ECSSN and NZSA webinar: Analysing haphazard Surveys with mild-ish Assumptions with Gordana Popovic

While statisticians generally recommend collecting data via probability sampling, applied researchers studying humans frequently use haphazard and other non-probability samples to conduct surveys, and statistical consultants are regularly asked to help analyse such data. Principled analysis of non-probability samples relies on having reference data from probability samples or censuses, to adjust for non-response bias by say weighting or covariate adjustment. What can be done without auxiliary information?

It is well known that odds ratios are invariant under certain types of selection bias, for example outcome dependent sampling, which is why we use them in case control studies. We extend these results to selection bias on both the outcome and predictor, by way of some assumptions about how these biases are related.

In a collaboration on racial bias in police searches from a haphazard Facebook survey, where no auxiliary data was available, we used these assumptions to estimate odds ratios. We discuss how the assumptions were communicated with researchers so that their plausibility could be assessed by domain experts.

Biography:

Gordana Popovic completed her PhD in statistics and statistical ecology at UNSW in 2007, and has been working as a statistical consultant at Stats Central, UNSW Sydney every since. Her focus is on mentoring junior applied researchers to take a principled approach to quantitative research though teaching and collaboration.