Germán Rodríguez

Multilevel Models
Princeton University
We conclude our analysis of the Snijders and Bosker data by letting the intercept and slopes depend on a school-level predictor. Here are the data one last time.

. use https://data.princeton.edu/pop510/snijders, clear (Scores in language test from Snijders and Bosker, 1999) . quietly sum iq_verb . gen iqvc = iq_verb - r(mean)

> library(foreign) > snijders <- read.dta("https://data.princeton.edu/pop510/snijders.dta") %>% + mutate(iqvc = iq_verb - mean(iq_verb))

A useful exploratory step is to plot the estimated intercept and slope residuals obtained in the last model against potential predictors at the school level. Here an obvious candidate is school SES. As usual we will center this variable. We also store the school-level standard deviation for later use.

. bysort schoolnr: gen first = _n==1 . sum schoolses if first // just one Variable | Obs Mean Std. Dev. Min Max -------------+--------------------------------------------------------- schoolses | 131 18.40458 4.433749 10 29 . gen sesc = schoolses - r(mean) . scalar sesc_sd = r(sd)

> ses <- (group_by(snijders, schoolnr) %>% slice(1))$schoolses > school <- data.frame(meanses = mean(ses), sdses = sd(ses)) > snijders <- mutate(snijders, sesc = schoolses - school$meanses)

We are now ready to fit the model. We specify all the fixed effects first, including a cross-level interaction to let both the intercept and slope depend on school-level SES. These are followed by the random effects, which consists of the constant and the slope for verbal IQ:

. mixed langpost iqvc sesc c.iqvc#c.sesc || schoolnr: iqvc /// > , mle covariance(unstructured) Performing EM optimization: Performing gradient-based optimization: Iteration 0: log likelihood = -7608.5763 Iteration 1: log likelihood = -7607.8425 Iteration 2: log likelihood = -7607.8383 Iteration 3: log likelihood = -7607.8383 Computing standard errors: Mixed-effects ML regression Number of obs = 2,287 Group variable: schoolnr Number of groups = 131 Obs per group: min = 4 avg = 17.5 max = 35 Wald chi2(3) = 1059.58 Log likelihood = -7607.8383 Prob > chi2 = 0.0000 ------------------------------------------------------------------------------- langpost | Coef. Std. Err. z P>|z| [95% Conf. Interval] --------------+---------------------------------------------------------------- iqvc | 2.515011 .0789096 31.87 0.000 2.360351 2.669671 sesc | .2410466 .0658997 3.66 0.000 .1118856 .3702077 | c.iqvc#c.sesc | -.0470625 .0174818 -2.69 0.007 -.0813263 -.0127988 | _cons | 40.71326 .2910762 139.87 0.000 40.14276 41.28376 ------------------------------------------------------------------------------- ------------------------------------------------------------------------------ Random-effects Parameters | Estimate Std. Err. [95% Conf. Interval] -----------------------------+------------------------------------------------ schoolnr: Unstructured | var(iqvc) | .1489215 .0932377 .0436548 .5080228 var(_cons) | 8.22376 1.367463 5.936506 11.39226 cov(iqvc,_cons) | -.8862652 .2848344 -1.44453 -.3280001 -----------------------------+------------------------------------------------ var(Residual) | 41.51683 1.295675 39.05346 44.13558 ------------------------------------------------------------------------------ LR test vs. linear model: chi2(3) = 215.75 Prob > chi2 = 0.0000 Note: LR test is conservative and provided only for reference.

> library(lme4) > cli <- lmer(langpost ~ iqvc*sesc + (1 + iqvc | schoolnr), + data = snijders, REML = FALSE) > cli Linear mixed model fit by maximum likelihood ['lmerMod'] Formula: langpost ~ iqvc * sesc + (1 + iqvc | schoolnr) Data: snijders AIC BIC logLik deviance df.resid 15231.677 15277.557 -7607.838 15215.677 2279 Random effects: Groups Name Std.Dev. Corr schoolnr (Intercept) 2.8677 iqvc 0.3859 -0.80 Residual 6.4434 Number of obs: 2287, groups: schoolnr, 131 Fixed Effects: (Intercept) iqvc sesc iqvc:sesc 40.71326 2.51501 0.24105 -0.04706

We find that the relationship between language scores and verbal IQ varies substantially from school to school, depending on the school's SES and unobserved factors.

For the average school at mean SES, the mean language score for a child with average verbal IQ is 40.71 and increases an average of 2.52 points per point (or 5.20 points per standard deviation) of verbal IQ.

The expected score for an average kid increases 0.24 points per point (or about one point per standard deviation) of school SES, but this effect has a standard deviation across schools of 2.87 points, so there remain substantial unobserved school effects.

The average gain in language scores per point of verbal IQ decreases 0.05 points per point (or about 0.2 points per standard deviation) of school SES. This effect has a standard deviation of 0.39 across schools. Comparatively speaking, SES explains a bigger share of the variation in slopes than in intercepts.

Finally, we note that the random intercept and slope have a negative correlation of -0.80. This means that schools than tend to show higher language scores for average kids also tend to show smaller gains in language scores per point of IQ.

Next we will compute the predicted random effects and regression lines and plot the results:

. predict yhat3, fitted . sort schoolnr iqvc . line yhat3 iqvc, connect(ascending) /// > title(Random Coefficient Model with SES) /// > xtitle(Verbal IQ (centered)) ytitle(Language Score) . graph export fig1lang3.png, width(500) replace (file fig1lang3.png written in PNG format)

> snijders <- mutate(snijders, yhat3 = predict(cli)) > ggplot(group_by(snijders, schoolnr), aes(iqvc, yhat3, group=schoolnr)) + + geom_line() + ggtitle("Random Coefficient Model with SES") > ggsave("fig1lang3r.png", width = 500/72, height = 400/72, dpi = 72)

The figure looks pretty much the same as in the previous analysis. We see that differences across schools in language scores are larger at low verbal IQs and smaller at high verbal IQs. This is all done at observed values of school SES. We could, of course, generate lines setting centered school SES to zero, to see what the model implies for schoools with average SES.

A better way to compare the effects of school SES and unobserved school characteristics is to plot the regression lines for 4 schools which represent different scenarios. We set two to have school SES one standard deviation above and one standard deviation below the mean. We then combine these with school effects on the intercept and slope also set one standard deviation above and below the mean.

Because of the high negative correlation between the slope and intercept, it makes more sense to combine a positive intercept residual with a negative slope residual, reflecting a "better" school where scores are higher and less dependent on verbal IQ. The contrast is provided by a negative intercept residual combined with a positive slope residual, reflecting a "worse" school where scores are lower and more dependent on verbal IQ.

Here are the four scenarios, using the standard deviations saved earlier and the parameter estimates from the last run.

. scalar a0 = _b[_cons] . scalar b0 = _b[iqvc] . scalar af = _b[sesc] * sesc_sd . scalar bf = _b[c.iqvc#c.sesc] * sesc_sd //_b[iqvcXsesc] * sesc_sd . scalar ar = exp(_b[lns1_1_2:_cons]) . scalar br = -exp(_b[lns1_1_1:_cons]) // negative sign to pair +ve with -ne . capture drop arit* // to prevent ambiguity with ar . local range range(-7.8 6.1) . twoway /// > function y=(a0+af+ar) + (b0+bf+br)*x, lpat(solid) lc(green) `range' /// > || function y=(a0+af-ar) + (b0+bf-br)*x, lpat(solid) lc(red) `range' /// > || function y=(a0-af+ar) + (b0-bf+br)*x, lpat(dash) lc(green) `range' /// > || function y=(a0-af-ar) + (b0-bf-br)*x, lpat(dash) lc(red) `range' /// > , title(Predicted Regression Lines Given School SES and RE) /// > xtitle(verbal IQ (centered)) ytitle(Language score) /// > legend(order(1 "+ses +re" 2 "+ses -re" 3 "-ses +re" 4 "-ses -re") /// > ring(0) pos(5)) . graph export fig2lang3.png, width(500) replace (file fig2lang3.png written in PNG format)

> b <- fixef(cli) > sr <- attr(VarCorr(cli)$schoolnr, "stddev") > af <- b[3] * school$sdses > bf <- b[4] * school$sdses > x <- seq(-7.8, 6.1, 0.2) > d <- data.frame(iqvc = x, + f1 = (b[1] + af + sr[1]) + (b[2] + bf - sr[2]) * x, + f2 = (b[1] + af - sr[1]) + (b[2] + bf + sr[2]) * x, + f3 = (b[1] - af + sr[1]) + (b[2] - bf - sr[2]) * x, + f4 = (b[1] - af - sr[1]) + (b[2] - bf + sr[2]) * x) > x = d$iqvc[1] - 0.5 > ggplot(d, aes(iqvc, f1)) + xlim(-10, 6) + ylab("language score") + + geom_line() + geom_line(aes(iqvc,f2)) + + geom_line(aes(iqvc,f3), linetype=2) + geom_line(aes(iqvc,f4), linetype=2) + + geom_text(aes(x, f1[1]),label="+ses +re", hjust=1) + + ge .... [TRUNCATED] > ggsave( "fig2lang3r.png", width = 500/72, height = 400/72, dpi = 72)

Compare first the top two lines, which are schools with SES one sd above (solid) and below (dashed) the mean, which happen to have favorable unobserved characteristics leading to higher scores and a shallower slope. We see differences in language scores by school SES at low verbal IQs, but these become smaller at higher verbal IQs and vanish at the top end.

Compare now the bottom two lines, which are schools with SES one sd above (solid) and below (dashed) the mean, which happen to have adverse unobserved characteristics leading to lower scores and a steeper slope. We see essentially the same story, with differences by school SES declining with child's verbal IQ.

Compare next the two solid lines, which represent schools with high SES with favorable (green) or adverse (red) conditions. The differences in language scores are larger than the differences by SES, and although they decline with verbal IQ, they never dissappear.

The same is true for the two dashed lines, which correspond to schools with low SES and favorable or adverse unobserved characteristics, which tell the same story but with generally lower language scores.

Graphs like these are essential to sort out the effects of observed and unobserved characteristics. They are a key component of my analysis of infant and child mortality in Kenya in the multilevel handbook.

Continue with 3-Level Models