你大概见过这种事：一条街上的小馆子，老板一问全是同一个县出来的。或者公司某个团队，简历翻一圈，一半人来自同一家前东家。人往哪儿去看上去是各人算各人的账，可账算出来的结果，扎堆得不像是各算各的。
这一页讲三件事：人为什么走，以及「哪儿穷就从哪儿走」为什么对不上数据；走了之后是不是就算「到了」；还有在两边长大的那代人身上发生了什么。
最省事的想法是一推一拉：这边活不下去把人推出去，那边工钱高把人拉过来，差越大走的人越多。这个说法一百多年前就有人写过，今天仍是新闻里默认的解释。
它的麻烦在于：最穷的国家并不是送出移民最多的国家。克莱门斯（Michael Clemens）2014 年梳理了四十五年的研究，讲得很清楚：一国人均收入从很低的位置往上走时，外流率通常先升，升到中高收入区间才开始往下掉。这条先升后降的曲线叫迁移驼峰（migration hump）。
原因不玄：走一趟要花钱——证件、路费、中介、撑到第一笔工资的那几个月。太穷的人不是不想走，是走不了；收入涨一点，走得起的人就多一点。于是「发展」在很长一段区间里不是外流的解药，先是它的燃料。这来自跨国面板数据的相关性，个别国家受政策和战争影响很大，别当铁律。
需求那一侧也得算。皮奥雷（Michael Piore）1979 年的《过客鸟》提出：富裕经济体自己会长出一类岗位——不稳定、没有阶梯、本地人做了觉得掉价；可对打算干几年就回去的人，这份活的意义是在老家兑现的，盖房、办婚事、供弟妹念书。不是哪儿缺人往哪儿流。
真正解释「扎堆」的是网络。墨西哥迁移研究计划（Mexican Migration Project）从 1982 年的四个社区起步，1987 年后年年做，挨家挨户问迁移史。马西（Douglas Massey）等人从中提炼出累积因果（cumulative causation）：每多一个人在外面站住脚，同村下一个人的成本就低一截——有地方落脚、有人带着找活、知道哪条路走得通。迁移生产出了自己的原因。
边界是同一批人自己查出来的：福塞尔与马西 2004 年用同一套数据看墨西哥的大城市样本，累积因果基本不成立——城里人际网本来就散。所以推拉模型不算错，是单位错了：它算国家的平均收入，走的却是具体的人，沿着具体的关系，而这些关系在村镇成网、在大城市不成网。
EN The tidiest story about why people leave is push and pull: conditions here push you out, wages there pull you in, and the bigger the gap the more people move. Someone wrote a version of this more than a century ago, and it is still the default explanation in the news. The trouble is that the poorest countries are not the ones that send the most migrants. In 2014 Michael Clemens reviewed forty-five years of research on this and set it out plainly: as a country's income per head rises from a low base, its emigration rate typically rises first, and only starts to fall once the country reaches the upper-middle-income range. That rise-then-fall curve is known as the migration hump. The reason is not mysterious. Leaving costs money — documents, fares, agents, and the months of living expenses before the first wage arrives. People who are very poor are not uninterested in leaving; they cannot afford it. A little more income means a few more people who can. So for a long stretch development is not the cure for emigration but its fuel. This comes from cross-country panel correlations, and individual countries are heavily shaped by policy and war, so do not read it as an iron law. The demand side matters too. Michael Piore's Birds of Passage (1979) argued that wealthy economies grow a particular class of jobs — unstable, with no ladder, jobs locals feel demeaned by. But for someone planning to work a few years and return, the meaning of that job is cashed out back home: a house built, a wedding held, younger siblings put through school. It is not that labour flows to wherever there is a shortage; particular kinds of jobs recruit people in particular situations. What really explains the clustering is networks. The Mexican Migration Project began with four communities in 1982 and has run annually since 1987, going house to house collecting migration histories. From it Douglas Massey and colleagues drew out cumulative causation: each person who establishes a foothold abroad lowers the cost for the next person from the same village — a place to sleep, someone to walk you to work, knowledge of which route works. Migration manufactures its own causes. The mechanism has a boundary, and the same researchers found it: Fussell and Massey, using the same data on Mexican urban areas in 2004, found cumulative causation largely absent there, because urban networks are already dispersed and one new emigrant confers little informational advantage. So the mechanism holds in towns and villages and not in large cities — that is not a discount on the finding, it is a statement of its range. Looked at again, push-pull is not wrong so much as wrongly scaled: it computes national average incomes, while what actually moves is specific people along specific relationships.
列维特（Peggy Levitt）做了一件当时少见的事：两头都住。她既在多米尼加共和国一个叫米拉弗洛雷斯的村子做田野，也在波士顿的牙买加平原区做——那里住着这个村子出去的人。2001 年成书《跨国村民》。
她记下的一个数字很说明问题：到 1994 年，村里 445 户人家中超过六成五在波士顿都会区有亲人。于是村里的公共事务要等波士顿那边表态才定得下来；而波士顿这边谁家孩子怎么管、谁跟谁在处对象，村里的舆论照样管得着。不是一个村子加一个移民社区，是同一个村子摊在两个地方。
回流的也不只是钱。列维特 1998 年提出社会汇款（social remittances）：回去的还有做法和标准——会该怎么开、家里谁说了算、什么算体面。钱有账可查，这些没有；但它改动的是当地人判断事情的尺子，往往比钱更难逆转。
钱的量级也不小：世界银行估算，2024 年流向中低收入国家的侨汇约六千八百五十亿美元，高于这些国家收到的外国直接投资，而这只算正规渠道。债与货币的人类学那期讲过，一笔钱往哪儿流从来不只是经济问题——侨汇同时是收入、是义务，也是一份在老家的位次。
给这类现象起名的是格利克·席勒（Nina Glick Schiller）等人 1992 年提出的跨国主义（transnationalism）：移民不是从一个容器搬进另一个容器，是在两边同时维持一个社会场。十年后她与维默（Andreas Wimmer）点出社会科学自己的毛病，叫方法论民族主义（methodological nationalism）——默认「国家 = 社会」是天然单位。统计按国界收、问卷按国界发，横跨国界的东西于是天然看不见：不是被忽略，是不在表格里。
反过来的批评来自同一阵营内部。波特斯（Alejandro Portes）等人调查哥伦比亚、多米尼加、萨尔瓦多移民，发现把跨国经营做成常规生计的只占一小部分，多数人还是把重心压在落脚地。所以：跨国主义作为「一种普遍生活方式」被高估了，作为「一类被旧框架漏掉的现象」是成立的。两句都得说。
EN Peggy Levitt did something unusual for her time: she lived at both ends. She did fieldwork in a Dominican village called Miraflores and also in Jamaica Plain in Boston, where people from that village had settled; the book, The Transnational Villagers, appeared in 2001. One figure she recorded says a great deal: by 1994, more than sixty-five per cent of the village's 445 households had relatives in the Boston metropolitan area. Village decisions waited on what people in Boston thought; meanwhile how a child was raised in Boston, or who was seeing whom, remained subject to village opinion. This is not a village plus a migrant enclave — it is one village spread across two places. What flows back is not only money. In 1998 Levitt proposed the idea of social remittances: practices and standards travel too — how a meeting should be run, who decides things at home, what counts as respectable and what counts as going too far. Money leaves a paper trail and these do not, but what they alter is the yardstick people use to judge things, and that is often harder to reverse than cash. The money is not small either. The World Bank estimates that remittances to low- and middle-income countries reached roughly $685 billion in 2024, more than those countries received in foreign direct investment. That counts only formal channels; cash carried by hand never enters the statistics, so the real figure is higher. As the piece on the anthropology of debt and money argued, where money goes and whether a debt can be discharged is never only an economic question, and remittances are a clear case: at once income, obligation, and a position in the hierarchy back home. The name for this came from Nina Glick Schiller and colleagues in 1992: transnationalism, the argument that migrants do not move from one container into another but maintain a social field across both. A decade later she and Andreas Wimmer went further and named a defect in the social sciences themselves — methodological nationalism, the default assumption that nation equals society equals the natural unit of analysis. Statistics are collected by border, surveys fielded by border, funding allocated by border; anything spanning borders becomes invisible, not because it is ignored but because there is no column for it. The counter-argument belongs here too, and it comes from inside the same camp. Alejandro Portes and colleagues, surveying Colombian, Dominican and Salvadoran immigrants, found that only a small share of respondents made cross-border business a regular livelihood; most put their weight on where they had landed. So: transnationalism is overstated as a general way of life, and sound as a category of phenomena the old framework missed entirely. Both sentences are true and both have to be said.
关于移民的下一代，二十世纪前半叶的主流预期是一条直线：一代比一代靠近主流，语言、职业、通婚依次跟上，时间自己会解决问题。这套说法出自芝加哥学派的种族关系循环，1964 年由戈登（Milton Gordon）整理成体系。
1993 年波特斯与周敏提出不同看法，叫分层同化（segmented assimilation）：不是一条线，是几条岔路。往哪条走，取决于落进什么样的劳动市场、身边有没有一个还站得住的社区、以及在别人眼里被归成哪一类人。
他们随后做了一项长期追踪：1992 年起在迈阿密与圣迭戈两地的学校里访了五千多名八、九年级的移民子女，一路追访到 2000 年代中期。一个和直觉相反的结果是：保住了父母语言、跟族裔社区还连着的孩子，学业和心理指标上平均并不吃亏，有些反而更好。他们管这叫选择性涵化（selective acculturation）。
为什么保住母语反而有用？不是语言有什么魔力，是父母的权威。孩子学得飞快、父母学得很慢，家里就出现角色倒置：孩子替父母看合同、打电话、跟老师谈话。会说话的那个人同时变成了拿主意的那个人，父母能给的约束和托底随之失效——不是不想给，是给不出去了。选择性涵化保住的是那条还能商量事的通道。
另一边的反驳同样有力。阿尔巴（Richard Alba）与倪志伟 2003 年的《重塑美国主流》指出：主流本身在扩张——今天算「主流」的东西里有大量是上一轮移民带进来又被吸收的；把下行那条路讲得太重，会低估长期数据里多数群体仍在向上的部分。至今有争论。
证据强度要说清：那项追踪覆盖两个都会区、特定年代、特定族裔构成，别搬到别处当规律。留下的是机制——涵化不是一根 0 到 1 的滑条，两套东西可以同时在身上，也可以同时接不稳。
还有一层，种族与族群那期讲过：边界是双向的。你怎么认自己是一回事，别人把你归成哪一类是另一回事，后者你改不了。第二代常报告的那种「两边都差着一点」，往往不是身份没长好，是两套归类同时在跑，而且对不齐。
EN For the first half of the twentieth century the dominant expectation about immigrants' children was a straight line: each generation moves closer to the mainstream, with language, occupation and intermarriage following in order, and time solves the problem by itself. Behind this lay the Chicago School's race relations cycle, systematised by Milton Gordon in 1964. In 1993 Alejandro Portes and Min Zhou proposed something different, segmented assimilation: not one line but several forks. Which one you take depends on what kind of labour market you land in, whether there is still a functioning community around you, and which category other people sort you into. They then ran a long-term study: from 1992 they surveyed more than five thousand eighth- and ninth-grade children of immigrants in schools in Miami and San Diego, following them into the mid-2000s. One result runs against intuition — children who kept their parents' language and stayed connected to an ethnic community did not on average do worse on academic and psychological measures, and on some did better. They call this selective acculturation. Why would keeping the mother tongue help? Not because language has any magic in it, but because of parental authority. Children learn the new language fast and parents slowly, and the household undergoes a role reversal: the child reads the contracts, makes the phone calls, talks to the teacher. The person who can speak becomes the person who decides. The limits and the backstop parents could provide stop working — not because they no longer want to provide them, but because they can no longer deliver them. What selective acculturation actually preserves is a channel through which things can still be discussed. The counter-argument is equally strong. Richard Alba and Victor Nee's Remaking the American Mainstream (2003) points out that the mainstream itself expands: a great deal of what now counts as mainstream was carried in by an earlier round of immigrants and absorbed. Emphasising the downward path too heavily understates how much of the long-run data still shows most groups moving up. The two sides frequently use different cuts of the same data, and the argument continues. On evidence strength: that study covers two metropolitan areas, a particular period, and a particular ethnic composition; do not transplant it anywhere as a law. What survives is the mechanism — acculturation is not a single slider from zero to one; both repertoires can be held at once, and both can fail to take hold at once. There is a further layer, made in the piece on race and ethnicity: boundaries run in two directions. How you identify yourself is one thing; which category other people sort you into is another, and that one is not yours to change. The feeling the second generation most often reports, of falling slightly short on both sides, is frequently not an identity that failed to form but two systems of classification running at once and failing to line up.