1/60
Looks like no tags are added yet.
population
entire group of individuals we want info on
population examples
entire school, kids in an orphange, etc.
census
collects data from every individual in a population
sample
subset of individuals in a population form which we actually collect data
convenience sample
choosing individuals that are easiest to reach
how do convenience samples cause bias
individuals usually like-minded (e.g. ppl arriving early at school)
voluntary response sample
people choose to respond
why are voluntary response samples biased
ppl with stronger opinions are more likely to reply (e.g. political surveys)
steps to describe bias
identify problem w/ design (e.g. sampling method or question wording)
explain how problem leads to under/overestimate
simple random sample/SRS
sample chosen such that every group has of n individuals in population has equal chance to be selected as sample
calculating SRS w/ technology
give every individual a numerical label from 1 → N
use MATH → probability → randInt(lower, upper, amt of #’s you want)
calculating SRS w/ table D
give every individual a numerical label from 1 → N
read consecutive digits from left to right off table D and ignore any value outside labeled range or that’s a duplicate
stratified random sample
group similar individuals into strata then do an SRS for each stratum and combine for full sample
cluster sample
group individuals that are close together (clusters) then take an SRS of clusters and sample all individuals in the cluster
pros of cluster sample
essentially convenience at random
an SRS means that every possible combo of ppl/items has
the same probability of being selected
smallest spread when graphing results means
best sampling variability
when comparing different ways to divide a sample
compare their spreads because smallest = best
why random sample?
avoids selection bias
allows for trustworthy inference abt population (bc of probability laws)
random samples have a margin of error that
limits size of likely error
why use larger random samples?
larger sample = better info/estimates compared to smaller samples
bias types
undercoverage
nonresponse
response bias
question wording
undercoverage def
some members of population can’t be chosen in sample
nonresponse def
when a specific person chosen can’t be contacted/refuses to answer
why is nonresponse different from voluntary response?
voluntary is self-elected but in nonresponse the person was randomly selected
response bias
systemic pattern of incorrect responses
question wording
most influential bias; e.g. extra context and double negatives
observational studies
can’t establish cause/effect; measures things but doesn’t influence anything
experiment
assigns a treatment to individuals and measures response; can measure cause/effect if well designed
confounding def
when 2 vars are associated so response can’t be separated btwn explanatory and other vars (usually in observational studies)
treatment def
conditions applied to individuals (e.g. coffee vs no coffee)
factors/lvls def
specific values of explanatory vars
how do treatments differ from lvls?
treatments are combos of lvls
experimental units def
physical entity/object/subject that receives a treatment independent of all other units in study e.g. households given medication
principles of well-designed experiment
comparison
random assignment
control
replication
comparison def
compares 2 or more treatments
random assignment def
use chance to assign experimental units to treatments; balances effects of confounding vars by making groups equivalent
control def
keep other variables that could affect response var constant for all groups
replication def
use enough experimental units per treatment so that diff treatment effects can be distinguished from chance
more people in a study/experiment =
more accurate and more representative
completely randomized design
treatments assigned to experimental units by chance; sometimes has a control group
placebo effect
response to dummy treatment
double-blind experiment
subjects and experimenter don’t know who got what treatment; prevents placebo effect + experimenter bias
causation determined by
statistically significant association
statistically significant means
observed effect is so large it couldn’t happen by chance
randomized block design
do random assignment within each block (essentially stratified random sample for experiments)
block def
group of experimental units known to have similar characteristics in some way that might affect response (assignment into blocks isnt random)
matched pairs design
randomized block but each block has a matching pair of similar experimental groups; could be one person who does both treatments in random order
randomized block design reduces
variation because it provides control (shrinks spread)
what percentage is considered statistically significant
<5% chhance of it happening
blocks are used to
strengthen conclusions (stronger responses within groups); sometimes rescale data to make comparison easier
random selection allows you to
generalize to a population
random assignment allows you to
determine causation
criteria for establishing causation
strong or consistent association
larger explanatory var values associated with stronger responses
cause has to come before effect
alleged cause has to be plausible
if they ask you to list possible treatments
LITERALLY LIST THEM ALL AND MULTIPLY LVLS TOGETHER TO GET NUMBER OF TREATMENTS
when designing an experiment, ALWAYS TRY TO
REDUCE VARIABILITY
random assignment doens’t
eliminate variables, just evens them
variability is like
noise that prevents radio signal (treatment effect) from going through
how to control for variability
blocking (eliminates source of variation within blocks)
increased sample size (better estimates but doesnt change underlying variability)
constant conditions (no artificial variability)
to prove a confounding variable
you have to link it to the response variable (impact on study outcome) and explanatory variable (impact on who gets it)
experimental units are the idea that
even if you measure individual responses at the end, the experimental unit is how you divide the treatments