Living Standards Measurement Survey 2003 (Wave 2 Panel) - Albania

Geographic coverage
---------------------------

National

Analysis unit
---------------------------

- Households
- Individuals

Kind of data
---------------------------

Sample survey data [ssd]

Sampling procedure
---------------------------

Sample size is 2,155 households

LSMS Sample Design

The LSMS design consisted of an equal-probability sample of housing units (HUs) within each of 16 explicit strata. These were selected in two stages. The first was to select - within strata - an agreed number of enumeration units (EAs) with probability proportional to number of HUs in the EA (according to 2001 Census data). The second stage was to select 8 HUs systematically from each selected EA. (Substitutes were used where necessary to ensure that 8 households were successfully interviewed in each EA, but I shall ignore that for current purposes.) Although probabilities within strata were (approximately) equal, probabilities varied greatly between the strata. Notably, the mountain region was heavily over-represented and the Central Rural region was under-represented in the sample.

Panel Survey Sample Design

The LSMS was so-designed, partly to enable separate analysis by broad strata (e.g. separate estimates for the mountain region). Regional analysis is much less important for the panel. The sample size will in any case be considerably smaller, so some regional sample sizes would inevitably be too small to permit robust estimation. The prime objective for the panel is to enable national-level estimates with the highest possible precision. To achieve this, the sample was structured in a way that minimises the overall variation in households' selection probabilities. In other words, the sample distribution over strata matched as closely as possible the population distribution.

- Statistical precision for national estimates is greatly improved, compared with the LSMS design. Design effects (under the assumption of equal stratum population variances) can be expected to be around 1.02 for the panel sample, compared with 1.28 for the LSMS sample. In other words, a panel sample of 1500 interviews would give precision equivalent to an equal-probability sample of 1172 households if it followed the LSMS distribution of households over strata, but gives precision equivalent to an equal-probability sample of 1471 households with the panel design. Precision is also further improved by retaining all 450 EAs in the sample, thus reducing the design effect due to the clustering (as mean responding sample size per cluster will reduce from 8.0 to around 3.3);
- The design was simple to implement as, within each stratum, the number of households to select was the same in each EA. (Note that sampling fractions have been expressed as a fraction of 8 for this reason);
- The sample size was set so as to make it likely that the number of achieved interviews would be between 1600 and 1700. Substitute households were not be used in the case of non-response.
Rather, all attempts were made to maximise the response rate. This enables the use of potentially powerful non-response weighting using the LSMS data.

Panel design

The Albanian panel survey sample was selected from households interviewed on the 2002 LSMS conducted by INSTAT with support from the World Bank. The sample size for the panel took approximately half the LSMS households and has re-interviewed these households annually in each of 2003 and 2004. The LSMS data collected in 2002 therefore constitute 'Wave 1' of the panel survey and giving three waves of panel data altogether. The fieldwork for Wave 3 was carried out in the spring of 2004.

The sample selected from the LSMS for the panel was designed to provide a nationally representative sample of households and individuals within Albania (see Appendix B for full description of the sample design and selection procedure). This differs from the LSMS where the sample was designed to be representative of each strata which broadly represented the main regions in Albania so that regional level statistics could be generated (Mountain, Central, Coastal, Tirana).

The panel also has no over-sampling as in the LSMS. This design was adopted as the smaller sample size for the panel would have made it more difficult to produce regionally representative samples and increased sampling error while over-sampling can introduce additional complications for analysis in the context of a panel. The panel data can be used for analysis broken down by strata to assess any differences between areas but should not be used to produce cross-sectional estimates at the regional level. The relatively small sample size for the panel must always be considered as cell sizes which are small have higher levels of error and can produce estimates which are less reliable. Panel surveys have a number of elements of which data users need to be aware when carrying out their analysis. The main features of the panel design are as follows:
- All members of Wave 1 households were designated as original sample members (OSMs) including children aged under 15 years.
- New members living with an OSM become eligible for inclusion in the sample
- All sample members are followed as they move address and any new members found to be living in their household included
- Sample members moving out of Albania are considered to be out of scope for that year of the survey (note that they remain potentially eligible for interview and it is possible they may return to a sample household at a future wave)
- From Wave 2, only household members aged 15 years and over are eligible for interview. As children turn 15, they become eligible for interview (This differs from the LSMS where the individual questionnaire collected some data on children under 15 from the mother or main carer).

The panel is essentially an individual level survey as individuals are followed over time regardless of the household they are living in at a given interview point. This is the key element of the panel design. Households change in composition over time as members move in and out, children are born and others die. New households are formed as people marry or children leave the parental home and households can disappear if all members die or all members move in different directions. The fact that households do not remain constant over time means that it is only possible to follow individuals over time, observing them in their household context at each interview point.

It should also be noted that a 'household' is not equivalent to a current address. A household may move to a new address but maintain the same composition. Similarly, an individual sample member may move between several addresses during the life of the survey. In this design, there is no substitution or recruitment of new households moving into addresses vacated by sample members.

Mode of data collection
---------------------------

Face-to-face [f2f]

Research instrument
---------------------------

Panel questionnaire content

The data for Wave 1 of the panel survey are the LSMS data so contains all the modules carried for the LSMS. To minimise respondent burden and help maintain response rates in the panel survey it was necessary to reduce the length and complexity of the LSMS questionnaire. However, it was also important to maintain comparability in question wording and response categories wherever possible as only variables which are comparable over time can be used for longitudinal analysis. The Wave 2 questionnaire is therefore a reduced version of the LSMS questionnaire with some additional elements that were required for the panel e.g. collecting details of people moving into and out of the household, and some new elements that had not been included on the LSMS. A cross-wave list of variables for Waves 1 and 2 shows which variables have been carried at both waves, which were carried at Wave 1 only and which at Wave 2 only (see ‘Variable Reconciliation LSMS_PANEL_final). The most notable changes were that the LSMS detailed consumption module was not collected at Wave 2 and the agriculture module was a reduced form compared to the LSMS.

The Wave 2 individual questionnaire contains some routing depending on whether or not the person is an original sample member interviewed on the LSMS or a new person who had joined the household since Wave 1. This is because some information only needs to be collected once e.g. place of birth and other information only needs to be updated on an annual basis. For example all qualifications were collected on the LSMS so for original members we only need to know if they have gained any new qualifications in the past year but for new members we need to ask about all qualifications. Users of the data need to be aware of this routing and in some cases may need to get information from an earlier wave if it was not collected at the current wave. Users are recommended to use the data in conjunction with the questionnaires so they are aware of the routing for different sample members.

地理覆盖范围
---------------------------
- 全国

分析单元
---------------------------
- 家庭
- 个人

数据类型
---------------------------
样本调查数据 [ssd]

抽样程序
---------------------------
样本规模为2,155户家庭
LSMS样本设计

LSMS设计包括在每个16个明确层中，对住房单位（HUs）的等概率抽样。这些层在两个阶段被选中。第一阶段是在层内选择一个商定的数量的人口普查单位（EAs），其选择概率与该EAs中住房单位数量成正比（根据2001年人口普查数据）。第二阶段是从每个选中的EAs中系统性地选择8个住房单位。在必要时使用替代品以确保每个EAs中成功采访到8户家庭，但在此目的下我将忽略这一点。尽管层内的概率（大约）相等，但层之间的概率差异很大。值得注意的是，山区在样本中过度代表，而中央农村地区代表性不足。

面板调查样本设计

LSMS的设计部分是为了能够按广泛的层进行单独分析（例如，为山区提供单独的估计）。对于面板来说，区域分析的重要性要小得多。无论如何，样本规模将显著较小，因此一些区域的样本规模不可避免地会太小，无法进行稳健的估计。面板的主要目标是能够以尽可能高的精度进行国家级估计。为了实现这一点，样本的结构被设计得最小化了家庭选择概率的整体变化。换句话说，样本在层上的分布尽可能接近人口分布。
- 与LSMS设计相比，国家估计的统计精度大大提高。在假设层人口方差相等的情况下，预期面板样本的设计效果约为1.02，而LSMS样本的设计效果约为1.28。换句话说，如果面板样本遵循LSMS在层上家庭分布的分布，则1500次访谈的样本将提供相当于1172户家庭等概率样本的精度，但如果采用面板设计，则提供相当于1471户家庭等概率样本的精度。保留所有450个EAs在样本中还可以进一步提高精度，从而减少由于聚类引起的设计效果（因为每个集群的平均应答样本量将从8.0减少到约3.3）；
- 该设计易于实施，因为在每个层内，每个EAs选择的家庭数量相同。（请注意，由于这个原因，抽样比例被表示为8的分数）；
- 样本规模被设定，以便实现的访谈数量有可能在1600到1700之间。对于非响应情况，不使用替代家庭。相反，尽一切努力提高应答率。这使能够使用LSMS数据的有力非响应加权。

面板设计

阿尔巴尼亚面板调查样本是从由INSTAT进行并由世界银行支持的2002年LSMS调查中采访的家庭中选择的。面板的样本规模大约是LSMS家庭的一半，并在2003年和2004年每年对这些家庭进行重新采访。因此，2002年收集的LSMS数据构成了面板调查的“第1波”，总共有三次面板数据。第3波的实地工作于2004年春季进行。

从LSMS中为面板选择的样本旨在提供阿尔巴尼亚全国范围内的家庭和个人样本（参见附录B中关于样本设计和选择程序的完整描述）。这与LSMS不同，LSMS的样本旨在代表阿尔巴尼亚的主要地区（山区、中央、沿海、地拉那），以便生成区域级别的统计数据。

面板也没有像LSMS那样的过度抽样。采用这种设计是因为面板较小的样本规模将使得产生代表区域的样本更加困难，并增加抽样误差，而过度抽样在面板分析的情况下可能引入额外的复杂性。面板数据可以用于按层拆分分析，以评估不同地区之间的差异，但不应用于产生区域级别的横截面估计。必须始终考虑面板的相对较小样本规模，因为单元格规模小具有更高的误差水平，并且可能产生不可靠的估计。在进行分析时，数据用户需要了解面板调查的几个要素。面板设计的主要特征如下：
- 第1波家庭的所有成员都被指定为原始样本成员（OSMs），包括15岁以下的孩子。
- 与OSM共同生活的新成员有资格加入样本
- 随着样本成员的搬家，所有样本成员都会被追踪，并且任何被发现居住在家庭中的新成员都会被包括在内
- 移出阿尔巴尼亚的样本成员被视为当年调查的范围之外（请注意，他们仍然可能具有被采访的资格，并且在未来的一波中他们可能返回到样本家庭）；
- 从第2波开始，只有15岁及以上的家庭成员有资格接受采访。随着孩子长大，他们有资格接受采访（这与LSMS不同，LSMS的个人问卷从母亲或主要照顾者那里收集了一些15岁以下孩子的数据）。

面板本质上是一项个人层面的调查，因为个人在给定采访点居住的任何家庭中都会随时间推移而被追踪。这是面板设计的关键要素。随着时间的推移，家庭组成会发生变化，因为成员的进出、孩子的出生和其他人的去世。当人们结婚或孩子离开父母的家时，会形成新的家庭，如果所有成员都去世或所有成员都搬向不同的方向，家庭可能会消失。家庭随时间推移不保持恒定的事实意味着只有在每个采访点观察他们在家庭背景下的个人时，才能随时间追踪个人。

还应注意的是，'家庭'并不等同于当前地址。家庭可能会搬到一个新地址，但保持相同的组成。同样，个人样本成员在调查期间可能会在几个地址之间移动。在此设计中，没有用新家庭替代或招募搬入样本成员留下的地址。

数据收集方式
---------------------------
面对面 [f2f]

研究工具
---------------------------
面板问卷内容

面板调查第1波的数据是LSMS数据，因此包含LSMS进行的所有模块。为了最小化受访者的负担并帮助维持面板调查的应答率，有必要减少LSMS问卷的长度和复杂性。然而，在尽可能的情况下，也必须保持问题措辞和响应类别的可比性，因为只有随着时间的推移可以比较的变量才能用于纵向分析。因此，第2波问卷是LSMS问卷的缩减版本，并包含一些面板所需的额外元素，例如收集搬入和搬出家庭的详细信息，以及一些LSMS中没有包含的新元素。1和2波变量的交叉列表显示了哪些变量在两波中都保留，哪些只在第1波保留，哪些只在第2波保留（见“变量协调LSMS_PANEL_final”）。最显著的变化是，第2波没有收集LSMS详细的消费模块，而农业模块与LSMS相比是简化的。
第2波的个人问卷包含一些路由，取决于这个人是否是在LSMS上采访的原始样本成员，或者自第1波以来加入家庭的新的个人。这是因为一些信息只需要收集一次，例如出生地和其他信息只需要每年更新一次。例如，所有资格都在LSMS上收集，因此对于原始成员，我们只需要知道他们在过去一年中是否获得了任何新的资格，但对于新成员，我们需要询问所有资格。数据用户需要了解这种路由，在某些情况下，如果它在本波中没有收集，他们可能需要从早期波次中获取信息。建议用户在使用数据时与问卷一起使用，以便了解不同样本成员的路由。