1/35
Looks like no tags are added yet.
Target population
The full group the survey wants to learn about
Sampling Frame
The list/ material from which the sample is actually drawn
Sampled population
The units that had a chance of selection from the sampling frame
Can differ from the target population (coverage error)
Probability Sample
Every unit has a known, non-zero chance of selection, so design weights and population inference are possible
Non-probability sample
Selection probabilities are unknown (web polls e.g.), so no design weights and no poulation inference.
Simple Random Sample (SRS)
Every possible Sample of size n is equally likely
Simple Random Sample With Replacement (SRSWR)
A unit can be drawn more than once:
Pr(selection) = 1 - (1-1/N)^n
Simple Random Sample Without Replacement
A unit can be drawn once only
Pr(selection) = n/N
Stratified Sampling
Divide the population into non-overlapping strata and sample independently
Proportionate stratification
Sampling fraction fh = nh/Nh is the same in every stratum
Disproportionate Stratificaiton
Sampling fractions differ across strata (e.g. oversampling a small group or large firms), needs weights
Cluster Sampling
Sample groups (clusters) of units, not individuals; only some clusters are selected
One-stage cluster sample
All units in each selected cluster are surveyed
Two-stage cluster sample
Select clusters (PSU), then sample units within each selected cluster
Multi-stage sample
Further random selection at more stages (e.g. EA, then household, then person)
Primary Sampling Unit (PSU)
The first-stage unit sampled (e.g. in household survey, the EA)
Enumeration Area (EA)
Small census geographic area used as the cluster/ PSU
Systematic Sampling
Pick a random start, then take very k-th unit form a list (k = N/n)
Implicit stratification
Sort the list by variable, then sample systematically so the sample spreads across that variable
Probability Proportional to Size (PPS)
Selection probability of a cluster is proportional to its size Nj
Equal Probability of Selection Method
Every unit in the population has the same selection probability, so weights ar all equal
Listing
Enumerating all dwellings/ household in a selected cluster so a second-stage sample can be drawn
Stratified two-stage design (SA hh surveys)
Strata (province x metro/non-metro x geotype), PPS selection of PSUs, then systematic selection of dwellings within PSUs
Design Weight
1/πi the inverse of the units probability of selection,
The number of population units that household represents
Non-response weight
Design weight x 1/(response rate of units group)
Respondents are scaled up to represent non-respondents
Unit non-response
A sampled unit gives no data at all
Item non-response
A responding unit skips particular questions
Missing Completely at Random (MCAR)
Missingness is unrelated to any observed or unobserved variable
Missing at Random (MAR)
Missingness depends on observed variables only and can be ignored after conditioning on them
Not Missing at Random (NMAR)
Missingness depends on the unobserved value itself, even after conditioning on observed variables
pweight
Stata probability weight, correct point estimate and with svyset, correct SEs
aweight
Analytic weight, rescaled to sum to one, correct point estimate but SEs not design-correct. Use only when a command wont take pweight
fweight
Frequency weight, each observation is replicated that many times (integer counts0
iweight
Imprtance weight, raw weight with no statisical interpretaitons, some commands need it like povdeco
svyset
Tells Stata the design (weights, PSUs, strata)l so svy: commands give design correct SEs
Nationally Representative
Probability sample from a frame covering the whole target population properly weighted, with Ses small enough to be useful