Lanlan Wang sent along this paper. Here’s the abstract:

We [Lanlan Wang and Peter Gordon] argue that levels of trust are shaped by formal and informal institutions. We discuss statistical test results that support this case. These utilize World Value Survey results and related data sources to examine in what way institutions or contextual factors affect individual trust levels. This investigation is the first, to our knowledge, to take into account micro-unit and macro-unit hierarchical structures and analyze the nature of context effects by using multilevel regression.

My comments:

1. I think the opening discussion of causal relations between trust and institutions would be clarified by using an intervention-based or potential-outcome approach to causal inference (something like the Rubin model). Then, instead of asking if “trust leads to better institutions–not the other way around,” you could consider possible interventions and consider how these would affect trust and affect institutions. I’m not saying that the intervention approach is the only way to think about causality, but it might clarify things here.

2. I’m surprised they don’t look at religious attendance, as well as denomination, as an input in their models.

3. They could do some rounding (see page 13). For example, 0.6173 could be rounded to 0.62 (actually, I’d just do 0.6). Also, I find sd’s more interpretable than variances, and I prefer to display them graphically (for example, see the Anova chapter in our new book).

4. I’m uncomfortable with the causal language on page 15, for example, “the context effect of Muslim, Orthodox and Catholic religions contributes to bringing down the level of trust.” I’d prefer to say that trust is lower in these countries, without attributing to the religion necessarily. I mean, I know about the Protestant ethic and all that, but still, I’d prefer to use associational language here.

5. I would make all the tables into graphs. Seriously. I mean it. For example, what do you learn by knowing that the “high range of the slope coefficient” for Muslim in model 20 is 0.5856?? My rule of thumb: don’t write any sentences that you don’t want anybody to read, and don’t put down any numbers that you don’t want anybody to read.

Anyway, this is important stuff, and it’s good to see multilevel modeling here. I hope this points the way to more focused studies of particular hypotheses.