Data

surveys
- careful with personal identification
- form brackets rather than actual values → less identifiable
- e.g.
company data
- data which the company has anyway
- e.g. amount of sick days
empirical data
- e.g. heart rate, skin conductivity

What to measure

measurement of changes
- across time and situation
- e.g. attitudes
Operationalization
- operational definition may vary
- take definition into account
conclusions should rely on many studies
- not just one, not just your own

anything with a social flavor in the experiment will be different in a different region
practical problem: there is little variance with the participants
always take into account if the participants of a study are relevant for your context

again Correlation vs Causation
Inference Problems
- time plots are not enough
  - a Counterfactual is needed
- also multiple samples are not better than single samples
  - more robust, just not solving the missing Counterfactual

Propensity Score Matching (PSM)
- matching 1:1 results from treatment and control group which have similar values
- e.g. in a sport study match athletes by sport and weight to analyse performance, could not compare a runner with a sumo fighter otherwise
Instrumental Variables (IV)
- variable is correlated with the treatment, not with the effect of the treatment
  - difference in treatment strength
- e.g. giving different doses of the same medicine to people in a medical study
Regression Discontinuity Design (RDD)
- clear cutoff across e.g. time but with very little difference
  - have to make sure that cutoff is not resulting in difference in sample
- e.g. difference between birthday at 31st August and 1st September is just 1 day - but one gets in the school system directly with 31st Aug and before, 1st Sep and later need to wait for next year
- e.g. Cross Border Deforestation
Difference in Difference (DiD)
- classic medical experiment structure - 2 samples, 1 receives treatment, 1 doesn’t
  - difference between treatment and placebo group is effect
- samples must be similar/equal by being large enough (n>40 each)
  - significance

Qualitative Research
- becomes more quantitative since the invention of LLMs (which is a good thing)
  - especially meta studies, analyzing sentiments
Correlation Study
Field Experiment
Survey
- Survey Design
Correlation Study
Lab Experiment