Re: [R] Large mixed & crossed-effect model looking at educational spending on crime rates with error messages
(And post in plain text) On Tue, Sep 3, 2019 at 10:34 AM Bert Gunter wrote: > > You should post this on the R-sig-mixed-models list, not here. When you do > so, use ?dput to include data, not cut and paste. > > Cheers, > Bert > > > On Tue, Sep 3, 2019 at 6:10 AM Ades, James wrote: > > > > > I posted my question at Stack Overflow, where it didn’t get much of a > > response, and I was pointed in this direction by Ben Bolker. I’m happy to > > send the whole dataset to anyone who wants but thought that it would be > > presumptuous to include an enormous dput() here. > > > > I’m looking at the effects of education spending per school district on > > crime rate (FBI crime data/UCR) within the cities and towns those school > > districts serve over a fifteen year period. The DV now has 203,410 > > observations of city/town crime data over those fifteen years. (I use that > > figure with some reticence, because there are so many moving parts and > > things to account for, but having employed over 100 datasets and hours > > passing through the code again, I think that figure is correct.) > > > > Cities are technically crossed with school district, in that one city > > might attend multiple school districts. This means that one city could have > > multiple values for expenditure per student. School districts, however, > > also overlap with counties. As if things weren’t complicated enough, cities > > are mostly nested within county (though there are cities that exist in two > > counties, but it’s not often, and it’s usually by a small amount). Given > > that each city/town has a distinct PLACE_ID, my understanding is that this > > could be represented as (1|PLACE_ID) + (1|STATE/COUNTY_ID) or > > (1|STATE/COUNTY_ID/PLACE_ID). > > > > I’m pretty familiar with mixed-effect models, and I’ve looked through > > clear and informative posts such as this one: > > https://stats.stackexchange.com/questions/228800/crossed-vs-nested-random-effects-how-do-they-differ-and-how-are-they-specified. > > I believe things would remain sane to include school district > > (full_district_id) as another crossed effect, as below: > > glmer.total <- glmer(CRIME_TOTAL ~ 1 + (year|PLACE_ID) + > > (1|STATE/COUNTY_ID) + (year|full_district_id), family = "poisson", control > > = glmerControl(optimizer = "nloptwrap", calc.derivs = FALSE), REML = FALSE, > > total.years, na.action = "na.omit") > > > > Variables (not included in this model, to keep things null and simple) are > > centered and logged: pop per city, pop.dens per city, year, unemployment > > rate per county, proportion children living in poverty per school district, > > per capita income per county, difference in those who voted democrat in > > presidential elections per county, log enforcement per city/town, centered > > expenditure per student/ 1000 (per school district). PLACE_ID corresponds > > to cities and towns, COUNTY_ID to counties, full_district_id to school > > districts, and state. > > > > First, if I try to run the full model, using UCSD’s supercomputer, I get > > the error that the job was killed, presumably because it got to a point > > where it consumed too much ram (I think 125mb). > > > > I then tried to create a small subsection of data with arrange(STATE, > > COUNTY_ID, PLACE_ID) and then slicing by the first ten states (up through > > Delaware), so that I have 26,599 values. If I run this null model with the > > above code, I get the following error: > > > > ``` > > Error in getOptfun(optimizer) : > > optimizer function must use (at least) formal parameters ‘fn’, ‘par’, > > ‘lower’, ‘control’ > > ``` > > > > Then I tried with the optimx, with these configurations: control = > > glmerControl(optimizer = "optimx”, > > optCtrl = list(method = "nlminb”, > > maxit=1, > > iter.max=1, > > eval.max=1, > > lower = c(0,0,0), > > upper = c(Inf,10,1))) > > > > and I received the following warning…since this is a null model, there > > aren’t any variables to really rescale.\ > > > > ``` > > Warning messages: > > 1: In nlminb(start = par, objective = ufn, gradient = ugr, lower = lower, > > : > > unrecognized control elements named ‘lower’, ‘upper’ ignored > > 2: In nlminb(start = par, objective = ufn, gradient = ugr, lower = lower, > > : > > unrecognized control elements named ‘lower’, ‘upper’ ignored > > 3: In checkConv(attr(opt, "derivs"), opt$par, ctrl = control$checkConv, : > > Model failed to converge with max|grad| = 0.00102386 (tol = 0.001, > > component 1) > > 4: In checkConv(attr(opt, "derivs"), opt$par, ctrl = control$checkConv, : > > Model is nearly unidentifiable: very large eigenvalue > > - Rescale variables? > > ``` > > > > I then tried more values (to 92,486, through Missouri). First, I tried the > > optimizer nloptr, and then I tried the optimix. I still received the same > > above errors. > > > > I’ve checked and rechecked everything, so I wanted to solicit advice, > > either for where I might be going wrong, or for what I could do to re
Re: [R] Large mixed & crossed-effect model looking at educational spending on crime rates with error messages
You should post this on the R-sig-mixed-models list, not here. When you do so, use ?dput to include data, not cut and paste. Cheers, Bert On Tue, Sep 3, 2019 at 6:10 AM Ades, James wrote: > > I posted my question at Stack Overflow, where it didn’t get much of a > response, and I was pointed in this direction by Ben Bolker. I’m happy to > send the whole dataset to anyone who wants but thought that it would be > presumptuous to include an enormous dput() here. > > I’m looking at the effects of education spending per school district on > crime rate (FBI crime data/UCR) within the cities and towns those school > districts serve over a fifteen year period. The DV now has 203,410 > observations of city/town crime data over those fifteen years. (I use that > figure with some reticence, because there are so many moving parts and > things to account for, but having employed over 100 datasets and hours > passing through the code again, I think that figure is correct.) > > Cities are technically crossed with school district, in that one city > might attend multiple school districts. This means that one city could have > multiple values for expenditure per student. School districts, however, > also overlap with counties. As if things weren’t complicated enough, cities > are mostly nested within county (though there are cities that exist in two > counties, but it’s not often, and it’s usually by a small amount). Given > that each city/town has a distinct PLACE_ID, my understanding is that this > could be represented as (1|PLACE_ID) + (1|STATE/COUNTY_ID) or > (1|STATE/COUNTY_ID/PLACE_ID). > > I’m pretty familiar with mixed-effect models, and I’ve looked through > clear and informative posts such as this one: > https://stats.stackexchange.com/questions/228800/crossed-vs-nested-random-effects-how-do-they-differ-and-how-are-they-specified. > I believe things would remain sane to include school district > (full_district_id) as another crossed effect, as below: > glmer.total <- glmer(CRIME_TOTAL ~ 1 + (year|PLACE_ID) + > (1|STATE/COUNTY_ID) + (year|full_district_id), family = "poisson", control > = glmerControl(optimizer = "nloptwrap", calc.derivs = FALSE), REML = FALSE, > total.years, na.action = "na.omit") > > Variables (not included in this model, to keep things null and simple) are > centered and logged: pop per city, pop.dens per city, year, unemployment > rate per county, proportion children living in poverty per school district, > per capita income per county, difference in those who voted democrat in > presidential elections per county, log enforcement per city/town, centered > expenditure per student/ 1000 (per school district). PLACE_ID corresponds > to cities and towns, COUNTY_ID to counties, full_district_id to school > districts, and state. > > First, if I try to run the full model, using UCSD’s supercomputer, I get > the error that the job was killed, presumably because it got to a point > where it consumed too much ram (I think 125mb). > > I then tried to create a small subsection of data with arrange(STATE, > COUNTY_ID, PLACE_ID) and then slicing by the first ten states (up through > Delaware), so that I have 26,599 values. If I run this null model with the > above code, I get the following error: > > ``` > Error in getOptfun(optimizer) : > optimizer function must use (at least) formal parameters ‘fn’, ‘par’, > ‘lower’, ‘control’ > ``` > > Then I tried with the optimx, with these configurations: control = > glmerControl(optimizer = "optimx”, > optCtrl = list(method = "nlminb”, > maxit=1, > iter.max=1, > eval.max=1, > lower = c(0,0,0), > upper = c(Inf,10,1))) > > and I received the following warning…since this is a null model, there > aren’t any variables to really rescale.\ > > ``` > Warning messages: > 1: In nlminb(start = par, objective = ufn, gradient = ugr, lower = lower, > : > unrecognized control elements named ‘lower’, ‘upper’ ignored > 2: In nlminb(start = par, objective = ufn, gradient = ugr, lower = lower, > : > unrecognized control elements named ‘lower’, ‘upper’ ignored > 3: In checkConv(attr(opt, "derivs"), opt$par, ctrl = control$checkConv, : > Model failed to converge with max|grad| = 0.00102386 (tol = 0.001, > component 1) > 4: In checkConv(attr(opt, "derivs"), opt$par, ctrl = control$checkConv, : > Model is nearly unidentifiable: very large eigenvalue > - Rescale variables? > ``` > > I then tried more values (to 92,486, through Missouri). First, I tried the > optimizer nloptr, and then I tried the optimix. I still received the same > above errors. > > I’ve checked and rechecked everything, so I wanted to solicit advice, > either for where I might be going wrong, or for what I could do to resolve > these error messages. > > I’ve provided a brief snippet of the data below (randomly pulling a number > of cities within counties of Arkansas, Arizona, and Alabama, as a dput. > > Thanks! > > > > > 231, 206, 935, 1070, 974, 1108, 1244, 1095, 1131, 1151, 1420, > 1316, 132
[R] Large mixed & crossed-effect model looking at educational spending on crime rates with error messages
I posted my question at Stack Overflow, where it didn’t get much of a response,
and I was pointed in this direction by Ben Bolker. I’m happy to send the whole
dataset to anyone who wants but thought that it would be presumptuous to
include an enormous dput() here.
I’m looking at the effects of education spending per school district on crime
rate (FBI crime data/UCR) within the cities and towns those school districts
serve over a fifteen year period. The DV now has 203,410 observations of
city/town crime data over those fifteen years. (I use that figure with some
reticence, because there are so many moving parts and things to account for,
but having employed over 100 datasets and hours passing through the code again,
I think that figure is correct.)
Cities are technically crossed with school district, in that one city might
attend multiple school districts. This means that one city could have multiple
values for expenditure per student. School districts, however, also overlap
with counties. As if things weren’t complicated enough, cities are mostly
nested within county (though there are cities that exist in two counties, but
it’s not often, and it’s usually by a small amount). Given that each city/town
has a distinct PLACE_ID, my understanding is that this could be represented as
(1|PLACE_ID) + (1|STATE/COUNTY_ID) or (1|STATE/COUNTY_ID/PLACE_ID).
I’m pretty familiar with mixed-effect models, and I’ve looked through clear and
informative posts such as this one:
https://stats.stackexchange.com/questions/228800/crossed-vs-nested-random-effects-how-do-they-differ-and-how-are-they-specified.
I believe things would remain sane to include school district
(full_district_id) as another crossed effect, as below:
glmer.total <- glmer(CRIME_TOTAL ~ 1 + (year|PLACE_ID) + (1|STATE/COUNTY_ID) +
(year|full_district_id), family = "poisson", control = glmerControl(optimizer =
"nloptwrap", calc.derivs = FALSE), REML = FALSE, total.years, na.action =
"na.omit")
Variables (not included in this model, to keep things null and simple) are
centered and logged: pop per city, pop.dens per city, year, unemployment rate
per county, proportion children living in poverty per school district, per
capita income per county, difference in those who voted democrat in
presidential elections per county, log enforcement per city/town, centered
expenditure per student/ 1000 (per school district). PLACE_ID corresponds to
cities and towns, COUNTY_ID to counties, full_district_id to school districts,
and state.
First, if I try to run the full model, using UCSD’s supercomputer, I get the
error that the job was killed, presumably because it got to a point where it
consumed too much ram (I think 125mb).
I then tried to create a small subsection of data with arrange(STATE,
COUNTY_ID, PLACE_ID) and then slicing by the first ten states (up through
Delaware), so that I have 26,599 values. If I run this null model with the
above code, I get the following error:
```
Error in getOptfun(optimizer) :
optimizer function must use (at least) formal parameters ‘fn’, ‘par’,
‘lower’, ‘control’
```
Then I tried with the optimx, with these configurations: control =
glmerControl(optimizer = "optimx”,
optCtrl = list(method = "nlminb”,
maxit=1,
iter.max=1,
eval.max=1,
lower = c(0,0,0),
upper = c(Inf,10,1)))
and I received the following warning…since this is a null model, there aren’t
any variables to really rescale.\
```
Warning messages:
1: In nlminb(start = par, objective = ufn, gradient = ugr, lower = lower, :
unrecognized control elements named ‘lower’, ‘upper’ ignored
2: In nlminb(start = par, objective = ufn, gradient = ugr, lower = lower, :
unrecognized control elements named ‘lower’, ‘upper’ ignored
3: In checkConv(attr(opt, "derivs"), opt$par, ctrl = control$checkConv, :
Model failed to converge with max|grad| = 0.00102386 (tol = 0.001, component
1)
4: In checkConv(attr(opt, "derivs"), opt$par, ctrl = control$checkConv, :
Model is nearly unidentifiable: very large eigenvalue
- Rescale variables?
```
I then tried more values (to 92,486, through Missouri). First, I tried the
optimizer nloptr, and then I tried the optimix. I still received the same above
errors.
I’ve checked and rechecked everything, so I wanted to solicit advice, either
for where I might be going wrong, or for what I could do to resolve these error
messages.
I’ve provided a brief snippet of the data below (randomly pulling a number of
cities within counties of Arkansas, Arizona, and Alabama, as a dput.
Thanks!
231, 206, 935, 1070, 974, 1108, 1244, 1095, 1131, 1151, 1420,
1316, 1321, 1414, 1484), full_district_id = c("0100240", "0100240",
"0100240", "0100240", "0100240", "0100240", "0100240", "0100240",
"0100240", "0100240", "0100240", "0100240", "0100240", "0100240",
"0100240", "0100240", "0100240", "0100240", "0100240", "0100240",
"0100240", "0100240", "0100240", "0100240", "0100240", "0100240",
"0100240", "0
