陈颖雪
城市群作为城镇化的主体形态，发展已经进入关键期，城际客运通道是支撑城市群一体化协调、健康发展的重要载体，为建立合理、高效的城际交通模式，需关注城市群旅客出行行为。探亲访友是城际三大出行目的之一，且与其他目的出行行为存在差异，因此需对其进行单独研究。
论文首先对城际探亲访友出行产生原因、行为现状、研究历程、国内外研究现状进行阐述，指出城际探亲访友行为与移民行为密切相关，出行份额大，出行特征独特，国内外对探亲访友出行行为研究除了针对出行方式选择行为建立MNL模型外，多限于出行行为调查数据统计描述，且欠缺以城市群为范围的探亲访友出行行为研究。其次,从城市群内涵、特征及城市群与交通的关系出发，指出探亲访友出行需求是由长三角城市群本质所引发，论证探亲访友出行行为研究对建立城市群合理交通模式的作用。城际探亲访友出行行为研究的前提是准确、深入的行为数据，国内相关官方数据积累不足，一些机构和学者的研究中数据调查方式简单且不重视抽样偏差，造成研究结论存在误导，亟需进行城际出行行为数据调查方式、样本抽样偏差界定及修正方面的研究。
论文针对城际出行的低频率特征，指出现场抽样和选择方案抽样是适合城际出行行为调查的有效且高性价比的抽样方法，但代价是产生两类抽样偏差：1）内生分层结构偏差（endogenous stratification bias），即样本与总体对于调查场所存在不同的访问概率分布；2）选择倾向度偏差（avidity bias），即由于样本中包含过多热心用户导致的选择意愿估计的向上偏差。通过总结复杂调查数据两类偏差修正方法的优缺点，提出适用于现场抽样和选择方案抽样的偏差修正方法。并把抽样偏差修正作为本研究的关键问题。
在此基础上，论文提出长三角城市群出行行为调查方案。从现场抽样效率、非集计模型样本容量要求、基于模型的统计推断方法在小样本研究中的优势等几个角度论证样本容量满足研究需求。从纯现场抽样样本特点论证样本外生变量结构合理性。结合抽样过程，分析现场抽样的两类偏差来源：在不同站场投入不同问卷量导致样本出行方式选择结构与总体不同，即样本内生分层结构偏差；不同出行频率旅客入样概率不同造成样本出行频率分布与总体不同，即选择行为倾向度偏差。由此，结合抽样设计过程，针对不同计量单位的分析指标分别计算抽样偏差修正所需的反转因子/抽样权重。
最后，基于现场抽样/选择方案抽样偏差修正后的调查数据分析长三角城市群探亲访友出行行为机理：
1）分移民、原住民对长三角城市群城际探亲、访友行为的出行频率、出行日期、出行方式、出行目的地等指标进行单变量统计描述和双变量相关分析。并引入拜访对象及拜访类型从根源上理解探亲访友出行行为。
2）针对未修正和抽样权重Ｗ人修正数据分别建立poisson模型、结构零为常数的ZIP（Zero-inflation poisson零膨胀的泊松）回归模型、结构零与“移民与否”有关的ZIP模型，利用模型拟合优度指标如似然值检验、离差、Vuong检验、命中率检验等对模型进行比选，提出抽样权重Ｗ人的加入修正了抽样数据的内生分层结构偏差和选择倾向度偏差，并帮助设定正确的探亲访友出行频率模型结构，抽样权重Ｗ人修正后的结构零与“移民与否（immi）”有关的ZIP模型能较好满足研究需求。利用该模型形式对探亲出行频率、访友出行频率单独建模，联动出行日期模型结果，分析城际探亲访友出行频次影响机理。
3）对比抽样偏差修正前后出行日期模型结果，论证抽样偏差修正必要性，后利用修正后的BL（binary logit）模型对每个时段居民是否进行探亲、访友出行进行建模，分析移民与否（immi）、职业（pro）、性别（sex）、年龄（age）、私家车拥有情况（car）、收入（income）等对探亲访友出行日期的影响。
4）在用似然比、命中率检验得出NL模型不显著优于MNL模型后，选择简单模型（MNL）模型进行进一步研究。对所建模型进行选择方案抽样数据偏差修正，一种方法是基于抽样权重/反转因子Ｗ人*次修正后的抽样数据建立非集计模型，另一种方法是将反转因子Ｗ人*次结合到模型的极大似然函数中。利用网络简单随机抽样问卷数据进行模型命中率检验，结果显示：两种修正方法下的非集计模型精度均能达到建模要求，两种修正方法下的模型释义大部分相似，但也存在少许差异。
关键词：城市群，探亲访友，出行行为，现场抽样/选择方案抽样，抽样偏差修正，ZIP（Zero-inflation poisson零膨胀泊松回归）模型，MNL模型
Urban Agglomeration is the inevitable trend of urbanization,intercity passenger corridor is important to Urban Agglomeration for supporting the coordinated and healthy development of urban agglomeration,in order to establishment a reasonable and efficient intercity transportation mode, urban agglomerations passenger travel behavior should be studied, Visiting Relatives and Friends(VFR) is one of the three intercity trip purpose, and its trip behavior is distinctive, therefore need to be studied separately.
Paper started with Urban Agglomeration content, features and its relationship with traffic, pointed out the need of VFR trips are caused by the nature of the Yangtze River Delta city group,Demonstration the effect of VFR travel behavior research for reasonable establishment of urban transport modes. Inter-city VFR travel causes, current situation and research status was analyzed,Pointed out that VFR travel behavior studied abroad is limited to statistical description of VFR behavior survey data, only MNL model established for VFR trip mode choice,study on Urban Agglomeration VFR travel behavior is deficiency.
Study on intercity travel behavior needs to rely on investigative techniques to obtain data, inter-city travel is a low-frequency behavior, require specific sampling method, so development of sampling theory and statistical inference methods are reviewed,including that on-site sampling and choice-based sampling is suitable for intercity travel behavior survey,the deviation caused by the sampling method is defined: endogenous stratification bias andavidity bias.By summing up the advantages and disadvantages of two types of error correction method, bias correction method for on-site sampling and choice-based sampling is proposed , and the sampling bias correction is a key issue in this study.
After that, survey programs and sample size, sample structure of Yangtze River Delta Urban Agglomeration travel behavior was introduced, from the aspect of on-site sampling efficiency, sample size for Disaggregate Model, advantages of Statistical Inference based on model to demonstrate sample size meeting research needs.From aspect of pure onsite sample demonstrates the same structure of exogenous variables features as the general to demonstrate sample structure reasonable.analyze source of two types sampling bias combined with the sampling process,Inversion factor / sampling weights for Sampling bias correction were calculated for the analysis of the different measurement units indicators.
Finally, analysis Yangtze River Delta city group VFR travel behavior mechanism based on sampling bias corrected survey data:
1)Study on univariate statistical description and Bivariate correlation analysis of the Yangtze River City Group VFR travel frequency, date of travel, travel mode, travel destinations.
2)poisson model, ZIP model with structure zero is constant, ZIP model with structure zero relates "immigrant or not (immi)" are building on uncorrected and corrected data.Goodness of fit test indicators such as likelihood ratio test, deviation test, Vuong test, hit ratios test were used to test the models, Confirming that adding sampling weights can help to set the correct model, ZIP model with structure zero relates "immigrant or not (immi)" with corrected data was testified can meet the demand accuracy of model building, so it was used to build VR(Visiting Relatives) and VF(Visiting Friends) travel frequency and analysis the behavior mechanism.
3)After necessity of sampling bias correction is confirmed,BL models was used to model resident visit relative or not, visit friends or not on each time, then impact of influence factors like immigrant or not (immi), occupation (pro), sex, age, private car ownership situation (car), income was studied based on the model built.
4)After verifying that NL model is not significantly better than the MNL modelusing the likelihood ratio testand hits ratio, the simpler model MNL model were chosen. Sampling bias were corrected by two means, one is to build model based on corrected data, the other is to incorporate sampling weights / reverse factorinto maximum likelihood function of the model, using random sampling from questionnaire data on network to do the hit ratio test, results show thatMNL model under two correction methods both can achieve modelaccuracyrequirements, the two model have similar interpretation of behavior mechanism, but there are also a little different.
Key Words: Yangtze River Delta Urban Agglomeration, Visiting Relatives and Friends（VFR）,travel behaviour, onsite sampling and choice-based sampling, sampling bias and correction, Zero-inflation poison model, MNL model
