范围远大于失业人数。这种成本还会在家庭内部传导，因为一个收入来源的丧失会改变每个成员的物质状况和情绪氛围，包括那些完全不在劳动力市场中的成员。这两点细化都意味着，失业渠道应当能在总体搜索行为中被检测到，而不仅仅是在受影响个体的记录中，这正是下文报告的中介检验所考察的内容。

基于以上分析，我们提出研究假设H2：

H2：城市失业率在空气质量达标天数比例与公众对抑郁情绪关注度之间的关系中起中介作用。

3.3 医疗劳动力的调节效应

医疗卫生人力规模作为公共卫生服务能力的核心指标，可能在空气质量与公众抑郁情绪关注之间的关系中发挥重要的调节作用。其理论基础可从以下两个维度加以阐述。

第一，健康保障与风险缓冲机制：充足的医疗卫生人员意味着更高的医疗卫生服务可及性和更强的问题应对能力（Aljuaid et al., 2022）。当居民生活在空气污染环境中时，对呼吸系统疾病等健康风险的担忧是引发焦虑和抑郁的重要路径。充足的医疗卫生人力资源能够通过提供及时的健康咨询、诊断服务和预防保健，显著削弱“污染—健康风险—心理压力”传导链条的强度。当居民感知到对其健康的制度性保护更强时，其面对环境危害的脆弱感和无助感相应减弱，从而削弱空气污染对心理健康的直接影响。

第二，资源互补与协同增强机制：清洁空气和优质医疗卫生服务可被视为生产“健康”这一最终产品的两类关键投入要素（Eckelman et al., 2018）。根据健康生产函数理论，不同要素之间可能存在互补效应。良好的空气质量可降低基础疾病的发病率，而丰富的医疗卫生人力资源则确保疾病得到高效治疗并推动更广泛的健康促进工作。两者的协同效应能够更大幅度地提升居民的整体健康水平和生活安全感，其综合心理健康效益可能超过各自独立效应之和（Phillips et al., 2005）。换言之，在医疗卫生服务水平较高的城市，空气质量改善所带来的边际心理健康效益可能更大，因为完善的医疗卫生体系能够更充分地将环境改善的“健康红利”转化为居民的感知安全感和幸福感。

基于上述分析，我们提出假设H3：

H3：城市医疗卫生人力规模正向调节空气质量对公众抑郁情绪关注的影响。

3.4 城市公共财政支出的调节效应

公共财政支出作为城市政府配置资源和履行公共职能的核心工具，是塑造城市社会经济发展环境和居民生活质量的关键宏观因素（Balynskaya and Vasilyeva, 2017）。更高的财政支出水平，尤其是投向公共服务和民生领域的支出，表明政府拥有更强的能力和资源来应对包括环境问题在内的各种社会挑战，从而可能缓冲外部环境压力对居民心理健康的负面影响。

根据Hobfoll（1989）的资源保存理论，个人或社区所拥有的资源越丰富，其抵抗和适应压力源的能力就越强。在城市层面，充裕的公共财政资源正是这样一种关键的宏观“保护性资源”，为实施有效干预和构建社会安全网提供了物质基础。具体而言，城市公共财政支出可能通过以下路径调节空气质量对公众抑郁情绪关注度的影响。

第一，环境治理与基础设施改善：更强的财政支出能力可以直接转化为更有效的环境治理投入（Wu et al., 2020），例如优化公共交通系统、推广清洁能源、扩大城市绿地——这些措施能够从根本上改善空气质量或减轻其不利影响。第二，公共服务供给与心理补偿：由财政资源支撑的优质公共服务，包括完善的社会保障体系、便捷的公共文化设施以及广泛覆盖的社区服务网络，能够为居民提供替代性的心理支持和休闲选择，从而部分补偿因空气质量差导致户外活动受限而造成的心理损失。第三，宏观经济稳定与压力缓解：强大的财政能力有助于维持良好的宏观经济环境和社会稳定（Ocampo, 2008），通过促进就业、稳定物价和提供社会救助，缓解因经济不确定性或失业风险而加剧的心理压力，从而削弱环境问题与心理困扰之间的关联。

基于以上分析，我们提出假设H4：

H4：城市公共财政支出水平显著调节空气质量达标天数比例对公众抑郁情绪关注度的影响。

4 模型构建与变量选择

4.1 模型构建

为检验空气质量与公众抑郁情绪关注度之间的关系，模型设定如下：

其中i表示城市，t表示年份；Dpre表示公众抑郁情绪关注度；Airq表示空气质量；X表示控制变量；μ表示城市固定效应，λ表示时间固定效应；ε表示随机误差项；α、β和ρ表示估计系数。

方程1采用双向固定效应最小二乘法在平衡面板上进行估计，观测单位为城市-年份：2011年至2023年13年间每年观测277个地级市，共得到3,601个观测值。未进行进一步的合并或平均。构建数据时唯一的聚合是将每日AQI和百度指数序列聚合到城市-年份层面，详见第4.4节。城市未被聚合为省份或区域，年份也未被归并为时期，因此在剔除每年的全国趋势后，每个系数均由城市内部随时间的变异识别。第6.2节的异质性分析遵循相同原则：将样本拆分并在每个子样本上重新估计相同的设定，而非将观测值合并，这正是表2、表3、表4中子样本观测数之和等于全样本的原因。

表2

区域异质性检验。

括号内为t统计量。***p < 0.01，**p < 0.05，*p < 0.1。

表3

人口流动异质性检验。

括号内为t统计量。***p < 0.01，**p < 0.05，*p < 0.1。

表4

城市资源异质性检验。

括号内为t统计量。***p < 0.01，**p < 0.05，*p < 0.1。

4.2 变量选择

4.2.1 被解释变量

抑郁情绪的公众关注度（Dpre）。借鉴互联网搜索行为分析的研究范式，本研究采用百度指数作为核心数据来源。百度指数是基于百度搜索引擎海量用户行为数据的分析平台，能够反映特定关键词在特定时间段和地理范围内的搜索强度，已被广泛用于追踪公共卫生、社会经济状况和公众情绪的动态趋势（Liang et al., 2025）。关于关键词选择：

首先，基于概念上的直接相关性和网络搜索中的常见表达，选取了三个核心关键词：“抑郁”（基本情绪表达）、“抑郁症”（临床疾病认知）和“抑郁症有哪些症状”（主动自查与信息探索）。这三个关键词分别对应公众关注度从一般情感到具体疾病认知的连续谱系，具有良好的内容效度。其次，针对每个关键词，检索了2011年1月1日至2023年12月31日期间每个地级行政区的每日指数（该平台报告的最细地理单元），并汇总为城市-年度总量。最后，将三个城市-年度总量以等权重相加，并对总和取对数。由于三个关键词的基础搜索量差异较大，表5将每个关键词分别纳入基线模型重新估计，以使研究结果不依赖于综合指标的权重设置。

表5

替代变量测量的稳健性检验。

括号内为t统计量。***p < 0.01，**p < 0.05，*p < 0.1。

同时，为验证百度指数作为抑郁情绪公众关注度代理指标的有效性，本研究从以下三个方面提供效度证据。

第一，内容效度：本研究选取的三个关键词（“抑郁”“抑郁症”和“抑郁症有哪些症状”）涵盖了从情绪表达到疾病认知再到主动求助的完整谱系，能够全面捕捉公众对抑郁相关话题的关注。

第二，效标关联效度：本研究将构建的百度指数抑郁情绪指标与《中国卫生健康统计年鉴》中的省级精神专科医院门诊就诊数据进行相关性检验，发现二者呈显著正相关（r = 0.412，p < 0.01），表明网络搜索行为能够在一定程度上反映实际心理健康服务需求。

第三，现有研究的支持：近年来，利用网络搜索数据追踪公众心理状态已成为行为流行病学的重要方法。Ayers et al. (2013) 表明，美国与澳大利亚的心理健康相关术语谷歌搜索量均呈现系统性的季节模式；Liang et al. (2025) 进一步验证了百度抑郁搜索指数与城市建成环境之间的关联。这些研究为本研究的测量策略提供了方法论支持。

需要指出的是，百度指数反映的是公众对抑郁话题的“关注度”，而非直接的“患病率”。然而，考虑到：（1）信息搜索行为往往由内在心理需求驱动；（2）经历情绪困扰的个体倾向于通过互联网寻求信息和帮助；（3）该指标能够规避传统调查中的社会期望效应，本研究认为将其作为人群层面抑郁情绪的代理指标是合理的。

4.2.2 解释变量

空气质量（Airq）。本研究使用中国生态环境部发布的官方监测数据。具体而言，以依据中国《环境空气质量标准》（GB 3095–2012）、按照《环境空气质量指数（AQI）技术规定（试行）》（HJ 633–2012）计算得出的空气质量指数（AQI）作为评价依据。AQI 是一个综合考虑六种主要污染物——细颗粒物（PM2.5）、可吸入颗粒物（PM10）、二氧化硫（SO2）、二氧化氮（NO2）、臭氧（O3）和一氧化碳（CO）——的复合指数，能够较为全面地反映一个城市的整体空气污染水平。

为刻画一个城市全年持续提供清洁空气这一基本环境公共品的能力，本研究并未简单采用年均 AQI，而是采用了一个更具政策含义和公众感知意义的指标：一年中空气质量达到或优于国家二级标准的天数比例。计算方法为：对每个城市，将全年每日 AQI 按等级分类，统计 AQI 值不超过 100（即空气质量为“优”或“良”）的天数，再除以该年有效监测天数。其公式表达为：

空气质量达标天数比例 = 全年 AQI ≤ 100 的天数 / 该年有效监测天数。

在 GB 3095–2012 AQI 监测全面实施之前的城市—年份，每日 API 监测覆盖范围仅限于重点环境保护城市，无法覆盖全部 277 个地级样本。因此，对于地方全面开展 AQI 报告之前的观测值，本研究直接采用从市级统计年鉴、统计公报和生态环境状况公报中提取的空气质量达到二级及以上的全年天数。达标天数比例则按所报告的全年达标天数除以该城市—年份的有效监测天数计算。在旧 API（GB 3095–1996）和新 AQI（GB 3095–2012）标准下，数值 ≤100 均对应空气质量达到二级及以上。第 4.4 节讨论了监测标准转换所导致的测量基准断裂。

4.2.3 中介变量

失业率（Ur）。以城镇登记失业人数年度值的自然对数衡量。数据来源于相应年份的《中国城市统计年鉴》“就业、失业与工资”部分。

4.2.4 调节变量

医疗卫生人力（Nmh）。以执业（助理）医师年度人数的自然对数衡量。数据来源于《中国城市统计年鉴》。该指标是评估区域医疗卫生人力资源可及性和公共卫生服务水平的基础性指标。

城市公共财政支出（Upe）。以地方一般公共预算支出年度值（单位：万元）的自然对数衡量。数据来源于《中国城市统计年鉴》“财政”部分。

4.2.5 控制变量

为提高研究精度，根据研究问题和数据可得性，并参考已有文献，选取了八个控制变量：（1）经济发展水平（人均地区生产总值，RGDP）；（2）产业发展水平（规模以上工业企业数量，Nid）；（3）对外开放水平（外商投资企业数量，Nfdi）；（4）固定资产投资总额（全社会年度固定资产投资总额，Tue）；（5）医疗水平（医院和卫生院数量之和，Shoh）；（6）信息技术水平（年末移动电话用户数，Mpu）；（7）失业保险覆盖情况（失业保险参保人数，Npe）；（8）环境水平（工业二氧化硫排放量，Tsoe）。

4.3 数据来源与描述性统计

本研究选取中国277个地级及以上城市作为样本，时间跨度为2011年至2023年。选择2011年作为样本起始年份的原因如下：第一，百度指数中的城市级数据自2011年起才相对完整，此前年份存在大量缺失数据；第二，中国环境空气质量监测网络自2012年起逐步完善，2011年数据可作为基准参考；第三，中国城镇化进程自2011年以来加快，城市层面的环境和心理健康问题日益突出，具有较强的研究价值。除百度指数中的抑郁情绪公众关注度指数、中国生态环境部发布的官方监测数据中的空气质量（对于全面报告AQI之前的城市-年份，由第4.2节所述官方公布达标天数补充），以及表6中使用的气象变量外，其他所有数据均来自相应年份的《中国城市统计年鉴》和城市统计公报。为避免极端值干扰实证结果，样本中所有连续变量均在1%和99%水平上进行了缩尾处理。相关变量的描述性统计见表表7：

表6

控制气象条件的稳健性检验。

括号内为t统计量。***p ≤ 0.01，**p ≤ 0.05，*p ≤ 0.1。第（1）列复现了表9第（5）列。控制变量为表9第（5）列的八个控制变量。

表7

描述性统计。

4.4 时间粒度与数据聚合

进入模型的变量来自原生频率差异很大的数据源，因此对齐程序被明确设定。生态环境部按小时和按日频率发布 AQI；本研究既不使用小时读数，也不使用年均值。对于在 HJ 633–2012 下具有完整逐日 AQI 监测的城市-年份，每个官方逐日 AQI 值都被分级；AQI ≤ 100 的天数计为达标天，年度达标天比例等于达标天数除以该日历年的有效监测天数。对于尚未全面开展本地 AQI 报告的城市-年份，我们使用年鉴和环境公报中官方发布的达到二级标准的天数年度数据来构建同一比例指标。百度指数按日频率针对三个关键词中的每一个以及每个地级行政区分别检索，日度值先汇总为城市-年度总量，然后将三个总量相加并取对数变换。其余每个变量在构造上都是年度变量，因为它来源于《中国城市统计年鉴》。

没有任何序列经过插值或升采样。对齐是朝相反方向进行的，即将两个高频序列聚合到年度水平，这是所有数据源共有的最粗频率。这一选择是刻意为之，而不仅仅是为了方便。将失业或财政支出等年度年鉴变量分解为月度或日度值，需要对数据无法支持的年内路径作出假设，并且会产生序列相关，而这种相关是插值规则的产物，而非数据的属性。向上聚合会舍弃年内变异，但不会引入此类产物，而且它与所关注的对象相匹配，因为这里所问的问题涉及一个城市在一年内持续提供清洁空气的能力，而不是它对个别污染事件的反应。贯穿始终的观测单位是城市-年份，且面板是平衡的。

空气质量序列的一个特征应当予以指出。中国从 API 向 AQI 体系的转变是分阶段进行的，而非同时进行。GB 3095–2012 于 2012 年发布；依据该标准的实时 AQI 报告于 2013 年 1 月 1 日在 74 个重点城市开始，2014 年 1 月 1 日扩展到 161 个城市，2015 年 1 月 1 日起扩展到所有地级及以上城市，该标准于 2016 年 1 月 1 日在全国生效。监测网络在同一时期扩张，因此转变前后的达标比例并非在完全相同的基础上测量，图 2 使由此产生的断点清晰可见。年份固定效应吸收了这一断点中所有城市共有的成分，识别依赖于给定年份内城市相对于全国路径的偏离，而非序列的水平。然而，由于各城市在不同年份切换，年份固定效应并不能吸收切换时间上的城市特定差异。为验证这一时间安排不会驱动结果，基线设定在 2015–2023 年重新估计，这是所有样本城市均依据 GB 3095–2012 报告 AQI 的窗口（2,493 个观测值）；空气质量的系数为 −0.403 [t = −6.004；表 8，第 (3) 列]，而全样本中为 −0.470，并且在 1% 水平上仍然显著。

表 8

样本调整稳健性检验。

括号内为 t 统计量。***p < 0.01，**p < 0.05，*p < 0.1。第（1）列剔除 2020–2022 年，第（2）列剔除四个直辖市，第（3）列将样本限定为 2015–2023 年，此时每个样本城市均按 GB 3095–2012 报告 AQI。

4.5 时间与区域格局

表 7 概括了混合样本中的变异，但未说明其在时间或空间上的形态。图 2 绘制了 2011–2023 年窗口内两个核心变量的总体及分区域情况，涵盖每年均有连续空气质量监测记录的 120 个城市。

图 (a) 给出总体情况。空气质量自 2014 年低点改善，但并非单调改善：达标天数占比均值从 2014 年的 0.674 升至 2021 年的峰值，2023 年为 0.822，而年均 AQI 从 94.8 降至 74.8。对抑郁情绪的关注并未与这一改善同步。它上升至 2019 年，2020 年回落，2021 年大致持平，此后下降。因此，两个总体序列并未以任何简单的反向方式共同变动，本文的核心论断无法从总体趋势中直接读出。在此期间，关注度主要受在全国层面发挥作用、与空气质量无关的力量主导：移动互联网使用的普及、心理疾病显著性的上升与污名化的下降，以及 COVID-19 冲击。这些正是年份固定效应所剔除的影响。

图 (b) 和 (c) 表明，总体情况掩盖了显著且持续的区域差异。2023 年达标天数占比从中部地区的 0.777 到东北地区的 0.867 不等，且自 2013 年起中部地区每年均为最低，东北和西部地区最高，差距没有持续缩小。区域内的离散程度实际上大于区域间。2023 年，区域内各城市达标天数占比的标准差平均为 0.101，而四个区域均值之间的极差仅为 0.090（这四个均值的标准差，即可比比较，仅约 0.04），当年城市层面的占比从 0.581 到 1.000 不等。因此，可用的横截面变异大多位于区域内的城市之间，而非区域之间，这也是分析在城市而非区域层面进行的原因之一。关注度在东部城市始终最高，在西部城市始终最低，这一差距与互联网普及率和城市规模的差异一致，并被城市固定效应吸收。两个变量对各区域的排序并不相同，这再次表明所关注的关系无法从水平值中恢复。

综合来看，这三个面板表明，无论是全国趋势还是跨地区排序，都不能用来评估我们所关注的关系，这一点值得明确指出，因为汇总序列容易让人得出相反的解读。下面使用的估计方法既不是在全国范围内比较污染年份与清洁年份，也不是比较污染地区与清洁地区。它是在同一年内，将一座城市与其自身相对于全国路径进行比较。表9第(1)列报告的正是这种比较，除两组固定效应外没有其他控制变量，系数为−0.176，在5%水平上显著。因此，某一年空气质量改善快于全国路径的城市，其对抑郁情绪的关注度会相对下降，尽管图2中的汇总序列显示，在空气质量改善的大部分时期内，关注度是上升的。

表9

基准回归结果。

括号内为t统计量。***p < 0.01，**p < 0.05，*p < 0.1。

5 实证分析

5.1 基准回归

表9报告了空气质量对公众抑郁情绪关注的基准回归结果。我们采用逐步加入控制变量的方法，以检验核心解释变量（Airq）估计的稳健性，所有列均控制了城市和年份固定效应。

The baseline regression results systematically test the relationship between air quality and public attention to depressive emotions. Within the panel model framework controlling for city and year fixed effects, regression results across all models robustly support research hypothesis H1—that the proportion of days meeting air quality standards is significantly negatively correlated with public attention to depressive emotions. Specifically, in Model (1) containing only the core explanatory variable, the estimated coefficient for air quality (Airq) is −0.176, significant at the 5% level. As urban characteristic variables including economic development, industrial scale, openness, investment, health services, information infrastructure, unemployment insurance and emissions are progressively introduced, the absolute value of the air quality coefficient rises from 0.176 to between 0.431 and 0.491, with significance rising to the 1% level. In the complete Model (5) incorporating all control variables, the coefficient for Airq stabilizes at −0.470 (95% confidence interval: −0.606 to −0.334), highly significant at the 1% level. This result indicates that, after controlling for other socioeconomic factors, each one percentage point increase in the proportion of days meeting urban air quality standards corresponds to an average decrease of approximately 0.47% in public attention to depressive emotions. From the trend of coefficient changes, the introduction of control variables not only did not weaken the core relationship but strengthened its robustness, suggesting that omitting urban socioeconomic characteristics may lead to underestimation of air quality’s mental health effects. The above results provide strong empirical support for the proposition that “clean air, as an important environmental public good, plays a fundamental role in maintaining public mental health,” and hypothesis H1 is verified.

Two aspects of the control set deserve comment. The sequence in which controls enter Table 9 does not affect the estimates in the full specification. Least squares with fixed effects is invariant to the order in which regressors are listed; what varies across columns (1) to (5) is the set of controls included, not their ordering, and any other entry sequence reproduces the same coefficients in the corresponding full specification; the intermediate columns would contain different subsets of controls under a different sequence, which is why they are presented as a sensitivity display rather than as alternative estimates. The sequence used here follows a substantive logic, moving from aggregate economic scale, through industrial and openness characteristics and investment, to public service and information infrastructure, and finally to social insurance and environmental pressure. Column (5), containing all eight controls, is the specification carried forward throughout the paper.

Several controls are themselves significantly associated with attention to depressive emotions, and the signs are largely as expected; because these controls enter in levels (yuan, counts and tonnes) rather than in logarithms, their coefficients are 0.000 or ±0.001 at three decimal places, so Table 9 conveys their sign and precision rather than their magnitude. Per capita GDP and mobile phone subscriptions both enter positively, consistent with search activity rising with income and with internet access rather than with mental health deteriorating as cities grow richer. This is a reporting channel, and it is one reason attention should not be read directly as prevalence. The count of hospitals and health centers enters negatively, consistent with better-provisioned cities routing health concerns into offline consultation. Industrial sulfur dioxide emissions also enter negatively once air quality is controlled for, which is initially counterintuitive; the most plausible reading is that, conditional on realized air quality, emission intensity proxies for a manufacturing-oriented employment structure and an older, less online population rather than for additional exposure. What matters for the present argument is that the coefficient on air quality is negative and significant in every column of Table 9 and, once any controls are included, remains between −0.431 and −0.491 and significant at the 1% level, so the finding does not hinge on which of these economic factors are included.

5.2 Robustness tests

5.2.1 Alternative measures for dependent and explanatory variables

Regressions were re-estimated using the logarithm of “depression” word frequency (Dpre1), logarithm of “depressive disorder” word frequency (Dpre2), and logarithm of “what are the symptoms of depression” word frequency (Dpre3) to measure public attention to depressive emotions; results are shown in Table 5, columns (1) through (3). Column (4) retains Dpre as the dependent variable and replaces the explanatory variable with the annual count of days meeting or exceeding the Grade II standard (Airq2), so the coefficient reported in the Airq / Airq2 row of that column is the coefficient on Airq2. The estimated coefficients for air quality are all significantly negative, indicating the baseline conclusions are robust.

5.2.2 Lagged effects of core variables

Following standard practice in this literature, the core explanatory variable was lagged by one to three periods and regressions re-estimated; results are shown in Table 1, columns (1) through (3). The impacts of one- to three-period lagged core explanatory variables on public attention to depressive emotions are all significantly negative, indicating that the influence of air quality persists beyond the contemporaneous year: although the magnitude declines from −0.816 at one lag to −0.523 at three lags, all three coefficients remain significant at the 1% level. Air quality can exert cumulative negative effects on public attention to depressive emotions over longer time series, providing indirect support for H1.

5.2.3 Sample adjustment tests

Sample adjustments were made from three aspects: the study window period, municipalities and the measurement basis of air quality. First, the major public health event that emerged at the end of 2019 caused negative shocks to global economic activities and had obvious impacts on public attention to depressive emotions, potentially leading to substantial estimation bias. The concern is specific rather than general. Lockdowns improved measured air quality while simultaneously altering search behavior, so the pandemic years could in principle produce a spurious association for reasons unrelated to the mechanism proposed here. The sign of that association cannot be established in advance: the most direct channel, in which cleaner air coincided with heightened pandemic-related anxiety, would work against the estimated negative effect, while local differences in the timing and stringency of restrictions could operate in either direction. This study therefore re-estimated regressions after excluding the 2020 to 2022 samples; Table 8, column (1) reports the relevant results. After excluding the pandemic years, the coefficient for air quality remains significantly negative at −0.478, against a baseline estimate of −0.470, indicating that the result is not driven by the pandemic period. The nationwide component of the COVID-19 shock is in any case absorbed by the year fixed effects that every specification includes.

Second, considering that municipalities differ substantially from other cities in economic and policy environments, with air quality and public attention to depressive emotions characteristics also potentially differing significantly, this study re-estimated regressions after excluding municipality samples; Table 8, column (2) reports the relevant results. The coefficient for air quality remains significantly negative with little change, demonstrating robust results.

This second test also speaks to the possible influence of the Tianjin Port explosion of August 2015, since Tianjin is one of the four municipalities dropped. The event leaves no detectable trace in annual city-level air quality. Tianjin’s compliant-day share was 0.463 in 2014, 0.609 in 2015 and 0.614 in 2016, and its mean annual AQI was 120.2, 103.7 and 104.1 in those 3 years, so 2015 records an improvement rather than a deterioration on the preceding year. That is what would be expected of a catastrophe which was severe but confined to one district and to a period of days, whereas the measure used here is an annual city-wide aggregate. Dropping Tianjin together with the other municipalities leaves the estimated coefficient at −0.495 and significant at the 1% level.

Third, because AQI reporting under GB 3095–2012 was phased in between 2013 and 2015 (Section 4.4), the sample is restricted to 2015–2023, the window in which every sample city reports under the new standard; Table 8, column (3) reports the results. The coefficient on air quality is −0.403 (t = −6.004), significant at the 1% level and of similar magnitude to the full-sample estimate of −0.470, so the phased transition between monitoring standards does not drive the baseline result.

5.2.4 Controlling for meteorological conditions

Weather enters the relationship in two ways. Wind, precipitation and atmospheric stability govern how emissions disperse and carry pollution across city boundaries, while sunshine duration and cloud cover are plausible direct influences on mood, so city-specific weather in a given year could confound the estimate. Annual meteorological variables were therefore constructed for every sample city, namely sunshine duration (Sun), mean temperature (Temp), precipitation (Prec), mean wind speed (Wind) and mean relative humidity (Hum), and added to the baseline specification. Table 6 reports the results. Column (1) reproduces the baseline estimate, column (2) adds sunshine duration, which captures the cloud-cover channel most directly, and column (3) adds all five meteorological variables. The coefficient on air quality is −0.393 (t = −6.032) in column (2) and −0.388 (t = −5.402) in column (3), compared with −0.470 in the baseline, so conditioning on weather attenuates the estimate by about 16 and 17%, respectively, but leaves it negative and significant at the 1% level. Sunshine duration itself enters negatively and significantly in both columns, consistent with a direct affective channel, yet adding it moves the air quality coefficient only modestly, which indicates that the baseline association does not simply proxy for cloud cover. The baseline conclusion is therefore robust to city-specific weather.

5.2.5 Spatially robust inference

The baseline standard errors treat city-year observations as independent conditional on the fixed effects, whereas the regression errors may be serially correlated within cities and, because pollution is transported between neighboring cities and search behavior diffuses through networks that cross administrative boundaries, spatially correlated across cities, so the precision of the estimate may be overstated. Table 10 therefore re-estimates the baseline specification with standard errors clustered at the city level (column 1) and at the province level (column 2), and with Conley (1999) spatial standard errors using distance cut-offs of 200 km and 500 km (columns 3 and 4). The point estimate is identical across columns by construction; the t-statistic on air quality lies between −4.879 (Conley, 500 km) and −6.314 (city clustering), and the coefficient remains significant at the 1% level in every column, so the baseline inference is not an artifact of serial or spatial correlation in the errors.

Table 10

Robustness of inference to spatial correlation.

t-statistics in parentheses. ***p ≤ 0.01, **p ≤ 0.05, *p ≤ 0.1. All columns estimate the specification of Table 9, column (5), so the coefficients are identical and only the standard errors differ. Columns (3) and (4) use Conley (1999) standard errors, which allow the errors of cities within the stated distance of each other to be correlated.

5.3 Endogeneity tests

5.3.1 Instrumental variable estimation

To address potential endogeneity issues from measurement error and omitted variables in the air quality variable, we introduce an instrumental variable for two-stage least squares estimation. Drawing on identification strategies commonly used in atmospheric physics and environmental economics, we select “the average proportion of days meeting annual air quality standards among geographically adjacent cities” as the instrumental variable. For target city i in year t, we calculate the simple arithmetic mean of the proportion of days meeting air quality standards (Airq) in year t for all its geographically adjacent cities (cities sharing borders), denoted as IV_Neighbor_Airq.

Since atmospheric pollutants (especially fine particulate matter PM2.5) exhibit significant regional transport characteristics, a city’s air quality is strongly spatially autocorrelated with surrounding areas’ air quality. Therefore, adjacent cities’ air quality is highly correlated with the focal city’s air quality, satisfying the relevance requirement for instrumental variables. Theoretically, other cities’ air quality primarily affects the focal city’s air quality through atmospheric physical processes (such as cross-regional pollutant transport), and should not directly impact focal city residents’ mental health through channels not mediated by the focal city’s air quality. Although inter-regional economic or social connections exist, this study’s model already controls for main socioeconomic characteristic variables of the focal city (such as per capita GDP, fixed asset investment, and industrial scale) as well as city and year fixed effects, which helps absorb potential confounding from regional common trends or associations. Therefore, this instrumental variable can be considered to satisfy the exclusion restriction. Relevant regression results are shown in Table 11.

Table 11

Instrumental variable endogeneity tests.

t-statistics in parentheses. ***p < 0.01, **p < 0.05, *p < 0.1.

From Table 11, the first-stage F-statistic on the excluded instrument (92.83) far exceeds the conventional threshold of 10, indicating a strong correlation between the instrumental variable and the explanatory variable. The LM statistic is 89.437 and significant at the 1% level, rejecting the null hypothesis of under-identification. The Wald F-statistic is 92.777, exceeding the Stock-Yogo critical value of 16.38 for a 10% maximal IV size (Stock and Yogo, 2005), rejecting the null hypothesis that the instrumental variable is weak. Therefore, the instrumental variable constructed in this study is valid and appropriate. In column (1), the first-stage regression coefficient for the instrumental variable is significantly positive at the 1% level; in column (2), the second-stage regression coefficient for air quality is significantly negative at the 1% level, indicating that after addressing potential endogeneity issues, the impact of air quality on public attention to depressive emotions remains evident, and the baseline conclusions are reliable.

5.3.2 Reverse causality and omitted variables

Two distinct threats are worth separating. Reverse causality would require that public attention to depressive emotions in a city changes that city’s air quality within the same year. A direct channel is hard to construct, though an indirect one is conceivable if deteriorating population mental health reduced economic activity and hence emissions; the bias that would produce is positive, pushing the estimated coefficient toward zero, so it would work against the estimated negative effect rather than in its favor and would make the reported estimate conservative rather than inflated. Three features of the design bear on this. The lag structure in Table 1 is the most informative, since attention in year t cannot influence air quality in years t-1, t-2 or t-3, yet the lagged coefficients remain negative and significant at −0.816, −0.544 and −0.523, and the one-period lag is larger in absolute value than the contemporaneous estimate. The instrument provides a second line of defense, because neighboring cities’ compliance shares are determined by their own emissions and by atmospheric transport rather than by the focal city’s search behavior. The sample adjustment tests provide a third, since the years in which a reverse channel is most plausible are precisely the pandemic years that column (1) of Table 8 excludes.

Omitted-variable bias is the more serious concern, and the fixed effects address only part of it. City fixed effects remove all time-invariant urban characteristics, among them climatic normals, terrain, industrial heritage and cultural attitudes toward mental illness. Year fixed effects remove nationwide shocks, including the common component of the 2013–2016 standard transition, the growth of mobile search and the pandemic. What survives is a city-specific, time-varying confounder that moves with air quality and independently with attention to depressive emotions. Local economic conditions are the obvious candidate, which is why per capita GDP, fixed asset investment, industrial scale and openness are controlled for directly, and why the pattern in Table 9 is informative: adding them moves the coefficient on air quality away from zero rather than toward it, from −0.176 without controls to between −0.431 and −0.491 once any of them are included, which indicates that omitting local economic conditions biases the estimate toward zero rather than inflating it. City-specific weather is the second candidate, since it governs pollutant dispersion and may also affect mood directly; adding sunshine duration and four further meteorological controls (Table 6) moves the coefficient only to −0.388, which remains significant at the 1% level. The instrumental variable estimate carries the strongest evidence on this point: the second-stage coefficient of −1.157 exceeds the fixed effects estimate in absolute value, which is the pattern expected when the fixed effects estimate is attenuated by measurement error in annual compliance shares, and not the pattern expected when it is inflated by an omitted local confounder. We nonetheless treat the instrumental variable results as corroborating the sign and significance of the fixed effects estimate rather than as supplying a preferred point estimate, since two-stage least squares is the less precise of the two and the exclusion restriction cannot be tested with a single instrument.

Two formal tests are reported in addition to the discussion above. The first is the Durbin–Wu–Hausman test (Hausman, 1978) in Table 11, which tests the exogeneity of air quality directly rather than by argument; the statistic is χ² = 5.23 (p = 0.022). Since the p-value is below 0.05, we reject the null hypothesis that air quality is exogenous at the 5% significance level, indicating that Airq is endogenous and that instrumental variable estimation is warranted. Read together with the instrumented coefficient of −1.157, which is larger in absolute value than the fixed-effects estimate of −0.470, the rejection indicates that the bias in the fixed-effects estimate runs toward zero. The baseline estimate is therefore conservative with respect to magnitude, which is why it is retained as the benchmark while the instrumental variable estimate is used to confirm the sign and significance of the effect. The second, reported in Table 12, is a two-step system GMM estimation (Arellano and Bond, 1991; Blundell and Bond, 1998) in which the lagged dependent variable is added and air quality is instrumented by its own lags. System GMM is informative here for two reasons the fixed-effects specification cannot address on its own: attention to depressive emotions is persistent, so an autoregressive term belongs in the model, and internal instruments do not require the exclusion restriction imposed on the neighboring-city instrument. The Arellano–Bond AR(2) test and the Hansen J test (Hansen, 1982) are reported so that instrument validity can be judged rather than assumed. In Table 12 the coefficient on air quality is −0.433 (t = −5.387), close to the fixed-effects estimate; the AR(2) test (p = 0.357) does not detect second-order serial correlation in the differenced errors, and the Hansen test (p = 0.203) does not reject the joint validity of the instruments. Taken with the lag structure in Table 1, the sample adjustments in Table 8 and the meteorological controls in Table 6, the design addresses reverse causality and omitted-variable bias with formal statistics as well as with argument.

Table 12

Additional endogeneity tests: two-step system GMM.

t-statistics in parentheses. ***p ≤ 0.01, **p ≤ 0.05, *p ≤ 0.1. Observations fall to 3,324 because one lag is used.

6 Further analysis

6.1 Mechanism analysis

6.1.1 Mediation effect

Based on the theoretical analysis above, to test the mediating effect of unemployment rate, this study follows the three-step mediation effect approach proposed by Baron and Kenny (1986), constructing models as follows:

where i denotes cities, t denotes years; Dpre represents public attention to depressive emotions; Airq represents air quality; Ur represents unemployment rate; X represents control variables; μ represents city fixed effects, λ represents time fixed effects; ε represents the random error term; α, β, γ, and ρ represent estimated coefficients.

Combining Equations 1, 2, and 3, the stepwise regression method is employed to test the mediating role of unemployment rate. Relevant regression results are shown in Table 13, sequentially presenting the total effect of air quality on public attention to depressive emotions (column 1), the impact of air quality on the mediating variable unemployment rate (column 2), and the direct effect of air quality and the mediating effect of unemployment rate after including the mediating variable (column 3). All models control for a series of urban socioeconomic variables and city and time fixed effects.

Table 13

Mediation effects.

t-statistics in parentheses. ***p < 0.01, **p < 0.05, *p < 0.1.

This study employs the stepwise regression method proposed by Baron and Kenny (1986) to test the mediating role of unemployment rate.

Step 1: Testing the total effect. As shown in Table 13, column (1), the regression coefficient of air quality on public attention to depressive emotions is −0.470, significant at the 1% level, indicating that air quality has a significant negative impact on public attention to depressive emotions—better air quality corresponds to lower public attention to depressive emotions. The significant total effect satisfies the prerequisite for mediation effect testing.

Step 2: Testing the impact of the independent variable on the mediating variable. As shown in Table 13, column (2), the regression coefficient of air quality on unemployment rate is −0.343, significant at the 1% level, indicating that air quality has a significant negative impact on unemployment rate—better air quality corresponds to lower urban unemployment rates. This result satisfies the second condition for mediation effect testing.

Step 3: Testing direct and mediation effects. As shown in Table 13, column (3), when both air quality and unemployment rate are included in the model, the regression coefficient of unemployment rate on public attention to depressive emotions is 0.485, significant at the 1% level, indicating that unemployment rate has a significant positive impact on public attention to depressive emotions. Meanwhile, the coefficient for air quality is −0.304, still significant at the 1% level, indicating that after controlling for unemployment rate, the direct effect of air quality on public attention to depressive emotions persists.

Synthesizing the three-step test results—significant total effect, significant impact of independent variable on mediating variable, significant impact of mediating variable on dependent variable with independent variable’s direct effect remaining significant—unemployment rate can be determined to play a partial mediating role between air quality and public attention to depressive emotions. Additionally, the Sobel (1982) test statistic is 4.954 and passes the significance test at the 1% level, further validating the robustness of the mediation effect. Because the Sobel test assumes that the indirect effect is normally distributed, a Monte Carlo confidence interval, which requires only the coefficient estimates and their standard errors (Preacher and Selig, 2012), was also computed. The indirect effect is −0.166 (−0.343 × 0.485) and accounts for 35.4% of the total effect, and the direct and indirect effects (−0.304 and −0.166) sum to the total effect of −0.470 in column (1). The 95% Monte Carlo confidence interval for the indirect effect, based on 2,000,000 draws, is [−0.237, −0.105], and the delta-method interval is [−0.232, −0.101]; both exclude zero. The above results confirm the transmission pathway of “air quality → unemployment rate → public attention to depressive emotions,” and research hypothesis H2 is supported. Regarding the underlying mechanisms, regions with better air quality often possess more advanced industrial structures and stronger economic vitality, capable of attracting quality enterprises and high-caliber talent, creating more employment opportunities, and thereby effectively reducing unemployment rates. Meanwhile, good environmental quality helps enhance comprehensive urban competitiveness, promoting service industry and high-tech industry development, further stabilizing the labor market. Unemployment not only signifies loss of economic income but also accompanies negative consequences including declining social status, loss of self-worth, and reduced social interactions—factors that significantly increase individuals’ risk of developing depressive emotions. Therefore, air quality indirectly maintains public mental health through stabilizing the employment market as a socioeconomic channel.

The mechanism deserves stating with more precision than the coefficients alone convey. The path runs from air quality to the local labor market and only then to psychological outcomes, and each step has its own empirical signature. The first is visible in column (2): a higher compliant-day share is associated with a significantly lower registered urban unemployment rate, the coefficient being −0.343 and significant at the 1% level. The sign and the significance are what the argument requires, and they match the two-sided argument developed earlier, in which polluted air simultaneously reduces effective labor supply through morbidity and absenteeism and reduces labor demand by weakening a city’s ability to retain mobile skilled workers and attract knowledge-intensive employers. The second step is visible in column (3): conditional on air quality, a higher unemployment rate is associated with significantly greater attention to depressive emotions, the coefficient being 0.485 and significant at the 1% level. Because this holds within cities and within years, it is not a comparison between structurally different labor markets but a response to change in a city’s own employment conditions.

What links the two steps psychologically is the breadth of the exposure. A rise in the city unemployment rate is experienced not only by those who lose work but by employed residents who read it as a signal about their own security, and by household members outside the labor force who absorb the material and emotional consequences. That is why a macro rate maps onto an aggregate attention measure rather than onto a count of affected individuals. It is also why the direct effect remains significant at −0.304 once the mediator is included: the labor market carries part of the environmental effect, but the appraisal and support mechanisms set out earlier operate on residents whose employment situation is unchanged, and those mechanisms are not routed through the labor market at all.

6.1.2 Moderation effects

Furthermore, moderation effect models were constructed to analyze the moderating roles of healthcare workforce size and urban public fiscal expenditure in the impact of air quality on public attention to depressive emotions. The specific models are as follows:

where Nmh represents healthcare workforce size, Upe represents urban public fiscal expenditure, γ and δ represent estimated coefficients, and other variables are defined as before. Equation 4 adds healthcare workforce size and its interaction term with air quality (Nmh × Airq) to the baseline model; Equation 5 adds urban public fiscal expenditure and its interaction term with air quality (Upe × Airq) to the baseline model.

Table 14, column (2) reports the test results for the moderating effect of healthcare workforce size. The main effect coefficient for air quality is −12.164 (p < 0.01), and the interaction term coefficient between healthcare workforce size and air quality is 0.868 (p < 0.01). Since the interaction term coefficient is positive, this means that as healthcare workforce size increases, the negative impact of air quality on public attention to depressive emotions weakens. In other words, in cities with adequate healthcare workforce allocation, the impact of air quality deterioration on public mental health can be buffered to some extent. Hypothesis H3 is therefore supported in its buffering form: healthcare provision attenuates the environmental effect rather than amplifying it. This is worth stating explicitly, because the two mechanisms proposed above carry opposite predictions. The health assurance mechanism implies attenuation, since institutional protection reduces the health anxiety through which pollution is converted into distress. The resource complementarity mechanism implies amplification, since clean air and healthcare would then be complements in producing health and the marginal psychological return to cleaner air would be larger where healthcare is better. The positive interaction coefficient indicates that the first mechanism dominates over the range of provision observed in these data, and the complementarity channel receives no support here. One clarification is needed for the magnitude. Because the interaction term is entered in levels, the main effect of −12.164 is the estimated effect of air quality when Nmh equals zero, which lies outside the observed range of healthcare provision and is not the quantity to be compared with the baseline. Evaluated at the sample mean of Nmh (13.387), the marginal effect of air quality is −0.544 (−12.164 ± 0.868 × 13.387), and it is that figure, rather than the uncentered main effect, that is comparable with the baseline estimate of −0.470. The linear interaction also implies that the marginal effect reaches zero at an Nmh value of about 14.01 (12.164/0.868), which lies inside the observed range of 10.349 to 16.323. Implied effects above that value turn positive, but they depend on the linear form of the interaction and should not be read as evidence that cleaner air raises attention to depressive emotions in the best-provisioned cities without a specification that relaxes that form.

Table 14

Moderation effects.

t-statistics in parentheses. ***p < 0.01, **p < 0.05, *p < 0.1.

Table 14, column (3) reports the test results for the moderating effect of urban public fiscal expenditure. Results show that the interaction term coefficient between public fiscal expenditure and air quality is 0.043, significantly positive at the 1% level, similarly indicating that public fiscal expenditure attenuates the relationship between air quality and public attention to depressive emotions—as fiscal expenditure levels increase, the negative impact of air quality on public attention to depressive emotions weakens. Evaluated at the sample mean of Upe (13.701), the marginal effect of air quality is −0.363 (−0.952 ± 0.043 × 13.701), and because the implied turning point (0.952/0.043 ≈ 22.1) lies well above the sample maximum of 18.930, the marginal effect remains negative throughout the observed range of fiscal expenditure, from −0.452 at the sample minimum to −0.138 at the sample maximum. Regarding the underlying mechanisms, higher public fiscal expenditure means local governments invest more resources in public services, social security, and infrastructure construction, which can effectively improve residents’ living environments and welfare levels, enhance their capacity to cope with external environmental pressures, and thereby buffer the impact of air pollution on public mental health. Research hypothesis H4 is verified.

In summary, both healthcare workforce size and urban public fiscal expenditure significantly attenuate the process through which air quality affects public attention to depressive emotions. In each case the interaction term is positive while the main effect is negative, so the environmental effect is weaker where public resources are more abundant. This finding reveals the important buffering function of public resource allocation in transmitting environmental health effects, providing empirical evidence for governments to mitigate the mental health harms of environmental pollution through optimizing healthcare resource distribution and increasing public fiscal investment.

6.2 Heterogeneity analysis

6.2.1 Regional heterogeneity

To examine regional differences in the impact of air quality on public attention to depressive emotions, this study divides sample cities into eastern, central, western, and northeastern regions according to the standard regional classification method of China’s National Bureau of Statistics. This classification not only reflects the gradient pattern of China’s macroeconomic development (eastern coastal leading development, central rising, western development, northeastern revitalization) but also encompasses systematic differences in industrial structure, environmental regulatory intensity, public service levels, and sociocultural characteristics. Conducting heterogeneity tests based on this regional framework helps reveal the different socioeconomic contexts and institutional environments on which environmental psychological effects depend. Specific regression results are shown in Table 2:

Table 2 reports regional heterogeneity test results for the impact of air quality on public attention to depressive emotions. From the regression coefficients, the impact of air quality on public attention to depressive emotions exhibits significantly differentiated characteristics across regions.

Specifically, the regression coefficient for air quality in the eastern region is −0.526, and −0.608 in the central region, both significant at the 1% level; the western region coefficient is −0.246, significant at the 5% level; the northeastern region coefficient is −0.255, not passing the significance test. In terms of impact intensity, the negative impact of air quality on public attention to depressive emotions is most pronounced in the central region, followed by the eastern region, relatively weaker in the western region, and not statistically significant in the northeastern region. Tests for the equality of coefficients across independent subsamples (Paternoster et al., 1998) qualify this ordering: the central-region coefficient is significantly larger in magnitude than the western one (z = −2.23, p = 0.026), the eastern coefficient exceeds the western one only at the 10% level (z = −1.70, p = 0.089), and the eastern and central coefficients do not differ from each other (p = 0.622). The northeastern point estimate is almost identical to the western one (−0.255 against −0.246, p = 0.972) and does not differ significantly from that of any other region, so its lack of significance is at least partly a matter of precision on 27 cities.

Regarding the underlying mechanisms, these regional differences can be explained from the following aspects. The central region is in a stage of rapid industrialization and urbanization advancement, with tensions between economic development and environmental protection being quite prominent; the public is more sensitive to air quality changes, and environmental quality fluctuations therefore have more significant impacts on their mental health. The eastern region has higher economic development levels, with generally higher education levels and health awareness among residents, holding higher expectations and requirements for environmental quality; the psychological gap effect from air pollution is relatively pronounced. The western region’s economic development is relatively lagging, with residents possibly more concerned with satisfying basic survival needs, resulting in relatively weaker psychological responses to environmental quality changes. The lack of statistical significance for air quality’s impact in the northeastern region may stem from the following reasons: First, the relatively small sample size (only 27 cities, 351 observations) limits statistical power; Second, the northeastern region is undergoing deep economic structural transformation and population outflow, with socioeconomic problems such as employment pressure, income decline, and public service contraction that residents face possibly having more dominant impacts on mental health, comparatively weakening the marginal effect of air quality; Third, the northeastern region has long winters with concentrated heating periods, with residents having strong psychological expectations and adaptability for winter air quality decline, possibly reducing their emotional sensitivity to air pollution. This result also suggests that in regions with greater socioeconomic pressure, environmental factors’ impacts on mental health may take a secondary position; future research can further explore the interactive effects of environmental and socioeconomic factors.

The above regional heterogeneity results indicate that the effects of air quality on public mental health are embedded in specific regional development stages and socioeconomic contexts. This finding provides important empirical evidence for formulating differentiated environmental governance and mental health intervention policies.

6.2.2 Population mobility heterogeneity

To further explore group heterogeneity in air quality’s psychological effects, this study conducts grouped tests based on cities’ population mobility directions. Specifically, using each city’s average annual net population migration rate during the sample period as the basis, cities are classified into “net population inflow cities” and “net population outflow cities.” This classification aims to capture systematic differences between these two types of cities in development momentum, population structure, and socioeconomic pressure. Net population inflow cities typically represent economically vibrant regions with abundant employment opportunities, attracting large numbers of migrant workers; while net population outflow cities often face challenges of sluggish economic growth and shrinking local employment opportunities, with youth labor outflow possibly exacerbating local socioeconomic pressures. Such structural differences may significantly moderate residents’ sensitivity and psychological response patterns to environmental quality. Specific results are shown in Table 3:

Table 3 reports heterogeneity test results based on population mobility direction. From the regression coefficients, the impact of air quality on public attention to depressive emotions exhibits significant differences between net population inflow and outflow cities. Specifically, the regression coefficient for air quality in net population inflow cities is −0.342, while for net population outflow cities it is −0.642, both significant at the 1% level. In terms of impact intensity, the negative impact of air quality on public attention to depressive emotions in net population outflow cities is notably stronger than in net population inflow cities, with the former’s absolute coefficient value approximately 1.88 times the latter’s. The difference between the two coefficients is statistically significant (z = −2.07, p = 0.039).

Regarding the underlying mechanisms, net population outflow cities often face development predicaments such as insufficient economic growth momentum, singular industrial structures, and limited employment opportunities. The continuous outflow of young and middle-aged labor force further weakens local social vitality and development potential, with remaining residents bearing greater economic pressure and social-psychological burdens. Against this background, air quality deterioration may become the “last straw” exacerbating their negative emotions, making residents in such cities exhibit higher psychological sensitivity to environmental quality changes. In contrast, net population inflow cities typically possess strong economic vitality and diversified employment opportunities; comprehensive public service systems and rich social support networks can buffer the impact of environmental pollution on residents’ mental health to some extent, thereby weakening the impact of air quality changes on public attention to depressive emotions.

The above results indicate that population mobility patterns reflecting urban development trajectories have significant moderating effects on air quality’s mental health effects, with residents of net population outflow cities being more vulnerable to environmental quality changes. This finding suggests that in advancing environmental governance, attention should be focused on mental health needs of residents in population-losing areas, organically combining environmental improvement with social support system construction to effectively alleviate mental health risks in such areas.

6.2.3 Urban resource heterogeneity

To further examine heterogeneity in air quality’s psychological effects across cities with different development trajectories, this study classifies sample cities into resource-based and non-resource-based categories according to the classification standards of the National Development and Reform Commission’s National Sustainable Development Plan for Resource-Based Cities. Resource-based cities are further subdivided into growth-stage, mature-stage, declining-stage, and regenerating-stage types based on their development stages and differing challenges faced. This classification aims to capture unique socioeconomic forms and environmental pressures shaped by resource endowments and industrial life cycles. Different types of resource-based cities have systematic differences in industrial structure, fiscal health status, environmental historical debts, and transformation pressures—structural characteristics that may profoundly moderate residents’ perceptions of and psychological responses to environmental quality. Specific results are shown in Table 4:

Table 4 reports heterogeneity test results based on urban resource endowment characteristics. From the regression coefficients, the impact of air quality on public attention to depressive emotions exhibits significant heterogeneity across different types of cities.

Specifically, within resource-based cities, the regression coefficient for air quality in mature-stage cities is −1.011, significant at the 1% level, with the most pronounced impact intensity; declining-stage cities have a coefficient of −0.784, growth-stage cities −0.764, both passing significance tests; regenerating-stage cities have a coefficient of −0.284, with relatively weaker impact. Non-resource-based cities have a coefficient of −0.356, significant at the 1% level. In terms of impact intensity, a gradient pattern of “mature > declining > growth > non-resource > regenerating” emerges. Tests for the equality of coefficients show which parts of this ordering are statistically reliable: the mature-stage and declining-stage coefficients are each significantly larger in magnitude than those of non-resource cities (z = −4.29, p ≤ 0.001; z = −2.18, p = 0.029) and of regenerating-stage cities (z = −4.05, p ≤ 0.001; z = −2.29, p = 0.022), whereas mature-stage and declining-stage cities do not differ from each other (p = 0.279), and the growth-stage estimate, based on 22 cities, is too imprecise to be distinguished from any other group.

Regarding the underlying mechanisms, these differences can be explained by each city type’s development stage and industrial characteristics. Mature-stage resource cities are in the peak period of resource exploitation, with highly concentrated extractive and heavy chemical industries, prominent environmental pollution problems, and residents’ long-term exposure to higher environmental risks, resulting in the strongest psychological sensitivity to air quality changes. Declining-stage resource cities face multiple pressures including resource depletion, industrial contraction, and economic downturn; residents must cope with historical environmental legacy issues while bearing transformation pains, making air pollution easily become an important trigger for negative emotions. Although growth-stage resource cities are in ascending development phases, the rapid intensification of resource exploitation is often accompanied by simultaneous environmental quality deterioration, with residents being more perceptive to environmental changes. Regenerating-stage resource cities have essentially completed industrial transformation and upgrading, with economic structures trending toward diversification and environmental governance achieving notable results; residents’ psychological states are relatively stable with lower sensitivity to air quality changes. Non-resource-based cities have relatively balanced industrial structures with more dispersed environmental pressures, and residents’ psychological responses are at moderate levels.

The above results indicate that urban resource endowment characteristics and development stages have significant moderating effects on air quality’s mental health effects. Residents of cities in resource exploitation peak periods and transformation difficulty periods are more susceptible to environmental quality fluctuations in their mental health. This finding provides important policy references for green transformation and mental health protection of resource-based cities, suggesting that policymakers should integrate environmental governance with mental health service system construction in advancing sustainable development of resource-based cities.

7 Conclusion and policy recommendations

Air pollution is an important environmental issue affecting global public health, and its relationship with mental health has attracted increasing academic attention. To explore the impact of urban air quality on public psychological well-being in China, this study utilizes panel data from 277 prefecture-level cities spanning 2011–2023, measuring environmental quality levels by the annual proportion of days meeting urban air quality standards and constructing public attention to depressive emotions indicators using the “Baidu Index,” breaking through traditional survey methods’ limitations in sample size and timeliness. On this basis, this study incorporates macro socioeconomic variables such as unemployment rates, healthcare workforce allocation, and public fiscal expenditure into the analytical framework, systematically examining the mechanisms through which air quality affects public attention to depressive emotions, and identifying heterogeneity characteristics across multiple dimensions including regional development stages, population mobility patterns, and urban resource endowments. The main conclusions are as follows:

First, air quality improvement has a significant positive effect on alleviating public attention to depressive emotions. Baseline regression results show that increases in the proportion of urban days meeting air quality standards can effectively reduce public online attention to depressive emotions, a conclusion that remains valid after multiple robustness tests including alternative variable measures, consideration of lagged effects, sample adjustments, controls for meteorological conditions, spatially robust inference, instrumental variable estimation and system GMM estimation. These results demonstrate that clean air, as a fundamental environmental public good, plays an irreplaceable role in maintaining public mental health.

Second, the impact of air quality on public attention to depressive emotions is embedded in complex socioeconomic systems, with both indirect transmission pathways and moderation by public resource allocation. Regarding transmission mechanisms, unemployment rate constitutes an important mediating channel—good air quality can enhance urban economic attractiveness, stabilize employment conditions, and thereby alleviate social-psychological pressure arising from accumulated unemployment risks, meaning environmental governance benefits can extend through the labor market to the mental health dimension. Regarding boundary conditions, healthcare workforce allocation and public fiscal expenditure play significant moderating roles: adequate healthcare resources provide institutional protection for residents against environmental health risks, while sufficient fiscal investment enhances residents’ capacity to resist external pressures through improving public service systems—both effectively buffering the negative psychological impacts of air pollution.

Third, the mental health effects of air quality exhibit significant structural differentiation across different types of cities. From the regional dimension, the effect is strongest in the central and eastern regions, whose coefficients do not differ significantly from each other, weaker in the western region, and imprecisely estimated and not statistically significant in the northeastern region. From population mobility patterns, residents of net population outflow cities are more sensitive to environmental quality changes in their psychological states. From resource endowment characteristics, residents of mature-stage cities in peak resource exploitation and declining-stage cities facing transformation difficulties are particularly affected by air quality fluctuations in their mental health.

Based on the above conclusions, this study proposes the following policy recommendations:

First, incorporate air quality governance into the overall strategic framework of public mental health promotion. Given clean air’s fundamental supporting role for public psychological well-being, governments at all levels should break from the traditional mindset of positioning air pollution control solely within the ecological environment domain, viewing it as a foundational measure for maintaining social-psychological health. When formulating environmental governance targets, beyond focusing on technical indicators such as pollutant concentrations, resident mental health improvement should be incorporated into the policy effectiveness evaluation system, promoting the formation of synergy mechanisms between environmental governance and mental health promotion.

Second, coordinate environmental governance and employment protection to fully release the social benefits of ecological improvement. Targeting the transmission pathway through which air quality affects public emotions via the labor market channel, development models mutually promoting environmentally friendly industries and high-quality employment should be actively cultivated. On one hand, enhance cities’ attractiveness to quality enterprises and high-caliber talent through improving regional environmental quality; on the other hand, ensure employment support during industrial green transformation to avoid localized unemployment risks from strengthened environmental regulations, ensuring that environmental improvement’s economic and social benefits reach the broad workforce. Simultaneously, strengthen the buffering and protective functions of healthcare resource allocation and public fiscal investment against environmental health risks, focusing on improving grassroots healthcare service accessibility, enhancing mental health professional allocation, and optimizing public fiscal expenditure structure to construct multi-level social support systems for residents.

Third, implement differentiated environmental mental health intervention strategies. Targeting the urban heterogeneity of air quality’s psychological effects, tailored policy combinations should be adopted for different types of cities. For the central and eastern regions, where the estimated effects are largest, advance environmental governance and mental health service system construction simultaneously, particularly in central cities that are still accelerating industrialization. For net population outflow cities, organically combine environmental improvement with revitalization development and social support network reconstruction to alleviate multiple psychological pressures on remaining residents. For mature-stage and declining-stage resource cities, pay high attention to residents’ psychological adaptation issues during green transformation, reducing the superimposed impact of environmental and economic dual pressures on residents’ mental health through improving social security and psychological assistance mechanisms during transition periods.

8 Limitations and future research

Three limitations should be acknowledged, each pointing to a specific extension.

The first concerns meteorology and temporal resolution. Weather enters the relationship twice. It determines measured air quality, since wind speed and direction, boundary-layer height, precipitation and temperature inversions govern how a given emission load disperses and routinely carry pollution across city boundaries, and it is a plausible direct influence on mood, since sunshine duration and cloud cover are associated with seasonal variation in depressive symptoms. Table 6 addresses this concern with annual city-level controls for sunshine duration, temperature, precipitation, wind speed and relative humidity, and the coefficient on air quality remains negative and significant at the 1% level. Annual aggregates cannot, however, capture the within-year coincidence of pollution episodes with particular weather conditions, and they omit boundary-layer height and inversion frequency, so the atmospheric dispersion channel and the direct affective channel are not fully separated. Matching daily air quality, weather and search data, and exploiting plausibly exogenous meteorological variation such as thermal inversions as an instrument, would help separate the two channels and is the most valuable extension of the present design.

The second concerns spatial structure. Both variables are spatially dependent: pollution is transported between neighboring cities, and search behavior diffuses through media and social networks that follow regional rather than administrative boundaries. Table 10 shows that inference is robust to this dependence, since the coefficient on air quality remains significant at the 1% level under city-clustered, province-clustered and Conley standard errors, and three further pieces of evidence indicate that the city-level results are consistent with the regional picture. The instrument is built from adjacent cities and is strongly relevant, which confirms the spatial dependence directly. The regional heterogeneity tests return negative and significant coefficients in the eastern, central and western regions, so no region contradicts the pooled result, and the northeastern estimate, although not significant, does not differ significantly from that of any other region. And the regional series in Figure 2 move broadly in parallel rather than diverging: from 2013 onward the central region records the lowest compliance share in every year, and attention remains highest in eastern cities and lowest in western cities throughout. Spatially robust standard errors correct inference but do not model spillovers, so a formal spatial Durbin specification with an explicit weight matrix would still be needed to estimate direct and spillover components separately and to establish whether a neighboring city’s air quality affects local attention beyond its effect through local air quality.

The third concerns the dependent measure. The Baidu Index records attention rather than prevalence, and search behavior depends on internet access, age structure and willingness to seek information about mental health. City fixed effects absorb the stable components of these differences and year fixed effects absorb their nationwide evolution, but city-specific changes in search propensity remain a source of measurement error. Linking the index to administrative records of psychiatric service utilization at city rather than provincial level would provide a sharper validation than the correlation reported above.

Statements

Data availability statement

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

Author contributions

CW: Conceptualization, Data curation, Methodology, Writing – original draft. HZ: Writing – original draft, Writing – review & editing, Conceptualization, Data curation, Project administration, Validation.

Funding

The author(s) declared that financial support was received for this work and/or its publication. This study was supported by the National Social Science Fund of China (14AJL015) and Chongqing Social Science Foundation (2020ZDJJ01).

Acknowledgments

Thanks to the National Social Science Fund of China (14AJL015) and Chongqing Social Science Foundation (2020ZDJJ01) for supporting this article.

Conflict of interest

The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Generative AI statement

The author(s) declared that Generative AI was used in the creation of this manuscript. The author(s) verify and take full responsibility for the use of generative AI in the preparation of this manuscript. Generative AI was used exclusively for language polishing and translation purposes—specifically, to improve grammar, syntax, and overall readability of the text. All substantive intellectual contributions, including the development of hypotheses, model specification, econometric estimation, robustness checks, and policy recommendations, were independently carried out by the author(s). The author(s) have carefully reviewed and edited all AI‑assisted content and assume full and sole responsibility for the accuracy, originality, and scientific integrity of the final manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

References

• 1

  AljuaidT. H.AlhrbiK. A.AlahmadiF. S.AlsharifR. M.AlanaziF. T. H.AljohaniM. D. S. (2022). The pivotal role and transformative impact of health insurance in enhancing healthcare accessibility, quality, and financial protection. Int. J. Health Sci.6, 2242–2263. doi: 10.53730/ijhs.v6nS10.15434

  • CrossRef
  • Google Scholar

• 2

  ArellanoM.BondS. (1991). Some tests of specification for panel data: Monte Carlo evidence and an application to employment equations. Rev. Econ. Stud.58, 277–297. doi: 10.2307/2297968

  • CrossRef
  • Google Scholar

• 3

  ArnstenA. F. T. (2015). Stress weakens prefrontal networks: molecular insults to higher cognition. Nat. Neurosci.18, 1376–1385. doi: 10.1038/nn.4087,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 4

  AyersJ. W.AlthouseB. M.AllemJ.-P.RosenquistJ. N.FordD. E. (2013). Seasonality in seeking mental health information on Google. Am. J. Prev. Med.44, 520–525. doi: 10.1016/j.amepre.2013.01.012,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 5

  BalynskayaN. R.VasilyevaA. G. (2017). Fiscal capacity of the city: the assessment of the influence on the sustainability of urban environment and the quality of living (the case of «second» cities of the Russian Federation). R-Economy3, 112–119. doi: 10.15826/recon.2017.3.2.013

  • CrossRef
  • Google Scholar

• 6

  BaronR. M.KennyD. A. (1986). The moderator–mediator variable distinction in social psychological research: conceptual, strategic, and statistical considerations. J. Pers. Soc. Psychol.51, 1173–1182. doi: 10.1037/0022-3514.51.6.1173,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 7

  BarrosJ. M.DugganJ.Rebholz-SchuhmannD. (2020). The application of internet-based sources for public health surveillance (infoveillance): systematic review. J. Med. Internet Res.22:e13680. doi: 10.2196/13680,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 8

  BeemerC. J.Stearns-YoderK. A.SchuldtS. J.KinneyK. A.LowryC. A.PostolacheT. T.et al. (2021). A brief review on the mental health for select elements of the built environment. Indoor Built Environ.30, 152–165. doi: 10.1177/1420326x19889653

  • CrossRef
  • Google Scholar

• 9

  BhuiK.NewburyJ. B.LathamR. M.UcciM.NasirZ. A.TurnerB.et al. (2023). Air quality and mental health: evidence, challenges and future directions. BJPsych Open9:e120. doi: 10.1192/bjo.2023.507,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 10

  BlockM. L.Calderón-GarcidueñasL. (2009). Air pollution: mechanisms of neuroinflammation and CNS disease. Trends Neurosci.32, 506–516. doi: 10.1016/j.tins.2009.05.009,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 11

  BlundellR.BondS. (1998). Initial conditions and moment restrictions in dynamic panel data models. J. Econom.87, 115–143. doi: 10.1016/s0304-4076(98)00009-8

  • CrossRef
  • Google Scholar

• 12

  BraithwaiteI.ZhangS.KirkbrideJ. B.OsbornD. P. J.HayesJ. F. (2019). Air pollution (particulate matter) exposure and associations with depression, anxiety, bipolar, psychosis and suicide risk: a systematic review and meta-analysis. Environ. Health Perspect.127:126002. doi: 10.1289/EHP4595,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 13

  BurrowsK.LuoJ.CaiY.Aschebrook-KilfoyB. (2025). Long-term PM2.5 exposure and mental health disparities: a prospective analysis of the all of us research program. Environ. Res.285:122691. doi: 10.1016/j.envres.2025.122691,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 14

  CaoZ.ZhouJ.LiM.HuangJ.DouD. (2023). Urbanites’ mental health undermined by air pollution. Nat. Sustain.6, 470–478. doi: 10.1038/s41893-022-01032-1

  • CrossRef
  • Google Scholar

• 15

  ConleyT. G. (1999). GMM estimation with cross sectional dependence. J. Econom.92, 1–45. doi: 10.1016/s0304-4076(98)00084-0

  • CrossRef
  • Google Scholar

• 16

  EckelmanM. J.ShermanJ. D.MacNeillA. J. (2018). Life cycle environmental emissions and health damages from the Canadian healthcare system: an economic-environmental-epidemiological analysis. PLoS Med.15:e1002623. doi: 10.1371/journal.pmed.1002623,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 17

  FolkmanS.LazarusR. S.GruenR. J.DeLongisA. (1986). Appraisal, coping, health status, and psychological symptoms. J. Pers. Soc. Psychol.50, 571–579. doi: 10.1037/0022-3514.50.3.571,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 18

  FriedE. I.FlakeJ. K.RobinaughD. J. (2022). Revisiting the theoretical and methodological foundations of depression measurement. Nat. Rev. Psychol.1, 358–368. doi: 10.1038/s44159-022-00050-2,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 19

  GBD 2019 Mental Disorders Collaborators (2022). Global, regional, and national burden of 12 mental disorders in 204 countries and territories, 1990-2019: a systematic analysis for the global burden of disease study 2019. Lancet Psychiatry9, 137–150. doi: 10.1016/S2215-0366(21)00395-3,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 20

  HahadO.BeutelM. E.GilanD. A.MichalM.DaiberA.MünzelT. (2020). Impact of environmental risk factors such as noise and air pollution on mental health: what do we know?Dtsch. Med. Wochenschr.145, 1701–1707. doi: 10.1055/a-1201-2155,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 21

  HannaR.OlivaP. (2015). The effect of pollution on labor supply: evidence from a natural experiment in Mexico City. J. Public Econ.122, 68–79. doi: 10.1016/j.jpubeco.2014.10.004

  • CrossRef
  • Google Scholar

• 22

  HansenL. P. (1982). Large sample properties of generalized method of moments estimators. Econometrica50, 1029–1054. doi: 10.2307/1912775

  • CrossRef
  • Google Scholar

• 23

  HausmanJ. A. (1978). Specification tests in econometrics. Econometrica46, 1251–1271. doi: 10.2307/1913827

  • CrossRef
  • Google Scholar

• 24

  HazlehurstM. F.HazlehurstM.NuriusP.HajatA. (2018). Individual and neighborhood stressors, air pollution and cardiovascular disease. Int. J. Environ. Res. Public Health15:472. doi: 10.3390/ijerph15030472,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 25

  HobfollS. E. (1989). Conservation of resources: a new attempt at conceptualizing stress. Am. Psychol.44, 513–524. doi: 10.1037/0003-066X.44.3.513,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 26

  HuangY.WangY.WangH.LiuZ.YuX.YanJ.et al. (2019). Prevalence of mental disorders in China: a cross-sectional epidemiological study. Lancet Psychiatry6, 211–224. doi: 10.1016/S2215-0366(18)30511-X,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 27

  JahodaM. (1982). Employment and Unemployment: A Social-Psychological Analysis. Cambridge: Cambridge University Press.

  • Google Scholar

• 28

  KhannaG.LiangW.MobarakA. M.SongR. (2025). The productivity consequences of pollution-induced migration in China. Am. Econ. J. Appl. Econ.17, 184–224. doi: 10.1257/app.20220655

  • CrossRef
  • Google Scholar

• 29

  KhosroradR.AvalH. E.NajafiM. L.RezaiH.DovlatabadiA.EstiriE. H.et al. (2022). The association of exposure to air pollution and depression in women; a cross-sectional study of a middle-income country. Environ. Res.215:114266. doi: 10.1016/j.envres.2022.114266,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 30

  LiX.FuH.WuY. (2020). Pollution mitigation, unemployment rate and wage inequality in the presence of agricultural pollution. China Econ. Rev.61:101425. doi: 10.1016/j.chieco.2020.101425

  • CrossRef
  • Google Scholar

• 31

  LiangY.HaoJ.ShenJ. (2025). Impact of environment in built-up areas on depressive mood: evidence from China. Appl. Geogr.176:103540. doi: 10.1016/j.apgeog.2025.103540

  • CrossRef
  • Google Scholar

• 32

  NtontisE.DruryJ.AmlôtR.RubinG. J.WilliamsR. (2020). What lies beyond social capital? The role of social psychology in building community resilience to climate change. Traumatology26, 253–265. doi: 10.1037/trm0000221

  • CrossRef
  • Google Scholar

• 33

  OcampoJ. A. (2008). “A broad view of macroeconomic stability,” in The Washington Consensus Reconsidered, (Oxford: Oxford University Press), 63–94.

  • Google Scholar

• 34

  PaternosterR.BrameR.MazerolleP.PiqueroA. (1998). Using the correct statistical test for the equality of regression coefficients. Criminology36, 859–866.

  • Google Scholar

• 35

  PaulK. I.MoserK. (2009). Unemployment impairs mental health: meta-analyses. J. Vocat. Behav.74, 264–282. doi: 10.1016/j.jvb.2009.01.001

  • CrossRef
  • Google Scholar

• 36

  PhillipsD. R.SiuO. L.YehA. G. O.ChengK. H. C. (2005). The impacts of dwelling conditions on older persons’ psychological well-being in Hong Kong: the mediating role of residential satisfaction. Soc. Sci. Med.60, 2785–2797. doi: 10.1016/j.socscimed.2004.11.027,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 37

  PreacherK. J.SeligJ. P. (2012). Advantages of Monte Carlo confidence intervals for indirect effects. Commun. Methods Meas.6, 77–98. doi: 10.1080/19312458.2012.679848

  • CrossRef
  • Google Scholar

• 38

  QiuX. Y.QiuX.ShiL.KubzanskyL. D.WeiY.CastroE.et al. (2023). Association of long-term exposure to air pollution with late-life depression in older adults in the US. JAMA Netw. Open6:e2253668. doi: 10.1001/jamanetworkopen.2022.53668,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 39

  SobelM. E. (1982). Asymptotic confidence intervals for indirect effects in structural equation models. Sociol. Methodol.13, 290–312. doi: 10.2307/270723

  • CrossRef
  • Google Scholar

• 40

  StockJ. H.YogoM. (2005). “Testing for weak instruments in linear IV regression,” in Identification and Inference for Econometric Models: Essays in honor of Thomas Rothenberg, eds. AndrewsD. W. K.StockJ. H. (Cambridge University Press), 80–108.

  • Google Scholar

• 41

  SunY.LvX.BiW.ZhangY.LuZ.GuoJ.et al. (2025). Genetic sensitivity moderates the effects of joint exposure to air pollution on depression risk: a genome-wide by environment interaction analysis. Environ. Pollut.385:127145. doi: 10.1016/j.envpol.2025.127145,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 42

  ThoitsP. A. (1995). Stress, coping, and social support processes: where are we? What next?J. Health Soc. Behav.35, 53–79. doi: 10.2307/2626957,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 43

  ThomsonE. M. (2019). Air pollution, stress, and allostatic load: linking systemic and central nervous system impacts. J. Alzheimer's Dis69, 597–614. doi: 10.3233/JAD-190015,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 44

  VentriglioA.BellomoA.di GioiaI.Di SabatinoD.FavaleD.De BerardisD.et al. (2021). Environmental pollution and mental health: a narrative review of literature. CNS Spectr.26, 51–61. doi: 10.1017/S1092852920001303,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 45

  World Health Organization (2021). WHO Global Air Quality Guidelines: Particulate Matter (PM2.5 and PM10), Ozone, Nitrogen Dioxide, Sulfur Dioxide and Carbon Monoxide. Geneva: World Health Organization.

  • Google Scholar

• 46

  WuL.MaT.BianY.LiS.YiZ. (2020). Improvement of regional environmental quality: government environmental governance and public participation. Sci. Total Environ.717:137265. doi: 10.1016/j.scitotenv.2020.137265,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 47

  ZengY.LinR.LiuL.LiuY.LiY. (2019). Ambient air pollution exposure and risk of depression: a systematic review and meta-analysis of observational studies. Psychiatry Res.276, 69–78. doi: 10.1016/j.psychres.2019.04.019,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 48

  ZhangZ.-Q.FuY.ZhaoY. L.HaoQ.LiZ. S.LiQ. Y.et al. (2026). sTREM2 mediates the associations of air pollution with depression and anxiety symptoms: the CABLE study. J. Affect. Disord.395:120665. doi: 10.1016/j.jad.2025.120665,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 49

  ZhangB. B.ZhangB.WuB.LiuJ. (2020). PM2.5 pollution-related health effects and willingness to pay for improved air quality: evidence from China's prefecture-level cities. J. Clean. Prod.273:122876. doi: 10.1016/j.jclepro.2020.122876

  • CrossRef
  • Google Scholar

• 50

  ZhaoX. L.ZhaoX.ZhouW.NeophytouA. M. (2025). Impact of continuously improved air quality on mental health: a quasi-experimental study based on longitudinal observations for Chinese adults. Environ Health3, 1020–1030. doi: 10.1021/envhealth.5c00033,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

Keywords

air quality, Baidu Index, depressive emotions, urban panel data, weather condition

Citation

Wang C and Zhang H (2026) The impact of air quality on public attention to depressive emotions: an empirical examination based on Chinese urban panel data. Front. Psychol. 17:1969885. doi: 10.3389/fpsyg.2026.1969885

Received

17 August 2026

Revised

17 September 2026

Accepted

30 September 2026

Published

16 October 2026

Volume

17 - 2026

Updates

Copyright

© 2026 Wang and Zhang.

This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

*Correspondence: Chao Wang, 372031795@qq.com

Disclaimer

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article or claim that may be made by its manufacturer is not guaranteed or endorsed by the publisher.